AI search is changing who gets found, not just how

For about twenty years, being found online meant one thing: appearing high in a list of links. The work was well understood, the feedback loop was visible in analytics and everyone in the industry had broadly the same mental model.

That model is now incomplete. A growing share of commercial questions get answered in a generated summary that sits above the list, cites a handful of sources and resolves the question before anyone clicks. Google began rolling out AI Overviews widely in 2024. ChatGPT, Perplexity, Gemini and Copilot all answer product and vendor questions directly, with varying degrees of attribution.

The important shift is not that answers got better. It is that the number of businesses a buyer encounters per question has collapsed.

The structural change

A traditional result page is a shortlist of ten. A generated answer is a shortlist of three or four and frequently one. If you were the seventh-best result, you were still discoverable by a motivated searcher. If you are the seventh-best source for a generated answer, you are not in the answer at all.

This produces a sharper distribution than classic search ever had. There is no equivalent of “page two” in an AI answer. You are named, or you are not.

Why your analytics will not tell you

This is the part most teams miss. When a buyer asks ChatGPT which vendors to consider and your competitor gets named, nothing appears in your analytics. There is no impression, no session, no bounce. The loss is real and completely invisible in the tools most marketing teams use to decide where to spend.

Some AI referral traffic is measurable, visits with referrers from chat.openai.com, perplexity.ai, gemini.google.com and similar do show up and segmenting them is worth doing. But referral traffic only counts the people who clicked through. It tells you nothing about the much larger set of conversations where your category was discussed and you were absent.

The only way to see that is to ask the question directly, repeatedly and record who gets named. That is a measurement discipline, not a reporting feature.

What actually seems to influence citation

Nobody outside these companies knows the ranking systems and the systems change frequently. Anyone claiming a definitive formula is guessing. But a few things are well supported by the platforms’ own documentation and by consistent observation:

Crawlability is a precondition. Different AI systems use different crawlers, GPTBot and OAI-SearchBot for OpenAI, PerplexityBot, ClaudeBot and Google-Extended for Google’s AI products. These are controlled separately from Googlebot in robots.txt. A site can rank perfectly well in Google search while being blocked from the systems that generate answers. This is a surprisingly common unforced error.

Extractable claims beat persuasive prose. A model composing an answer needs a specific, attributable statement. “We’re the trusted leader in customer experience” is unquotable. “Response time is measured from first contact and our published target is four business hours” is quotable. Pages written to be skimmed by a human in a hurry tend to contain very little a model can lift.

Entity consistency matters more than it used to. If your company name, focus and location are described differently across your site, your profiles and third-party listings, a model has a weaker basis for asserting anything specific about you. Structured data and consistent public descriptions are how you reduce that ambiguity.

Third-party corroboration carries weight. Being described by credible independent sources appears to matter considerably more in generated answers than a well-optimised page on your own domain. You cannot self-declare your way into being the recommended vendor.

What this does not mean

It does not mean classic SEO is finished. For most businesses, conventional organic and paid search still drive the large majority of commercial traffic and most of the signals that improve AI visibility also improve ordinary rankings. Treating AI visibility as a replacement rather than an additional layer is a way to lose revenue while chasing a trend.

It also does not mean any of this can be guaranteed. The same prompt returns different sources to different users, at different times, in different regions. Expect to read trends over months, not results over weeks and be sceptical of anyone who promises otherwise.

Where to start

Measure before you optimise. Write down the fifteen to twenty questions a real buyer would ask before choosing someone in your category. Run them across the assistants your customers actually use. Record who gets named, how you are described when you do appear and whether that description is accurate.

That document is your baseline. Almost everything useful follows from it and the exercise usually surfaces at least one factual error about your own business that is worth fixing immediately.

If you would rather not run it yourself, that measurement is exactly what our AI visibility audit produces.

Further reading: How AI assistants decide which local businesses to recommend and How to get your business recommended by ChatGPT.