Blog · Data and trends

How much traffic AI search actually sends

Published 1 April 2026 · Updated 12 October 2026 · 9 min read

How much traffic AI search actually sends
Photograph: David Becker / unsplash

The short answer: On the sites whose logs and analytics we have reviewed, AI assistants send a small share of sessions, and the figure in your analytics can be wrong in both directions at once. It is inflated by agent fetches that never carried a human, and deflated by apps that strip the referrer so the visit lands in Direct. We have no cross-industry figure we can source, and neither does anyone quoting one. The volume is small by design. A chat answer is built to end the question, not to pass it on.

That last sentence is the whole mechanism. Everything else here is arithmetic and plumbing.

The funnel has four gates, and three of them close

Work forward from a prompt and the size of the channel stops being surprising.

A person asks something. The assistant decides whether it needs the live web at all, and for a large share of turns it does not, because the model answers from what it already holds. If it does retrieve, it fetches a handful of documents and keeps a few passages. It then writes an answer, and attaches citations to part of what it wrote. Finally the reader decides whether to click one.

Four gates, and each one multiplies. Say half of turns trigger retrieval, your page is among the sources on a tenth of those, it survives into the visible citation list half the time, and one reader in ten clicks. The compound rate is one visit per four hundred prompts. Change any assumption and the order of magnitude barely moves, because the multiplication dominates. Those four rates are illustrative, not measured: we have no source for any of them, and they are there to show how the compounding works. This is why the channel can be genuinely important and still account for a sliver of sessions.

One Google surface has the click gate measured. Pew Research Center tracked the browsing of 900 US adults on an opt-in panel across 68,879 Google searches in March 2025, of which 12,593 produced an AI summary. Users clicked a result link on 8 per cent of visits where a summary appeared, against 15 per cent where none did. Those are different sets of searches, not the same queries with and without a summary, so read it as a difference between two populations of visits rather than a controlled comparison. It is also one engine, one country and one month, and it says nothing about click behaviour in assistant chat. See AI search statistics worth quoting for the rest of the defensible set.

The first counting error: fetches dressed as visits

Open your server logs and you will find far more AI activity than your analytics shows. Most of it is not traffic.

OpenAI's crawler documentation, read again on 15 October 2026, documents four agents with separate jobs. GPTBot gathers training data. OAI-SearchBot builds the index behind ChatGPT's search. ChatGPT-User fetches a page because a person in the product asked something that needed it, and the same page says that because those actions are initiated by a user, robots.txt rules may not apply. The fourth, OAI-AdsBot, checks landing pages submitted as ads and visits nothing else. Of the four, only ChatGPT-User involves a live human, and even then the human is reading the assistant's answer, not your page. Other engines document their own agents differently, so check each one's page rather than assuming this split.

The volume on the crawl side grew fast, at least where it has been measured. Cloudflare's crawler analysis of traffic across its own network, covering May 2024 to May 2025, recorded GPTBot's share of AI crawler traffic rising from 5 per cent to 30 per cent, and its raw request volume up 305 per cent year on year. That is one network's view, not the web's, and none of it is a session. A dashboard that counts those hits as AI visibility is counting robots, and agent traffic is not bot traffic draws the line properly.

The practical damage is a ratio nobody computes: fetches per referred session. Pull one month of logs, count requests from the assistant user agents, count sessions whose referrer was an assistant domain, and divide. On every site we have looked at it comes out far above one, often by orders of magnitude, but we have not run that count on a sample large or consistent enough to quote a figure, so we are not quoting one. A high ratio is not a failure. It is the shape of the channel, and knowing yours stops you from reading a crawl spike as a demand signal.

The second counting error: visits dressed as Direct

The undercount is quieter and larger.

GA4's default channel group, as documented on 12 October 2026, now carries an AI Assistant channel: medium ai-assistant, set when the referrer matches Google's own list of AI assistants. Referral catches media of referral, app or link. Direct catches the rest, source (direct) with medium (none) or (not set), whenever the browser hands over no referrer at all. That last case is the one that matters here. A desktop app, an iOS or Android app, a link opened in an in-app browser, a copied URL pasted into a new tab: in our own testing those can arrive with no referrer, while the web clients at chatgpt.com and perplexity.ai set one. Our testing covered a few surfaces, not the whole set, and the list Google matches against is not published in full, so treat any AI Assistant row as a partial count rather than the channel.

So your reported AI referral number is best read as a floor rather than an estimate. Two ways to put a better bound on it. Add a stable query parameter to the URLs you most want cited, then watch for it arriving with no referrer. Or run a one-question self-reported attribution field at signup and compare the share who name an assistant against the share your analytics credits. The gap between those two is the size of your blind spot. We have seen it come out wide, but we have no sample behind a multiplier, so measure yours rather than borrowing ours. Attribution when there is no click goes further into what survives that gap.

Google adds its own wrinkle, and it is a wrinkle about clicks specifically. Its documentation on AI features is explicit that appearances inside AI Mode, or inside an AI Overview, roll up into the same Search Console totals as every other result, filed under the Web search type. Alongside that, Search Console does report these two surfaces on their own. The Generative AI performance report, announced on 3 June 2026 and rolled out to all sites worldwide on 31 August 2026, covers AI Overviews and AI Mode and breaks them down by page, country, device and date. What it gives you is impressions. There is still no clicks-by-surface row to filter on, so a chart claiming to isolate your AI Overviews clicks is modelling them rather than reading them, and the question to ask is what the model is.

Small channel, different buyers

Volume is the wrong single number anyway, which is the part worth taking to a finance meeting.

Somebody who clicks a citation has already read an answer. They are checking a claim, looking for the detail the summary dropped, or verifying that your company is real. That is a later stage of intent than a cold organic click, and it tends to show up as a higher conversion rate on a much smaller base. Do AI referrals convert better covers how to test that on your own data instead of borrowing someone's blog figure, which matters because most published conversion comparisons come from tooling companies measuring their own funnel, where the audience is unusually self-selecting.

There is also traffic the channel sends that never touches your domain at all. When an assistant answers from a forum thread about you, the citation points elsewhere and the reader may go straight to your pricing page by typing it. That path lands in Direct, and nothing in your analytics will attribute it back. Why Reddit gets cited so much, and what to do about it explains why so much of the citation weight sits on pages you do not own.

The check, in one afternoon

Pick a single month. From server logs, count requests by assistant user agent, separating the crawl agents from the user-initiated one. From analytics, count sessions whose referrer host is an assistant domain. From your CRM or signup form, count self-reported mentions of an assistant. Put the three counts side by side.

You now have a floor from analytics, a ceiling from self-report, and a crawl figure that belongs in a different column entirely. Report the range, say which gate each number came through, and refuse to average them. The self-check has the rest of the ten things worth testing before anyone buys a tool.

Honest limits

We have no defensible cross-industry figure for AI referral share, and this post deliberately quotes none. What we have is what we have seen in client logs and analytics during audits, a handful of sites rather than a sample anyone could call representative, with no common measurement window, so none of it is stated here as a number. The four rates in the funnel arithmetic are illustrative and sourced to nothing. The Pew figures are one engine, one country, one month and an opt-in panel of 900 people, and they describe Google AI Overviews, not assistant chat, where we know of no comparable public measurement. We also cannot size the Direct leakage from the outside at all: only your own tagging and self-report can bound it, and absence of a referrer in your logs tells you nothing about how many people read an answer about you and never arrived.

FAQ

Is AI search traffic worth chasing if the share is tiny? Often, because the comparison is not against your whole site. It is against the cost of the fixes, most of which are access and markup work you wanted done regardless, and the downstream value of being named in an answer the buyer reads without clicking at all.

Why does my Direct traffic keep rising? Referrer-stripped arrivals from assistant apps are one plausible cause, alongside the usual suspects of app links and privacy settings. We cannot tell you which it is on your site, and neither can anyone else from the outside. The way to separate them is tagged URLs on heavily cited pages, not a theory.

Can I see AI Overviews clicks in Search Console? Clicks, no. Impressions, yes. Google's documentation puts those appearances inside the same overall totals as everything else, filed under the Web search type, with no clicks-by-surface split to filter on. Since 3 June 2026 Search Console has also carried a separate Generative AI performance report, available to all sites worldwide since 31 August 2026, which reports impressions for AI Overviews and AI Mode on their own.

Should I count crawler hits as AI visibility? No. A fetch means a system read you, not that a person did. Keep the crawl log as an access check, which is what it is good for, and keep it out of the traffic number.

Sources

Related reading

AI search statistics worth quotingDo AI referrals convert betterAgent traffic is not bot trafficWhy Reddit gets cited so much, and what to do about itRunning your own AI search studyWhat changed in AI search this year