Blog · Data and trends

How buyers are actually using AI search

Published 9 June 2026 · Updated 8 October 2026 · 7 min read

How buyers are actually using AI search
Photograph: atelierbyvineeth ... / unsplash

The short answer: Buyers use AI search to compress the middle of the funnel. They ask a loose question, get a handful of named options, then spend three or four follow-up turns adding constraints until the list is short enough to act on. Very few of those turns end in a click. The decision that matters to you happens inside the conversation, before anyone reaches your site, and it is decided by whether your constraints are stated anywhere a machine can read.

That description is less dramatic than the headlines and more useful. Nobody has abandoned Google. What has changed is which part of the buying process happens in a chat window.

The measured behaviour, not the vibe

Start with the one dataset we know of that watched real people instead of asking them. Pew Research Center analysed the browsing of 900 US adults who had agreed to share their activity, over March 2025, covering 68,879 Google searches. An AI summary appeared on 18 per cent of them. On those pages, users clicked a traditional result on 8 per cent of visits, against 15 per cent where no summary appeared. Clicks on the sources cited inside the summary landed at 1 per cent. Browsing ended on 26 per cent of summary pages, against 16 per cent without.

Read that carefully, because the usual reading is wrong. The click halved. It did not vanish. A quarter of sessions ended, which means three quarters continued somewhere, and the continuation is the part worth owning. We collect the numbers that survive this kind of scrutiny in AI search statistics worth quoting.

The assistant side of the picture comes from OpenAI's own analysis of ChatGPT usage, published through the National Bureau of Economic Research in September 2025, which classified a representative sample of conversations with an automated pipeline. Three categories, practical guidance, seeking information and writing, made up close to 80 per cent of conversations. Non-work messages rose from 53 per cent of usage to more than 70 per cent across the period studied. The paper puts adoption at roughly a tenth of the world's adults by July 2025.

Practical guidance and seeking information are the two buckets a buyer sits in. They are also the two buckets where a brand name is a legitimate part of a helpful answer.

The prompt is a brief, not a keyword

A keyword was two or three words and carried no conditions. A buying prompt carries most of the deal: team size, budget ceiling, the tool it has to sit next to, the country, the compliance regime, the deadline. "Project tool" was a keyword. "Something my eleven-person team can run without an admin, that exports to Jira, under 2,000 dollars a year" is a brief.

Google's AI features documentation says that AI Overviews and AI Mode may use a query fan-out technique, issuing multiple related searches across subtopics and assembling the answer from the results. That is Google documenting its own two surfaces. No other engine publishes a comparable description, so treat fan-out as a Google mechanism rather than a law of the category. Where it applies, the consequence is that your page is never competing for the whole question. It is competing, separately, for each clause. If your pricing lives in a PDF and your integration list renders only after JavaScript runs, you are absent from two of the fan-out subqueries while still ranking perfectly well for your brand name. That gap between brand visibility and clause visibility recurs in the audits we run, which cover a self-selected set of sites that came to us rather than a sample of the web.

The terminology argument around this work matters far less than the mechanism. If the labels are still being debated in your team, AEO vs GEO vs SEO: what the three terms actually mean settles it in one read.

Where brands actually get dropped

In the prompt ladders we run during audits, a shape repeats. The loose first question returns a handful of names, usually the obvious ones, and each added constraint shortens the list. We do not keep a counted tally of how far it shortens, so take the shape and not a number. Elimination, rather than ranking, is what we can see happening. We cannot see why a name is dropped, because these surfaces do not expose their reasoning. What we can see is that the names surviving a constraint turn tend to be the ones whose answer to that constraint is written down somewhere fetchable.

Four ways we have watched brands fall out of a list they started in.

Silence on a constraint. No published price band, no stated minimum seat count, no named region. In our ladders, the unstated constraint is the one a brand tends not to survive.

Contradiction across your own pages. A pricing page saying one thing and a comparison page saying another gives the model two versions and no tiebreaker, so it hedges or moves on.

Staleness with a visible date. A visibly old date on an integrations page gives a reader, human or machine, a reason to discount it. We have not run a controlled test of dated against undated pages, so this one is a judgement, not a measurement.

Inaccessibility. The passage exists and nothing can fetch it. Run Can AI crawlers actually read your site before you blame your copy, because a fetch problem looks exactly like a content problem from the outside.

The click you do get is a different animal

Google's own documentation on AI features states that clicks originating from results pages with AI Overviews tend to be higher quality, with users more likely to spend longer on the destination. Treat that as a vendor claim about its own surface, because it is one. It is consistent with what we see in the analytics of a small number of client sites, where AI-sourced sessions are fewer and longer. That is a handful of properties we happen to have access to, not a sample anyone could audit, and we give the direction rather than a figure because we have not measured it to a standard worth quoting. Either way it is the reason raw session counts are a bad way to judge this channel. The fuller argument sits in Do AI referrals convert better.

Measurement got slightly easier in June 2026, when Google Search Console began reporting generative AI performance separately. Separate reporting means you can finally tell whether a traffic decline is a ranking problem or a surface problem. It does not tell you what was said about you, and that remains the gap no analytics product closes.

A test you can run this week

Pick six prompts a real buyer would type, not six that flatter you. Then run the ladder.

  1. Ask the loose version. Record every brand named, in order.
  2. Add the constraint you win on. Note whether you survive.
  3. Add the constraint you are quietest about, usually price or a specific integration.
  4. Ask the engine directly which source told it that.
  5. Repeat the whole ladder on a second engine, and again two days later.

You are looking for the turn where you disappear, and then for the sentence on your own site that should have kept you in. Often there is no such sentence. Writing it is usually an afternoon of work. The self-check lists ten more tests in the same spirit, none of which need tooling.

Honest limits

The Pew figures are US browsing from March 2025, one country and one engine, and AI Overviews have changed since. The OpenAI analysis is a vendor reporting on its own product, with no independent replication, and it covers all conversations rather than purchase intent specifically. Neither dataset separates a buyer from a student, so the step from those numbers to "this is how buyers behave" is our inference and not a finding. The ladder behaviour described here comes from the prompt ladders inside our own audits, which are not logged as a counted dataset, which is why no turn-by-turn figure appears in this post. The dated-page claim is untested, and the fan-out mechanism is documented by Google for Google's own surfaces.

We also cannot connect a mention to revenue with any honesty. These surfaces pass little or no referrer, so the link between being named in turn three and a form submission six days later is inference, not measurement. Anyone presenting that link as a tracked number is presenting a model.

FAQ

Are buyers replacing Google with ChatGPT? Mostly not. The observed pattern is additive: an assistant for the research and comparison phase, a conventional search for navigation and for checking a specific claim. Plan for both surfaces rather than betting on one.

How many follow-up turns should I plan for? Enough to assume the buyer will add constraints. In the ladders we run, the list shortens with each constraint added, so the constraint you never published is the one that costs you.

Does a low click rate mean AI search is not worth the work? It means the click is the wrong unit. The mention is the impression, and the measurable outcome is whether you survive the constraint turns with an accurate sentence attached.

Can I see what a full assessment of this looks like? The sample audit report walks through the structure of one, including how prompt ladders and evidence are presented. It is a sample, not a client document.

Sources

Related reading

AI search statistics worth quotingDo AI referrals convert betterAEO vs GEO vs SEO: what the three terms actually meanCan AI crawlers actually read your siteCitation churn, and how fast the cited set changes