Skip to content

Article

Getting named in an AI answer is a retrieval problem

ChatGPT, Perplexity and Google AI Overviews name brands based on retrieval: what they can fetch, identify and trust. What you control and how to measure it.

Published

Ask an AI answer engine for the best supplier in your category and you don’t get ten blue links. You get a paragraph, and your brand is either in it or it isn’t.

There’s no page two. No “we’re sixth, so we still get some clicks”. Named or absent.

Before the engine writes a word of that paragraph, it fetches. It pulls passages from pages it can reach, works out which companies those passages are about, and decides which of them to believe. A brand gets named only if it survives all three steps – and each step can be measured.

Retrieval is the step where the engine fetches before it writes

Retrieval is the stage where an answer engine collects candidate passages from an index or the live web, before the language model composes the answer. ChatGPT, Perplexity and Google’s AI Overviews all lean on some version of it.

Your brand has to pass three tests inside that stage. Can the engine fetch the page? Can it tell who the page is about, meaning the entity (a company, product or person it can tell apart from lookalikes)? And does the same claim show up in more than one place?

One concrete detail: Google says AI Overviews and AI Mode may use “query fan-out”, issuing multiple related searches across subtopics and data sources (Google Search Central, AI features documentation, last updated 10 December 2025). The answer is assembled from several searches, so you can be fetched by one and missed by another.

How each engine ranks the passages it fetches isn’t documented well enough to reverse-engineer. Treat any claim of a secret formula as a guess.

What a brand can influence

Four things sit inside your control, and the first is plain access. Google says a page needs to be indexed and eligible to be shown in Search with a snippet, and that there are no additional requirements or special optimisations for AI Overviews or AI Mode (same Google documentation, December 2025).

Access is set per crawler, and the crawlers do different jobs. OpenAI says a site that disallows OAI-SearchBot will not be shown in ChatGPT search answers, while GPTBot relates to training (OpenAI crawler documentation, checked October 2026, undated). Perplexity recommends allowing PerplexityBot, which it says is used to surface and link sites in its results (Perplexity bot documentation, checked October 2026, undated). Check all three rules in your robots.txt before you do anything else.

The second lever is entity identity: one name, one description of what you do, the same facts everywhere. The third is consistency across the sources an engine might pull from – your site, directories, profiles, third-party write-ups. My working assumption, an informed guess rather than anything the engines confirm, is that an engine trusts a claim more when independent sources agree on it. If you sell in several markets, that consistency has to survive the language change, which is one more reason international SEO is not translation.

The fourth is shape. Engines lift passages, not pages. A section that opens with its answer and makes sense on its own is easier to lift than one that needs three earlier paragraphs.

What a brand can’t control

You can’t control the model’s training data. Whatever it absorbed about your category before you fixed your site is baked in, and no robots.txt edit reaches it.

You can’t control which sources an engine prefers on a given day, because the engine doesn’t tell you. And you can’t control variance: ask the same question five times in fresh sessions and read the five answers side by side. Whether they agree is the first thing worth knowing about your category.

None of this makes the work pointless. It makes it a measurement problem with a ceiling, which is a better thing to manage than a mystery.

How to measure it

Build a fixed question set: the questions a buyer in your market actually asks, plus a few that name you directly. Start with around 25. Ask every question several times, on every engine you care about, from the same setup, and log the date and the full answer.

Then count two things. The share of answers that name you, and how accurately they describe you. A mention with the wrong product line is not a win.

A hypothetical, to show the arithmetic: 25 questions, 4 runs each, 3 engines gives 300 answers. If 30 of them name you, your share is 10 percent. The number matters less than the direction across the next few runs, and than which sources the engine cites when it does name you.

Treat it as a recurring system, not a one-off audit someone heroically assembles once. A check that only happens when one person remembers is hero mode, and it’s no way to track a moving target.

Classic search and AI answer engines compared

Classic search AI answer engines
What the user gets A ranked list of links One written answer, sometimes with cited sources
What you compete for A position A mention inside the answer
Where your control ends Your page and links to it The model’s training data and the engine’s source choices
How you measure Rankings, clicks, impressions Share of answers naming you, and their accuracy

For deeper cuts on this topic, see the AI search and GEO section.

The bottom line

Being named in an AI answer is decided before the answer is written. The engine fetches, identifies and weighs, and you can influence each of those steps through access, a clear identity, consistent facts across sources and passages built to be lifted. You can’t influence the rest, and nobody who promises otherwise has shown you the evidence.

A note on self-promotion: NON.agency is the agency I founded in 2022, and LLM visibility is part of the work we do (more on the NON.agency project page). The method above doesn’t need us. A spreadsheet and an afternoon will get you the first reading. If you want to know who’s behind it, I’ve put fifteen years of running search operations on the about page.

One thing is still unknown. There is no public data on how stable citation patterns are over time. Until there is, a single snapshot is a reading, not a result, and the only honest approach is to keep asking the same questions and watch what moves.

Questions people ask

What is retrieval in AI search?

Retrieval is the step where an AI answer engine collects candidate passages from an index or the live web before it writes its answer. If your page isn't fetched, identified and trusted at that step, it can't be named.

How do I check whether ChatGPT or Perplexity names my brand?

Write a fixed set of buyer questions, ask each one several times on each engine, and log every answer with its date. Then count how many answers name you and whether the description is correct.

Do I need special optimisation to appear in Google AI Overviews?

Google says there are no additional requirements beyond being indexed and eligible to show a snippet in Search. The work is the usual technical and content hygiene, plus clarity about who you are.

Can I guarantee my brand gets cited by AI engines?

No. You can raise the odds by being fetchable, identifiable and consistent across sources, but the model's training data and the engine's source preferences sit outside your control.

Missing visibility in Google and AI?Write to me.

Write to me