Frequently asked questions

Organic LLM traffic (oLLM) is human visitors who reach an online store by clicking an organic link that ChatGPT — or another LLM platform — shows when someone expresses purchase intent. It's real people clicking through, not the AI crawlers or bots that scrape sites for training. The channel was effectively born in August 2024, when ChatGPT began adding outbound product links to its answers, creating a genuinely new way for customers to arrive at an e-commerce website.

Yes. Throughout the study period these links were generated by the model itself, based on what it judged most relevant — they were not paid placements or ads, and no retailer could buy its way into them. We call it organic LLM traffic (oLLM) precisely to separate it from any paid LLM traffic (pLLM), such as the sponsored placements and ad formats platforms have since begun testing.

We treat oLLM as a distinct new channel and benchmark it head-to-head against the eight channels retailers already budget for: organic and paid search, direct, email, referral, affiliate, paid social, and “other”. It behaves differently enough from each of them — in who it brings, how they engage, and whether they buy — that it earns its own line in your analytics rather than being folded into referral or direct.

No, and that's deliberate. This is a descriptive study of how oLLM actually performs across hundreds of websites, not a how-to guide for getting recommended by ChatGPT. We flag generative engine optimization (GEO) as an important but still-open question: it isn't yet clear how a retailer can reliably raise its exposure in AI answers, its click-through, and its conversions, or how those steps relate. That's a frontier the paper points to, not one it solves.

No — we measure organic ChatGPT referrals only. Features such as in-chat Instant Checkout, agentic shopping assistants, and paid ad placements all arrived after our observation window closed, and each could materially change the picture. We note them as developments to watch, but they are outside what this study measures, so we don't make claims about them.

Very little, so far. One year after launch, oLLM accounted for under 0.2% of all sessions in our data — roughly 200× smaller than Google organic search.

Treat the eye-catching growth percentages with care. They're real, but they start from a tiny base, so a small absolute increase looks enormous in percentage terms. In our data, oLLM's traffic volume stayed low and fairly flat over the year. The meaningful improvement is in quality — conversion rate has been climbing — rather than in the sheer amount of traffic, which has not yet taken off.

Below Google, above paid social. After adjusting for the website, device and time of year, organic search converts about 13% higher than oLLM, and most other channels higher still — affiliate links, for instance, convert about 86% more often. The one channel oLLM clearly beats is paid social, which converts roughly 53% lower. So oLLM sits in the mid-to-low part of the ranking, not the runaway winner some single-site reports suggest.

Partly. The signals are mixed: oLLM visitors bounce less than most channels, which suggests the link was relevant to what they wanted — but they also view fewer pages, spend less time on site, and ultimately convert below most channels. The picture is of a focused, task-oriented visit rather than either idle browsing or high-conviction buying. High intent is real, but more limited than the popular narrative claims.

No — if anything the opposite. Average order values from oLLM tend to be lower than from other channels, and they have been falling over the study period even as conversion rates rise. That decline partly offsets the conversion gains, which is why the overall value of an oLLM session improves only modestly over time rather than climbing steeply.

On revenue per session — which combines how often a visit converts with how much it spends — oLLM beats only paid social and trails every other channel. It is improving gradually as conversion rises, but rising conversion is partly cancelled out by falling order values, so the value of an oLLM session is climbing slowly rather than catching up quickly to search or email.

They bounce less than most channels — a good sign that the AI sent them somewhere relevant — but they don't browse more. oLLM visitors view fewer pages and spend less time on site than visitors from search, email, or referral. The combination points to an efficient, in-and-out visit: people arrive with a specific goal from the chat rather than settling in to explore the catalogue.

Two reasons we document directly. First, many of those striking figures come from a single site or a small handful, which turn out to be outliers; one widely cited 15.9% conversion rate sits well above the second-highest site in our entire sample (12.9%). Second, measuring a channel this tiny on daily data, sometimes with minimum-transaction filters, can mechanically inflate the number. Once the sparse data is handled properly, the apparent advantage shrinks and the gaps to other channels reappear.

ChatGPT, overwhelmingly — it accounts for more than 90% of all LLM-referred sessions in our data. The others are small by comparison: Perplexity around 4%, Gemini around 3%, Copilot around 2%, with Deepseek and Grok barely registering. Because ChatGPT so dominates the channel today, our main analysis focuses on it specifically, rather than blending platforms that behave and convert differently.

Our headline results are ChatGPT-only, because it makes up over 90% of the traffic and blending in tiny, possibly different platforms would muddy the comparison. We did re-run the full analysis with all LLM platforms included as a robustness check, and the core findings held — so the focus on ChatGPT simplifies the story without distorting it.

Complex, high-consideration purchases. On websites selling complex products, oLLM's share of traffic is about 4.6 times higher than on simple-product sites; its conversion there can even exceed organic search, while its revenue per session moves much closer to search without overtaking it. For cheap, routine items the advantage largely vanishes. High-complexity categories include vehicles, finance, and business services; low-complexity ones include news, sports, and entertainment.

Because that's where a conversational interface genuinely helps — synthesising reviews, specifications, and trade-offs for a decision that needs real comparison and deliberation. For a cheap, simple product, the extra step of leaving the chat and navigating to a store adds friction without adding much value, so shoppers gain little from routing the purchase through an assistant. The harder the decision, the more the AI's guidance is worth.

We can't observe individuals, but at the website level the pattern is clear: oLLM's share of traffic is about 3.8 times higher where a site's audience skews tech-savvy, and about 5.5 times higher where it skews younger. In other words, shopping through ChatGPT today is concentrated among more digitally proficient consumers — the same groups that adopt new tools first.

Three forces likely weigh on it. The platforms are still early on commerce features — product catalogues, live prices and availability, reliable deep links. Companies, seeing small volumes, haven't optimised their sites for AI entry, which creates a chicken-and-egg problem: weak results discourage the very investment that would improve them. And consumers are still learning to trust and use LLMs for buying — only about 2% of ChatGPT conversations even involve a purchasable product.

Yes, on conversion. Over the first year the share of oLLM visits ending in a purchase rose steadily, while traditional channels stayed flat apart from the usual holiday peak. But average order value fell over the same period, so the two effects partly cancel: the value of an oLLM session is improving, just moderately rather than dramatically.

Probably not soon. When we project the trends forward, oLLM's conversion rate keeps closing the gap with organic search but does not reach parity within the next year. Revenue per session — the fuller measure of value — only reaches organic-search levels under the most optimistic of our forecasting scenarios. The direction is encouraging; the pace is gradual.

Two different things, for now. This study measures referral clicks from AI assistants like ChatGPT — people who leave the assistant and land on your site. That's not the same as AI Overviews, the AI summaries Google shows inside its own results, which are a separate, in-search experience. On our evidence, these assistant referrals remain niche, roughly 200× smaller than Google organic search. Beyond what we measured, the broader industry context is that the two may converge — as assistants build wider indexes of the web and search engines fold in AI Overviews and an “AI mode”, the line between assistant and search engine is likely to blur. That convergence is market context, not something our data tests.

If the goal is conversions, the honest answer today is: watch it closely, but don't yet reallocate budget to it. Volumes are small and revenue per session trails most channels, so it won't move your numbers in the short term. But the traffic is high-quality in the right categories and steadily improving, so the smart move is to track it as its own channel now — so you can see the trend on your own site before deciding to invest.

This is where the near-term upside concentrates. In complex categories, oLLM already pairs comparatively strong conversion and revenue per session with low bounce rates, making it a credible high-intent complement to organic and paid search, even at small volumes. The practical step is to ensure product pages serve shoppers who arrive mid-decision from an LLM conversation: clear specifications, comparisons, and answers to the questions an assistant would have raised.

Expect only modest conversion gains for now. For routine, low-complexity products, oLLM's traffic share and conversion are weak, so channel-specific optimisation is unlikely to pay off much yet. It's reasonable to keep investment minimal and revisit as the platforms roll out shopping features — instant checkout, agents — that may suit simple, repeat purchases better than today's click-out-to-a-store experience.

Twelve months of first-party Google Analytics data — August 2024 to July 2025 — from 973 e-commerce websites with about $20 billion in combined annual revenue. Within that we observe over 50,000 transactions traced to ChatGPT referrals, set against 164 million from traditional channels. Because it's read directly from each site's own analytics, we can see which channel delivered each visitor, whether the visit converted, and how much revenue followed. The data was provided through Grips Intelligence.

All 973 are e-commerce websites — businesses that sell online — but they span far more than classic retail. The sample covers 24 product categories across every continent, and both B2C and B2B: from fashion, beauty and general retail to vehicles, finance, business services, travel and more. That breadth is deliberate — it lets us test whether oLLM behaves the same way across very different kinds of online businesses, rather than in one narrow sector.

Grips Intelligence is a market-intelligence company that supplied the underlying dataset and covered the computing costs for the study. Grips did not commission, design, or approve the research, and its business model rests on accurate, neutral data rather than flattering any platform. The representativeness of its data has been checked in prior peer-reviewed work against trusted benchmarks such as Similarweb, Shopify's quarterly reports, and the U.S. E-Commerce Census.

Yes, quite likely. We use last-click attribution — the industry standard — which credits whichever channel delivered the final click before a purchase. So when a shopper discovers a product in ChatGPT but then opens a search engine or the retailer's site to verify before buying, the sale is credited to that later channel, not to ChatGPT. This means our figures, if anything, understate oLLM's true influence higher up the purchase funnel.

No. The platforms analysed — OpenAI, Google and others — had no role whatsoever. The data and computing came from Grips Intelligence, whose own business depends on accurate, unbiased market intelligence and which did not commission the study, shape its questions, choose its methods, interpret its results, or approve the manuscript before submission. The authors declare no financial interest in making any platform look better or worse.