How AI assistants actually pick products

AI assistants recommend a small fraction of the products they know about, and the difference is third-party evidence. being in the training data gets you recognized, not named. what correlates with actually being recommended is what the internet says about you, Reddit threads and backlinks above all. and each assistant reads a different internet, so the work is per assistant, not generic.

Recognition is not recommendation

ChatGPT can know your product exists and still never say your name. a 2026 Okara analysis found that of the products ChatGPT recognized, only around 3% were organically recommended in answers. the other 97% sit in a strange purgatory: the model can describe them if you ask directly, but when a user asks the open question, “what should i use for this”, they are simply never in the answer.

that gap is the whole game. most GEO advice is about getting known, submitting everywhere, publishing volume, stuffing your name into text. but recognition is nearly free and nearly worthless. the scarce thing is being one of the two or three names an assistant volunteers, and that is earned differently.

What correlates with being named

third-party evidence correlates with recommendation more strongly than anything you publish yourself. a MaxAEO study of 1.2M AI citations found Reddit presence correlated with being recommended at +0.40, and backlinks at +0.32. correlation is not causation, and nobody outside the model labs can prove the mechanism. but the direction is consistent across every study i have read: the products that get named are the products the rest of the internet talks about.

it makes sense from the assistant’s side. its job is to give a defensible answer, and a product that real people discuss in retrievable text is a safer recommendation than one that only describes itself. your own site sets up the answer. other people’s text closes it.

worth being precise about the backlink half of that finding, because it is easy to mishear as permission to buy links. the correlation is with the kind of links that come attached to real coverage: a review that links you, a comparison that includes you, a directory that describes you. those carry text an assistant can retrieve. a purchased link on a page nobody reads carries nothing, and the spam risk of it lands on you either way.

Different assistants read different internets

each assistant over-indexes its own sources, so a generic GEO effort under-serves all of them. Profound’s citation research shows the split clearly: Perplexity leans hard on Reddit, while ChatGPT leans on Wikipedia-style reference sources. same question, different retrieval diet, different winners.

assistant	over-indexes	what that means for you
Perplexity	Reddit threads and community discussion	earn honest presence where your users already post. the method is the SAFE reddit playbook
ChatGPT	Wikipedia-style reference sources	get described accurately in neutral, reference-shaped pages: comparisons, directories, encyclopedic writeups

the practical move is to decide which assistant your buyers actually use, and weight the work toward its diet. a consumer app whose users live in ChatGPT needs reference-shaped coverage. a technical product whose buyers research in Perplexity needs to exist in the right threads. doing a bit of everything, aimed at nobody in particular, is how you end up in the recognized-but-never-named 97%.

What a small product does with this

pick one assistant, learn its diet, and feed it deliberately. list the ten questions your buyers ask, run them through that assistant, and look at what gets cited: those domains and formats are your target list, not a generic checklist from a GEO thread. then build the two levers in order, your own answer-shaped pages first because they are fully in your control, third-party presence second because it compounds slower but weighs more.

and rerun those ten questions monthly, in fresh sessions, logging which products get named. the answers move as the assistants retune, and a lightweight log is the difference between knowing your position and guessing it.

none of this is fast, and anyone selling you a shortcut into the recommendation set is selling the recognized-but-ignored tier with better packaging.

the full method, both levers end to end, lives in the AI search playbook. and if you want this mapped onto your product, the breakdown call is free. Let’s talk Growth.