I Expected Better Stores To Get Recommended More Often. The Data Said Otherwise.

Over the last few months, I've been running a research project around a question that seems increasingly important: As AI becomes a shopping advisor, what actually determines which brands get recommended?

Most conversations around AI search focus on visibility.

Can ChatGPT find you? Can Gemini cite you? Can Perplexity retrieve your content?

Those are useful questions. But they aren't the question I was interested in.

I wanted to understand something deeper: Why does AI recommend one brand instead of another?

So we decided to measure it.

The Experiment

We collected 20,000 AI-generated product recommendations across five e-commerce categories: Beauty, Supplements, Coffee, Pets and Home & Living

In total:
20,000 recommendations
1,490 brands
100 shopping intents

For every brand, we measured two things:

• How often AI recommended it
• How AI-ready the store actually was

Our assumption seemed obvious.

Better stores should get recommended more often.

Right?

The First Surprise

They didn't.

Across every category, the relationship between store quality and recommendation frequency was close to zero.

Beauty: r = 0.17

Supplements: r = -0.015

Coffee: r = 0.019

Pets: r = -0.366

Home & Living: r = 0.108

Not once did we find a meaningful positive relationship.

That was unexpected.

If recommendation frequency isn't explained by store quality, then what explains it?

The Obvious Explanation: Fame

The next theory was simple.

Maybe AI just recommends famous brands.

That would make sense.

Famous brands have:

• more mentions
• more reviews
• more content
• more media coverage
• more backlinks

So we analyzed the 200 most-recommended brands and measured several public fame signals:

• Wikipedia readership
• Number of language editions
• Article depth
• Brand-name characteristics

Then we compared those metrics against recommendation frequency.

The result:

Store quality explained 2.1% of recommendation frequency.

Public fame explained 24.9%.

Fame mattered roughly twelve times more than store quality.

At that point I thought we had our answer.

But we didn't.

The Bigger Surprise

Even fame explained only about a quarter of recommendation behavior.

Three quarters remained unexplained.

That led to an uncomfortable possibility:

Maybe recommendations are mostly random.

So we tested that too.

Every shopping question was repeated twenty times.

If recommendations were random, the winners should constantly change.

They didn't.

The same brands kept showing up.

Again and again.

Across categories, the same brand occupied the #1 position between 78% and 91% of repeated runs.

Which means the unexplained portion isn't random.

It's stable.

There appears to be a system.

We just don't fully understand it yet.

What I Think Is Happening

Most of today's AI optimization industry focuses on visibility.

Can AI see you?

Can AI retrieve you?

Can AI cite you?

Those are valid questions.

But our findings suggest recommendation may operate on a completely different layer.

Visibility determines whether a brand enters the candidate set.

Recommendation determines whether it gets chosen.

Those are not the same thing.

And based on what we've measured so far, recommendation behavior appears much harder to explain than visibility.

One Interesting Pattern

Another finding that surprised me:

The more commoditized a category becomes, the more likely AI is to recommend a marketplace instead of a brand.

In Beauty, almost every recommendation was a brand.

In Home & Living, more than 20% of recommendations pointed to retailers and marketplaces.

In other words, AI increasingly says: Go to Amazon.

instead of: Buy Brand X.

That feels important.

Open Question

At this point, I don't think the most interesting question in AI search is: How do I get cited?

I think the more interesting question is: What makes AI consistently choose one brand over another?

Store quality barely explains it.

Fame explains more, but not most.

Yet recommendation behavior remains remarkably stable.

I'm curious what others think.

If AI recommendation systems aren't primarily optimizing for store quality, what signals do you think they're actually responding to?