There is a very specific moment happening inside marketing teams right now. Someone finally gets around to checking whether AI knows the company exists. They open ChatGPT, type something close to what a buyer would ask, and the company name comes back in the answer.
Relief all round. The AI thing is handled. Back to the campaign calendar.
Then we run that same question across five engines and the company appears in exactly one of them.
This is not a fluke and it is not a broken setup. Visibility in one engine tells you almost nothing about visibility in any other, and quietly assuming otherwise has become one of the more expensive habits in B2B marketing.
The short answer: AI visibility does not transfer between engines. ChatGPT, Claude, Gemini, Perplexity and Google AI Overviews each read different sources, update on different schedules, and weigh trust differently. A company can be the most-cited source in one and completely absent from three others. Measuring one engine tells you about that engine and nothing else.
Five Engines, Five Completely Different Answers
It is tempting to talk about "AI" as though it were one thing. Your buyers do not experience it that way. Some of them live in ChatGPT. Your technical buyers are increasingly in Claude. Perplexity has taken a real bite out of research-stage queries. Gemini shows up by default for anyone already inside Google''s world, and AI Overviews reach an enormous number of people who never made a decision to use an AI product at all.
That is five separate products, each making its own independent decision about whom to credit. They differ in what they pull from, how recently they pulled it, how much weight they hand a given source, and how often they refresh what they think they know. Ask the same question, the same way, on the same Tuesday afternoon, and you will routinely get different companies named back.
Which is why a single "AI visibility score" is not a summary of anything useful. It is an average of five different realities, and averaging is precisely the operation that destroys the thing you needed to look at.
A brand can be the single most-cited source in one engine and completely absent from three others. One blended number is designed, whether anyone intended it or not, to hide exactly that.
Why They Disagree
Four differences do most of the work here.
They are reading different things. Every engine has its own view of the web, assembled at its own moment, by its own crawler, under its own rules. A page sitting comfortably inside one engine''s working set may not exist inside another''s at all.
They are running on different clocks. Some reflect a change to your site within days. Others carry an older picture of your company around for months. This is why repositioning shows up so unevenly. One engine has already learned the new category, another is still confidently describing the business you stopped being two years ago.
They trust different sources. One engine leans on a trade publication comparison piece. Another prefers a community thread. Another goes to documentation. The source that wins a question in one place can be completely irrelevant in the next.
Some of them are not looking anything up. That last one is the big one, and almost nobody selling AI visibility talks about it.
The Two Streams, and Why Most Tools Only See One
Every AI answer arrives by one of two routes.
In the first, the engine goes and searches the web right now, reads what it finds, and writes you a summary. Live retrieval. This is what everyone pictures, and it is the only route a mention-tracking tool can observe, because it is the only one that leaves footprints a scraper can follow.
In the second, the engine does not search at all. It answers from what the model already knows, straight out of training. Nothing was fetched. Nothing was checked.
Your position on those two can be completely different, and the gap between them tells you something no single number ever will. We call that gap Stream Divergence, and we treat it as a durability read.
If you are strong when the engine searches and gone when it does not, your visibility is rented. It survives only as long as the engine keeps looking things up, and it is exposed to any change in how aggressively a given product decides to retrieve. If you hold up under both, you own something considerably more durable.
This also happens to be the surface that agent workflows increasingly run on. As more software talks to models directly instead of through a chat window, the question stops being what the engine can find and becomes what it already believes.
Visibility that only exists during live retrieval disappears the second the engine stops searching. You will not get a warning.
The Failure Is Usually the Same Underneath
Engines behave differently, but the reasons a brand fails to get cited sit on one of four layers, and those layers cascade. A break at the top makes everything below it irrelevant, which is why the order you repair them in matters far more than the length of the to-do list.
Access. Can the crawler physically reach the page? A server throwing a 403 at AI crawlers closes the door before a single word of content matters. This is much more common than it should be, usually because somebody blocked AI crawlers during a panic about scraping in 2024 and nobody has looked at it since.
Structure. Can the engine parse what it finds? A page that renders entirely through JavaScript, with no schema, forces the engine to guess. A significant number of modern marketing sites are effectively a blank page to anything that does not execute scripts, and plenty of AI crawlers do not.
Semantics. Does it describe you correctly? This is where brands get filed under a category they abandoned years ago, and it is the most damaging of the four because it is invisible from the outside. You are showing up. Just in the wrong answers, described in the wrong terms, to buyers who read that and conclude you are not what they need.
Authority. Does the engine trust the source enough to actually quote it? Usually the honest answer is that it trusts your competitor''s self-authored comparison page more, because your competitor bothered to write one and you did not.
Working out which layer is broken turns "we seem to be under-cited" into a specific, ranked, fixable cause. Without that diagnosis you are publishing more content and hoping, which is an expensive way to hope.
What Measuring This Properly Actually Involves
The honest version is not clever. It is just laborious, which is exactly why so many tools approximate it instead and call the approximation a score.
You take the questions your buyers genuinely type at the moment they are choosing a vendor. Not keyword proxies. The real ones. You run each of them on every engine. You run each one at least three times, spaced at least a day apart, because these systems are not deterministic and a single run is an anecdote wearing a lab coat. You run each on both routes, retrieval and memory. And you keep the prompt, the full answer, and the exact list of cited sources for every single run.
Then you freeze that question set and run the identical panel again next month.
Freezing is the part people skip and it is the part that makes the whole thing mean something. If the questions quietly drift, you can no longer tell genuine improvement from a friendlier list of questions.
What comes out the other side is a separate reading per engine: where you turn up, how you get described, who gets named in your place, and whether any of it survives the engine deciding not to search.
What You Do With the Answer
Two things usually become obvious within about ten minutes of looking at it.
The first is that the gaps are not evenly spread, so the work should not be either. If you are absent from two engines and strong in three, the two are where the return is, and the reasons they are weak are almost always specific and fixable rather than a vague shortage of content.
The second is that a decent amount of what you need already exists on your own site, sitting in the wrong place. This is the most common finding in our audits and the cheapest thing to act on. Companies regularly have a genuinely good answer to a buying question parked on a page where nobody is asking that question, while a competitor owns the page where it actually gets asked. Moving it is an edit measured in days, not a content programme measured in quarters.
The Bit Worth Remembering
Checking one engine and feeling reassured belongs in the same category as checking one keyword and declaring the SEO healthy. It is not a small sample. It is one out of five systems that are all making up their own minds independently.
The question worth answering is not whether AI knows who you are. It is which engines name you, which ones name somebody else instead, what they say about you when they do bother, and whether any of it holds when the engine stops looking things up.
That is what our AEO and GEO work measures, and the free visibility check runs your real buyer questions across all five engines so you can see the spread yourself.