Most writing about AI visibility treats "the AI" as one thing with one set of rules. It is not. Six assistants answer the same question in six different ways, from six different retrieval pipelines, and they routinely disagree about which brands deserve a sentence.
This post is what each of them actually does, why the disagreement is the useful part, and what determines whether your name ends up in the answer rather than the footnotes.
Naming is not ranking
Start with the distinction everything else rests on.
A search result is a position in an ordered list, retrieved from an index. It is stable enough that checking once tells you something true.
Being named is different. The assistant assembles prose from passages it retrieved, plus whatever the model already knows from training, and decides which brands earn a mention in the sentence a person reads. Those are separate events. A page can rank first for a query and never be named in the answer above it. A brand can be named in an answer where none of its own pages were fetched.
That second case is worth sitting with. The model has parametric knowledge, so a brand with enough presence in its training data gets named without anyone reading its site today. Pleasant, and precarious: nothing you publish this quarter is holding that position up, and you will not see it slip until it has.
We report named and cited as separate states for exactly this reason.
Six pipelines, described plainly
ChatGPT runs its own search product. OpenAI documents its crawlers separately by purpose, with OAI-SearchBot handling search retrieval and GPTBot covering training. Blocking one does not block the other, and teams routinely block the wrong one.
Claude added web search in 2025, reached through the API. In our measurements Claude cites fewer sources per answer than the others and is noticeably more willing to say it does not know, which makes an absence there less informative than it looks. On many category prompts it names nobody at all, and we record that as no answer rather than as a loss.
Perplexity is built around citation. It documents PerplexityBot and its other agents publicly, and it attributes more heavily than anything else on this list. That makes it the clearest read you can get on whether your pages are actually being retrieved, as opposed to whether your brand is merely known.
Gemini draws on Google's index and sits inside products a large part of the world already has open, which makes it the assistant most likely to answer a question the asker never framed as a search.
Google AI Overviews and Google AI Mode are two surfaces, not one. Google launched AI Mode as its own tab and has kept extending it. Both are, in Google's own framing, rooted in the core Search ranking and quality systems rather than running a parallel ranking system. They still disagree. We have measured the two naming different brands for the same prompt on the same day, which is why we report every engine separately and never publish a blended score.
The gate before the decision
Before any of this matters there is a binary question: can the assistant reach you at all.
For Google, generative features require a page to be indexed and eligible to appear with a snippet. Google's guidance is that nosnippet, data-nosnippet, max-snippet and noindex restrict how your content is featured in AI experiences. One legacy directive left on a template can disqualify a whole section of a site from being named anywhere.
For the rest it is robots.txt. Google publishes its full crawler list including Google-Extended, and the robots.txt specification is the mechanism. Four lines of audit, rarely done, and it is the difference between being considered and being invisible.
You are answering questions nobody typed
The retrieval set is wider than the prompt.
Google generates a set of related queries alongside the one a person typed and fetches results for all of them, then assembles the answer from that broader pool. The practical consequence is that a page answering one narrow question thoroughly can be pulled into an answer for an adjacent question it never targeted, and a page targeting a broad head term can miss entirely because nothing on it answers any of the specific sub-questions.
This is also why prompt selection is a methodological decision rather than a setup step, and why our methodology treats a rewritten prompt as a new series rather than a continuation of the old one.
So what actually gets you named
From what we see in real answers, in rough order of leverage:
- Retrievability. Binary, cheap to fix, and the reason a surprising number of brands are absent.
- A passage that answers the question cleanly. Assistants lift. A page that answers "how much does X cost" in one clear paragraph gets used; the same answer split across a table and three footnotes does not.
- How third parties describe you. The pages cited in your category are mostly not yours. Roundups, comparison posts and forum threads brief the model before your homepage does, which is why citation tracking is a separate measurement here rather than a sub-metric of mentions.
- Entity clarity. A brand name that is also a common phrase is harder for a model to resolve to one company. This is unglamorous and it compounds.
Notice what is absent from that list. There is no schema trick. Google states plainly that structured data is not required for generative AI search and there is no special markup to add. Anyone selling you AEO markup for AI Overviews is selling something Google says does not exist.
How to tell any of this apart
You cannot diagnose a naming problem from one check, because assistants generate fresh each time. Two runs an hour apart can name different brands with nothing having changed.
What works is dull and effective. Fix a prompt set. Run it on each assistant on a schedule. Store every answer in full with its citations, so that any figure you later quote can be opened and checked. Report per engine with the denominator attached.
For one engine end to end, the how to rank in ChatGPT answers guide walks through the six steps in order, starting with the robots.txt check most guides skip. For the definition underneath all of this, see what AI visibility is. And if you would rather watch it than run it, every plan here carries all six engines from the first day, because the engine you are losing on is rarely the one you would have picked.
The disagreement between assistants is not noise to be averaged away. It is the entire diagnostic.