Does AI search actually send traffic to your website?

AI search does send traffic to your website, and the volume is small enough that most teams conclude the channel is broken when it is working normally. An answer engine is built to resolve the question in the answer. A click is what happens when the answer is incomplete, when the reader wants to check a source, or when the next step genuinely lives on your site. Designing for click volume from a surface built to remove clicks is a losing setup.

Scale sets the ceiling. Similarweb's 2026 Generative AI Landscape report counted an average of 770.7 million referral visits a month from AI platforms worldwide between June 2025 and May 2026, up 117.4% year over year, and still put generative AI at roughly a fifth the size of search. A channel can more than double and stay small next to the one it is being measured against, which is why a business comparing this month's assistant referrals with its organic search sessions will always be disappointed.

The influence, though, is not small, and it does not show up as a session. A buyer asks an assistant which vendors handle their problem, reads the answer, and arrives on your site a week later by typing your name into a browser. That visit is recorded as direct or branded search. The assistant decided which three names were in the running and left no trace in your analytics that it did.

Two different questions are therefore worth keeping apart. "How much traffic does AI search send" has a real answer and it is a modest one. "How often does an engine name us when a buyer asks about our category" is the question that decides pipeline, and it needs a measurement your analytics does not provide. Our guide to tracking AI visibility across engines covers the second one.

How much do AI Overviews reduce clicks?

AI Overviews reduce clicks substantially, and two independent measurements now put a number on it. The Pew Research Center study published in July 2025 tracked the real browsing of more than 900 United States adults across 68,879 Google searches made during March 2025 and found that users clicked a link on 8% of visits where an AI summary appeared, against 15% of visits where none did.

Clicks on the summary itself were rarer still. The same study recorded a click on a link inside the AI summary on 1% of visits, which is the finding most worth sitting with, because being cited inside an AI Overview is the outcome an entire industry is now selling. Pew also found that users ended their browsing session after a search with a summary 26% of the time, against 16% without one.

Ahrefs measured the same effect from the ranking side in March 2025, comparing 150,000 keywords that carried an AI Overview against 150,000 informational keywords that did not, and reported a 34.5% lower click-through rate for the top-ranking page. Its follow-up analysis on December 2025 data put the reduction at 58% for the top result and 50.8% for position two.

Study	What it measured	Scale	Result
Pew Research Center, July 2025	Real browsing behaviour with and without an AI summary	900+ US adults, 68,879 searches, March 2025	Clicked a link on 8% of visits with a summary, 15% without
Ahrefs, March 2025	Click-through rate for the top-ranking page	300,000 keywords: 150,000 with an AI Overview, 150,000 without	34.5% lower click-through rate
Ahrefs follow-up, December 2025 data	Click-through rate for the top-ranking page	300,000 keywords, re-measured	58% reduction at position one, 50.8% at position two
Ahrefs own analytics, June 2025	Conversion quality of AI referrals	One site, 30 days	0.5% of visits produced 12.1% of signups

Is AI search traffic worth anything if the volume is that low?

AI search traffic is worth a great deal per visit, which is the part the volume numbers hide. Ahrefs published its own 30-day analytics in June 2025 showing AI search sources accounting for 0.5% of visits to its site while producing 12.1% of signups over the same period, a gap its team described as roughly 23 times the conversion rate of ordinary organic traffic.

Read that finding carefully before repeating it. Ahrefs measured one company, one product and one conversion event, in a category where its buyers are unusually heavy users of AI tools. The useful claim is not the multiple. The useful claim is the shape: a visitor who arrives after an assistant has already compared you against alternatives has done most of the evaluation before landing, so the first visit sits much later in the buying process than a first visit from a search result.

Selection explains the effect. To reach you through an assistant, a person had to ask a real question, read an answer that named you, and choose your link over the others in the citation list. Each of those steps filters out browsing. What arrives is small, late and pre-qualified, and it converts like a referral from a trusted colleague rather than like a visit from a keyword.

For a business the operational consequence is a reporting change rather than a strategy change. Assistant referrals need their own segment, judged on conversion rate and pipeline, never on session volume. A channel delivering half a percent of visits and a tenth of signups looks like a rounding error in a traffic report and like a priority in a revenue report, and only one of those two reports is telling the truth about it.

How do you see AI traffic in Google Analytics and Search Console?

Seeing AI traffic takes deliberate setup in Google Analytics, because assistant visits arrive as ordinary referrals rather than as a search channel. Build a segment on the referring hostnames and keep it in one place: chatgpt.com, perplexity.ai, gemini.google.com, copilot.microsoft.com and claude.ai cover most of the volume today, and the list needs revisiting each quarter as assistants change how they link out.

Expect the segment to undercount. Some assistants strip the referrer, some route clicks through their own redirect, and a citation read on a phone often produces a branded search on a desktop an hour later. Undercounting is not a reason to skip the segment, but it is a reason to avoid treating its number as the size of the channel.

Search Console now covers part of the gap on Google's own surfaces. Google introduced generative AI performance reports on 3 June 2026 and, according to Search Engine Land's coverage of the global rollout, completed the worldwide release on 31 August 2026. The report shows impressions in AI Overviews, AI Mode and generative features in Discover, broken out by page, country, device and date.

What that report does not show matters as much as what it does. There are no clicks, no click-through rate, no queries and no average position, and the data is absent from the Search Analytics API and the BigQuery bulk export, so a monthly manual CSV is currently the only export. It is also Google only, which means it says nothing about ChatGPT, Claude or Perplexity. Our guide to automating AI visibility reporting covers what to run alongside it.

How do you verify that a session really came from AI search?

You can verify the part of AI referral traffic that arrives labelled, and you cannot verify the rest, so read the number your analytics reports as a floor rather than as the size of the channel. The labelled part is a session whose referrer names an assistant host, or whose landing URL carries a tracking parameter the assistant added when it linked out. Everything else lands in Direct or in branded search and is indistinguishable from the rest of it: the assistant that sends no referrer header, the click routed through a redirect, and the reader who saw your name in an answer on a phone and looked you up on a laptop an hour later.

A verification procedure that survives being challenged in a meeting has four steps and takes an afternoon. First, build the analytics segment on assistant referring hostnames described in the section above, and write down both the host list and the date you set it, because the list goes stale as assistants change how they link out. Second, check the same period in your server logs and not only in analytics, since a log records requests that a page tag can miss. Third, rule out the obvious false positive: a tagged URL that a search engine has indexed still carries its tag when somebody clicks it from the search results, so a session whose source says an assistant while its referrer says a search engine is not an assistant click. Fourth, report engaged sessions rather than raw sessions, for the reason in the next section.

How big is the part of AI search traffic you cannot see?

The part of AI search traffic you cannot see is the larger part, and there is a neutral measurement of its shape rather than a guess. Cloudflare publishes a crawl-to-refer ratio on Radar, the number of HTML pages a platform's crawlers fetch for every visit it sends back, and reported Anthropic at roughly 70,900 page requests per referral for the week of 19 to 26 June 2025, with Mistral at the other end of the range sending about ten referrals for every page it fetched (David Belson and Sam Rhea, The crawl before the fall of referrals, Cloudflare, 1 July 2025, read at source 6 November 2026). Cloudflare attaches its own caveat to it, that Claude's native app sends no referrer header, so a ratio of that shape is partly a count of missing labels rather than of missing visits, and the live figure moves: our page on whether to block AI crawlers works through what the ratio does and does not justify.

What you should not do is infer the unlabelled half from a rise in direct traffic. That inference has been tested and it failed. Kevin Indig compared GA4 data across 12 mostly B2B sites, excluding homepages, pricing and login pages, and reported that from January to September 2026 total Direct rose 140.9% while engaged Direct moved by -1.0% on monthly counts, and that across 632 tests of AI visibility against Direct and branded search at lags of zero to six weeks he could not detect hidden AI traffic without self-reported attribution (Does AI drive more direct traffic?, Growth Memo, 19 October 2026, read at source 6 November 2026). Total Direct rising while engaged Direct stays flat is the signature of automated traffic rather than of buyers, which is why the fourth step above is the one people skip, and whether your direct traffic is hidden AI traffic goes through the rest of that study.

So report the labelled sessions as labelled sessions, with the host list and the dates that produced them printed next to the number. The one change that would close the gap is a question on your own forms asking new leads where they first heard of you. Self-reported attribution is the only method the study above leaves standing.

What should you do if AI answers the question instead of sending a click?

When AI answers the question instead of sending a click, the response is to compete for the answer rather than for the session. Being the source a model quotes puts your name and your framing in front of the buyer at the moment they are deciding who is worth considering. That is the same job a first-page ranking used to do, delivered without a visit, and it is worth having even when nobody clicks.

Three things follow for how pages get written. Answer the buyer's question in its own words in the first sentence under the heading, because that sentence is the unit an engine lifts. Keep each section able to stand alone, since a reader arriving from a citation never saw the section above. Put a named source in the same paragraph as every number, because an unsourced figure reads as invented to a model checking whether a page supports a claim.

Then change what you report on. Track how often each engine names you on the questions your buyers actually ask, which competitors get named instead, and which of your pages the engines quote when they do cite you. Those three measures move when your work is landing. Sessions from assistants will stay small while all three improve, and a team grading itself on sessions will conclude a working program has failed.

The channel does still convert, so keep the conversion path short for the few who arrive. A reader who follows a citation is close to a decision and is the wrong person to route through a newsletter signup. Our AI citation checklist covers the page-level work, and the SIGNALS framework explains which dimensions carry the most weight when an engine picks between two pages that both answer the question.

What else do people ask about AI search traffic?