AI search

What is AI search optimization?

The documented part, the measured part, and the part the industry invented.

What is AI search optimization?

AI search optimization is the practice of trying to influence whether a business or a page gets named, retrieved, or cited by AI assistants such as ChatGPT, Google AI Overviews, Perplexity and Claude. It has no first-party checklist behind it. Google states in its own documentation that there are no additional requirements to appear in AI Overviews or AI Mode, and three of the four major systems have published nothing at all about how they select sources. What exists is a small body of measured research, a much larger body of vendor assertion, and several platform statements that directly contradict tactics sold under this name.

Four different outcomes wear the same label

The phrase "AI visibility" is used for at least four things that have different causes and different levers. A model can name your business from its own weights with no citation attached. It can retrieve your page and cite it. It can retrieve your page, use a fact from it, and credit a different source. Or a reader can click through to you.

These get averaged into a single number by almost every tool that sells a dashboard. They should not be. Being named from memory is a function of how much was written about you before the training cutoff. Being cited in a live answer is a function of what the retrieval layer surfaced this morning.

Seer Interactive analysed 804,491 AI responses across 1,926 brands, 15,783 prompts and four platforms in March 2026, and proposed that the model picks the brands first from parametric memory and then goes looking for sources that support the choice it already made. They supported it with six behavioural tests across roughly 362,000 responses. If that ordering is right, a large share of on-page citation tactics are operating downstream of a decision that was already taken.

That is a hypothesis with real data behind it, not a settled mechanism, and it should be read that way.

What Google has published, including the parts vendors skip

Google Search Central publishes a guide to optimizing for generative AI features that contains an explicit mythbusting section. It is the single most useful primary document in this subject and it is quoted less often than blog posts about it.

The guide states: "There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary." Its summary tells readers to "Prioritize effective SEO strategies over 'AEO/GEO hacks'" and names the specific hacks: chunking content, creating unnecessary AI text files such as llms.txt, and pursuing inauthentic mentions.

Elsewhere the same guidance says you do not need to write in a specific way just for generative AI search, and that there is no requirement to break content into tiny pieces for AI to understand it.

Danny Sullivan, Google's Search Liaison, has said on record that SEO for AI is still SEO. That is a Google position, not an independent one, and it applies to Google's surfaces rather than to ChatGPT or Perplexity. It is still the most direct platform statement anyone has.

What each platform has actually published about source selection

The useful question is not what a pipeline diagram says. It is which company has published which claim, and where the rest is inference.

Published versus inferred versus unpublished, per system, as of 2026-09-03
SystemDocumented by the ownerNot published
Google AI Overviews / AI ModeQuery fan-out; rooted in core Search ranking; no extra requirements; no special schema neededRanking weights, candidate set sizes, passage selection
ChatGPTDecides when to search; rewrites the query into targeted searches; may issue follow-up queries; Bing and Shopify named among providersSource-selection criteria, weighting, full provider list
ClaudeBrave Search listed as a Web Search subprocessor; an approximate user_location parameterEverything about how sources are chosen. There is no local business data layer documented at all
PerplexityCrawler names and behaviourAny ranking or selection formula

The datasets do not contain the businesses being sold this

This is the finding that changed how we write about the subject, and we have not seen it stated anywhere else.

SOCi's 2026 Local Visibility Index covers 2,751 multi-location brands and more than 350,000 locations. Yext analysed 8.7 million Google search results across 2,500 US ZIP codes. BrightLocal's ChatGPT source study ran 800 manual searches across 20 verticals in 20 major US cities. Natzir Turrado, inspecting ChatGPT's internal business data, found coverage thinning sharply outside dense metro areas.

A single-location roofer in a town of fifteen thousand people appears in none of those samples. The advice derived from them is nevertheless sold to exactly that business, usually without the sample being mentioned.

That does not make the findings wrong. It makes their transfer to a one-location trade business an untested assumption, and an untested assumption is what it should be called.

What can be measured, and what is noise

SparkToro and Gumshoe ran 2,961 brand-recommendation prompts through ChatGPT, Claude and Google's AI surfaces with 600 volunteers between November and December 2025. The odds of the same brand list appearing twice came out under 1 in 100. The same list in the same order was closer to 1 in 1,000. Methodology and raw data were published.

The direct consequence is that "we moved you to position two in ChatGPT" is a claim about a quantity that does not hold still long enough to be a position. A visibility rate measured across many runs is defensible. A rank is not.

On the traffic side, Pew Research Center observed roughly 70,000 searches from 900 US adults in March 2025. Users clicked a result on 15 percent of visits with no AI summary present and 8 percent when one was. One percent clicked a link inside the summary. That is the only fully independent, non-commercial measurement in this field.

Google added generative AI performance data to Search Console in June 2026. It is the only free first-party instrument a site owner has, and the industry is busy selling third-party trackers instead.

The evidence

What is actually established, and how

Sorted by how strong the evidence is, not by how convenient it is.

ClaimBasis
Google documents that there are no additional requirements and no special optimizations necessary to appear in AI Overviews or AI Mode.Documented by the platform
Google names content chunking, AI-specific text files such as llms.txt, and inauthentic mention seeking as tactics to skip for Google Search.Documented by the platform
ChatGPT rewrites a user query into one or more targeted search queries and may issue further queries after reviewing results.Documented by the platform
Identical prompts rarely return identical brand lists. SparkToro measured under 1 in 100 across 2,961 runs.We measured this
Click-through was 15 percent without an AI summary and 8 percent with one, in Pew browsing data covering roughly 70,000 searches.We measured this
Citation selection may follow brand selection rather than drive it, per Seer Interactive across 804,491 responses.Inference, not documentation
Whether any published local AI finding transfers to a single-location business outside a dense metro.Not publicly documented
Limits

What nobody can currently tell you

Stated because the alternative is implying a certainty that does not exist.

  • Nobody outside these companies knows the source-selection criteria for ChatGPT, Claude or Perplexity. Every named pipeline stage and candidate count in circulation is inference presented as architecture.

  • No dataset in this field samples single-location businesses in low-density areas, so nothing published can be said to apply to them.

  • Whether being cited leads to any commercial outcome is unmeasured. Nobody has connected an AI recommendation to a transaction.

  • Whether the post-hoc citation hypothesis is correct. It has supporting behavioural tests from one vendor and no independent replication.

  • What ChatGPT does when it answers without searching at all. Almost the entire literature assumes retrieval happened.

Corrections

What people get wrong about this

  • AI search optimization is a new discipline with its own ranking factors.

    What is actually the case

    No platform publishes ranking factors for AI answers. The tables of weighted factors circulating for ChatGPT and Perplexity are invented, and Google states its ordinary search guidance is what applies to its own AI features.

  • You can hold a position in an AI assistant the way you hold a position in search results.

    What is actually the case

    The most rigorous independent test of stability found repeat prompts almost never produce the same ordered list. A stable position is not a property these systems have.

  • The staged retrieval pipelines described in vendor guides are documented architecture.

    What is actually the case

    Candidate counts, named stages and passage-level re-ranking steps appear in no platform documentation. Some vendors label them as hypothesis at the source, and the label gets stripped as the description is copied.

Verify it

How to check this on your own site

You should not have to take our word for any of it.

  1. Run one commercial prompt about your category ten times, logged out, fresh session, and record every business named each time. The variation between runs is the honest ceiling on what any tracking dashboard can tell you.
  2. Open the generative AI performance data in Google Search Console and compare it against whatever a third-party visibility tool reports for the same period.
  3. Request your own pages with a crawler user agent and compare the response byte length against a browser request, to confirm nothing is being served a thinner page.
Questions people actually ask

Questions we get asked constantly

Is AI search optimization the same thing as GEO or AEO?

They name roughly the same activity. GEO comes from a 2024 academic paper, AEO predates the current wave and carried over from featured snippets and voice search, and the distinctions drawn between them in vendor content are mostly branding.

Does any of this replace ordinary SEO?

Not on Google's surfaces, where Google states its normal guidance applies. For ChatGPT and Perplexity nobody has published selection criteria, so there is no documented basis for claiming a replacement discipline exists.

What should a small local business actually do first?

Confirm the crawlers can reach the site and that the facts on it are correct and consistent. Those are cheap, verifiable and not contingent on any contested theory of how these systems rank.

Sources

Where this came from

Every factual claim above traces to one of these. Each entry says what it supports and the date it was read, because platform documentation changes without notice.

Keep reading

Related

How generative engine optimization differs from SEO Does schema markup help AI search? How ChatGPT finds local businesses Research methodology

Want this measured on your own site rather than explained?

AI Search Optimization