For about a day, the Selix homepage said “AI picks three winners in your category.” Five cards underneath it read 3 brands named. A closing line read there are three names, and you are either one of them or you are not.
None of it was measured. We are writing this down because we would rather be the company that publishes its own correction than the company that quietly edits the page.
Where the number came from
It came from a mockup. Early on we built an example answer for a coilover query, and the example happened to name three brands. That example became the design. The design became a headline. The headline became a claim about how AI behaves — and nobody, including us, had checked whether it was true.
This is the ordinary way a false number gets into marketing. Not a decision to exaggerate. A placeholder that nobody re-examined once it started reading like a fact.
The question that caught it
Our founder read the page and asked, roughly: why do we keep saying AI picks three — is that a real thing?
So we went looking. The published GEO research measures whether a brand gets named, not how many get named per answer. Omniscient’s study of 5,323 model outputs reports recommendation rates by awareness stage — 19%, 28%, 79% — and says nothing about counts. Their later analysis of 23,387 citations maps sources, not list lengths. We could not find a study anywhere that supports “three.”
So the claim came off the site the same day.
What it says now
The headline reads “AI names a short list. Be on it.” The cards no longer all say three — they read four, six, three, five, four, because real answers do not return a fixed count and pretending otherwise was the tell.
What survived is everything that was true without a number: the list is short, it is finite, there is no page two, and you cannot see your own absence. That last one is the actual product thesis, and it never needed the three.
The rule we wrote afterwards
We now separate two kinds of number, and treat them completely differently.
Illustrative product data — the figures in a demo store, an example scorecard, a sample report. Invented, clearly in service of showing how the interface works, and nobody can check them because they describe a fictional shop. Normal, and low risk.
Claims about how AI behaves — how many brands get named, how often, which engine does what, how many shoppers ask. These describe the world. Anyone can test them immediately. These never ship without a source, and never get invented for rhythm.
The three-brands line failed because it started life in the first category and quietly migrated to the second.
The part that stings
Selix is a company whose entire pitch is that you should know, rather than assume, what the models say about you. We shipped an assumption about what the models say. The irony is not lost on us.
It also points at something we should have been doing from the start. We are in a position to actually measure this — run the check across a few hundred real categories, count how many brands each answer names, and publish the distribution. Then the headline could carry a number nobody else has, with first-party data underneath it. That work is not done yet. When it is, it goes in the AI Shopping Index, publicly, whatever it says.
If you spot a claim on this site you think we cannot back, tell us. We would genuinely rather hear it from you than from a prospect.