We captured 20,000 AI product recommendations across five ecommerce categories and matched every one to the real store behind the brand. The market is busy asking whether AI can see a brand. We asked a harder question: when AI recommends one brand over another, what is it actually responding to?
Most of the AI search market is converging on one number: are you visible. Are you mentioned, are you cited, do you show up. That is a real and important shift, and the data behind it is stark. But being mentioned by an AI system and being recommended by one are becoming two very different outcomes. A brand can be in the candidate set and never make the shortlist. The question we set out to answer is what moves a brand from one to the other, and whether the work brands are being told to do, making their stores cleaner and more machine-readable, actually moves that needle at all.
For every category we measured two things per brand: how often AI recommended it (Recommendation Frequency™) and how AI-ready its store actually is (AI Commerce Score™, our 0 to 100 measure of how readable, trustworthy, and machine-legible a store is). Then we correlated them. If readiness drove recommendation, these bars would climb. They do not.
| Category | Recs | Brands | On-index | Marketplace | r | Link |
|---|---|---|---|---|---|---|
| Beauty | 4,000 | 238 | 57.6% | 0.4% | +0.170 | None |
| Supplements | 4,000 | 367 | 33.5% | 5.5% | -0.015 | None |
| Coffee | 4,000 | 228 | 49.8% | 1.8% | +0.019 | None |
| Pets | 4,000 | 405 | 19.8% | 7.1% | -0.366 | Weak negative |
| Home & Living | 4,000 | 271 | 60.7% | 20.9% | +0.108 | None |
On-index is the share of recommendations that map to a single-brand store we measure. Marketplace is the share that went to retailers and marketplaces, which we separate out and exclude from the brand-level correlation. r is the Pearson correlation between Recommendation Frequency™ and AI Commerce Score™ across single-brand stores.
A second pattern fell out of the data. The share of recommendations going to marketplaces rather than single brands rises sharply with how commoditized and everyday the category is. In beauty, almost everything is a brand. In home and living, one in five recommendations is a retailer.
This matters because a marketplace is not a brand. When AI recommends Amazon or Wayfair, it is not choosing a product, it is deferring the choice. We keep these out of the brand analysis entirely. The escalation from 0.4 percent in beauty to 20.9 percent in home and living is itself a finding: in broad, undifferentiated categories, AI increasingly hands the decision back to a marketplace rather than naming a brand.
If readiness does not drive recommendation, what does? The pattern across all five categories points one way. The brands AI recommends most are the ones it saw most during training. Clinique and SkinCeuticals in beauty. NOW Foods in supplements. Peet's in coffee. Blue Buffalo and Purina in pets. IKEA, West Elm, and Pottery Barn in home. These are not the best-built stores. Several are not even in our index, and the ones that are sit deep in the AI Invisible band. They win because they are famous.
We call this Recommendation by Memory™: the model reaches into its parametric memory and returns the names it has seen most often. The opposite, Recommendation by Understanding™, is when a system actually retrieves, reads, and evaluates a store before recommending it. Today commerce runs almost entirely on memory. That is exactly why a famous brand with a broken store still wins, and a modern brand with an excellent store still loses.
There is real value in visibility measurement, and the traffic data behind it is sobering. But visibility is the first layer, not the last. Knowing you appear in 30 percent of answers does not tell you why, does not tell you what moved you into the shortlist, and does not tell you what to change. Our data shows the most common assumption, that a cleaner and more readable store earns more recommendations, simply is not true today at the level of recommendation frequency. So the actionable question is not whether AI sees you. It is what makes AI choose you, and that is the question this research program exists to answer.
The next step is direct. We are adding measurable proxies for brand fame, search demand, mention volume, brand age, to quantify exactly how much of recommendation is explained by fame rather than readiness. If readiness sits at roughly zero and fame explains the rest, that is the clearest possible map of where the leverage actually is. That is the Recommendation Intelligence Framework™, and it is built on data, not slides.