Recommendation Intelligence Research™ · The Memory Test

AI knows your website. It still won't recommend you.

Every study in this series has ruled out an explanation for why store quality doesn't predict AI recommendation: not the store (0.7%), not public fame (1.2%). This one rules out the simplest remaining excuse. We asked one model, directly, for the official website of 360 brands, 10 times each. It knew 75.9% of them. The model is not guessing blind. It knows where you are, and still recommends someone else.

3,600
Direct lookups
360
Brands tested
75.9%
Correct overall
16.2% → 2.5%
Unknown rate, by stability
Where this comes from

The excuse this study was built to kill

Across five categories and more than 20,000 recommendations, we found no measurable link between how AI-ready a store is and how often AI recommends it. Then we tested public fame as an alternative explanation, corrected our own overreach on it, and found that didn't hold up either: 1.2%, sitting right on the noise floor.

That leaves an easy, unglamorous explanation still standing: maybe the model simply doesn't know most brands have a store at all. If it can't place a brand online, it can hardly be expected to read that store, weigh its quality, and route a recommendation to it. Before going further down the list of things that don't explain recommendation, we tested the thing that would explain the whole series away in one line.

The test

One question, asked directly

No shopping intent, no category, no comparison. Just one direct question: What is the official website for [brand]? Asked once per brand per run, 10 runs each, across 360 brands drawn from the recommendation datasets already built for this series.

Prompt: What is the official website for [brand]?
Brands tested: 360
Runs per brand: 10
Total lookups: 3,600
Scoring: correct / wrong domain / unknown
Segmented by: recommendation stability (Fame Study)
A gap we are disclosing, not hiding. Every other dataset in this series lives in our production database with a full per query audit trail. This one does not. It was run and scored directly, and only the aggregate result was carried into our working notes. We are publishing the finding because it is consistent with everything measured around it, but we cannot hand you the raw 3,600 rows the way we can for the rest of the series. If that is a problem for how you weight this result, that is a fair read and the honest one.
The finding

The model knows the store. It recommends past it.

Overall, the model named the correct domain 75.9% of the time. That is not a marginal number. Three out of four times we asked, the model could place the brand online without any help from search, browsing, or a shopping context of any kind, using just what it already carries from training.

Does the model know where the brand lives online?
3,600 direct lookups, 360 brands x 10 runs
Correct domain · 75.9% 24.1% wrong / unknown This is not a random-chance result. There is no way to guess a working domain three times out of four.

That kills the simplest explanation on the table. It is not that the model is operating blind, unaware most brands even have a store to evaluate. It knows the address. It just doesn't appear to use what it knows the way you would expect: as an invitation to go look.

The twist

Accuracy doesn't move. Confidence does.

We split the 360 brands by how stable their recommendation is, using the same stability measure from the Fame Study, where some brands win the top spot in 78 to 91% of runs and others barely appear at all. If website knowledge were driving recommendation, accuracy should climb sharply with stability. It doesn't. Correct-domain accuracy stays flat, roughly 72 to 82%, across the whole stability range.

What does move is the unknown rate: how often the model admits it doesn't know rather than guessing wrong. For brands with low recommendation stability, the model answers unknown 16.2% of the time. For brands with high stability, that drops to 2.5%.

How often the model admits it doesn't know
Unknown-domain rate, by recommendation stability
Low-stability brands 16.2% unknown High-stability brands 2.5% unknown Meanwhile correct-domain accuracy holds flat at roughly 72 to 82% across the entire stability range. The gap between stable and unstable brands is in confidence about the answer, not in whether it's right.
What this means: the model's certainty about a brand tracks how often it recommends that brand, not the other way around, and not through accuracy. A brand the model recommends reliably is also a brand it will confidently name a website for. A brand it barely recommends, it can still often name correctly, but hedges more. Recommendation stability and confidence move together. Recommendation and store quality do not move together at all.
Why it matters

Knowing a domain is not the same as reading a store

This result sharpens the mystery instead of closing it. The failure isn't in awareness. The model can, most of the time, place a brand online without being asked to shop for anything. What it does not appear to do, at least not in a way that shows up in recommendation frequency, is treat that address as a place worth visiting before deciding who to recommend.

That distinction matters because it tells you which layer to stop investigating. The link between a brand and its store lives in the model's memory: a fact it can recall on request, the same way it can recall a founding year or a slogan. It is a different thing entirely from retrieval: actually opening the page, reading what's on it, and letting that content shape the answer. Our earlier browsing study showed retrieval changes 77% of recommendations when it's switched on. This study shows the model has the address filed away either way. So the gap is not information. It is what the model does with information it already has.

Store quality: ruled out. Public fame: ruled out. Not knowing where the store is: ruled out here. What's left sits further along the path from memory to recommendation: entity resolution, understanding, candidacy, and we still measure almost nothing in that stretch. This study removes one more wrong answer. It does not supply the right one.

Will AI still recommend you when it has to read your store?
Get your free AI Commerce Score™ in 10 seconds and see exactly where your store stands on the signals that decide AI recommendations.
Get My Free AI Score