# ChatGPT Referrals to E-Commerce Websites: How Do LLMs Compare Against Traditional Channels? **Maximilian Kaiser**¹ · **Christian Schulze**² ¹ University of Hamburg; Grips Intelligence — research@maximiliankaiser.org ² Frankfurt School of Finance & Management — email@christian-schulze.de *Forthcoming, Marketing Science. Final manuscript March 2026.* > [!info] Full manuscript > Complete text, converted from the authors' LaTeX source (March 2026 manuscript) — all sections, tables, figures, and the online appendix. Table and figure numbering follows this manuscript (tables 1–4, A.1–A.8, B.1–B.12, D.1–D.3; figures 1–26, A.1, B.1–B.5). Note: the older SSRN-preprint PDF numbers some appendix figures differently (it keeps the AOV/RPS robustness graphs in the body as Figures 15–16), so figure numbers 15+ won't match that preprint. For the curated key-findings digest, see the summary note. SSRN preprint: [abstract 5585812](https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5585812). Original PDF: Attachments/Kaiser_Schulze_2025-12_ChatGPT-referrals-ecommerce.pdf. **Related:** pLLM - Paid Ads in ChatGPT · E-Commerce Conversion Rates · GRIPS Intelligence Data · GenAI Optimization · RPS - Revenue per Session · oLLM paper - Alternative Explanations for greater CR but lower AOV · oLLM Practitioner Article - HBR or MIT Sloan · llms.txt Adoption and resulting oLLM Traffic · 2026-05-26_kaiser-schulze-ollm-research-catchup --- **Abstract** We investigate organic Large Language Model traffic (oLLM) versus traditional digital channels in e-commerce. Analyzing 12 months of first-party data from 973 websites with \$20 billion combined revenue, we examine over 50,000 transactions from ChatGPT referrals alongside 164 million transactions from traditional channels. Using regression models that account for data sparsity, we assess financial metrics (conversion rate, average order value, revenue per session) and engagement metrics (bounce rate, session duration, page views). Results are consistent across extensive robustness checks. One year after launch, oLLM exhibits conversion rates and revenue per session above paid social but below all other traditional channels. Product complexity moderates the effects: oLLM’s financial outcomes and traffic shares are stronger in complex product categories. Engagement metrics show favorable bounce rates but lower session duration and page views. Temporal analysis shows increasing conversion rates but declining average order values, yielding only moderate revenue-per-session gains over time. Cross-website analyses support growing consumer LLM proficiency as the underlying mechanism. The descriptive study positions oLLM as a new and developing channel. With low volumes and modest revenue per session, oLLM currently serves niche informational needs of proficient consumers and does not yet function as a broad conversion channel. *Keywords: large language models (LLMs), electronic commerce, channel analytics, performance marketing, digital economy* # Introduction A new digital channel emerged in August 2024 when ChatGPT, the dominant Large Language Model (LLM) platform, began providing organic outgoing links that direct users to relevant e-commerce websites upon expression of purchase intent. For instance, consumers searching for an espresso machine receive personalized product comparisons and country-specific purchase links (Figure 1). As of the time of this study, these links are fully organic and not advertiser-driven. ![](assets/figs/Kaiser_Schulze_oLLM_fig_1.png) ***Figure 1.** Product recommendations for espresso machines provided by ChatGPT, featuring outgoing links to online retailers (“Buy” links on the right).* This organic LLM traffic (oLLM) enters a landscape long shaped by organic and paid search (Ghose and Yang 2009; Li et al. 2016; Reisenbichler et al. 2022) but also referrals, email, affiliate, paid social, and direct channels (Manchanda et al. 2006; Bleier and Eisenbeiss 2015; Haan, Wiesel, and Pauwels 2016; Wies, Bleier, and Edeling 2023; SimilarWeb 2024). Channels differ in where and how they match consumers to sellers along the purchase funnel. Their lower-funnel effectiveness is commonly measured by conversion rates in academia (Ghose and Yang 2009; Rutz and Bucklin 2011) and practice (Google 2024). How oLLM compares remains an open question. Several mechanisms suggest oLLM may generate superior outcomes. LLMs access richer contextual information than traditional channels (Soviero et al. 2024), which have been constrained by privacy regulation (Johnson, Shriver, and Du 2020; Miller and Skiera 2024; Aridor et al. 2025). They synthesize information across attributes, reviews, and contextual factors (Y. Yu et al. 2024) and can reason across dimensions that keyword-based search cannot (Al-Hasan et al. 2024; Luo et al. 2019; P. Yu et al. 2025). Conversational interfaces further enable preference clarification and iterative personalization (Jannach 2022; Al-Hasan et al. 2024), potentially reducing cognitive burden and increasing convenience and persuasion (Steyvers and Kumar 2024; Salvi et al. 2025). Early industry evidence is consistent with this logic: a study of 100 websites reports 6.7% conversion rate for oLLM versus 3.9% for organic search (ThoughtMetric 2025); another in-depth study reports 15.9% versus 1.8% (Seer Interactive 2025); and value per session from oLLM is estimated to be 4.4× higher than organic search (Semrush 2025). Other evidence points in the opposite direction. Perceived recommendation usefulness may be hampered by inaccuracies in current LLM outputs (Search Engine Land 2025), declining response quality at scale (Huang and Rust 2025), and interface characteristics that can increase cognitive load (Nguyen, Sidorova, and Torres 2022). Consumer adoption of LLM-based shopping remains early-stage: technology anxiety and trust moderate adoption intentions (Foroughi et al. 2025), adoption and use correlate with digital sophistication and education (Yang et al. 2025), and only 2.1% of ChatGPT conversations involve purchasable products (Chatterji et al. 2025), reflecting delayed emergence of behavioral shifts (Padilla et al. 2025). Empirical findings consistent with these indicators include Adobe (2025) reporting oLLM conversion rates 9% lower than “non-AI” channels, and SALT (2025) finding engagement levels 27% below organic search in most categories. Understanding oLLM has broader implications. Changes in search costs and decision quality may influence consumer welfare (Anderson and Renault 2006; Lynch and Ariely 2000; Brown and Goolsbee 2002; Ursu, Zhang, and Honka 2022; Dinerstein et al. 2018; Ellison and Ellison 2009). Recommendation bias may differ from advertiser-influenced channels (OpenAI 2025; Reuters 2025).[^3]As deep-link entry grows, website design may need to adapt and retail media may become less important (Wroe 2024; eMarketer 2024). More generally, LLMs are reshaping online behavior (Padilla et al. 2025; Gholami et al. 2026) and fueling expectations of disruption to e-commerce channels (The Economist 2023). Our descriptive study is the first large-scale empirical analysis of oLLM relative to traditional channels. Using 12 months of data (August 2024–July 2025) from 973 e-commerce websites with \$20 billion in annual revenue, we observe more than 50,000 oLLM transactions and 164 million transactions from other channels. Through direct access to companies’ Google Analytics accounts provided by Grips Intelligence,[^4] we observe the channel that brought each consumer to the e-commerce website and their subsequent purchase behavior. Importantly, because these channel assignments rely on last-click attribution, we cannot capture upper-funnel contributions and will understate channels’ roles when they primarily serve discovery functions (Li and Kannan 2014; Li et al. 2016; Haan, Wiesel, and Pauwels 2016; Berman 2018). Still, last-click metrics remain the industry standard and, under current data constraints, provide the most comparable basis for cross-website channel benchmarking. We interpret all results with these limitations in mind. We compare oLLM with traditional channels across financial metrics (conversion rate, average order value, and revenue per session) and engagement metrics (bounce rate, session duration, and page views) using both model-free and regression approaches. We test robustness across alternative data processing choices, website samples, LLM platforms, and timeframes. We also explore differential effects across websites that differ in product category complexity and consumers’ LLM proficiency (Figure 2). ![](assets/figs/Kaiser_Schulze_oLLM_fig_2.png) ***Figure 2.** Conceptual model* Two patterns emerge. First, oLLM exhibits conversion rates and revenue per session above paid social but below all other traditional channels. Product complexity moderates the effects: oLLM’s financial outcomes and traffic shares are stronger in complex product categories, pointing to limited perceived usefulness for simpler products. Engagement metrics reveal relatively favorable bounce rates but lower session duration and page views. Second, temporal analysis shows increasing conversion rates. However, declining average order values partially offset these improvements, resulting in moderate revenue-per-session uplift over time. Cross-website analyses reveal that this pattern is consistent with growing consumer LLM proficiency as an underlying mechanism. The paper proceeds as follows: We present data, metrics, and methodology, followed by descriptive evidence and regression results comparing oLLM with traditional channels. We then examine robustness, temporal dynamics, and heterogeneity in oLLM patterns across websites. We conclude with summary of key insights and limitations. # Data, Metrics, and Models ## Traffic Volume of oLLM and Focus on ChatGPT One year after launch, oLLM accounts for less than 0.2% of all traffic in our dataset (Figure 3), about 200 times smaller than Google’s organic search. This reflects oLLM’s low volume in the broader market (SE Ranking 2025; Ahrefs 2025). ![](assets/figs/Kaiser_Schulze_oLLM_fig_3.png) ***Figure 3.** Organic LLM traffic (from ChatGPT only) compared to traditional channels (log scale)* ChatGPT dominates the oLLM channel, accounting for over 90% of observed sessions[^5] from LLM platforms (Figure 4). Other platforms (Perplexity 4.1%, Gemini 2.6%, Copilot 2.1%, Deepseek 0.07%, Grok 0.02%) show negligible volume. Our main analyses therefore examine ChatGPT traffic only, but we include all platforms in a robustness check (Section 4). ![](assets/figs/Kaiser_Schulze_oLLM_fig_4.png) ***Figure 4.** Observed traffic from large language models in our data (absolute scale)* ## Dataset We analyze first-party e-commerce data from Google Analytics across 973 websites. Table 1 shows overlapping 3-, 6-, and 12-month data blocks. Our main analyses use the 6-month dataset, capturing most oLLM sessions while avoiding potential idiosyncratic introduction effects. We aggregate data weekly to account for sparse observations. We replicate all analyses using alternative timeframes (3 and 12 months) and aggregation periods (daily, monthly) in Section 4. **Table 1. Overview of Datasets** | Metric | 12 Months (robustness) | 6 Months (main) | 3 Months (robustness) | |:--|--:|--:|--:| | Date Range Start | 2024-07-28 | 2025-02-02 | 2025-04-27 | | Date Range End | 2025-08-02 | 2025-08-02 | 2025-08-02 | | Number of Websites | 973 | 973 | 953 | | Total Sessions | 10,513,177,108 | 6,010,804,542 | 3,271,538,643 | | Total Revenue | \$20,602,399,030 | \$11,485,304,150 | \$6,301,491,260 | | Total Transactions | 164,875,690 | 92,826,266 | 51,391,206 | | Total oLLM Transactions | 50,251 | 44,437 | 33,680 | | Total oLLM Sessions | 4,915,779 | 4,149,187 | 2,944,179 | | Total oLLM Revenue | \$6,965,894 | \$6,082,834 | \$4,594,709 | | Mean Sessions per Website/week | 238,627 | 254,189 | 253,509 | | Mean Revenue per Website/week | \$467,631 | \$485,698 | \$488,298 | | Mean Transactions per Website/week | 3,742 | 3,925 | 3,982 | | Mean Conversion Rate per Website/week | 3.30% | 3.27% | 3.31% | Note: Weekly aggregated data. oLLM refers to traffic from ChatGPT referrals. ### Coverage Following Similarweb (2025), our dataset contains websites from 24 categories. The largest is “E-Commerce and Shopping” (including retailers and marketplaces) with 1.2 billion sessions, followed by “Lifestyle” (including Fashion, Beauty and Cosmetics) with 1.1 billion sessions. Detailed category descriptions and statistics appear in Online Appendix A.1. The dataset covers all continents, with the Americas (2.7 billion sessions) and Europe (2.1 billion sessions) representing the majority. 49 countries contribute at least 10 million sessions each. Detailed statistics appear in Online Appendix A.2. ## Metrics We focus our analyses on financial metrics: conversion rate (CR), average order value (AOV), and revenue per session (RPS). CR is the most widely used financial performance metric in e-commerce, capturing the share of website sessions that result in a transaction. We also measure AOV, because the probability of purchase tends to be higher when the product is cheaper, meaning CR and AOV are confounded. RPS captures the joint effect of purchase likelihood and spending per purchase, and gives meaningful intuition on the value of sessions from an advertising perspective. Table 2 provides descriptives for oLLM and eight traditional channels. **Table 2. Channel Volume Summary Statistics (February–August 2025)** | Channel | Obs | Sessions (M) | Revenue (\$M) | |:---------------|-------:|-------------:|--------------:| | oLLM | 30,232 | 4.1 | 6.1 | | Paid Search | 44,800 | 1,530.0 | 3,190.4 | | Organic Search | 47,244 | 1,427.3 | 2,595.4 | | Direct | 47,247 | 1,269.6 | 2,695.3 | | Other | 46,941 | 830.5 | 1,581.5 | | Referral | 46,789 | 353.5 | 605.2 | | Paid Social | 27,019 | 337.5 | 145.5 | | Email | 36,544 | 215.9 | 513.5 | | Affiliate | 13,476 | 42.5 | 152.6 | Notes: (M) = millions, observations refer to week-website-device-channel data points. We investigate engagement through bounce rate (BR), session duration (SD), and page views (PV). While engagement does not directly translate to financial results, it indicates traffic quality. High BR (visitors leaving without another click) typically signals poor fit between website content and consumer interest. Complete variable definitions appear in Online Appendix A.3. ## Modeling Approach Sparse oLLM observations require careful modeling. At the week/website/device level, CRs may derive from few sessions and AOVs from single transactions. Our models address two challenges: (a) datapoints vary substantially in underlying sessions, and (b) ratio metrics exhibit overdispersion. For CR and BR, we estimate quasibinomial models with dispersion parameters. For AOV, RPS, SD, and PV, we estimate weighted linear models with bounded, website-relative weights to reduce aggregation-induced heteroskedasticity and cap leverage. All models use: $$\text{logit}\!\left(p^{\text{CR}}_{it}\right) = \alpha_{c(i,t)} + \gamma_i + \delta_{d(i,t)} + \mu_{m(t)},$$ where $\alpha_{c}$ are channel fixed effects, $\gamma_i$ website fixed effects, $\delta_d$ device fixed effects, and $\mu_m$ month fixed effects. The linear specifications follow the same fixed-effects structure, with outcomes modeled in levels rather than log-odds. Complete specifications appear in Online Appendix A.4. # Results ## Model-Free Evidence Model-free comparisons suggest lower CR, AOV, and RPS for oLLM relative to traditional channels (Figures 5 and 6). However, these distributions reflect sparse data: the median oLLM observation contains zero transactions. Without accounting for underlying session counts or website heterogeneity, such comparisons prove misleading—a pattern that persists across engagement metrics (Figures 7 and 8). ![](assets/figs/Kaiser_Schulze_oLLM_fig_5.png) ***Figure 5.** Model-free evidence on CR by channel* ***Figure 6.** Model-free evidence on AOV and RPS by channel* ![](assets/figs/Kaiser_Schulze_oLLM_fig_6_1.png) *(a) AOV* ![](assets/figs/Kaiser_Schulze_oLLM_fig_6_2.png) *(b) RPS* ![](assets/figs/Kaiser_Schulze_oLLM_fig_7.png) ***Figure 7.** Model-free evidence on BR by channel* ***Figure 8.** Model-free evidence on PV and SD by channel* ![](assets/figs/Kaiser_Schulze_oLLM_fig_8_1.png) *(a) PV* ![](assets/figs/Kaiser_Schulze_oLLM_fig_8_2.png) *(b) SD* ## Regression Results Table 3 reports regression estimates with channel fixed effects $\alpha_c$ measuring differences between each traditional channel and oLLM (the reference category). We highlight in **bold** coefficients with $p < 0.05$ and favoring oLLM. **Table 3. Regression Results: 6 Months Data** | | CR | AOV | RPS | BR | SD | PV | |:--|--:|--:|--:|--:|--:|--:| | Affiliate | 0.621\*\*\* (0.060) | 24.7\*\*\* (3.9) | 3.57\*\*\* (0.12) | 0.217\*\*\* (0.035) | -12.2\* (6.8) | 0.642\*\*\* (0.057) | | Paid Search | 0.370\*\*\* (0.058) | 4.08 (3.10) | 1.95\*\*\* (0.08) | -0.156\*\*\* (0.033) | 31.2\*\*\* (4.5) | 0.905\*\*\* (0.038) | | Other | 0.366\*\*\* (0.059) | 5.51\* (3.15) | 2.19\*\*\* (0.08) | 0.883\*\*\* (0.033) | 46.1\*\*\* (4.7) | 0.622\*\*\* (0.039) | | Direct | 0.282\*\*\* (0.058) | 13.5\*\*\* (3.1) | 3.07\*\*\* (0.08) | 0.321\*\*\* (0.033) | 5.86 (4.56) | 0.692\*\*\* (0.038) | | Email | 0.276\*\*\* (0.059) | 10.5\*\*\* (3.3) | 2.69\*\*\* (0.09) | 0.181\*\*\* (0.033) | 55.4\*\*\* (5.1) | 0.876\*\*\* (0.042) | | Referral | 0.218\*\*\* (0.059) | 12.6\*\*\* (3.2) | 1.61\*\*\* (0.08) | 0.202\*\*\* (0.033) | 59.8\*\*\* (4.8) | 0.776\*\*\* (0.040) | | Organic Search | 0.121\*\* (0.058) | 1.83 (3.08) | 1.01\*\*\* (0.08) | -0.136\*\*\* (0.033) | 14.6\*\*\* (4.5) | 0.571\*\*\* (0.037) | | Paid Social | -0.760\*\*\* (0.059) | 5.42 (3.49) | -0.718\*\*\* (0.091) | 0.360\*\*\* (0.033) | -131\*\*\* (5) | -0.413\*\*\* (0.045) | | Fixed Effects: Website, Device, Month | Yes | Yes | Yes | Yes | Yes | Yes | | $R^2$ \[pseudo-$R^2$\] | \[0.739\] | 0.472 | 0.284 | \[0.711\] | 0.457 | 0.155 | | Observations | 340,292 | 289,524 | 340,292 | 340,292 | 340,292 | 340,292 | | | \* p<0.1; \*\* p<0.05; \*\*\* p<0.01 | | | | | | Notes: All coefficients relative to oLLM baseline. Standard errors in parentheses. ### Financial Outcomes Figure 9 displays channel fixed effects as log-odds coefficients. To facilitate interpretation, we convert these to percentage differences in conversion likelihood. Among all channels, paid social and organic search exhibit the smallest gaps relative to the oLLM baseline (shown as red line). Paid social sits substantially below oLLM (coefficient: -0.760), corresponding to a 53% lower likelihood of conversion. Organic search modestly exceeds oLLM (coefficient: 0.121), translating to a 13% higher conversion likelihood. Other channels show larger differences: affiliate’s coefficient of 0.621 corresponds to an 86% higher likelihood of conversion relative to oLLM. These regression-adjusted estimates substantially narrow the gaps indicated by model-free evidence. ![](assets/figs/Kaiser_Schulze_oLLM_fig_9.png) ***Figure 9.** Regression results for CR by channel (log-odds)* AOV estimates exhibit wider confidence intervals, yielding statistically significant differences for only four of eight channels (Figure 10). Affiliate shows the largest difference at \$24.7 higher than oLLM. Differences for organic search, paid search, paid social, and other are not statistically significant. ![](assets/figs/Kaiser_Schulze_oLLM_fig_10.png) ***Figure 10.** Regression results for AOV by channel* RPS results show tighter confidence intervals (Figure 11). oLLM exceeds paid social but falls significantly below remaining channels. ![](assets/figs/Kaiser_Schulze_oLLM_fig_11.png) ***Figure 11.** Regression results for RPS by channel* ### Engagement Outcomes BR is the only metric where oLLM shows more favorable outcomes than most channels (Figure 12), though organic and paid search—channels optimized for low BR—still show better rates. Email shows 20% higher BR than oLLM; organic search shows 13% lower. ![](assets/figs/Kaiser_Schulze_oLLM_fig_12.png) ***Figure 12.** Regression results for BR by channel* PV analysis shows oLLM below all channels except paid social, while SD places oLLM far ahead of paid social and comparable to affiliate and direct channels (Figure 13). ***Figure 13.** Regression results for PV and SD by channel* ![](assets/figs/Kaiser_Schulze_oLLM_fig_13_1.png) *(a) PV* ![](assets/figs/Kaiser_Schulze_oLLM_fig_13_2.png) *(b) SD* # Robustness The main models aggregate data at the weekly level, apply no minimum observation thresholds, cover 973 websites, analyze ChatGPT as the only oLLM channel, and span 6 months. We test the robustness of our findings with respect to all of these specifications. Table 4 summarizes observations and model fit. Notably, the $R^2$ values of the CR model remain stable despite substantial variation in sample size. **Table 4. Robustness Checks Overview** | Specification | Robustness Concern Addressed | R² | Observations | |:--------------------------------|:---------------------------------------|------:|-------------:| | Main Model | – (reference model) | 0.739 | 340,292 | | | | | | | Daily Aggregation | Effect of sparse observations for oLLM | 0.689 | 3,896,527 | | Monthly Aggregation | Effect of sparse observations for oLLM | 0.750 | 149,628 | | Min 10 Sessions | Effect of sparse observations for oLLM | 0.739 | 320,021 | | Min 100 Sessions | Effect of sparse observations for oLLM | 0.740 | 280,190 | | Min 1000 Sessions | Effect of sparse observations for oLLM | 0.742 | 188,376 | | Winsorized | Outliers | 0.727 | 340,292 | | | | | | | Top 25% oLLM Sessions | Selection of websites in the data | 0.700 | 91,385 | | Top 25% Total Transactions | Selection of websites in the data | 0.738 | 91,942 | | | | | | | ChatGPT plus other LLMs | Operationalization of oLLM | 0.746 | 408,342 | | ChatGPT plus mobile app traffic | Operationalization of oLLM | 0.741 | 395,430 | | | | | | | 12 Months | Stability of the effects over time | 0.723 | 614,714 | | 3 Months | Stability of the effects over time | 0.727 | 188,038 | Notes: R² values shown for CR regression model denote McFadden pseudo-$R^2$ values. Figure 14 reports CR channel fixed effects across all specifications. The core finding is robust: oLLM exhibits lower CR than all traditional channels except paid social. Differences relative to organic search become statistically insignificant in five specifications (monthly aggregation, minimum 100 sessions, top 25% of oLLM traffic, oLLM including mobile app, and the 3-month window) and exhibit a non-significant sign reversal in one specification (minimum 1,000 sessions). Effect sizes vary systematically: the top-25%-revenue subset shows consistently larger gaps, whereas the minimum-1,000-sessions specification yields smaller effects. Nonetheless, the overall pattern remains consistent. ![](assets/figs/Kaiser_Schulze_oLLM_fig_14.png) ***Figure 14.** Illustration of robustness check results for CR by channel* AOV, RPS, and engagement metrics exhibit similar robustness, with minor variation in effect sizes. Details on all tests are available in Online Appendix B. # Temporal Dynamics of oLLM Outcomes The preceding analyses document oLLM’s lower CR and RPS relative to most traditional channels. Yet oLLM remains a young and rapidly developing channel, with both the underlying technology and consumer adoption still maturing. This raises a natural generalizability question: do the observed gaps persist, narrow, or widen over the study period? To address this question, we examine temporal trends in CR, AOV, and RPS. Channel fixed effects across 12-, 6-, and 3-month windows already suggest gradual improvement. We now model these trajectories explicitly, using multiple trend specifications to assess robustness. This analysis cannot isolate specific mechanisms driving temporal changes but can characterize the direction and pace of oLLM’s trajectory relative to established channels. ## Model-free Evidence on Time Trend Model-free evidence (Figure 15) shows CR increases over time for oLLM, while traditional channels remain stable except for seasonal peaks in November and December. Mean weekly AOV (Figure 16) exhibits substantially more volatility, especially for oLLM, and appears to decrease over the observation period. RPS remains stable for traditional channels but shows systematic increase for oLLM (Figure 17). ![](assets/figs/Kaiser_Schulze_oLLM_fig_15.png) ***Figure 15.** CR by channel over time* ![](assets/figs/Kaiser_Schulze_oLLM_fig_16.png) ***Figure 16.** AOV by channel over time* ![](assets/figs/Kaiser_Schulze_oLLM_fig_17.png) ***Figure 17.** RPS by channel over time* ## Regression Analysis with Time Trend Given data sparsity challenges with model-free evidence for oLLM, we estimate regression models mirroring main specifications but including trend variables. Trend variable specification substantially impacts predictions, so we test four alternatives: linear trend, centered (linear) model, Gompertz model (asymmetric S-curve), and Sigmoid model (logistic with hard upper limit). Complete specifications appear in Online Appendix C.1. ## Projection of Financial Metrics To facilitate interpretation, we create population estimates for all websites across time, allowing us to project oLLM outcomes in the next year. While trend continuation is highly uncertain, projections may offer insight into development speed necessary for oLLM to converge with traditional channels. Figure 18 illustrates slow CR increase for traditional channels and comparatively steeper increase for oLLM. Three models consistently predict CR for oLLM will approach organic search but not reach parity within 12 months. The linear time model shows exponential growth due to the logit transformation of probabilities. ![](assets/figs/Kaiser_Schulze_oLLM_fig_18.png) ***Figure 18.** Prediction of CR development per channel* Figure 19 shows substantial downward trend for oLLM AOV, which counteracts the positive development in CR and raises the question, whether the CR improvement is only an artifact of lower AOVs. ![](assets/figs/Kaiser_Schulze_oLLM_fig_19.png) ***Figure 19.** Prediction of AOV development per channel* For RPS, we find oLLM gradually produces more valuable sessions. Similar to CR predictions, the linear model is more optimistic about oLLM outcomes, while the other three models consistently predict oLLM will not converge with the next best channel, organic search, within 12 months (Figure 20). ![](assets/figs/Kaiser_Schulze_oLLM_fig_20.png) ***Figure 20.** Prediction of RPS development per channel* # Cross-Website Variation in oLLM Outcomes The preceding analyses document oLLM’s lower CR, AOV, and RPS relative to all traditional channels except paid social, as well as changes over time. Multiple mechanisms could explain these patterns, ranging from trust deficits and privacy concerns to differences in search intent and funnel position. We focus on two mechanisms grounded in prior research and amenable to empirical exploration with our data. First, LLM recommendations may offer limited usefulness due to inaccuracies or interface friction (Search Engine Land 2025; Nguyen, Sidorova, and Torres 2022). Second, consumers are still developing proficiency in using LLMs (Chatterji et al. 2025; Foroughi et al. 2025; Yang et al. 2025). To investigate these conjectures, we exploit cross-sectional variation across[^6] the 973 websites in our dataset. These analyses are descriptive rather than causal: they offer suggestive evidence on potential mechanisms but remain vulnerable to alternative explanations and should not be interpreted as definitive proof. ## Product Complexity We examine how oLLM outcomes vary with product complexity, under the assumption that LLMs are more useful when purchases require substantial information synthesis and deliberation. For simple products with straightforward decisions, conversational assistance may add little incremental value while introducing friction. We proxy product complexity using website categories classified by three LLMs (Claude, Gemini, ChatGPT). Heavy industry and engineering, business and consumer services, vehicles, law and government, finance, and jobs and career are categorized as high-complexity categories, while reference materials, news and media, adult, sports, and arts and entertainment are categorized as low-complexity (see Table A.2 in the Online Appendix for complete rankings). Websites in high-complexity categories exhibit 4.6 times higher oLLM traffic shares on average, consistent with our usefulness assumption and classification. Across all three financial metrics, oLLM outcomes are more favorable on high-complexity websites (see Online Appendix D for detailed results). When product complexity is high, oLLM CR exceeds paid social, referral, email, organic search, and direct channels (Figure 21); AOV differences relative to other channels narrow (Figure 22); and RPS moves closer to organic and paid search levels (Figure 23). Taken together, these patterns are consistent with (currently) limited usefulness for simple products as one potential mechanism behind oLLM’s weaker financial metrics in Section 3. However, this interpretation does not align with the temporal pattern in Section 5, which shows decreases in AOV. A caveat is that, although all conversions in our data are financial transactions, what constitutes a conversion can still vary across product categories (see Table A.2 in the Online Appendix), so the complexity split may partly reflect these category-level differences in conversion type rather than differences in LLM assistance per se. ![](assets/figs/Kaiser_Schulze_oLLM_fig_21.png) ***Figure 21.** CR by channel (log-odds), split by websites’ category complexity* ![](assets/figs/Kaiser_Schulze_oLLM_fig_22.png) ***Figure 22.** AOV by channel, split by websites’ category complexity* ![](assets/figs/Kaiser_Schulze_oLLM_fig_23.png) ***Figure 23.** RPS by channel, split by websites’ category complexity* ## LLM Proficiency We assess whether consumer LLM proficiency moderates channel outcomes using two website-level proxies: the share of “technophiles” among visitors (from Google’s affinity segments) and average visitor age. Websites in the top technophile quartile exhibit 3.8 times greater oLLM traffic share than those in the bottom quartile. Similarly, websites with younger visitors show 5.5 times greater oLLM traffic share, supporting our proficiency proxies. Figure 24 shows that CR gaps between oLLM and traditional channels are substantially smaller when LLM proficiency is high; oLLM CR even exceeds organic search in both analyses. In contrast, Figure 25 indicates that higher proficiency is associated with larger AOV gaps, meaning oLLM generates lower order values than traditional channels. Although the technophile split yields relatively imprecise AOV estimates, the age split produces clearly significant differences. Finally, Figure 26 shows mixed RPS patterns: oLLM gains on paid and organic search among high-technophile websites, and shows improved positioning relative to direct, affiliate, and other among younger-visitor websites (regressions in Online Appendix D). ***Figure 24.** CR by channel (log-odds), split by consumer LLM proficiency proxies* ![](assets/figs/Kaiser_Schulze_oLLM_fig_24_a.png) *(a) Technophile split* ![](assets/figs/Kaiser_Schulze_oLLM_fig_24_b.png) *(b) Age split* ***Figure 25.** AOV by channel, split by consumer LLM proficiency proxies* ![](assets/figs/Kaiser_Schulze_oLLM_fig_25_a.png) *(a) Technophile split* ![](assets/figs/Kaiser_Schulze_oLLM_fig_25_b.png) *(b) Age split* ***Figure 26.** RPS by channel, split by consumer LLM proficiency proxies* ![](assets/figs/Kaiser_Schulze_oLLM_fig_26_a.png) *(a) Technophile split* ![](assets/figs/Kaiser_Schulze_oLLM_fig_26_b.png) *(b) Age split* These patterns mirror Section 5, where rising CR accompanies declining AOV,[^7] yielding only moderate RPS improvements. The parallel cross-sectional and temporal findings indicate that growing consumer LLM proficiency might contribute to oLLM’s development over time. # Key Insights and Limitations Our research examines organic LLM traffic (oLLM) relative to traditional digital channels using 12 months of data from 973 e-commerce websites with over 50,000 oLLM transactions. Two key patterns emerge. **Current positioning.** One year after launch, oLLM exhibits a higher conversion rate (CR) and revenue per session (RPS) than paid social, but lower CR and RPS than all other traditional channels. Engagement metrics show comparatively favorable bounce rates (BR) for oLLM, but fewer page views (PV) and intermediate session duration (SD). Extensive robustness checks and the geographic and category diversity of the dataset support the stability and external relevance of these findings. Online Appendix E discusses why results may differ from industry reports. Gaps between oLLM and traditional channels are more pronounced in low-complexity product categories. In contrast, oLLM shows more favorable relative outcomes and approaches the RPS of organic search for high-complexity categories, which may benefit more from the extensive context, superior synthesis, and adaptive conversational interface LLMs can offer. While exploratory, these findings indicate that oLLM’s comparatively weak financial metrics may reflect limitations in how users currently use and value LLMs as a shopping tool. **Temporal trajectory.** Over the first twelve months, CR and RPS increased, while average order value (AOV) declined. Projections suggest that the CR gap relative to traditional channels will continue to narrow, offset by a widening AOV gap. Thus, RPS is expected to improve only moderately, reaching parity with organic search—the next-ranked channel—only under the most aggressive forecasting scenario. This trade-off—higher conversion but lower order value—is also evident when segmenting websites by visitor LLM proficiency. Websites serving more LLM-proficient consumers display the same combination of elevated CR and reduced AOV. The robustness of this dual pattern across temporal and cross-sectional analyses points to growing consumer LLM proficiency as a likely mechanism behind the observed evolution of oLLM. ## Implications #### For retailers. oLLM currently offers the greatest value for retailers operating in high-complexity categories where consumers demand extensive comparison and guidance. In these settings, oLLM referrals combine comparatively strong CR and RPS with favorable BR, suggesting that they can serve as a high-intent complement to organic and paid search, even if volumes remain small. For retailers focused on low-complexity, routine purchases, current low traffic shares and weaker CR and RPS imply limited short-term gains from channel optimization alone. #### For LLM platforms. Our analyses suggest that lower usefulness in low-complexity product categories aligns with the observed gap between oLLM and traditional channels. LLM platforms have launched several initiatives since the end of our observation period targeting this limitation, including instant checkout and agentic shopping. Early research indicates that agentic shopping tools may prove particularly useful for routine purchases, given a substantially higher share of observed shopping-related queries than with general-purpose LLMs (Yang et al. 2025). Such platform-based agents will compete with retailer-embedded alternatives like Amazon’s Rufus, which benefit from keeping users in familiar on-site interfaces while providing optional conversational support. Further, recent announcements on the introduction of paid advertising formats within ChatGPT introduce strategic considerations that go beyond our organic-traffic data. While oLLM is unpaid and therefore attractive for companies concerned about Return on Advertising Spend (ROAS), paid LLM traffic (pLLM) has the potential to redefine LLMs’ role in the channel ecosystem (Hermann, Puntoni, and Schweidel 2025). Depending on implementation, sponsored placements could reduce friction for simple purchases through one-tap offers or shift the balance between organic guidance and promotional content in ways that negatively affect usefulness and trust and thus also oLLM’s positioning. ## Limitations and future research All analyses in this paper are descriptive rather than causal. As such, the analysis of oLLM metrics is vulnerable to confounding factors, including systematic differences across channels’ users and usage situations. The time-trend analysis presented in Section 5 can only provide an indication of future channel outcomes. LLM platforms, retailer strategies, and consumer behaviors evolve rapidly. While our twelve-month observation period captures the initial maturation phase, the documented patterns may not persist. Disruptive developments, such as major interface redesigns, widely adopted checkout features, or paid advertising rollouts, could fundamentally alter these trajectories. Similarly, the exploratory heterogeneity analyses in Section 6 should only be seen as initial indications of potential mechanisms. The regression framework controls for website, device, and time effects, but the analysis remains descriptive, and the product complexity and proficiency measures are constructed at the website level rather than at the level of individual products or consumers. Understanding whether the observed patterns reflect channel-specific user behavior, LLM-induced behavioral changes, or unobserved website characteristics requires further investigation. Finally, our analyses rely on last-click attribution, which underestimates oLLM’s contribution to the purchase funnel (Berman 2018; Li and Kannan 2014; Li et al. 2016). Future work using multi-touch attribution or experiments on individual websites could clarify the role of oLLM in the upper funnel and its interaction with traditional channels and supplement our broad study with more granular, in-depth insights. Beyond these limitations, our study raises new questions on the emerging role of oLLM in the digital channel mix—including whether treating oLLM as merely another channel understates its unique ability to shape consumer decision processes through its conversational interface. In Online Appendix F, we outline promising directions for future research on how LLMs may reshape digital commerce, channel strategy, and consumer welfare. # Funding, Author Affiliations and Competing Interests **Funding.** This research was supported by Grips Intelligence through the provision of proprietary data and computing resources. No direct financial support was provided to the authors. No additional third-party funding was received. **Author Affiliations and Competing Interests.** The first author is employed by Grips Intelligence, a market intelligence company serving multiple clients, including various technology platforms. Grips Intelligence’s business model depends on delivering accurate, unbiased market intelligence; it has no vested interest in making any particular platform, including Google, OpenAI, or others analyzed in this study, appear more or less favorable. The company’s involvement was limited to: (1) providing access to data, (2) covering computing costs for data processing, and (3) offering technical guidance on data collection and segmentation methodology. Grips Intelligence did not commission the study. The company had no role in formulating the research questions, designing the study, selecting analytical methods, interpreting results, or drafting the manuscript, and it did not approve the manuscript prior to submission. The authors declare no other financial or non-financial competing interests. # Online Appendix A — Additional Data and Variable Definitions ## A.1 Category Volume Details Table A.1 provides a detailed overview of the size and relevance of oLLM across all 24 top-level categories in our dataset. The table shows sessions and revenue for all channels combined, as well as the percentage share attributable to oLLM traffic. **Table A.1. Category Volume Summary Statistics (February–August 2025)** | Category | Sessions (M) | Revenue (\$M) | oLLM % Sessions | oLLM % Revenue | |:--|--:|--:|--:|--:| | E-commerce and Shopping | 1,201.31 | 1,915.33 | 0.05 | 0.03 | | Lifestyle (Fashion, Cosmetics) | 1,127.32 | 2,026.85 | 0.02 | 0.02 | | Business, Cons. Services | 617.84 | 780.24 | 0.10 | 0.10 | | Home and Garden | 531.37 | 1,410.57 | 0.03 | 0.03 | | Consumer Electronics | 465.61 | 758.11 | 0.20 | 0.12 | | Vehicles | 441.12 | 300.18 | 0.03 | 0.05 | | Science and Education | 249.85 | 197.41 | 0.24 | 0.12 | | Arts and Entertainment | 227.42 | 797.40 | 0.07 | 0.14 | | Finance | 227.41 | 223.95 | 0.09 | 0.06 | | Travel and Tourism | 184.67 | 1,279.32 | 0.10 | 0.06 | | Games | 163.22 | 151.66 | 0.05 | 0.02 | | Health | 163.10 | 350.79 | 0.04 | 0.04 | | Sports | 136.05 | 417.42 | 0.02 | 0.02 | | Food and Drink | 104.21 | 328.88 | 0.02 | 0.02 | | Pets and Animals | 66.14 | 209.35 | 0.02 | 0.04 | | Heavy Industry | 41.32 | 108.46 | 0.03 | 0.07 | | Law and Government | 23.01 | 58.53 | 0.14 | 0.05 | | Gambling | 8.50 | 85.01 | >0.00 | >0.00 | | News and Media | 7.84 | 1.60 | 0.12 | 0.24 | | Hobbies and Leisure | 7.44 | 34.66 | 0.06 | 0.07 | | Community and Society | 6.91 | 31.98 | 0.19 | 0.10 | | Adult | 6.39 | 15.96 | 0.01 | 0.01 | | Jobs and Career | 2.09 | 1.07 | 2.06 | 2.22 | | Reference Materials | 0.67 | 0.60 | 0.02 | 0.04 | Note: Sessions and Revenue shown in millions for all channels combined. oLLM Share shows percentage of total volume attributed to oLLM channel. **Table A.2. Product Category Definitions and Conversion Logic** | **Category** | **Description** | **Example of Conversion** | **Example Domains in Category** | |---|---|---|---| | Reference Materials | Information lookup, maps, encyclopedias, dictionaries | • Premium subscription • Ad-free app upgrade | wikipedia.org, dictionary.com, google.com/maps, britannica.com | | News and Media | News content, journalism, current events | • News subscription • Article unlock fee | cnn.com, bbc.com, nytimes.com, theguardian.com, reuters.com | | Adult | Content intended for adult audiences only (18+) | • Monthly fan subscription • Pay-Per-View video | pornhub.com, xvideos.com, xnxx.com, xhamster.com, onlyfans.com | | Sports | Sports news, team info, stats, athletic content | • News subscription • Pay-Per-View match pass | espn.com, cricbuzz.com, marca.com, espncricinfo.com | | Arts and Entertainment | Streaming services, music, visual/performing arts | • Monthly streaming fee • Digital movie rental | youtube.com, netflix.com, spotify.com, twitch.tv, imdb.com | | Community and Society | Social connections, dating, faith, community orgs | • Premium dating features • Virtual awards | facebook.com, reddit.com, match.com, tinder.com, meetup.com | | Food and Drink | Cooking recipes, grocery delivery, restaurants | • Food delivery order • Meal kit box purchase | allrecipes.com, ubereats.com, doordash.com, foodnetwork.com | | Games | Gaming platforms, info, communities, esports | • Game download purchase • Streamer subscription | twitch.tv, roblox.com, steampowered.com, ign.com | | Hobbies and Leisure | Recreational activities, creative pursuits, photography | • Digital pattern download • Stock photo credits | pinterest.com, etsy.com, flickr.com, instructables.com, alltrails.com | | Lifestyle | Fashion retailers, apparel, beauty, personal care | • Apparel purchase • Gift set purchase | shein.com, nike.com, macys.com, zara.com, sephora.com | | Pets and Animals | Pet care, supplies, animal content, communities | • Pet food purchase • Grooming fee | chewy.com, petco.com, petsmart.com, akc.org, petmd.com | | Gambling | Online casinos, sports betting, poker, lottery | • Wager placement • Lottery ticket | bet365.com, draftkings.com, fanduel.com, pokerstars.com | | E-commerce and Shopping | Marketplaces, retailers, price comparison | • Retail goods purchase • Gift card purchase | amazon.com, temu.com, aliexpress.com, ebay.com, walmart.com | | Travel and Tourism | Travel planning, booking platforms, accommodation | • Hotel booking • Flight ticket | booking.com, tripadvisor.com, agoda.com, trip.com, airbnb.com | | Science and Education | Educational resources, academic institutions | • Course certificate fee • Digital textbook | wikipedia.org, coursera.org, khanacademy.org, duolingo.com | | Home and Garden | Home improvement, furniture, gardening, DIY | • Home decor purchase • Furniture purchase | ikea.com, homedepot.com, lowes.com, wayfair.com, harborfreight.com | | Consumer Electronics | Devices, gadgets, smartphones, personal tech | • Smartphone purchase • Accessory purchase | samsung.com, apple.com, bestbuy.com, huawei.com, hp.com | | Health | Healthcare, wellness, medical info, pharma | • Telehealth copay • Home test kit | webmd.com, mayoclinic.org, nih.gov, cvs.com, walgreens.com | | Jobs and Career | Job search, employment websites, HR platforms | • Premium profile fee • Salary report | indeed.com, linkedin.com, glassdoor.com, ziprecruiter.com | | Finance | Banking, investing, insurance, payment platforms | • Insurance premium • Wire/Transfer fee | paypal.com, chase.com, bankofamerica.com, coinbase.com | | Law and Government | Government services, legal resources, portals | • Passport fee • Document copy fee | usa.gov, irs.gov, gov.uk, uscis.gov, whitehouse.gov | | Vehicles | Automotive marketplaces, manufacturers, car info | • Vehicle history report • Accessory purchase | cars.com, carmax.com, autotrader.com, tesla.com, ford.com | | Business and Consumer Services | B2B services, shipping, real estate, professional | • Premium subscription • Transaction fee | indeed.com, linkedin.com, paypal.com, fedex.com, zillow.com | | Heavy Industry and Engineering | Manufacturing, construction, agriculture, energy | • Replacement part • Safety webinar fee | caterpillar.com, ge.com, boeing.com, deere.com, siemens.com | Note: Categories and sample domains according to SimilarWeb. Domains and conversion events are examples only. Categories sorted from “most simple” to “most complex” as rated by three LLMs (Claude, Gemini, ChatGPT). ## A.2 Continent and Country Volume Details ### Continent-Level Coverage Table A.3 summarizes traffic volume across continents. The dataset covers global e-commerce activity, with the majority of traffic stemming from the Americas and Europe. **Table A.3. Continent Volume Summary Statistics (February–August 2025)** | Continent | Sessions (M) | Revenue (\$M) | oLLM % Sessions | oLLM % Revenue | |:--|--:|--:|--:|--:| | Americas | 2,713.87 | 4,206.87 | 0.06 | 0.06 | | Europe | 2,134.33 | 5,202.14 | 0.06 | 0.06 | | Asia | 858.51 | 784.57 | 0.11 | 0.06 | | Oceania | 278.08 | 925.98 | 0.05 | 0.04 | | Africa | 57.70 | 30.12 | 0.25 | 0.09 | Note: oLLM share shows percentage of total volume attributed to ChatGPT Referrals. Local currency is converted with daily average conversion rates to US dollars. ### Country-Level Coverage Our dataset provides broad geographic coverage at the country level. Table A.4 presents detailed statistics for all countries contributing at least 10 million sessions to our dataset. A total of 49 nations meet this threshold, demonstrating the global reach of our data. **Table A.4. Country-Level Summary (6 Months)** | Country | Sessions (M) | Revenue (\$M) | oLLM % Sessions | oLLM % Revenue | |:--|--:|--:|--:|--:| | United States | 1,800.26 | 3,237.66 | 0.04 | 0.06 | | Brazil | 696.81 | 757.23 | 0.07 | 0.04 | | United Kingdom | 457.15 | 1,051.88 | 0.04 | 0.03 | | France | 398.80 | 844.59 | 0.05 | 0.05 | | India | 247.27 | 134.66 | 0.11 | 0.03 | | Australia | 226.82 | 857.65 | 0.06 | 0.04 | | Germany | 207.46 | 742.43 | 0.07 | 0.05 | | Italy | 190.39 | 191.42 | 0.04 | 0.08 | | Netherlands | 181.34 | 669.00 | 0.07 | 0.08 | | Thailand | 121.75 | 114.65 | 0.04 | 0.04 | | Poland | 120.85 | 307.97 | 0.07 | 0.04 | | South Korea | 101.73 | 86.92 | 0.06 | 0.04 | | Canada | 95.65 | 143.60 | 0.12 | 0.08 | | Spain | 81.88 | 164.12 | 0.11 | 0.12 | | Saudi Arabia | 60.34 | 84.04 | 0.05 | 0.02 | | Sweden | 57.37 | 144.53 | 0.04 | 0.06 | | New Zealand | 50.53 | 66.32 | 0.05 | 0.03 | | Japan | 46.96 | 53.87 | 0.14 | 0.15 | | Colombia | 46.24 | 17.39 | 0.07 | 0.04 | | Greece | 40.32 | 46.37 | 0.05 | 0.07 | | Belgium | 40.14 | 105.78 | 0.08 | 0.15 | | Vietnam | 39.63 | 16.67 | 0.15 | 0.07 | | Turkey | 37.09 | 28.59 | 0.11 | 0.08 | | Finland | 35.58 | 45.65 | 0.03 | 0.04 | | Israel | 30.16 | 116.38 | 0.08 | 0.04 | | Switzerland | 29.53 | 98.04 | 0.09 | 0.08 | | Mexico | 27.24 | 19.48 | 0.16 | 0.17 | | Portugal | 27.11 | 109.27 | 0.11 | 0.05 | | Ireland | 27.10 | 37.65 | 0.06 | 0.08 | | Russia | 25.72 | 31.86 | 0.03 | 0.01 | | United Arab Emirates | 24.98 | 36.47 | 0.18 | 0.06 | | Denmark | 24.92 | 98.21 | 0.06 | 0.06 | | Romania | 23.72 | 47.49 | 0.11 | 0.13 | | Austria | 22.62 | 108.63 | 0.07 | 0.07 | | China | 21.37 | 21.54 | 0.05 | 0.01 | | Malaysia | 20.98 | 4.53 | 0.17 | 0.07 | | Ukraine | 20.44 | 17.79 | 0.13 | 0.21 | | Hungary | 20.32 | 69.02 | 0.06 | 0.06 | | Philippines | 17.40 | 11.20 | 0.25 | 0.35 | | Norway | 16.75 | 33.29 | 0.06 | 0.10 | | Indonesia | 16.58 | 27.13 | 0.28 | 0.02 | | Czechia | 15.88 | 40.24 | 0.10 | 0.13 | | Slovakia | 15.73 | 35.72 | 0.05 | 0.08 | | Egypt | 14.02 | 5.86 | 0.21 | 0.05 | | South Africa | 13.04 | 6.61 | 0.23 | 0.13 | | Pakistan | 12.94 | 4.00 | 0.32 | 0.27 | | Bulgaria | 12.32 | 16.48 | 0.08 | 0.07 | | Croatia | 11.78 | 96.77 | 0.06 | 0.03 | | Chile | 10.17 | 4.10 | 0.20 | 0.13 | | Taiwan | 8.49 | 8.07 | 0.18 | 0.15 | Note: Omitted 141 countries with lower coverage. ## A.3 Complete Variable Definitions Table A.5 provides comprehensive definitions for all dependent and independent variables used in our regression models. This includes focal financial metrics (CR, AOV, RPS), secondary engagement metrics (BR, SD, PV), and control variables. **Table A.5. Variable Definitions** | **Variable Type** | **Variable Name** | **Definition** | |---|---|---| | Dependent Variables: Focal | Conversion Rate (CR) | Most widely used financial metric in e-commerce, defined as $(\text{Number of Transactions} / \text{Number of Sessions})$ per time unit, if $(\text{Number of Sessions} > 0)$ | | | Average Order Value (AOV) | Captures size of the typical shopping basket in USD, defined as $(\text{Total Revenue} / \text{Number of Transactions})$ per time unit, if $(\text{Number of Transactions} > 0)$ | | | Revenue Per Session (RPS) | Combines CR and AOV into one metric, defined as $(\text{Total Revenue} / \text{Number of Sessions})$ per time unit, if $(\text{Number of Sessions} > 0)$ | | Dependent Variables: Secondary | Bounce Rate (BR) | Focal engagement metric, measures share of visitors that leave target website without further action, capturing content relevance. | | | Session Duration (SD) | Total time spent on websites from entry to exit, measured in seconds. Reflects engagement depth and deliberation time in purchase decisions. | | | Page Views (PV) | Measures number of distinct pages viewed during website sessions, indicates relevance of the destination website. | | Independent Variable: Focal | Channel | Categorical variable identifying visitor source, with oLLM (i.e., organic referral traffic from ChatGPT) serving as baseline comparison category in regression analysis. Other channels include Affiliate, Direct, Email, Organic Search, Paid Search, Paid Social, Referral, and Other. | | Independent Variables: Controls | Website | E-commerce website identifier (website fixed effect) capturing shop-specific characteristics, as well as industry effects. | | | Device | Device used by website visitors, accounting for device-specific conversion and engagement patterns. | | | Month | Month fixed effects capturing seasonal patterns and temporal trends. | ## A.4 Complete Technical Model Specifications This section provides complete technical specifications for all regression models used in our analysis, including notation, data structure, and detailed model equations. ### Notation and Data Structure Reflecting the data structure, we index websites, weeks, channels, device types, and months as shown in Table A.6. | Symbol | Description | |:----------------------|:-------------------------------------| | $i \in \{1,\dots,I\}$ | Website | | $t \in \{1,\dots,T\}$ | Week | | $c(i,t)$ | Channel (e.g., oLLM, organic search) | | $d(i,t)$ | Device (desktop or mobile) | | $m(t)$ | Calendar month of observation | **Table A.6. Indices** Table A.7 summarizes notation for derived metrics and outcome variables. | Symbol | Description | |:-------------------------------------------------------------|:--------------------------------------------------------------------| | $n^{\text{sess}}_{it}$ | Number of sessions for $i,t$ | | $y^{\text{txn}}_{it}$ | Number of transactions for $i,t$ | | $y^{\text{rev}}_{it}$ | Total revenue for $i,t$ | | $y^{\text{bounce}}_{it}$ | Number of bounces for $i,t$ | | $y^{\text{SD}}_{it}$ | Total session duration for $i,t$ (seconds) | | $y^{\text{PV}}_{it}$ | Total page views for $i,t$ | | $\text{CR}_{it}= y^\text{txn}_{it} / n^\text{sess}_{it}$ | Conversion rate for $i,t$ | | $\text{AOV}_{it}= y^\text{rev}_{it} / y^\text{txn}_{it}$ | Average order value for $i,t$ (defined for $y^{\text{txn}}_{it}>0$) | | $\text{RPS}_{it}= y^\text{rev}_{it} / n^\text{sess}_{it}$ | Revenue per session for $i,t$ (includes zeros) | | $\text{BR}_{it}= y^{\text{bounce}}_{it}/ n^\text{sess}_{it}$ | Bounce rate for $i,t$ | | $\text{SD}_{it}= y^{\text{SD}}_{it} / n^\text{sess}_{it}$ | Average session duration for $i,t$ | | $\text{PV}_{it}= y^{\text{PV}}_{it} / n^\text{sess}_{it}$ | Average pageviews per session for $i,t$ | **Table A.7. Outcomes** Note: Outcome variables are only defined for $n^{\text{sess}}_{it} > 0$. We include several fixed effects, with notation summarized in Table A.8. The list contains a time-trend variable (equal to the count of months since the start of the observation period) that is used in Section 5. | Symbol | Description | |:------------|:-------------------------------------------------------------------------------| | $\alpha_c$ | Channel fixed effects | | $\gamma_i$ | Website fixed effects | | $\delta_d$ | Device fixed effects | | $\mu_m$ | Month fixed effects | | $\beta^X_0$ | Intercept for outcome $X \in \{\text{AOV}, \text{RPS}, \text{SD}, \text{PV}\}$ | | $trend_t$ | Linear monthly time trend | **Table A.8. Fixed Effects and Parameters** ### Model Specification for CR and BR By definition, $\text{CR}_{it}, \text{BR}_{it} \in [0,1]$ for all $i,t$, as they measure the probability of success (a transaction or a bounce) for a given number of trials (sessions). Both metrics exhibit overdispersion in our data. For CR, the variance of the transaction counts, $\mathrm{Var}(y^{\text{txn}}_{it})$, exceeds that implied by a standard binomial model. Similarly, BR shows variance in bounce counts, $\mathrm{Var}(y^{\text{bounce}}_{it})$, that cannot be adequately captured by standard binomial assumptions. We therefore employ quasibinomial specifications with dispersion parameters $\theta^{\text{CR}}$ and $\theta^{\text{BR}}$ for both metrics: $$\begin{aligned} y^{\text{txn}}_{it} \mid n^{\text{sess}}_{it}, p^{\text{CR}}_{it}, \theta^{\text{CR}} &\sim \text{QuasiBinomial}\!\left(n^{\text{sess}}_{it},\, p^{\text{CR}}_{it}, \theta^{\text{CR}}\right), \\ \text{logit}\!\left(p^{\text{CR}}_{it}\right) &= \alpha_{c(i,t)}^{CR} + \gamma_i^{CR} + \delta_{d(i,t)}^{CR} + \mu_{m(t)}^{CR}, \\ \mathrm{Var}\!\left(y^{\text{txn}}_{it}\right) &= \theta^{\text{CR}} \cdot n^{\text{sess}}_{it} \cdot p^{\text{CR}}_{it} \cdot \left(1 - p^{\text{CR}}_{it}\right), \\ y^{\text{bounce}}_{it} \mid n^{\text{sess}}_{it}, p^{\text{BR}}_{it}, \theta^{\text{BR}} &\sim \text{QuasiBinomial}\!\left(n^{\text{sess}}_{it},\, p^{\text{BR}}_{it}, \theta^{\text{BR}}\right), \\ \text{logit}\!\left(p^{\text{BR}}_{it}\right) &= \alpha_{c(i,t)}^{BR} + \gamma_i^{BR} + \delta_{d(i,t)}^{BR} + \mu_{m(t)}^{BR}, \\ \mathrm{Var}\!\left(y^{\text{bounce}}_{it}\right) &= \theta^{\text{BR}} \cdot n^{\text{sess}}_{it} \cdot p^{\text{BR}}_{it} \cdot \left(1 - p^{\text{BR}}_{it}\right). \end{aligned}$$ ### Model Specification for AOV, RPS, SD, and PV The remaining outcome metrics, $\text{AOV}_{it}, \text{RPS}_{it}, \text{SD}_{it}, \text{PV}_{it}$, are modeled with standard linear specifications. We begin with the monetary variables: $$\begin{aligned} \text{AOV}_{it} &= \beta^{\text{AOV}}_{0} + \alpha^{\text{AOV}}_{c(i,t)} + \gamma^{\text{AOV}}_{i} + \delta^{\text{AOV}}_{d(i,t)} + \mu^{\text{AOV}}_{m(t)} + \varepsilon^{\text{AOV}}_{it}, \\ \varepsilon^{\text{AOV}}_{it} &\sim \mathcal{N}(0,\sigma^2_{\text{AOV}}), \\ \text{RPS}_{it} &= \beta^{\text{RPS}}_{0} + \alpha^{\text{RPS}}_{c(i,t)} + \gamma^{\text{RPS}}_{i} + \delta^{\text{RPS}}_{d(i,t)} + \mu^{\text{RPS}}_{m(t)} + \varepsilon^{\text{RPS}}_{it}, \\ \varepsilon^{\text{RPS}}_{it} &\sim \mathcal{N}(0,\sigma^2_{\text{RPS}}). \end{aligned}$$ By definition of $\text{AOV}_{it}$, we exclude observations with zero transactions. In contrast, $\text{RPS}_{it}$ includes realized zeros, as observations with zero revenue are informative about channel quality. For the engagement metrics, we specify SD and PV per session analogously: $$\begin{aligned} \text{SD}_{it} &= \beta^{\text{SD}}_{0} + \alpha^{\text{SD}}_{c(i,t)} + \gamma^{\text{SD}}_{i} + \delta^{\text{SD}}_{d(i,t)} + \mu^{\text{SD}}_{m(t)} + \varepsilon^{\text{SD}}_{it}, \\ \varepsilon^{\text{SD}}_{it} &\sim \mathcal{N}(0,\sigma^2_{\text{SD}}), \\ \text{PV}_{it} &= \beta^{\text{PV}}_{0} + \alpha^{\text{PV}}_{c(i,t)} + \gamma^{\text{PV}}_{i} + \delta^{\text{PV}}_{d(i,t)} + \mu^{\text{PV}}_{m(t)} + \varepsilon^{\text{PV}}_{it}, \\ \varepsilon^{\text{PV}}_{it} &\sim \mathcal{N}(0,\sigma^2_{\text{PV}}). \end{aligned}$$ ### Weighting Scheme All linear models are estimated with weights to handle heteroskedasticity from aggregation, as averages based on more sessions (or transactions in the case of AOV) are systematically less noisy. Unweighted OLS would give too much weight to small, volatile cells, which would particularly impact channels with sparse observations, such as oLLM. For website $i$ and week $t$, let $s_{it}$ be sessions, $\bar s_i=\frac{1}{T_i}\sum_t s_{it}$, $\sigma_{s,i}=\sqrt{\frac{1}{T_i-1}\sum_t (s_{it}-\bar s_i)^2}$, $z_{it}=(s_{it}-\bar s_i)/\sigma_{s,i}$, and define a bounded, site-relative weight $$w_{it}=\frac{1}{1+\exp(-z_{it})}\in(0,1).$$ This monotone proxy for inverse variance (i) favors more precise cells, (ii) caps leverage so that no website dominates, and (iii) avoids raw-size bias from very large retailers. A simpler alternative would be to weight by $$w_{it}=\sqrt{sessions_{it}}.$$ This squareroot scheme concentrates weight on the largest websites and their most-visited channels (Figure A.1): the largest 12.5% of websites carry approximately one-third of total weight for each channel, while the largest 50% of websites carry over three-quarter. In contrast, the sigmoid approach balances weight more evenly across websites while preserving within-website variation that reflects session counts. ![](assets/figs/Kaiser_Schulze_oLLM_fig_A_1.png) ***Figure A.1.** Cumulative Weight Comparison* We *do not* weight the quasibinomial CR and BR models because $n_{it}^{\text{sess}}$ already enters the likelihood and additional weights would therefore double-count precision. # Online Appendix B — Detailed Robustness Check Results This section presents detailed regression results on all robustness checks, covering data processing choices, observed websites, observed LLM platforms, and observed timeframes. It further contains result graphs on financial metrics AOV and RPS, as well as on engagement metrics BR, PV, and SD. ## Robustness Regressions: Data Processing Choices We vary the aggregation scheme by testing daily and monthly data as alternatives to weekly aggregation. Daily aggregation increases granularity but exacerbates sparsity for oLLM. Monthly aggregation reduces sparsity and improves estimates of ratio metrics, at the cost of fewer observations and potentially wider confidence intervals. We then apply filters requiring minimum observations of 10, 100, or 1000 per week/website/device/channel combination. These thresholds increase data robustness but disproportionately reduce oLLM observations, potentially biasing the sample toward larger websites. Finally, we winsorize dependent variables at the 2.5% and 97.5% levels to reduce outlier influence. **Table B.1. Regression Results: Daily Aggregated Data** | | CR | AOV | RPS | BR | SD | PV | |:--|--:|--:|--:|--:|--:|--:| | Affiliate | 0.449\*\*\* (0.025) | 22.9\*\*\* (4.3) | 3.62\*\*\* (0.06) | 0.267\*\*\* (0.014) | -15.9\*\*\* (3.6) | 0.674\*\*\* (0.038) | | Paid Search | 0.401\*\*\* (0.025) | 3.59 (3.98) | 2.13\*\*\* (0.05) | -0.168\*\*\* (0.013) | 18.1\*\*\* (2.7) | 0.937\*\*\* (0.029) | | Other | 0.384\*\*\* (0.025) | 8.98\*\* (4.01) | 2.55\*\*\* (0.05) | 0.892\*\*\* (0.013) | 39.2\*\*\* (2.8) | 0.794\*\*\* (0.029) | | Direct | 0.345\*\*\* (0.025) | 12.8\*\*\* (4.0) | 3.25\*\*\* (0.05) | 0.352\*\*\* (0.013) | -3.77 (2.74) | 0.733\*\*\* (0.029) | | Email | 0.300\*\*\* (0.025) | 10.6\*\*\* (4.1) | 3.05\*\*\* (0.05) | 0.144\*\*\* (0.013) | 63.8\*\*\* (2.9) | 0.971\*\*\* (0.031) | | Referral | 0.258\*\*\* (0.025) | 13.2\*\*\* (4.0) | 1.77\*\*\* (0.05) | 0.208\*\*\* (0.013) | 34.5\*\*\* (2.8) | 0.759\*\*\* (0.030) | | Organic Search | 0.156\*\*\* (0.025) | 2.02 (3.98) | 1.21\*\*\* (0.04) | -0.126\*\*\* (0.013) | -3.46 (2.72) | 0.561\*\*\* (0.029) | | Paid Social | -0.750\*\*\* (0.025) | 4.71 (4.18) | -0.371\*\*\* (0.051) | 0.278\*\*\* (0.013) | -130\*\*\* (3) | -0.346\*\*\* (0.032) | | Fixed Effects: Website, Device, Month | Yes | Yes | Yes | Yes | Yes | Yes | | $R^2$ \[pseudo-$R^2$\] | \[0.689\] | 0.170 | 0.149 | \[0.639\] | 0.262 | 0.052 | | Observations | 3,896,527 | 2,953,982 | 3,896,527 | 3,896,527 | 3,896,527 | 3,896,527 | | | \* p<0.1; \*\* p<0.05; \*\*\* p<0.01 | | | | | | Notes: All coefficients relative to oLLM baseline. Standard errors in parentheses. **Table B.2. Regression Results: Monthly Aggregated Data** | | CR | AOV | RPS | BR | SD | PV | |:--|--:|--:|--:|--:|--:|--:| | Affiliate | 0.449\*\*\* (0.109) | 18.4\*\*\* (3.7) | 3.20\*\*\* (0.13) | 0.267\*\*\* (0.064) | -1.49 (9.87) | 0.627\*\*\* (0.045) | | Paid Search | 0.401\*\*\* (0.108) | 1.00 (2.88) | 2.01\*\*\* (0.09) | -0.168\*\*\* (0.062) | 45.2\*\*\* (6.5) | 0.951\*\*\* (0.030) | | Other | 0.384\*\*\* (0.108) | 3.99 (2.93) | 2.36\*\*\* (0.09) | 0.892\*\*\* (0.062) | 66.3\*\*\* (6.8) | 0.721\*\*\* (0.031) | | Direct | 0.345\*\*\* (0.108) | 10.1\*\*\* (2.9) | 2.93\*\*\* (0.09) | 0.352\*\*\* (0.062) | 14.6\*\* (6.6) | 0.693\*\*\* (0.030) | | Email | 0.300\*\*\* (0.108) | 8.45\*\*\* (3.06) | 2.46\*\*\* (0.10) | 0.144\*\* (0.062) | 72.5\*\*\* (7.3) | 0.931\*\*\* (0.033) | | Referral | 0.258\*\* (0.108) | 10.8\*\*\* (3.0) | 1.63\*\*\* (0.09) | 0.208\*\*\* (0.062) | 66.3\*\*\* (7.0) | 0.787\*\*\* (0.032) | | Organic Search | 0.156 (0.108) | -1.71 (2.86) | 1.07\*\*\* (0.09) | -0.126\*\* (0.062) | 27.0\*\*\* (6.5) | 0.599\*\*\* (0.030) | | Paid Social | -0.750\*\*\* (0.109) | 2.48 (3.26) | -0.666\*\*\* (0.104) | 0.278\*\*\* (0.062) | -117\*\*\* (8) | -0.381\*\*\* (0.035) | | Fixed Effects: Website, Device, Month | Yes | Yes | Yes | Yes | Yes | Yes | | $R^2$ \[pseudo-$R^2$\] | \[0.750\] | 0.645 | 0.366 | \[0.681\] | 0.432 | 0.384 | | Observations | 149,628 | 132,458 | 149,628 | 149,628 | 149,628 | 149,628 | | | \* p<0.1; \*\* p<0.05; \*\*\* p<0.01 | | | | | | Notes: All coefficients relative to oLLM baseline. Standard errors in parentheses. **Table B.3. Regression Results: 6 Months – Min 10 Sessions** | | CR | AOV | RPS | BR | SD | PV | |:--|--:|--:|--:|--:|--:|--:| | Affiliate | 0.612\*\*\* (0.062) | 24.1\*\*\* (4.0) | 3.61\*\*\* (0.11) | 0.225\*\*\* (0.036) | -36.1\*\*\* (7.5) | 0.528\*\*\* (0.060) | | Paid Search | 0.361\*\*\* (0.061) | 3.41 (3.28) | 1.94\*\*\* (0.08) | -0.148\*\*\* (0.034) | 0.465 (5.432) | 0.763\*\*\* (0.043) | | Other | 0.357\*\*\* (0.061) | 4.88 (3.33) | 2.20\*\*\* (0.08) | 0.891\*\*\* (0.034) | 14.2\*\* (5.6) | 0.454\*\*\* (0.044) | | Direct | 0.273\*\*\* (0.061) | 12.8\*\*\* (3.3) | 3.05\*\*\* (0.08) | 0.329\*\*\* (0.034) | -25.4\*\*\* (5.4) | 0.546\*\*\* (0.043) | | Email | 0.267\*\*\* (0.061) | 9.63\*\*\* (3.46) | 2.59\*\*\* (0.09) | 0.189\*\*\* (0.034) | 28.3\*\*\* (5.9) | 0.739\*\*\* (0.047) | | Referral | 0.209\*\*\* (0.061) | 12.0\*\*\* (3.4) | 1.62\*\*\* (0.08) | 0.210\*\*\* (0.034) | 30.0\*\*\* (5.7) | 0.639\*\*\* (0.045) | | Organic Search | 0.112\* (0.061) | 1.14 (3.27) | 0.999\*\*\* (0.079) | -0.128\*\*\* (0.034) | -16.5\*\*\* (5.4) | 0.429\*\*\* (0.043) | | Paid Social | -0.770\*\*\* (0.062) | 4.97 (3.66) | -0.631\*\*\* (0.092) | 0.368\*\*\* (0.034) | -158\*\*\* (6) | -0.548\*\*\* (0.049) | | Fixed Effects: Website, Device, Month | Yes | Yes | Yes | Yes | Yes | Yes | | $R^2$ \[pseudo-$R^2$\] | \[0.739\] | 0.472 | 0.362 | \[0.711\] | 0.481 | 0.182 | | Observations | 320,021 | 287,409 | 320,021 | 320,021 | 320,021 | 320,021 | | | \* p<0.1; \*\* p<0.05; \*\*\* p<0.01 | | | | | | Notes: All coefficients relative to oLLM baseline. Standard errors in parentheses. **Table B.4. Regression Results: 6 Months – Min 100 Sessions** | | CR | AOV | RPS | BR | SD | PV | |:--|--:|--:|--:|--:|--:|--:| | Affiliate | 0.539\*\*\* (0.074) | 25.6\*\*\* (5.4) | 3.20\*\*\* (0.15) | 0.252\*\*\* (0.041) | -47.3\*\*\* (10.9) | 0.401\*\*\* (0.045) | | Paid Search | 0.288\*\*\* (0.073) | 4.20 (4.81) | 1.50\*\*\* (0.13) | -0.121\*\*\* (0.039) | -12.4 (9.4) | 0.659\*\*\* (0.039) | | Other | 0.284\*\*\* (0.073) | 6.29 (4.84) | 1.68\*\*\* (0.13) | 0.918\*\*\* (0.039) | -1.20 (9.47) | 0.172\*\*\* (0.039) | | Direct | 0.200\*\*\* (0.073) | 13.9\*\*\* (4.8) | 2.63\*\*\* (0.13) | 0.356\*\*\* (0.039) | -40.2\*\*\* (9.4) | 0.412\*\*\* (0.039) | | Email | 0.193\*\*\* (0.073) | 9.81\*\* (4.95) | 1.93\*\*\* (0.14) | 0.216\*\*\* (0.039) | 12.0 (9.8) | 0.579\*\*\* (0.040) | | Referral | 0.137\* (0.073) | 13.0\*\*\* (4.9) | 1.27\*\*\* (0.13) | 0.237\*\*\* (0.039) | 6.10 (9.54) | 0.492\*\*\* (0.039) | | Organic Search | 0.039 (0.073) | 2.15 (4.79) | 0.569\*\*\* (0.132) | -0.101\*\*\* (0.039) | -30.9\*\*\* (9.3) | 0.317\*\*\* (0.038) | | Paid Social | -0.843\*\*\* (0.074) | 6.38 (5.10) | -0.944\*\*\* (0.141) | 0.395\*\*\* (0.039) | -165\*\*\* (10) | -0.683\*\*\* (0.041) | | Fixed Effects: Website, Device, Month | Yes | Yes | Yes | Yes | Yes | Yes | | $R^2$ \[pseudo-$R^2$\] | \[0.740\] | 0.486 | 0.410 | \[0.712\] | 0.513 | 0.489 | | Observations | 280,190 | 267,602 | 280,190 | 280,190 | 280,190 | 280,190 | | | \* p<0.1; \*\* p<0.05; \*\*\* p<0.01 | | | | | | Notes: All coefficients relative to oLLM baseline. Standard errors in parentheses. **Table B.5. Regression Results: 6 Months – Min 1000 Sessions** | | CR | AOV | RPS | BR | SD | PV | |:--|--:|--:|--:|--:|--:|--:| | Affiliate | 0.447\*\*\* (0.119) | 39.2\*\*\* (11.9) | 2.65\*\*\* (0.31) | 0.293\*\*\* (0.059) | -21.1 (16.2) | 0.481\*\*\* (0.076) | | Paid Search | 0.196\* (0.118) | 22.0\* (11.3) | 1.43\*\*\* (0.29) | -0.086 (0.057) | 10.3 (15.2) | 0.915\*\*\* (0.071) | | Other | 0.190 (0.118) | 24.8\*\* (11.3) | 1.35\*\*\* (0.29) | 0.955\*\*\* (0.057) | -3.05 (15.30) | 0.193\*\*\* (0.072) | | Direct | 0.101 (0.118) | 29.5\*\*\* (11.3) | 2.25\*\*\* (0.29) | 0.392\*\*\* (0.057) | -21.8 (15.2) | 0.547\*\*\* (0.071) | | Email | 0.088 (0.119) | 24.1\*\* (11.4) | 1.40\*\*\* (0.30) | 0.255\*\*\* (0.058) | 28.2\* (15.5) | 0.639\*\*\* (0.072) | | Referral | 0.045 (0.119) | 28.5\*\* (11.3) | 1.13\*\*\* (0.30) | 0.278\*\*\* (0.058) | 13.2 (15.4) | 0.580\*\*\* (0.072) | | Organic Search | -0.053 (0.118) | 19.2\* (11.2) | 0.553\* (0.292) | -0.066 (0.057) | -4.59 (15.22) | 0.586\*\*\* (0.071) | | Paid Social | -0.937\*\*\* (0.119) | 27.5\*\* (11.5) | -0.799\*\*\* (0.299) | 0.431\*\*\* (0.058) | -141\*\*\* (16) | -0.550\*\*\* (0.073) | | Fixed Effects: Website, Device, Month | Yes | Yes | Yes | Yes | Yes | Yes | | $R^2$ \[pseudo-$R^2$\] | \[0.742\] | 0.491 | 0.479 | \[0.713\] | 0.705 | 0.676 | | Observations | 188,376 | 185,057 | 188,376 | 188,376 | 188,376 | 188,376 | | | \* p<0.1; \*\* p<0.05; \*\*\* p<0.01 | | | | | | Notes: All coefficients relative to oLLM baseline. Standard errors in parentheses. **Table B.6. Regression Results: Winsorized Data** | | CR | AOV | RPS | BR | SD | PV | |:--|--:|--:|--:|--:|--:|--:| | Affiliate | 0.636\*\*\* (0.058) | 29.9\*\*\* (1.4) | 3.27\*\*\* (0.05) | 0.208\*\*\* (0.032) | 9.91\*\*\* (2.32) | 0.673\*\*\* (0.016) | | Paid Search | 0.382\*\*\* (0.057) | 18.5\*\*\* (1.1) | 2.35\*\*\* (0.03) | -0.161\*\*\* (0.031) | 55.2\*\*\* (1.5) | 0.913\*\*\* (0.010) | | Other | 0.362\*\*\* (0.057) | 19.7\*\*\* (1.1) | 2.23\*\*\* (0.03) | 0.829\*\*\* (0.031) | 27.9\*\*\* (1.6) | 0.388\*\*\* (0.011) | | Direct | 0.278\*\*\* (0.057) | 28.7\*\*\* (1.1) | 3.14\*\*\* (0.03) | 0.306\*\*\* (0.031) | 26.4\*\*\* (1.6) | 0.674\*\*\* (0.010) | | Email | 0.286\*\*\* (0.057) | 19.3\*\*\* (1.2) | 2.51\*\*\* (0.03) | 0.174\*\*\* (0.031) | 43.7\*\*\* (1.7) | 0.885\*\*\* (0.012) | | Referral | 0.173\*\*\* (0.057) | 24.4\*\*\* (1.2) | 1.84\*\*\* (0.03) | 0.191\*\*\* (0.031) | 56.7\*\*\* (1.6) | 0.764\*\*\* (0.011) | | Organic Search | 0.130\*\* (0.057) | 16.6\*\*\* (1.1) | 1.56\*\*\* (0.03) | -0.138\*\*\* (0.031) | 39.5\*\*\* (1.5) | 0.617\*\*\* (0.010) | | Paid Social | -0.757\*\*\* (0.057) | 12.3\*\*\* (1.3) | -0.218\*\*\* (0.036) | 0.352\*\*\* (0.031) | -107\*\*\* (2) | -0.387\*\*\* (0.012) | | Fixed Effects: Website, Device, Month | Yes | Yes | Yes | Yes | Yes | Yes | | $R^2$ \[pseudo-$R^2$\] | \[0.727\] | 0.788 | 0.493 | \[0.729\] | 0.739 | 0.625 | | Observations | 340,292 | 289,524 | 340,292 | 340,292 | 340,292 | 340,292 | | | \* p<0.1; \*\* p<0.05; \*\*\* p<0.01 | | | | | | Notes: All coefficients relative to oLLM baseline. Standard errors in parentheses. ## Robustness Regressions: Observed Websites We restrict the sample to two subsets: the top 25% of websites by oLLM sessions, testing whether results are driven by sites with low oLLM traffic, and the top 25% by transactions, assessing whether larger websites exhibit systematically different oLLM outcomes. The 40% overlap between subsets indicates that current oLLM traffic may disproportionately favor smaller websites, despite LLMs reportedly favoring short tail sources in other settings (Padilla et al. 2025; Gholami et al. 2026). **Table B.7. Regression Results: Top 25% oLLM** | | CR | AOV | RPS | BR | SD | PV | |:--|--:|--:|--:|--:|--:|--:| | Affiliate | 0.704\*\*\* (0.101) | 39.8\*\*\* (4.1) | 3.10\*\*\* (0.12) | 0.145\*\* (0.061) | -18.5 (11.6) | 0.447\*\*\* (0.040) | | Paid Search | 0.322\*\*\* (0.098) | 7.72\*\* (3.07) | 1.38\*\*\* (0.08) | -0.143\*\* (0.057) | 32.6\*\*\* (7.8) | 0.734\*\*\* (0.027) | | Other | 0.439\*\*\* (0.098) | 8.10\*\*\* (3.13) | 1.47\*\*\* (0.09) | 0.935\*\*\* (0.057) | 35.2\*\*\* (8.0) | 0.209\*\*\* (0.027) | | Direct | 0.237\*\* (0.098) | 15.9\*\*\* (3.0) | 1.67\*\*\* (0.08) | 0.326\*\*\* (0.056) | -2.82 (7.69) | 0.361\*\*\* (0.026) | | Email | 0.202\*\* (0.099) | 12.2\*\*\* (3.4) | 1.66\*\*\* (0.09) | 0.238\*\*\* (0.057) | 32.0\*\*\* (8.7) | 0.434\*\*\* (0.030) | | Referral | 0.174\* (0.099) | 8.28\*\*\* (3.19) | 1.09\*\*\* (0.09) | 0.210\*\*\* (0.057) | 53.7\*\*\* (8.3) | 0.441\*\*\* (0.028) | | Organic Search | 0.102 (0.098) | 2.41 (3.01) | 0.607\*\*\* (0.080) | -0.121\*\* (0.056) | 19.3\*\* (7.5) | 0.481\*\*\* (0.026) | | Paid Social | -1.05\*\*\* (0.10) | 13.9\*\*\* (3.7) | -0.391\*\*\* (0.099) | 0.261\*\*\* (0.057) | -148\*\*\* (9) | -0.816\*\*\* (0.032) | | Fixed Effects: Website, Device, Month | Yes | Yes | Yes | Yes | Yes | Yes | | $R^2$ \[pseudo-$R^2$\] | \[0.700\] | 0.719 | 0.342 | \[0.724\] | 0.510 | 0.604 | | Observations | 91,385 | 76,686 | 91,385 | 91,385 | 91,385 | 91,385 | | | \* p<0.1; \*\* p<0.05; \*\*\* p<0.01 | | | | | | Notes: All coefficients relative to oLLM baseline. Standard errors in parentheses. **Table B.8. Regression Results: Top 25% Revenue** | | CR | AOV | RPS | BR | SD | PV | |:--|--:|--:|--:|--:|--:|--:| | Affiliate | 0.865\*\*\* (0.150) | 29.8\*\*\* (9.6) | 6.35\*\*\* (0.28) | 0.203\*\* (0.086) | 8.01 (16.27) | 0.724\*\*\* (0.035) | | Paid Search | 0.632\*\*\* (0.149) | 1.16 (7.96) | 3.37\*\*\* (0.20) | -0.159\* (0.083) | 29.9\*\* (12.0) | 1.03\*\*\* (0.03) | | Other | 0.685\*\*\* (0.149) | 7.21 (8.10) | 3.71\*\*\* (0.21) | 0.893\*\*\* (0.083) | 95.6\*\*\* (12.4) | 0.260\*\*\* (0.027) | | Direct | 0.535\*\*\* (0.149) | 9.68 (7.99) | 4.81\*\*\* (0.21) | 0.356\*\*\* (0.083) | -43.8\*\*\* (12.1) | 0.562\*\*\* (0.026) | | Email | 0.494\*\*\* (0.149) | 5.14 (8.37) | 4.13\*\*\* (0.22) | 0.224\*\*\* (0.084) | 25.3\* (13.1) | 0.652\*\*\* (0.028) | | Referral | 0.503\*\*\* (0.149) | 9.05 (8.24) | 2.67\*\*\* (0.22) | 0.221\*\*\* (0.084) | 22.3\* (12.9) | 0.619\*\*\* (0.028) | | Organic Search | 0.422\*\*\* (0.149) | 7.28 (7.96) | 2.33\*\*\* (0.20) | -0.143\* (0.083) | 25.6\*\* (12.0) | 0.897\*\*\* (0.026) | | Paid Social | -0.505\*\*\* (0.150) | 5.56 (8.80) | -0.989\*\*\* (0.231) | 0.338\*\*\* (0.084) | -151\*\*\* (14) | -0.647\*\*\* (0.029) | | Fixed Effects: Website, Device, Month | Yes | Yes | Yes | Yes | Yes | Yes | | $R^2$ \[pseudo-$R^2$\] | \[0.738\] | 0.423 | 0.338 | \[0.727\] | 0.386 | 0.619 | | Observations | 91,942 | 81,983 | 91,942 | 91,942 | 91,942 | 91,942 | | | \* p<0.1; \*\* p<0.05; \*\*\* p<0.01 | | | | | | Notes: All coefficients relative to oLLM baseline. Standard errors in parentheses. ## Robustness Regressions: Observed LLM Platforms We explore two alternative definitions of oLLM. First, we expand oLLM to include not only ChatGPT, but also Perplexity, Gemini, Copilot, Deepseek, and Grok. Second, we expand oLLM to include not only mobile and desktop website traffic, but also mobile *app* traffic from ChatGPT. **Table B.9. Regression Results: 6 Months Data – All LLMs** | | CR | AOV | RPS | BR | SD | PV | |:--|--:|--:|--:|--:|--:|--:| | Affiliate | 0.647\*\*\* (0.055) | 26.0\*\*\* (4.1) | 4.10\*\*\* (0.10) | 0.202\*\*\* (0.001) | 2.91 (6.22) | 0.855\*\*\* (0.054) | | Paid Search | 0.392\*\*\* (0.054) | 4.50 (3.30) | 2.31\*\*\* (0.06) | -0.178\*\*\* (0.001) | 52.6\*\*\* (3.8) | 1.12\*\*\* (0.03) | | Other | 0.389\*\*\* (0.054) | 6.31\* (3.35) | 2.51\*\*\* (0.06) | 0.857\*\*\* (0.001) | 63.9\*\*\* (4.0) | 0.814\*\*\* (0.035) | | Direct | 0.312\*\*\* (0.054) | 16.1\*\*\* (3.3) | 3.45\*\*\* (0.06) | 0.305\*\*\* (0.001) | 30.5\*\*\* (3.9) | 0.890\*\*\* (0.034) | | Email | 0.290\*\*\* (0.054) | 11.3\*\*\* (3.5) | 2.99\*\*\* (0.07) | 0.162\*\*\* (0.001) | 72.9\*\*\* (4.4) | 1.12\*\*\* (0.04) | | Referral | 0.239\*\*\* (0.054) | 13.2\*\*\* (3.4) | 1.91\*\*\* (0.07) | 0.174\*\*\* (0.001) | 79.1\*\*\* (4.1) | 1.00\*\*\* (0.04) | | Organic Search | 0.159\*\*\* (0.054) | 2.88 (3.28) | 1.39\*\*\* (0.06) | -0.154\*\*\* (0.001) | 33.8\*\*\* (3.8) | 0.767\*\*\* (0.033) | | Paid Social | -0.776\*\*\* (0.055) | 6.42\* (3.72) | -0.339\*\*\* (0.076) | 0.356\*\*\* (0.001) | -116\*\*\* (5) | -0.219\*\*\* (0.041) | | Fixed Effects: Website, Device, Month | Yes | Yes | Yes | Yes | Yes | Yes | | $R^2$ \[pseudo-$R^2$\] | \[0.746\] | 0.398 | 0.263 | \[0.710\] | 0.436 | 0.146 | | Observations | 408,342 | 333,657 | 408,342 | 408,342 | 408,342 | 408,342 | | | \* p<0.1; \*\* p<0.05; \*\*\* p<0.01 | | | | | | Notes: All coefficients relative to oLLM baseline. Standard errors in parentheses. **Table B.10. Regression Results: 6 Months Data – Incl. App** | | CR | AOV | RPS | BR | SD | PV | |:--|--:|--:|--:|--:|--:|--:| | Affiliate | 0.547\*\*\* (0.041) | 26.0\*\*\* (3.3) | 3.05\*\*\* (0.12) | 0.170\*\*\* (0.025) | -7.67 (5.98) | 0.609\*\*\* (0.051) | | Paid Search | 0.312\*\*\* (0.040) | 6.16\*\*\* (2.37) | 1.52\*\*\* (0.07) | -0.193\*\*\* (0.023) | 34.7\*\*\* (3.6) | 0.878\*\*\* (0.031) | | Other | 0.282\*\*\* (0.040) | 10.3\*\*\* (2.4) | 1.73\*\*\* (0.07) | 0.850\*\*\* (0.023) | 48.1\*\*\* (3.8) | 0.624\*\*\* (0.033) | | Direct | 0.213\*\*\* (0.040) | 15.8\*\*\* (2.4) | 2.68\*\*\* (0.07) | 0.277\*\*\* (0.023) | 12.5\*\*\* (3.7) | 0.658\*\*\* (0.032) | | Email | 0.201\*\*\* (0.040) | 12.4\*\*\* (2.6) | 2.42\*\*\* (0.08) | 0.137\*\*\* (0.023) | 61.6\*\*\* (4.2) | 0.872\*\*\* (0.036) | | Referral | 0.147\*\*\* (0.040) | 17.0\*\*\* (2.5) | 1.19\*\*\* (0.08) | 0.159\*\*\* (0.023) | 62.2\*\*\* (3.9) | 0.754\*\*\* (0.034) | | Organic Search | 0.036 (0.040) | 2.18 (2.35) | 0.545\*\*\* (0.070) | -0.179\*\*\* (0.023) | 19.1\*\*\* (3.6) | 0.529\*\*\* (0.031) | | Paid Social | -0.849\*\*\* (0.041) | 7.09\*\* (2.85) | -1.13\*\*\* (0.09) | 0.318\*\*\* (0.023) | -124\*\*\* (4) | -0.417\*\*\* (0.038) | | Fixed Effects: Website, Device, Month | Yes | Yes | Yes | Yes | Yes | Yes | | $R^2$ \[pseudo-$R^2$\] | \[0.741\] | 0.442 | 0.233 | \[0.717\] | 0.461 | 0.175 | | Observations | 395,430 | 331,870 | 395,430 | 395,430 | 395,430 | 395,430 | | | \* p<0.1; \*\* p<0.05; \*\*\* p<0.01 | | | | | | Notes: All coefficients relative to oLLM baseline. Standard errors in parentheses. ## Robustness Regressions: Observed Timeframes We re-estimate models on 3- and 12-month datasets. The 12-month dataset provides longer horizon but may capture idiosyncratic introduction effects. The 3-month dataset contains fewer observations but reflects the most recent period of relative stability. **Table B.11. Regression Results: 3 Months Data** | | CR | AOV | RPS | BR | SD | PV | |:--|--:|--:|--:|--:|--:|--:| | Affiliate | 0.522\*\*\* (0.072) | 30.7\*\*\* (3.4) | 3.29\*\*\* (0.15) | 0.221\*\*\* (0.041) | -35.3\*\*\* (9.2) | 0.545\*\*\* (0.056) | | Paid Search | 0.309\*\*\* (0.070) | 9.78\*\*\* (2.59) | 1.73\*\*\* (0.10) | -0.135\*\*\* (0.039) | 16.2\*\*\* (6.0) | 0.772\*\*\* (0.036) | | Other | 0.359\*\*\* (0.070) | 10.5\*\*\* (2.6) | 2.04\*\*\* (0.10) | 0.871\*\*\* (0.039) | 26.9\*\*\* (6.3) | 0.393\*\*\* (0.038) | | Direct | 0.220\*\*\* (0.070) | 20.6\*\*\* (2.6) | 2.92\*\*\* (0.10) | 0.287\*\*\* (0.039) | -6.85 (6.08) | 0.572\*\*\* (0.037) | | Email | 0.233\*\*\* (0.071) | 17.7\*\*\* (2.8) | 2.50\*\*\* (0.11) | 0.192\*\*\* (0.039) | 42.5\*\*\* (6.8) | 0.781\*\*\* (0.041) | | Referral | 0.162\*\* (0.071) | 18.2\*\*\* (2.7) | 1.36\*\*\* (0.10) | 0.202\*\*\* (0.039) | 46.1\*\*\* (6.4) | 0.640\*\*\* (0.039) | | Organic Search | 0.076 (0.070) | 5.53\*\* (2.58) | 0.768\*\*\* (0.095) | -0.145\*\*\* (0.039) | 3.92 (5.96) | 0.480\*\*\* (0.036) | | Paid Social | -0.904\*\*\* (0.071) | 10.2\*\*\* (3.0) | -0.973\*\*\* (0.114) | 0.372\*\*\* (0.039) | -147\*\*\* (7) | -0.531\*\*\* (0.043) | | Fixed Effects: Website, Device, Month | Yes | Yes | Yes | Yes | Yes | Yes | | $R^2$ \[pseudo-$R^2$\] | \[0.727\] | 0.663 | 0.312 | \[0.727\] | 0.457 | 0.247 | | Observations | 188,038 | 159,659 | 188,038 | 188,038 | 188,038 | 188,038 | | | \* p<0.1; \*\* p<0.05; \*\*\* p<0.01 | | | | | | Notes: All coefficients relative to oLLM baseline. Standard errors in parentheses. **Table B.12. Regression Results: 12 Months Data** | | CR | AOV | RPS | BR | SD | PV | |:--|--:|--:|--:|--:|--:|--:| | Affiliate | 0.450\*\*\* (0.058) | 21.3\*\*\* (3.6) | 3.56\*\*\* (0.08) | 0.267\*\*\* (0.033) | -1.07 (5.51) | 0.704\*\*\* (0.038) | | Paid Search | 0.402\*\*\* (0.057) | 1.60 (3.08) | 2.16\*\*\* (0.06) | -0.169\*\*\* (0.032) | 38.2\*\*\* (3.8) | 0.992\*\*\* (0.027) | | Other | 0.384\*\*\* (0.057) | 5.66\* (3.11) | 2.57\*\*\* (0.06) | 0.891\*\*\* (0.032) | 58.5\*\*\* (4.0) | 0.779\*\*\* (0.027) | | Direct | 0.345\*\*\* (0.057) | 11.0\*\*\* (3.1) | 3.21\*\*\* (0.06) | 0.351\*\*\* (0.032) | 12.1\*\*\* (3.9) | 0.763\*\*\* (0.027) | | Email | 0.300\*\*\* (0.057) | 8.98\*\*\* (3.21) | 2.84\*\*\* (0.06) | 0.143\*\*\* (0.032) | 68.7\*\*\* (4.2) | 0.984\*\*\* (0.029) | | Referral | 0.259\*\*\* (0.057) | 11.9\*\*\* (3.1) | 1.81\*\*\* (0.06) | 0.208\*\*\* (0.032) | 60.1\*\*\* (4.0) | 0.830\*\*\* (0.028) | | Organic Search | 0.156\*\*\* (0.057) | -0.494 (3.067) | 1.23\*\*\* (0.06) | -0.127\*\*\* (0.032) | 18.2\*\*\* (3.8) | 0.624\*\*\* (0.026) | | Paid Social | -0.750\*\*\* (0.057) | 2.45 (3.37) | -0.468\*\*\* (0.069) | 0.277\*\*\* (0.032) | -118\*\*\* (4) | -0.324\*\*\* (0.031) | | Fixed Effects: Website, Device, Month | Yes | Yes | Yes | Yes | Yes | Yes | | $R^2$ \[pseudo-$R^2$\] | \[0.723\] | 0.419 | 0.289 | \[0.658\] | 0.405 | 0.196 | | Observations | 614,714 | 526,432 | 614,714 | 614,714 | 614,714 | 614,714 | | | \* p<0.1; \*\* p<0.05; \*\*\* p<0.01 | | | | | | Notes: All coefficients relative to oLLM baseline. Standard errors in parentheses. ## Robustness Graphs: Financial Metrics beyond Conversion Rate AOV robustness checks exhibit greater variation in effect sizes (Figure B.1), with two non-significant sign changes for organic search (monthly aggregation and 12-month timeframe, see Tables B.2 and B.12). Results remain largely consistent with the main model, which found non-significant differences between oLLM and four traditional channels. ![](assets/figs/Kaiser_Schulze_oLLM_fig_B_1.png) ***Figure B.1.** AOV by channel - All specifications* RPS robustness checks shown in Figure B.2 confirm the main analysis. In contrast to the CR results, RPS for oLLM remains significantly below organic search across all 12 specifications. The top-25%-revenue specification again exhibits systematically larger gaps, mirroring the patterns observed for CR in Figure 14. ![](assets/figs/Kaiser_Schulze_oLLM_fig_B_2.png) ***Figure B.2.** RPS by channel - All specifications* ## Robustness Graphs: Engagement Metrics Figures B.3 through B.5 present comprehensive robustness checks for engagement metrics (BR, PV, SD) across all specifications tested. ![](assets/figs/Kaiser_Schulze_oLLM_fig_B_3.png) ***Figure B.3.** BR by channel - All specifications* ![](assets/figs/Kaiser_Schulze_oLLM_fig_B_4.png) ***Figure B.4.** PV by channel - All specifications* ![](assets/figs/Kaiser_Schulze_oLLM_fig_B_5.png) ***Figure B.5.** SD by channel - All specifications* # Online Appendix C — Time Trend Technical Specifications ## C.1 Complete Time Trend Regression Equations This section provides complete regression specifications for all time trend models used in Section 5. We present equations for all dependent variables with time trend interactions, including the four alternative trend specifications tested: linear, centered, Gompertz, and Sigmoid. ### Trend Specifications for CR and BR Models The quasibinomial CR models are specified with four alternative trend transformations: #### Linear trend: $$\begin{aligned} \text{logit}\!\left(p^{\text{CR}}_{it}\right) &= \alpha_{c(i,t)}^{CR} + \gamma_i^{CR} + \delta_{d(i,t)}^{CR} + \mu_{m(t)}^{CR} \\ &\quad + \text{trend}_t + \alpha_{c(i,t)}^{CR} \cdot \text{trend}_t. \end{aligned}$$ We note that this translates into a seemingly exponential effect for the quasibinomial regressions due to the logit transformation for predictions, yet remains a standard linear effect in the regular regressions. #### Centered transformation , where $\text{trend}^{\mathbb{C}}_t = \frac{\text{trend}_t - \text{median}(\text{trend})}{\text{sd}(\text{trend})}$: $$\begin{aligned} \text{logit}\!\left(p^{\text{CR}}_{it}\right) &= \alpha_{c(i,t)}^{CR} + \gamma_i^{CR} + \delta_{d(i,t)}^{CR} + \mu_{m(t)}^{CR} + \\ & \; +\;\text{trend}^{\mathbb{C}}_t +\, \alpha_{c(i,t)}^{CR} \cdot \text{trend}^{\mathbb{C}}_t , \end{aligned}$$ This results in a simple intercept shift for the linear regression models and reduces the perceived exponential effect for the quasibinomial models. #### Gompertz transformation , where $\text{trend}^{\mathbb{G}}_t = -\exp\!\left(-\frac{\text{trend}_t}{\max(\text{trend})}\right)$: $$\begin{aligned} \text{logit}\!\left(p^{\text{CR}}_{it}\right) &= \alpha_{c(i,t)}^{CR} + \gamma_i^{CR} + \delta_{d(i,t)}^{CR} + \mu_{m(t)}^{CR} + \\ & \; +\;\text{trend}^{\mathbb{G}}_t +\, \alpha_{c(i,t)}^{CR} \cdot \text{trend}^{\mathbb{G}}_t , \end{aligned}$$ #### Sigmoid transformation , where $\text{trend}^{\mathbb{S}}_t = \frac{1}{1 + \exp(-\text{trend}^{\text{scaled}}_t)}$ and $\text{trend}^{\text{scaled}}_t$ is the standardized $\text{trend}_t$: $$\begin{aligned} \text{logit}\!\left(p^{\text{CR}}_{it}\right) &= \alpha_{c(i,t)}^{CR} + \gamma_i^{CR} + \delta_{d(i,t)}^{CR} + \mu_{m(t)}^{CR} + \\ & \; +\;\text{trend}^{\mathbb{S}}_t +\, \alpha_{c(i,t)}^{CR} \cdot \text{trend}^{\mathbb{S}}_t , \end{aligned}$$ ### Trend Specifications for Linear Models (AOV and RPS) For the linear models, the trend specifications are analogous. We provide the complete equations below: #### Average Order Value (AOV): $$\begin{aligned} \text{AOV}_{it} &= \beta^{\text{AOV}}_{0} + \alpha^{\text{AOV}}_{c(i,t)} + \gamma^{\text{AOV}}_{i} + \delta^{\text{AOV}}_{d(i,t)} + \mu^{\text{AOV}}_{m(t)}+ \text{trend}_t +\, \alpha^{\text{AOV}}_{c(i,t)} \cdot \text{trend}_t + \varepsilon^{\text{AOV}}_{it}, \\ \varepsilon^{\text{AOV}}_{it} &\sim \mathcal{N}(0,\sigma^2_{\text{AOV}}). \end{aligned}$$ #### Revenue per Session (RPS): $$\begin{aligned} \text{RPS}_{it} &= \beta^{\text{RPS}}_{0} + \alpha^{\text{RPS}}_{c(i,t)} + \gamma^{\text{RPS}}_{i} + \delta^{\text{RPS}}_{d(i,t)} + \mu^{\text{RPS}}_{m(t)}+ \text{trend}_t +\, \alpha^{\text{RPS}}_{c(i,t)} \cdot \text{trend}_t + \varepsilon^{\text{RPS}}_{it}, \\ \varepsilon^{\text{RPS}}_{it} &\sim \mathcal{N}(0,\sigma^2_{\text{RPS}}). \end{aligned}$$ All other definitions remain unchanged from the base model specifications in Online Appendix A.4. The four alternative trend transformations (linear, centered, Gompertz, Sigmoid) can be applied to any of the linear models above by replacing $\text{trend}_t$ with the appropriate transformation. # Online Appendix D — Heterogeneity Regressions: Quartile Splits This section presents regression results underlying Figures 21 – 26. Each column presents a separate regression on a subset of websites, based on websites’ “product category complexity”, “share of technophiles” (based on Google’s affinity segmentation), and “average user age”. **Table D.1. Channel Effects by Product Category Complexity (Quartile Split)** | | CR: Bottom Quartile | CR: Top Quartile | AOV: Bottom Quartile | AOV: Top Quartile | RPS: Bottom Quartile | RPS: Top Quartile | |:--|--:|--:|--:|--:|--:|--:| | Affiliate | 0.272\* (0.145) | 0.244\*\* (0.103) | 50.4\*\*\* (12.6) | 64.6\*\*\* (13.9) | 3.35\*\*\* (0.36) | 3.61\*\*\* (0.27) | | Paid Search | 0.198 (0.135) | 0.056 (0.093) | 40.5\*\*\* (8.9) | 4.02 (9.77) | 2.29\*\*\* (0.20) | 1.09\*\*\* (0.15) | | Other | 0.305\*\* (0.135) | 0.610\*\*\* (0.094) | 18.5\*\* (9.0) | 6.58 (10.00) | 2.36\*\*\* (0.21) | 2.06\*\*\* (0.16) | | Direct | 0.449\*\*\* (0.135) | -0.345\*\*\* (0.093) | 41.1\*\*\* (8.9) | 24.5\*\* (9.7) | 3.53\*\*\* (0.20) | 2.39\*\*\* (0.15) | | Email | 0.199 (0.136) | -0.317\*\*\* (0.095) | 42.1\*\*\* (9.3) | 14.6 (10.7) | 2.89\*\*\* (0.22) | 2.08\*\*\* (0.18) | | Referral | 0.142 (0.136) | -0.398\*\*\* (0.094) | 39.5\*\*\* (9.1) | 23.5\*\* (10.1) | 1.44\*\*\* (0.21) | 1.62\*\*\* (0.16) | | Organic Search | 0.287\*\* (0.135) | -0.374\*\*\* (0.093) | 22.8\*\*\* (8.8) | 4.66 (9.67) | 1.28\*\*\* (0.20) | 0.404\*\*\* (0.151) | | Paid Social | -0.846\*\*\* (0.138) | -1.31\*\*\* (0.10) | 27.1\*\*\* (9.7) | 17.3 (12.5) | -0.192 (0.225) | -1.26\*\*\* (0.20) | | $R^2$ \[pseudo-$R^2$\] | \[0.680\] | \[0.782\] | 0.523 | 0.281 | 0.222 | 0.349 | | Observations | 24,816 | 85,281 | 20,946 | 69,402 | 24,816 | 85,281 | | Fixed Effects: Device, Website, Month | Yes | Yes | Yes | Yes | Yes | Yes | | | \* p<0.1; \*\* p<0.05; \*\*\* p<0.01 | | | | | | Notes: All coefficients relative to oLLM baseline. Standard errors in parentheses. **Table D.2. Channel Effects by Technophile Share (Quartile Split)** | | CR: Bottom Quartile | CR: Top Quartile | AOV: Bottom Quartile | AOV: Top Quartile | RPS: Bottom Quartile | RPS: Top Quartile | |:--|--:|--:|--:|--:|--:|--:| | Affiliate | 0.855\*\*\* (0.223) | 0.547\*\*\* (0.110) | 14.9\*\*\* (4.1) | 65.1\*\*\* (25.1) | 4.18\*\*\* (0.24) | 4.27\*\*\* (0.47) | | Paid Search | 0.888\*\*\* (0.221) | -0.076 (0.105) | 5.49 (3.41) | 17.9 (19.3) | 2.44\*\*\* (0.17) | 0.971\*\*\* (0.298) | | Other | 0.573\*\*\* (0.221) | -0.379\*\*\* (0.105) | 7.01\*\* (3.44) | 10.9 (19.7) | 2.63\*\*\* (0.17) | 2.26\*\*\* (0.31) | | Direct | 0.728\*\*\* (0.221) | 0.174\* (0.105) | 11.6\*\*\* (3.4) | 32.0 (19.5) | 2.90\*\*\* (0.17) | 3.31\*\*\* (0.31) | | Email | 0.686\*\*\* (0.221) | 0.272\*\* (0.107) | 1.05 (3.53) | 36.1\* (20.9) | 3.01\*\*\* (0.18) | 2.68\*\*\* (0.34) | | Referral | 0.680\*\*\* (0.221) | 0.214\*\* (0.105) | 13.5\*\*\* (3.5) | 29.0 (20.0) | 1.19\*\*\* (0.18) | 1.50\*\*\* (0.32) | | Organic Search | 0.630\*\*\* (0.221) | -0.426\*\*\* (0.105) | 4.73 (3.40) | 20.8 (19.2) | 1.43\*\*\* (0.17) | 0.642\*\* (0.295) | | Paid Social | -0.224 (0.222) | -1.39\*\*\* (0.11) | 3.29 (3.64) | 21.9 (23.3) | -0.119 (0.190) | -1.61\*\*\* (0.37) | | $R^2$ \[pseudo-$R^2$\] | \[0.666\] | \[0.840\] | 0.688 | 0.185 | 0.235 | 0.200 | | Observations | 41,719 | 40,321 | 37,026 | 33,380 | 41,719 | 40,321 | | Fixed Effects: Device, Website, Month | Yes | Yes | Yes | Yes | Yes | Yes | | | \* p<0.1; \*\* p<0.05; \*\*\* p<0.01 | | | | | | Notes: All coefficients relative to oLLM baseline. Standard errors in parentheses. **Table D.3. Channel Effects by Customer Age (Quartile Split)** | | CR: Bottom Quartile | CR: Top Quartile | AOV: Bottom Quartile | AOV: Top Quartile | RPS: Bottom Quartile | RPS: Top Quartile | |:--|--:|--:|--:|--:|--:|--:| | Affiliate | -0.087 (0.088) | 0.505\*\* (0.220) | 30.7\*\*\* (6.2) | 12.6\* (6.9) | 2.11\*\*\* (0.23) | 4.52\*\*\* (0.38) | | Paid Search | 0.048 (0.083) | 0.545\*\* (0.219) | 12.9\*\*\* (4.3) | -12.5\*\* (5.9) | 1.99\*\*\* (0.13) | 1.98\*\*\* (0.27) | | Other | 0.090 (0.083) | 0.165 (0.219) | 20.9\*\*\* (4.3) | -11.6\* (6.0) | 1.68\*\*\* (0.14) | 3.21\*\*\* (0.28) | | Direct | -0.035 (0.083) | 0.531\*\* (0.219) | 26.1\*\*\* (4.2) | -0.791 (5.975) | 2.53\*\*\* (0.13) | 4.55\*\*\* (0.27) | | Email | -0.091 (0.084) | 0.428\* (0.219) | 22.1\*\*\* (4.7) | -5.17 (6.17) | 3.07\*\*\* (0.15) | 2.88\*\*\* (0.29) | | Referral | 0.049 (0.084) | 0.521\*\* (0.219) | 24.8\*\*\* (4.4) | -6.22 (6.09) | 1.77\*\*\* (0.14) | 1.90\*\*\* (0.28) | | Organic Search | -0.310\*\*\* (0.083) | 0.291 (0.219) | 15.9\*\*\* (4.2) | -16.9\*\*\* (5.9) | 0.918\*\*\* (0.127) | 1.43\*\*\* (0.27) | | Paid Social | -1.13\*\*\* (0.09) | -0.283 (0.219) | 8.36 (5.22) | -7.26 (6.46) | 0.544\*\*\* (0.164) | -0.943\*\*\* (0.312) | | $R^2$ \[pseudo-$R^2$\] | \[0.792\] | \[0.779\] | 0.684 | 0.650 | 0.435 | 0.241 | | Observations | 41,229 | 42,957 | 33,981 | 37,060 | 41,229 | 42,957 | | Fixed Effects: Device, Website, Month | Yes | Yes | Yes | Yes | Yes | Yes | | | \* p<0.1; \*\* p<0.05; \*\*\* p<0.01 | | | | | | Notes: All coefficients relative to oLLM baseline. Standard errors in parentheses. # Online Appendix E — Reconciling Differences with Prior Studies This appendix discusses potential explanations for why our findings differ from some industry reports and from expectations based on prior research. We examine both methodological factors and substantive mechanisms that may account for these discrepancies. ## Potential Explanations for Differences vis-à-vis Industry Study Results Two recent industry studies report substantially higher CR for oLLM than for organic search, in direct contrast to our findings. While we cannot definitively reconcile these results, two factors likely contribute. ### Website-level heterogeneity Seer Interactive (2025) analyze a single website and report an oLLM CR of 15.9%. Our broader sample indicates that such outcomes are atypical. We observe one website with an even higher oLLM CR of 26.0%, but this is a clear outlier; the second-highest oLLM CR drops to 12.9%, suggesting that exceptionally high CRs are rare rather than representative. The multi-site study by ThoughtMetric (2025) reports a 6.7% average oLLM CR across 100 websites, yet in our data only seven websites achieve oLLM CRs of 6.7% or higher. Given this breadth, sample selection alone seems unlikely to explain the divergence. ### Methodological upward bias Differences in analytic practice between industry reports and academic work may also matter. Industry analyses often operate on daily aggregates, which present serious challenges when evaluating a channel like oLLM that represents less than 0.2% of total sessions (see Section 2.4). The sizable gaps between our model-free evidence and our regression estimates (Section 3) indicate that these challenges can produce genuinely misleading interpretations, and do not simply reflect academic overcautiousness. While ThoughtMetric (2025) provides limited methodological transparency, improper aggregation or data filtering practices—such as applying minimum transaction thresholds—might introduce upward bias in reported oLLM CR. ## Potential Explanations for Differences vis-à-vis Research Study Expectations Some prior research has posited that oLLM should provide stronger financial and engagement outcomes than traditional channels due to greater context, better synthesis, and a superior interface. Our evidence indicates this is not (yet) realized. We highlight three contributing factors. ### Platform architecture and commercial optimization gaps LLM technology remains early-stage. Platforms have prioritized core model quality and general UX over commerce-specific capabilities. Because only 2.1% of conversations involve purchasable products (Chatterji et al. 2025), features critical for retail, such as comprehensive product catalogs, current price and availability, and reliable deep-linking, may not be fully built out (Semrush 2025). Consistent with this view, oLLM-driven traffic volumes have been stagnant. Although BR is comparatively favorable (suggesting relevance when clicks occur), oLLM lags mature channels on session depth (PV) and duration (SD). ### Website infrastructure and investment coordination Given low volumes and relatively weak financial outcomes to date, many retailers likely have not fully optimized landing pages, site structure, or merchandising for oLLM. This creates a coordination problem: weak outcomes depress incentives to invest, and limited investment sustains weak outcomes. Our evidence is consistent with such dynamics. Improvements in CR coincide with declining AOV, suggesting early optimization might have focused on completion probability rather than basket expansion. The fact that large sites are not (yet) leading in oLLM effectiveness further suggests that dedicated optimization is not currently an industry priority. ### Consumer adoption and channel integration While LLMs have reshaped online behavior with respect to many major websites (Gholami et al. 2026), shopping via LLMs remains nascent: only a small fraction of conversations concern purchasable products (Chatterji et al. 2025). Users may not view oLLM as a default entry point for shopping, in particular for low-complexity products (see Section 6.1). Early-stage familiarity and trust plausibly induce verification behavior. Users confirm recommendations via search engines or retailer sites—which, under last-click attribution, assigns conversions to other channels. Engagement patterns align with this interpretation: low BR on LLM-provided links indicates relevance, yet shorter SD and fewer PVs imply high task orientation; combined with declining AOV, these signals suggest users are still building proficiency and developing habits (see Section 6.2, but also Padilla et al. (2025)). # Online Appendix F — Directions for Future Research Ours is the first large-scale empirical study describing financial and engagement outcomes of organic LLM traffic as an e-commerce channel. Following an empirics first approach, it provides solid empirical evidence, but also raises many important questions for future research. ## Outcome Gap Mechanisms and Improvement Trajectories Our finding that oLLM shows improving CRs alongside declining AOVs warrants further investigation. Future research should expand on our cross-website analysis of consumers’ LLM proficiency and product category complexity, and examine more deeply whether this pattern reflects platform learning effects, consumer learning effects, or systematic differences in the types of products oLLM facilitates. Investigation of these mechanisms could inform predictions about convergence trajectories and assess whether current outcomes represent temporary growing pains or structural limitations of LLMs. ## Platform Friction and Cross-Platform Comparison Our cross-website analysis in Section 6.1 indicates that oLLM outcomes might reflect perceived tradeoffs between LLM usefulness and convenience. The current oLLM experience requires consumers to leave the LLM interface and navigate to a retailer website, introducing friction that may partly explain weaker financial metrics for simpler products. Different approaches to reducing this friction—such as in-platform shopping agents, specialized shopping LLMs, and on-site conversational assistants—offer different tradeoffs in convenience and usefulness. Understanding which friction points matter most for conversion and how they can be overcome represents a promising research direction. Comparative analyses examining different providers (ChatGPT, Claude, Gemini, Perplexity, Grok, Deepseek) could further reveal how interface and algorithm differences affect consumer-product matching. ## Generative Engine Optimization and Retailer Visibility oLLM’s growing volume and improving financial metrics justify companies’ increasing interest in Generative Engine Optimization (GEO), yet many aspects of this nascent field remain unexplored. Our understanding along the entire funnel is limited: it is unclear how companies can increase their LLM exposure, click-through, and subsequent conversions, and how these stages are interrelated. Additionally, the coordination problem between weak oLLM outcomes and limited retailer optimization suggests research opportunities examining how retailers should modify website architecture and merchandising strategies for direct product-page entry (Chen et al. 2025). ## Moderating Factors in Channel Outcomes Future research should investigate contextual factors explaining current gaps. Competition intensity represents a promising avenue, as high-competition categories may benefit more from preference elicitation. Brand strength effects present competing predictions worth investigating, as oLLM recommendations may exhibit less advertising bias, but may also favor brands that are more prominent in retrieved and training data (Chen et al. 2025; Padilla et al. 2025). ## Consumer Adoption and Market Implications Research examining how consumers develop LLM proficiency and shopping habits in LLM environments would inform understanding of adoption dynamics. The finding that only a small fraction of LLM conversations concern purchasable products (Chatterji et al. 2025) suggests substantial room for growth. Investigation of oLLM’s impact on broader market outcomes deserves attention despite current low volumes. Key areas include the introduction of advertising in LLM platforms (pLLM) and potential disruption to the retail media ecosystem as consumers bypass traditional homepage advertising. ## References - Adobe. 2025. “Adobe Analytics: Traffic to U.S. Retail Websites from Generative AI Sources Jumps 1,200%.” Adobe Blog (March 17). . - Ahrefs. 2025. “63% of Websites Receive AI Traffic (New Study of 3,000 Sites).” Ahrefs Blog (February 6). . - Al-Hasan, Tamim Mahmud, Aya N. Sayed, Fayçal Bensaali, Yassine Himeur, Iraklis Varlamis, and George Dimitrakopoulos. 2024. “From Traditional Recommender Systems to GPT-Based Chatbots: A Survey of Recent Developments and Future Directions.” *Big Data and Cognitive Computing* 8 (4): 36. . - Anderson, Simon P., and Régis Renault. 2006. “Advertising Content.” *American Economic Review* 96 (1): 93–113. . - Aridor, Guy, Yeon-Koo Che, Brett Hollenbeck, Maximilian Kaiser, and Daniel McCarthy. 2025. “Evaluating the Impact of Privacy Regulation on e-Commerce Firms: Evidence from Apple’s App Tracking Transparency.” *Management Science*. . - Berman, Ron. 2018. “Beyond the Last Touch: Attribution in Online Advertising.” *Marketing Science* 37 (5): 771–92. . - Bleier, Alexander, and Maik Eisenbeiss. 2015. “Personalized Online Advertising Effectiveness: The Interplay of What, When, and Where.” *Marketing Science* 34 (5): 669–88. . - Brown, Jeffrey R., and Austan Goolsbee. 2002. “Does the Internet Make Markets More Competitive? Evidence from the Life Insurance Industry.” *Journal of Political Economy* 110 (3): 481–507. . - Chatterji, Aaron, Thomas Cunningham, David J. Deming, Zoe Hitzig, Christopher Ong, Carl Yan Shan, and Kevin Wadman. 2025. “How People Use ChatGPT.” Working Paper 34255. NBER Working Paper. National Bureau of Economic Research. . - Chen, Mahe, Xiaoxuan Wang, Kaiwen Chen, and Nick Koudas. 2025. “Generative Engine Optimization: How to Dominate AI Search.” *arXiv Preprint arXiv:2509.08919*. - Dinerstein, Michael, Liran Einav, Jonathan Levin, and Neel Sundaresan. 2018. “Consumer Price Search and Platform Design in Internet Commerce.” *American Economic Review* 108 (7): 1820–59. . - Ellison, Glenn, and Sara Fisher Ellison. 2009. “Search, Obfuscation, and Price Elasticities on the Internet.” *Econometrica* 77 (2): 427–52. . - eMarketer. 2024. “Retail Media Ad Spending Forecast H1 2024.” Insider Intelligence / eMarketer. . - Foroughi, Behzad, Mohammad Iranmanesh, Elaheh Yadegaridehkordi, Jun Wen, Morteza Ghobakhloo, Murali Senali, and Norzalita Annamalai. 2025. “Factors Affecting the Use of ChatGPT for Obtaining Shopping Information.” *International Journal of Consumer Studies* 49 (1): e70008. . - Gholami, Samira, Cristiana Firullo, Cristobal Cheyre, and Alessandro Acquisti. 2026. “Beyond Search: LLM Adoption and Web Traffic Concentration.” *SSRN*, February. . - Ghose, Anindya, and Sha Yang. 2009. “An Empirical Analysis of Search Engine Advertising: Sponsored Search in Electronic Markets.” *Management Science* 55 (10): 1605–22. . - Google. 2024. “Google Analytics 4 Dimensions and Metrics.” Google. . - Haan, Evert de, Thorsten Wiesel, and Koen Pauwels. 2016. “The Effectiveness of Different Forms of Online Advertising for Purchase Conversion in a Multiple-Channel Attribution Framework.” *International Journal of Research in Marketing* 33 (3): 491–507. . - Hermann, Erik, Stefano Puntoni, and David A. Schweidel. 2025. “Conversational AI: The Next Frontier of Digital Platform Monetization.” *SSRN*, October. . - Huang, Ming-Hui, and Roland T. Rust. 2025. “The GenAI Future of Consumer Research.” *Journal of Consumer Research* 52 (1): 4–17. . - Jannach, Dietmar. 2022. “Evaluating Conversational Recommender Systems: A Landscape of Research.” *Artificial Intelligence Review*. . - Johnson, Garrett A., Scott K. Shriver, and Shaoyin Du. 2020. “Consumer Privacy Choice in Online Advertising: Who Opts Out and at What Cost to Industry?” *Marketing Science* 39 (1): 33–51. . - Li, Hongshuang, and P. K. Kannan. 2014. “Attributing Conversions in a Multichannel Online Marketing Environment: An Empirical Model and a Field Experiment.” *Journal of Marketing Research* 51 (1): 40–56. . - Li, Hongshuang, P. K. Kannan, Siva Viswanathan, and Abhishek Pani. 2016. “Attribution Strategies and Return on Keyword Investment in Paid Search Advertising.” *Marketing Science* 35 (6): 831–48. . - Luo, Xueming, Siliang Tong, Zheng Fang, and Zhe Qu. 2019. “Machines Vs. Humans: The Impact of Artificial Intelligence Chatbot Disclosure on Customer Purchases.” *Marketing Science* 38 (6): 937–47. . - Lynch, John G., and Dan Ariely. 2000. “Wine Online: Search Costs Affect Competition on Price, Quality, and Distribution.” *Marketing Science* 19 (1): 83–103. . - Manchanda, Puneet, Jean-Pierre Dubé, Khim Yong Goh, and Pradeep K. Chintagunta. 2006. “The Effect of Banner Advertising on Internet Purchasing.” *Journal of Marketing Research* 43 (1): 98–108. . - Miller, Klaus M., and Bernd Skiera. 2024. “Economic Consequences of Online Tracking Restrictions: Evidence from Cookies.” *International Journal of Research in Marketing* 41 (2): 241–64. . - Nguyen, Quynh N., Anna Sidorova, and Russell Torres. 2022. “User Interactions with Chatbot Interfaces Vs. Menu-Based Interfaces: An Empirical Study.” *Computers in Human Behavior* 128: 107093. . - OpenAI. 2025. “Improved Shopping Results from ChatGPT Search.” OpenAI Help Center. . - Padilla, Nicolas, H. Tai Lam, Anja Lambrecht, and Brett Hollenbeck. 2025. “The Impact of LLM Adoption on Online User Behavior.” *SSRN*, August. . - Reisenbichler, Martin, Thomas Reutterer, David A. Schweidel, and Daniel Dan. 2022. “Frontiers: Supporting Content Marketing with Natural Language Generation.” *Marketing Science* 41 (3): 441–52. . - Reuters. 2025. “OpenAI Rolls Out New Shopping Features with ChatGPT Search Update.” . - Rutz, Oliver J., and Randolph E. Bucklin. 2011. “From Generic to Branded: A Model of Spillover in Paid Search Advertising.” *Journal of Marketing Research* 48 (1): 87–102. . - SALT. 2025. “Do Users Really Show Higher Intent When They Click Through from an LLM to a Website?” Accessed May 12, 2025. . - Salvi, Francesco, Manoel Horta Ribeiro, Riccardo Gallotti, and Robert West. 2025. “On the Conversational Persuasiveness of GPT-4.” *Nature Human Behaviour*. . - SE Ranking. 2025. “Which Platforms Lead in AI Traffic in 2025? ChatGPT, Perplexity, Gemini & More.” SE Ranking Blog (August 27). . - Search Engine Land. 2025. “ChatGPT Search Vs. Google: A Deep Dive Analysis of 62 Queries.” . - Seer Interactive. 2025. “Case Study: 6 Learnings, 1 Site — How Traffic from ChatGPT Converts.” Accessed June 3, 2025. . - Semrush. 2025. “We Studied the Impact of AI Search on SEO Traffic: Our Findings.” Accessed July 21, 2025. . - SimilarWeb. 2024. “The 2024 Marketing Benchmark Report.” Report (PDF), Accessed December 2, 2025. . - Similarweb. 2025. “Industry Category.” Similarweb Help Center, Accessed December 2, 2025. . - Soviero, Beatriz, Daniel Kuhn, Alexandre Salle, and Viviane P. Moreira. 2024. “ChatGPT Goes Shopping: LLMs Can Predict Relevance in eCommerce Search.” In *Advances in Information Retrieval: Proceedings of ECIR 2024*, 3–11. Lecture Notes in Computer Science. Springer. . - Steyvers, Mark, and Aakriti Kumar. 2024. “Three Challenges for AI-Assisted Decision-Making.” *Perspectives on Psychological Science* 19 (5): 722–34. . - The Economist. 2023. “Is Google’s 20-Year Search Dominance about to End?” . - ThoughtMetric. 2025. “The Conversion Advantage: Why ChatGPT Traffic Outperforms Google Search for e-Commerce.” Accessed June 17, 2025. . - Ursu, Raluca M., Qianyun Zhang, and Elisabeth Honka. 2022. “Search Gaps and Consumer Fatigue.” *Marketing Science* 42 (1): 110–36. . - Wies, Simone, Alexander Bleier, and Alexander Edeling. 2023. “Finding Goldilocks Influencers: How Follower Count Drives Social Media Engagement.” *Journal of Marketing* 87 (3): 383–405. . - Wroe, Simon. 2024. “How Walmart Became an Advertising Powerhouse.” *Financial Times*, August. . - Yang, Jeremy, Noah Yonack, Kate Zyskowski, Denis Yarats, Johnny Ho, and Jerry Ma. 2025. “The Adoption and Usage of AI Agents: Early Evidence from Perplexity.” *arXiv Preprint*, December. . - Yu, Peiyang, Zeqiu Xu, Jiani Wang, and Xiaochuan Xu. 2025. “The Application of Large Language Models in Recommendation Systems.” . - Yu, Yue, Wei Ping, Zihan Liu, Boxin Wang, Jiaxuan You, Chao Zhang, Mohammad Shoeybi, and Bryan Catanzaro. 2024. “RankRAG: Unifying Context Ranking with Retrieval-Augmented Generation in LLMs.” In *Advances in Neural Information Processing Systems (NeurIPS 2024)*. . [^3]: LLM platforms had not yet introduced affiliate-like systems or ads during our study period (Hermann, Puntoni, and Schweidel 2025). [^4]: The representativeness of Grips’ data relative to e-commerce in general has been previously assessed by comparing it to trusted proprietary datasets (e.g., Similarweb) and verifiable public ones (e.g., Shopify Quarterly Reports, United States E-Commerce Census). For details, see Online Appendix E in Aridor et al. (2025). [^5]: A session refers to a continuous period of website activity by a consumer, potentially comprising multiple page views. Sessions typically end after 30 minutes of inactivity. A single consumer may generate multiple sessions on the same website within a day. [^6]: Our data does not allow us to segment by product or customer characteristics *and* channels simultaneously. We therefore analyze differences across websites rather than within websites. [^7]: The declining order values could reflect increased purchase targeting that reduces exploratory browsing or heightened price consciousness from LLM transparency about alternatives.