Here’s an experiment that ruins a lot of good afternoons. Ask ChatGPT to name the best tools in your category and screenshot the answer. Now ask Gemini the same question, word for word. Then Perplexity. Then Claude.
Four answers. Rarely the same four names.
Most teams find this out by accident. Someone in sales mentions a prospect who “asked ChatGPT and we came up first,” and everyone feels good for a day. Then a colleague runs the identical question in Gemini and watches a competitor take the top slot while the brand doesn’t appear at all. The reflex is to decide one of the models is broken. Neither is. They’re just reading from different libraries.
A model’s answer is built from two things: what it absorbed while it was trained, and what it pulls in live the moment you ask. Both of those differ from one model to the next, and the difference compounds.
Start with training. Every model was raised on a different diet of text, frozen at a different moment. One finished reading the web last spring; another kept going into the autumn. One swallowed an enormous volume of developer forums; another leaned harder on books and news. So before a single live search happens, each model already carries a slightly different picture of who matters in your category, shaped by which corners of the internet it happened to read most of.
Then comes retrieval, where the gap widens. When a model reaches out to the live web, it doesn’t reach into the same place. Gemini is a Google product and grounds its answers in Google’s index, which means the brands that have quietly done well in ordinary search tend to carry over. Perplexity wears its sources on its sleeve, pulling fresh pages and community threads and listing them in the margin. ChatGPT’s search behaves differently again. The upshot is simple and inconvenient: each model is standing in a different aisle of a different library, and it can only recommend the books on the shelf in front of it.
This is why “optimize for AI” is such a quietly dangerous phrase. It treats a crowd of very different systems as a single audience. Optimize for the one that happens to be open in your own browser and you’ve optimized for an audience of one.
And that’s the version of the mistake that hides best. A founder uses ChatGPT all day, sees the brand surface nicely there, and decides the AI-visibility problem is handled. But the founder isn’t the buyer. The buyer might live in Gemini because it’s already wired into the email and documents they use every day, or in Perplexity because they like seeing the receipts. You cannot read your coverage off the one model in your own tabs. That isn’t data. It’s a sample size of one, collected by the least objective person in the building.
So what actually travels across all of them?
Not tricks. Tricks are local. Each model patches its own quirks on its own schedule, and anything you reverse-engineered this month is rented, not owned. What travels is the boring, durable stuff: a clear and consistent account of what you do, repeated in enough credible places that every library ends up with its own copy. If your story shows up in ordinary search, Gemini finds it. If it shows up in the open conversations people actually have, Perplexity finds it. If it was established and repeated long enough to make it into the training mix, it’s already baked into the next model before that model searches for anything at all.
Breadth is the strategy. You are not optimizing for a model. You are trying to become the kind of brand that’s hard to leave out, no matter which shelf you’re read from. That’s a slower game than chasing one engine’s formatting preferences, and it’s the only one that survives the next model release.
The practical first move costs you an afternoon and a little ego. Take your five most important questions, the ones a buyer would actually type, and run each one through four models. Write down who appears where. The pattern in that grid will tell you more than any dashboard: it shows you which libraries already carry you, and which ones have never heard your name. The blanks are your map.
Because the question was never whether AI knows your category. It knows your category fine. The question is whether every version of the answer includes you, or just the one you happened to check.