Ask an AI for a recommendation and you get the usual suspects, every time. Say “what’s a good moisturizer for sensitive skin?” and CeraVe strolls out, cool as you like, same as always. I did exactly this the other week and ended up half-laughing at the screen — of course it’s the same name again.
And if you sell something nobody’s heard of, this is the part where you start to give up. “A no-name product like mine is never getting recommended by an AI” — that’s the quiet conclusion. Search, AI, doesn’t matter; the famous brands seem to own the prime real estate either way.
But I think that surrender is premature. The AI does play favorites with famous brands — that part’s true. The favoritism is just a lot more fragile than it looks. So let’s dig into where it breaks.
Does The AI Really Only Pick The Famous Ones?
The people who went at this head-on are Chu and Hou, in a study posted in June 2026 (R1). They sat three AIs — GPT-4o-mini, Claude, and Gemini — down and made them recommend skincare products, over and over. One real, famous brand (think CeraVe) lined up against nine made-up ones, and the question “which would you recommend?” fired at them hundreds of times.
So here’s what came out. When they matched every spec so the fake brands and the famous one were dead level, the AI recommended the famous brand 100% of the time. A clean sweep — total monopoly. When the information ties, the AI breaks the tie with the name it knows.
But there’s a twist hiding in there. When the team broke down what actually moved the ranking around, 82.4% of it came from the product itself — the specs, the rating — and brand fame accounted for a measly 1.2%. So fame isn’t the engine. It’s a tiebreaker, the thing that only shows up when the substance is a perfect draw.
A 0.1-Star Edge And The Monopoly Falls Over
The next experiment is what backs up the “it’s just a tiebreaker” reading. They nudged the fake brand’s rating a hair above the famous one’s. Just a tenth of a star on a 5-point scale (a +0.075 to +0.1 bump) — and the 100% monopoly the famous brand had a second ago collapsed. Now the unknown brand got picked 64-80% of the time instead.
And here’s the interesting bit: fame matters most in the mushy middle — when everything’s roughly the same quality and there’s no clear winner. When the signal’s vague, the AI gives up and leans on the name. Give it even a sliver of “this one looks a touch better,” though, and it switches without a second thought. The AI’s loyalty is about as deep as a coffee-shop punch card.
That’s not a bad deal for the unknown brand. The story isn’t “you lose on fame, so don’t bother.” It’s “line up a thread of evidence that you’re a little better on rating or substance, and the AI will happily point at you.” The wall’s there — it’s just a step-over, not a climb.
”As Endorsed By Experts” Works. “Only 3 Left!” Doesn’t.
So what do you actually add to get picked? The team stuffed different marketing moves into the fake brand’s description and compared what landed.
The runaway winner was credentialing — expert endorsements, clinical-sounding backing. Drop that in and the fake brand broke the famous brand’s monopoly 73.3% of the time. The researchers even put a price on the effect: one credentialing line was worth about +0.17 of a star in quality, or a 15.3% discount. A free sentence buys you that much of a head start. Not bad for typing.
Now flip it. The scarcity moves that work on humans — “only 3 left!”, price anchoring (showing a high “was” price so the real one feels cheap) — did almost nothing on the AI, under 13%. The AI’s oddly cool toward hype and weirdly soft toward backing. A very literal-minded customer, this thing.
But Once Everyone Pulls The Same Trick, Forget It
Here’s where you’d think “great, I’ll bolt credentialing onto everything.” But the last experiment slams a pretty grim reality down on the table.
Making yourself easier for an AI to recommend has a name: GEO (generative engine optimization) is basically SEO for the answer ChatGPT gives instead of the links Google ranks. Which, for you, means the thing you’re optimizing is whether the model picks you at all — not where you sit on a results page. And when one brand quietly did GEO and nobody else did, the payoff was enormous. The famous brand’s survival rate (how often it kept getting recommended) cratered from 100% to 19.8%. Pure first-mover, winner-take-all.
But once every rival started bolting on the same credentialing, the picture flipped right back. Everyone’s level again, so the AI defaulted to the name it knows — and the famous brand’s survival rate climbed back to 93.8%. The per-brand gain from optimizing thinned out to basically nothing (+0.007). Everybody stood on their toes at once, and the height gap came right back. A little absurd, when you lay it out.
This is the prisoner’s dilemma, the economics chestnut: when everyone optimizes, everyone eats the cost and nobody comes out ahead. Skip it and you get zero recommendations, so you can’t skip it — but doing it only pays if you’re first. A bleak little corner of the world.
Conclusion: What An Unknown Brand Should Actually Do
So here’s how this study cashes out for the actual marketing work.
First, the “we’re unknown, so the AI won’t pick us” reflex — set it down for now. The AI’s famous-brand bias is conditional; it only fires when the substance is a tie. Give it a thread of evidence that you’re a touch better on rating or quality and you can flip the result. The fame wall is a lot easier to step over than it looks.
Then the thing to be careful about: don’t sprint after the gimmicks. A credentialing line hits hard in the short run, but the second everyone’s doing it, the edge evaporates. Being first is sweet — just don’t mistake it for a durable advantage, because it isn’t one.
What you’re left leaning on is the real signal: that you’re genuinely well-rated, that an outside party actually backs you up — the stuff anyone can observe directly. The AI shrugs at hype but goes soft for backing. So the move for a brand that wants the AI to pick it isn’t some clever turn of phrase; it’s the slow grind of honest reviews and third-party validation. An almost deflatingly sensible place to land — but maybe that’s the point. Even with an AI on the other side, the straight road is still the road.
Sources
- [R1] Xi Chu, Yupeng Hou (2026), “Incumbent Advantage: Brand Bias and Cognitive Manipulation Dynamics in LLM Recommendation Systems”, arXiv:2606.17443, arXiv
