An analysis of 20,000 AI-generated product recommendations across e-commerce.
For the last year, most conversations around AI optimization have focused on visibility.
Can AI crawl your website? Can AI retrieve your content? Can AI cite your brand?
Those are useful questions.
But while analyzing AI recommendation behavior across ecommerce, we kept running into a different question:
Why does AI consistently recommend certain brands over others?
To investigate, we analyzed 20,000 AI-generated product recommendations across five e-commerce categories:
- Beauty
- Supplements
- Coffee
- Pets
- Home & Living
In total, the dataset included:
- 20,000 recommendations
- 1,490 brands
- 5 commerce categories
What we found surprised us.
Hypothesis #1: Better Stores Get More Recommendations
The first assumption seemed obvious.
If a store is easier for AI systems to understand, process and evaluate, it should receive more recommendations.
To test this, we compared recommendation frequency against our AI Commerce Score™, a framework designed to evaluate:
- machine readability
- semantic structure
- content depth
- technical implementation
- AI readiness
The expectation was simple: Higher quality stores should receive more recommendations.
The data said otherwise.
Store quality explained only:
2.1% of recommendation frequency.
Not a weak relationship.
Almost no relationship at all.
That immediately raised another question.
If AI isn't recommending brands because they have better stores...
What is driving recommendations?
Hypothesis #2: Fame
The next explanation was public fame.
Popular brands have:
- more reviews
- more backlinks
- more media coverage
- more mentions
- more content
Perhaps AI simply recommends brands that humans already recognize.
To test this, we analyzed the 200 most-recommended brands from our dataset.
For each brand, we collected public fame signals including:
- Wikipedia readership
- Number of language editions
- Article depth
- Brand name characteristics
We then compared those metrics against recommendation frequency.
Results
Store Quality: 2.1%
Public Fame: 24.9%
Most recommendation behavior: Still unexplained
Public fame explains significantly more recommendation behavior than store quality, but most recommendation behavior remains unexplained.
Fame mattered.
Much more than store quality.
But it still failed to explain most recommendation outcomes.
The Most Recommended Brands Aren't Better Stores
We then split the dataset into two groups:
- Top 50 most-recommended brands
- Bottom 50 least-recommended brands
If recommendation frequency reflected store quality, the difference should have been obvious.
It wasn't.
| Metric | Top 50 | Bottom 50 |
|---|---|---|
| Recommendation Frequency | 30.9% | 5.0% |
| AI Commerce Score™ | 50.8 | 49.8 |
The recommendation gap was massive.
The quality gap was almost nonexistent.
The most recommended brands receive more than six times as many recommendations despite nearly identical store quality scores.
The winners weren't better stores.
They were simply recommended more often.
Maybe Recommendations Are Random?
At this point, a reasonable explanation would be: AI recommendation behavior is mostly random.
So we tested that.
Every shopping query was repeated twenty times.
If recommendation behavior were unstable, we would expect different winners across repeated runs.
Instead, we observed the opposite.
The same brands kept appearing.
Again.
And again.
And again.
Across categories, the top-ranked brand remained in first place between:
78% and 91% of runs.
Recommendation outcomes remain highly stable across repeated runs.
This was perhaps the most surprising result in the entire study.
Because it means the unexplained portion of recommendation behavior is not random.
The system appears remarkably stable.
We simply don't understand it yet.
What This Suggests
When we combine all findings, we observe four things:
- Store quality explains very little recommendation behavior.
- Public fame explains significantly more.
- Most recommendation behavior remains unexplained.
- Recommendation outcomes remain highly stable.
Taken together, these findings suggest there may be another layer operating beneath traditional AI visibility metrics.
Visibility answers: Can AI see a brand?
Recommendation answers: Will AI choose a brand?
Those are fundamentally different problems.
A Possible Recommendation Layer
One interpretation of these findings is that recommendation systems operate on a deeper decision layer than current AI visibility tools measure.
A layer beneath:
- rankings
- citations
- retrieval
- visibility
A layer that influences trust, selection and recommendation.
We don't yet know what variables define that layer.
But the data suggests it exists.
And understanding it may become increasingly important as AI systems play a larger role in commercial decision-making.
Open Questions
This study answered one question.
It created several more.
If store quality explains 2.1%...
And fame explains 24.9%...
What explains the remaining 73%?
Potential candidates might include:
- trust signals
- entity relationships
- training data exposure
- recommendation reinforcement effects
- semantic authority
- citation networks
- brand familiarity patterns
- factors we haven't identified yet
At this stage, we don't know.
But that's exactly what we're investigating next.
Methodology
Dataset:
- 20,000 AI-generated recommendations
- 1,490 brands
- 5 ecommerce categories
Categories:
- Beauty
- Supplements
- Coffee
- Pets
- Home & Living
Measured variables:
- Recommendation frequency
- AI Commerce Score™
- Wikipedia readership
- Language editions
- Article depth
- Brand name characteristics
Repeated-run testing:
- 20 repeated recommendation runs per query
Final Thought
Most AI optimization discussions today focus on visibility.
Our data suggests recommendation behavior may be a separate problem entirely.
Visibility determines whether AI can find you.
Recommendation determines whether AI chooses you.
And those may turn out to be two very different systems.
Top comments (0)