AI does not simply see a brand and recommend it. Somewhere between discovery, retrieval, understanding, evaluation and selection, a decision gets made. Atom Foundry is researching what happens at each stage, one study at a time.
What we know. What we have measured. What we still don't understand.
This page is not a claim that Atom Foundry has solved AI recommendation. It is the opposite. Here is the system we are trying to understand. Here is the evidence we have collected so far. Here are the questions we have not solved yet. Every new study we publish becomes a new piece of this map, not just a new article.
This is a working research model, not a claim about a fixed AI pipeline. Real systems may skip stages, run several at once, or loop back before answering.
Before a shopping question is ever asked, a language model already carries some notion of which brands exist and how they tend to be positioned. That prior knowledge comes from training, not from the live conversation, and it can shape an answer before retrieval or evaluation even begin.
Brand prominence appears to matter, a little. Wikipedia fame alone explains 1.2% of recommendation frequency. Widening to four public signals, cross-lingual reach, domain age, and recent media volume, raises that to just 11.2% combined, still far under the 61.4% the model's own past behavior explains about itself. Measured store quality explains almost none of it. A separate test asked directly what a model claims to know about a brand, then checked whether any of it shows up in a real recommendation: 75.8% of possessed facts never appear (95% CI 69.9-81.6%, n=1,392 fact checks across 9 brands). Knowing about a brand and using what it knows are two different things, and that gap does not shrink for brands that already win more often (r=-0.04, n=9). A causal follow-up tested whether that gap is fixable: inject one real, verified fact directly into context for 4 brands that barely get recommended today, as if retrieval had just succeeded. Average mention-rate lift across the cohort was 77.9 percentage points, but every brand showed the lift outpacing the rate the specific fact actually got cited, a Linkage Gap signature, more evidence the bottleneck is getting information in front of the model at generation time, not whether the model is capable of reasoning with it once it's there.
What exactly enters a model's memory, how strongly a brand is represented there, and why a possessed fact gets pulled into one generated answer but not another. Four easily-measurable public-footprint signals, and simple win-rate, have now been ruled out as the main answer; the actual retrieval-at-generation-time mechanism remains genuinely open. The fact-injection test simulates a successful retrieval step rather than measuring how often real retrieval actually surfaces the right fact, that's a separate, harder question.
When a model has live web access, it can supplement whatever it already knows with real time retrieval. A single switch, search on or off, turns out to matter more than almost anything else we have tested.
77% of product recommendations changed when web search was switched on, same model, same questions, the only change was retrieval.
Retrieval is not a minor adjustment to AI recommendations. In many categories, it can fundamentally change the answer, and how much it changes the answer differs by category.
There is a difference between AI being able to read your information and AI understanding what that information means. A store may clearly describe itself as a performance running brand, while a model interprets it primarily as a fashion brand.
Category understanding, product understanding, attribute understanding, brand positioning, intent alignment, and the semantic relationships between them.
A brand cannot be recommended if it never becomes a candidate.
Getting considered and getting selected are different problems, governed by different things. Our full population study identified one store factor that consistently separates brands that are ever recommended from brands that never are, and it holds up under every check we ran.
We went further and tested it causally. A brand with zero history, zero reviews, and zero press was invented from scratch and pitted against an entrenched category leader across 36 shopping intents. It won 0 of 360 runs. Giving it a review score changed that immediately, it started winning 53.1% of comparisons. Press mentions and sales-volume claims barely moved the needle, 1.9% and 0.3%. A separate causal test manipulated hidden upstream context instead of store-side evidence: a single fabricated line in the system message, invisible to the shopper, claiming they'd just seen an ad for a brand. Candidacy jumped to 97-100% under nearly any hidden-context condition, but winner rate depended heavily on which one, topping out at 88% for the fabricated ad-exposure claim, ahead of a condition that combined that same claim with a real, verified product fact.
Once a handful of candidates are inside the consideration set, the model has to compare them somehow. By price? By quality signals? By reviews? By authority? By familiarity? By product fit? By evidence? By category relevance?
We handed the model one comparison fact at a time between two genuinely close brands. A better star rating decided the pick in 160 of 160 runs, 100%, zero exceptions across 16 contested pairs. A longer, more specific spec list won 81.9% of runs. A lower price and faster availability won only 60.6%, barely above a coin flip.
How these signals interact when they conflict, for example a cheaper product with a worse rating, and whether the same hierarchy holds across categories.
This is where Recommendation Intelligence™ begins. We measure Recommendation Frequency™, Recommendation Share™, Recommendation Position™, the reasons a model gives for its pick, and the differences between the brand that wins and the brands that don't.
We can measure who gets recommended and how often. The mechanism behind the selection itself is not fully explained.
We can also measure what reasons a model gives for its recommendation. We do not yet know whether those stated reasons are the actual causal drivers of the selection, or a justification constructed after the fact.
Every recommendation has a winner and, quietly, a brand that lost. We can already observe both sides of that outcome. A first, direct test of three of the candidate signals below, rating, specificity, and message format, together in the same comparison, found rating alone decides the winner 90.8% of the time. A named follow-up isolated authority: alone, a disclosed third-party mention decided the winner 85.2% of the time, the second-strongest single signal measured in this series, but crossed against rating in the same comparison its own lift shrank to about 1.5 points, a same-direction point estimate too small for the sample to call statistically confirmed. A sibling study then isolated brand familiarity, a bare, unsourced claim of broad recognition with no named source: alone it decided the winner 79.8% of the time, but crossed against rating its own lift shrank to just 0.2 points, smaller still and equally unconfirmed. Neither residual effect survives a properly clustered significance test at the level of independent purchase intents, so both are best read as small, consistent directions worth more data, not settled findings. Which of the remaining signals matter, and how they interact with rating, is still open.
What separates the brand AI selects from the brand it leaves behind?
A model might recommend a brand today, but will it recommend the same brand again tomorrow, in the next run of the same question, or even three questions later in the same conversation? We measure recommendation stability across two different axes: between sessions, whether the same winner keeps winning weeks apart, and within a single conversation, whether an early mention survives to the final answer.
Between sessions, recommendation behavior is more stable than expected, even in places where drift would seem likely. Within a single conversation, it's much less stable: an explicitly favorable introduction survives to the final pick only 49.5% of the time across 200 four-turn conversations, ranging from 100% for the strongest brands down to 0% for the weakest, and a brand's real-world baseline recommend rate only loosely predicts which group it falls into.
A model can say it recommends Brand A. But how strong are the reasons behind that pick? How much would the answer change with slightly different phrasing? How close are the alternatives, and what is the actual margin between the first choice and the second?
Seven independent methods now test one piece of this: whether the model's hedged-vs-assertive language tracks its real internal confidence. It mostly doesn't (r=0.03), and for the most-confident pick type, the language runs backwards. Recommendation Confidence™ is the metric we expect to keep growing out of answering the rest.
Recommendation is not purchase.
A model may recommend a brand, and the person on the other end might still click somewhere else, change their mind, compare more options, abandon the purchase entirely, or buy from a competitor anyway. How an AI recommendation translates into actual commercial behavior is the final, and least understood, link in this chain.
Every unanswered question is a research opportunity.
Grouped by which stage of the decision path each study speaks to, not by publish date.
We're still discovering what happens between being seen and being chosen.
Illustrative example · single-site signal for atomfoundry.dev.
View full signals →