AI recommendations are replacing traditional search clicks because the answer, and often a shortlist of named brands, now arrives before the click. On a classic results page a buyer had to open several sites to build a shortlist; in an AI answer the assistant has already done that reading and names a few options. Where it has been measured, the effect is large: in Pew's browsing data people clicked a traditional result in 8% of visits when an AI summary appeared, against 15% without one, and a randomized experiment found that AI Overviews cut outbound organic clicks by 39.8% where they appeared. The click is not disappearing everywhere, but on the questions where an answer appears, being named in it increasingly decides whether you are considered at all.
How the discovery loop changed
For two decades the loop was the same: type a query, scan a list of links, open several pages, and build your own shortlist from what those pages said. The click was where the research happened, so the click was what search marketing measured.
An AI answer moves the research in front of the click. The engine retrieves pages, lifts the passages that answer the question, and composes a response that often names specific brands. Google describes AI Mode as issuing multiple related searches across subtopics and data sources at once before it writes the answer, a technique it calls query fan-out. By the time the buyer reads the response, the comparison has been done for them. The click that remains is more often a check or a purchase than research.
| Classic results page | AI answer | |
|---|---|---|
| Who reads the pages | The buyer, one tab at a time | The engine, before the buyer sees anything |
| Where the shortlist forms | In the buyer's head, after several clicks | In the answer text |
| What the remaining click is for | Research | Checking a name, or buying |
| What a brand competes for | A position among ten links | One of a few names, and a citation |
| Where it shows up in analytics | Organic search sessions | Often nowhere: no click, or a later direct visit |
Three consequences follow, and the rest of this article is about them:
- The shortlist forms inside the answer, so being named matters more than ranking for the head term.
- What an engine can lift from your pages, and confirm on other sites, decides whether you are named.
- Measurement has to move to the answer side, because a recommendation that produces no click leaves no trace in analytics.
What the evidence says about AI answers and search clicks
The clearest evidence comes from Google Search, because that is where researchers can compare searches with and without an AI answer on the same page.
- Pew Research Center tracked browsing by 900 U.S. adults in March 2025. With an AI summary on the page, users clicked a traditional result in 8% of visits, against 15% without one. They clicked a link inside the summary in 1% of visits, and ended their browsing session on 26% of pages with a summary, against 16% without.
- A randomized experiment by Saharsh Agarwal and Ananya Sen hid AI Overviews for a random group of users. Where an AI Overview appeared, outbound organic clicks fell 39.8% and zero-click searches rose 34.5%, while sponsored clicks were unaffected. It is a working paper, so treat the exact figures as provisional, but random assignment makes it the strongest evidence that the AI answer itself causes the drop rather than merely appearing on different kinds of query.
- Ahrefs found that the presence of an AI Overview now correlates with a 58% lower average clickthrough rate for the top-ranking page, up from 34.5% in its study a year earlier.
The Ahrefs re-run also shows how far the click on the top result has fallen where an answer sits above it. For keywords with an AI Overview, position-one clickthrough rate went from 0.073 in December 2023 to 0.016 in December 2025, while informational keywords without one stood at 0.039. That is a correlation across keywords rather than a controlled test, but it points the same way as the randomized study.
Our own site shows the same shape, though we cannot attribute it to AI answers. Over three months insightwonder.com received 45,793 impressions and 82 clicks in Google Search Console, and 309 queries containing "best" or "top" produced 7,909 impressions and 0 clicks. Those are exactly the comparison questions an AI answer now tends to settle on the results page. We read that as an illustration, not as proof.
For a side-by-side reading of Google's own figures against these studies, see what Google's AI search numbers do and do not show.
Why AI recommendations replace clicks, not search
The honest version of this article's title carries a qualifier. AI answers are replacing clicks on the queries where they appear, not replacing search as a whole. Google says AI Mode is driving an incremental increase in Search queries overall, and in Pew's study 18% of Google searches produced an AI summary, so most of the searches it observed showed no AI summary at all.
Traffic sent by AI assistants is also growing from a small base. Adobe reported that traffic to U.S. retail sites from generative AI tools in the 2025 holiday season rose 693.4% on the year before, while noting that the base of users remains modest. A large percentage on a modest base is a direction of travel, not a replacement for the channels you already have.
Attention to the answer, on the other hand, is already broad. In a Pew survey in February 2026, 60% of U.S. adults said they read the AI summaries at the top of search results, and 30% said they do not. Whether or not anyone clicks, the answer is being read, and that is why what it says about you matters.
What gets a brand named in the answer
If the shortlist forms inside the answer, the work is to be eligible for it. None of this guarantees a mention, because assistants decide what to name answer by answer. What you control is whether the engine can find, read and trust what it needs.
- Let the right agents in. OpenAI's documentation describes three agents with different jobs: GPTBot gathers content that may train its models, OAI-SearchBot builds the index behind ChatGPT's search features, and ChatGPT-User fetches a page because a person asked for it. One rule that blocks all three removes you from ChatGPT's search answers. The AI crawler access guide covers the other vendors.
- Serve the content without JavaScript. Most AI crawlers do not execute JavaScript, so a page whose text appears only after client-side rendering is an empty shell to them, and your own analytics will never tell you.
- Write passages that survive being lifted, and put evidence in them. The original GEO paper, accepted to KDD 2024, tested rewriting strategies on a benchmark of 10000 queries. It reports that such methods can boost visibility by up to 40% in generative engine responses; adding citations to sources, quotations and statistics performed best, while keyword stuffing offered little to no improvement. The authors also found the effect varies by domain, so read it as a direction, not a forecast.
- Be corroborated somewhere other than your own site. A model needs the same fact from independent places before it will state it confidently, and a claim that exists only on your own domain is the weakest evidence you can offer. Third-party corroboration covers how to find and correct those sources.
The selection step itself, and why one brand is named over another, is covered in how AI decides which brands to name.
How to measure a recommendation that leaves no click
A buyer who reads your name in an answer and later types your address arrives as direct traffic, if they arrive at all. Nothing in that visit points back to the answer. So the recommendation has to be measured where it happens: in the answers themselves.
- Fix a question set. Write the questions a buyer would ask without knowing you exist, keep them stable between runs, and record the intent of each. Building a prompt set explains how.
- Exclude questions that contain your brand name. Ask whether a named company is any good and the assistant will name it. In one of our own runs 13 of 15 generated questions contained the brand name, and the score read 13% when the honest value was zero.
- Report each engine separately. In our own benchmark one model named 3.5 providers per answer and another 2.1. An average across them describes neither situation.
- Keep cited and recommended apart. Your page can be cited as a source while the answer recommends a competitor. Count only sources marked in the answer text: counting everything the engine retrieved inflated our own citation figures by two to four times before we fixed it.
- Use Google's own report for Google. Search Console's generative AI performance report shows where your pages appear in AI Overviews and AI Mode, by page and country. It reports impressions, not clicks.
What being cited is and is not worth
A citation link inside an AI answer does not give back the click it replaced. Pew found that people clicked a link inside an AI summary in 1% of visits. Seer Interactive found that on informational queries, being cited in an AI Overview brought +120% more organic clicks per impression than not being cited, but still trailed queries without an AI Overview by -38%. Being cited beats not being cited. It does not restore the old traffic.
That is the real shift. The value of appearing in an AI answer is mostly that you are on the shortlist the buyer reads. Whatever that is worth shows up, if at all, in places analytics does not credit to the answer, which is why the answer side has to be measured directly and AI referral sessions treated as one signal among several rather than the scoreboard.
If you would rather not run those checks by hand, InsightWonder's AI visibility tracking runs a fixed question set across engines and reports each one separately.
Sources
- Pew Research Center, Google users are less likely to click on links when an AI summary appears (July 2025) and survey on reading AI summaries (June 2026).
- Saharsh Agarwal and Ananya Sen, field experiment on AI Overviews, SSRN working paper (April 2026).
- Ahrefs, AI Overviews reduce clicks, updated study (February 2026).
- Seer Interactive, AIO impact on Google CTR, 2026 update (April 2026).
- Google, AI Mode and query fan-out (March 2025), Q2 2026 earnings remarks and Search Generative AI performance reports (June 2026).
- Adobe, 2025 holiday shopping season (January 2026).
- Aggarwal et al., GEO: Generative Engine Optimization, KDD 2024.
- OpenAI, Overview of OpenAI crawlers.
An earlier version of this article relied on figures from aggregator sites. Every number we could not trace to the organisation that produced it was removed, along with three case studies we could not verify; the two figures kept from that version, the GEO paper's up to 40% and Adobe's 693.4%, were checked at the source.