DEV Community: Daniel Pokorný

# A stranger tried to break my study. Here's what happened.

Daniel Pokorný — Tue, 14 Jul 2026 17:04:04 +0000

I was about to publish a clean finding. Then someone asked a better question, and the number that looked like it settled the argument turned out to be a trap.

Last week I published a controlled experiment showing that, on the same AI model, simply turning web search on changed 77% of product recommendations.

Same model. Same 50 buying prompts. One variable.

The follow-up article was already written.

I was going to break the results down by category, show that Pets changed 88% while Fitness changed 61%, and tell a neat story:

The more fragmented the market, the more browsing overrides what the model remembers.

Then a data engineer named Rami read the study and asked a question I couldn't answer.

The objection

His question was simple.

How do you know that's the category and not your prompts?

Generic prompts like:

Best dog food

should rely on retrieval much more heavily than specific prompts like:

Best cat litter for odor control.

If some categories happened to contain mostly generic prompts while others contained mostly specific ones, then what looked like a category effect might simply be a prompt-design effect.

He was right.

I couldn't rule it out.

So I paused the publication.

The trap

I tagged every one of the 50 prompts as either generic or specific and reran the analysis.

At first glance, it looked like I had my answer.

Generic prompts: 74.2% changed
Specific prompts: 80.0% changed

Specific prompts changed more.

Exactly the opposite of what Rami predicted.

I could have published that result, declared the objection resolved, and moved on.

It would have been the wrong conclusion.

Simpson's Paradox strikes

Once I looked inside each category, the pattern disappeared.

Some categories showed generic prompts changing more.

Others showed specific prompts changing more.

There was no consistent direction.

The pooled difference wasn't measuring causation.

It was measuring composition.

The categories that happened to contain most of my specific prompts (Coffee, Pets, Wellness) were also the categories that already had the largest browsing effect.

Specific prompts looked more volatile simply because of where they happened to appear.

This is a textbook example of Simpson's Paradox.

Ironically, the single number that looked like it answered the criticism was actually the most misleading statistic in the study.

The real test

To isolate the effect properly, prompt specificity has to be held constant.

So I compared categories within each prompt type.

Generic prompts only

Category	Recommendation change
Coffee	89.8%
...	...
Fitness	62.4%

Spread: 27 percentage points

Specific prompts only

Category	Recommendation change
Pets	94.4%
...	...
Fitness	52.6%

Spread: 42 percentage points

Fitness remains the lowest.

Coffee and Pets remain the highest.

The category effect survives after controlling for prompt style.

Looking for the confound

I also measured whether categories containing more generic prompts systematically produced larger recommendation changes.

The correlation was:

r = -0.56

If prompt composition were actually driving the result, this correlation should have been strongly positive.

Instead, it was moderately negative.

That doesn't prove the confound doesn't exist.

But it certainly doesn't support it either.

Where the study is still weak

I'd rather point out the weaknesses myself than hope nobody notices.

Some categories have very few prompts.

For example:

Coffee contains only one generic prompt
Fitness contains only one specific prompt
Electronics contains no specific prompts

That means some category-level estimates rely on very small samples.

Likewise, the correlation of -0.56 across only ten categories isn't statistically significant.

At that sample size, significance would require roughly |r| > 0.63.

So this is evidence of a direction.

Not proof.

There's another limitation.

The prompt mix wasn't intentionally balanced.

It emerged naturally during the study.

A future experiment should deliberately balance generic and specific prompts across every category.

That's already on my roadmap.

So the honest conclusion isn't:

"There is no confound."

It's simply:

I looked for the confound, found evidence pointing the other way, and the category effect still survives after stratification.

The second hole

Rami raised another criticism.

The original analysis had no statistical treatment.

Fair enough.

Fortunately, that's easy to improve.

A naïve confidence interval assumes every brand-intent pair is independent.

They aren't.

Brands cluster within buying intents.

So instead I ran a cluster bootstrap, resampling the 50 buying intents rather than individual brand observations.

The result stayed remarkably stable.

Estimated recommendation change: 77%
95% confidence interval: 74%–80%

The naïve interval had been 75%–79%.

Accounting for clustering widened it by roughly 1.6×.

The wider interval is the honest one.

That's the version I'm publishing.

Why I'm writing about this

The interesting part isn't that the finding survived.

The interesting part is that the statistic which looked like it settled the debate was the one most likely to fool me.

Had I published:

"Specific prompts change more. Objection handled."

it would have sounded confident.

It would have been easy to quote.

And it would have been wrong.

Nobody would have questioned it.

It probably would have lived online for years.

Instead, a stranger asked a better question than I had asked myself.

That turned out to be far more valuable than simply having my original conclusion confirmed.

Full study

Full study and methodology:

https://atomfoundry.dev/research/web-search-changes-ai-recommendations

# I Flipped One Toggle and 77% of the AI's Product Recommendations Changed

Daniel Pokorný — Sun, 12 Jul 2026 20:07:38 +0000

How much does web search change what an AI recommends?

I measured it using a fixed set of e-commerce buying questions.

The result surprised me.

Correction, 15 July 2026. An earlier version of this piece claimed that changing the model moved only about 6% of recommendations, and concluded that which model you ask barely matters.

That number was derived rather than measured. I subtracted two overlap figures from two different comparisons and treated the difference as the model effect. It is not the same quantity, and it was wrong.

Measured directly, gpt-4o-mini without search against gpt-4o without search, 66.9% of the recommended brands changed. Not 6%.
One caveat on that figure: the two runs were collected twelve days apart, so it mixes the model change with whatever drifted in between, and I cannot separate them from this data.

The search comparison does not have that problem, both conditions ran together, so the 77% figure is unaffected and stands.

The corrected picture: swapping the model rewrites roughly two thirds of the recommendations. Turning on browsing rewrites roughly three quarters. There is no such thing as "what AI recommends." There is only what a particular model, in a particular configuration, at a particular moment, recommends...

I'm leaving the mistake visible rather than editing it out, because the way it happened is the useful part: a derived number looked like a measured one, and it agreed with the story I wanted to tell.

Why I Ran This Experiment

A few weeks ago I published research showing something unexpected.

Across thousands of AI recommendations, store quality explained almost nothing about why brands were recommended.

Brand popularity explained much more.

Yet almost 73% of recommendation behavior remained unexplained.

After publishing the study, data engineer Rami suggested an interesting hypothesis.

Maybe part of that "unexplained" behavior isn't mysterious at all.

Maybe it's simply the retrieval layer.

His suggestion was simple:

Run the exact same prompts twice.

Once with web search enabled.
Once with web search disabled.

Then compare the recommendations.

So I did.

Experiment Setup

Model:
GPT-4o

Conditions:
• Browsing ON
• Browsing OFF

Dataset:
50 buying prompts

Categories:
Pets
Beauty
Supplements
Coffee
Fashion
Fitness
Food
Home
Wellness
Electronics

Runs:
10 per prompt per condition

Metric:
Unique recommended brands

To isolate the effect of browsing from the effect of model size, I also repeated the experiment using GPT-4o-mini.

The Result

The overlap between GPT-4o with browsing enabled and GPT-4o without browsing was only:

23%

That means:

77% of the recommended brands changed simply by enabling web search.

Same model.

Same prompts.

Same methodology.

One toggle.

What Changed?

With browsing disabled, GPT-4o recommends mostly from its internal memory.

Those recommendations tend to favor well-known brands the model likely encountered frequently during training.

With browsing enabled, the model relies heavily on live retrieval.

Many memory-based recommendations disappear and are replaced by brands discovered during search.

These weren't small ranking adjustments.

They were largely different recommendation lists.

Was It Really Search?

A reasonable objection would be:

Maybe GPT-4o simply recommends different brands than GPT-4o-mini.

So I isolated that variable.

Model difference (Browsing OFF)

Moving from GPT-4o-mini to GPT-4o changed only about:

6% of recommendations

Search difference (Same model)

Turning browsing ON changed:

77% of recommendations

The retrieval layer had a dramatically larger impact than model size.

It Depends on the Category

Browsing didn't affect every category equally.

Category	Recommendations Changed
Pets	88%
Fitness	61%

The pattern appears consistent.

Markets dominated by a handful of famous brands showed smaller changes.

Markets with many niche brands showed much larger changes.

Pets has consistently produced the strongest effects across every study I've run.

I still don't fully understand why.

The Bigger Black Box

Rami's original hypothesis had another part.

He suggested that much of the unexplained variance may live inside the model's training history.

Questions like:

Which sources mentioned a brand?
How often?
In what context?
With what sentiment?

Those signals aren't observable today.

I can measure what gets recommended.

I cannot yet measure why a model internally trusts one brand more than another.

That remains one of the biggest black boxes in AI recommendation systems.

Why This Matters

Many companies treat "AI visibility" as a single objective.

I don't think it is.

There are at least two different systems at work.

Memory

Long-term exposure during training.

This rewards brands that are already widely known.

Retrieval

Live information gathered at query time.

This rewards brands that are discoverable and well represented across the web.

Those are different mechanisms.

And they often produce different recommendations.

Key Takeaway

If your optimization strategy only targets one of these layers, you're missing the other.

Depending on whether an AI assistant uses browsing, retrieval may completely replace memory-driven recommendations.

In this experiment, enabling browsing changed 77% of the recommended brands.

That's too large to ignore.

Methodology

50 fixed ecommerce buying prompts
10 executions per prompt
GPT-4o with browsing ON
GPT-4o with browsing OFF
GPT-4o-mini comparison
Recommendation overlap measured using distinct recommended brands per buying intent

Browsing implementations evolve over time, so treat the percentages as a snapshot rather than a universal constant.

The important finding isn't the exact number.

It's the magnitude of the effect.

I'm continuing this research through Atom Foundry, where I study how AI systems understand, evaluate, and recommend ecommerce brands.

If you spot weaknesses in the methodology or have ideas for improving the experiments, I'd genuinely like to hear them.

# AI Recommends by Fame. But Fame Doesn't Explain Most Recommendations.

Daniel Pokorný — Mon, 22 Jun 2026 12:20:00 +0000

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.

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.

# The State of AI Recommendations Across Commerce 2026

Daniel Pokorný — Fri, 19 Jun 2026 12:05:00 +0000

Most businesses assume that if AI systems can understand their products, they will eventually recommend them.

Our latest research suggests the relationship may not be that simple.

Over the past several months, we analyzed 20,000 AI recommendations across five e-commerce categories:

Beauty
Supplements
Coffee
Pets
Home & Living

The goal was straightforward: To understand whether store readiness actually predicts recommendation behavior.

Or put differently: Do AI systems recommend the businesses that are best prepared for AI?

The answer surprised us.

The Assumption

Most discussions about AI visibility start with a reasonable belief.

If a business improves its structure, content, product information, trust signals, and machine readability, AI systems should be more likely to recommend it.

This assumption has fueled a growing industry around AI optimization, AI visibility, AI readiness, and AI commerce infrastructure.

But assumptions are not evidence.

We wanted to measure recommendation behavior directly.

What We Measured

For each category, we collected thousands of recommendations generated by AI systems in response to high-intent shopping questions.

We then compared recommendation frequency against each brand's AI Commerce Score™.

The expectation was simple: Higher scores should lead to more recommendations.

Instead, we found something very different.

Figure 1. AI Commerce Score™ and Recommendation Frequency™

Based on 20,000 AI recommendations across five ecommerce categories, recommendation frequency showed little to no meaningful relationship with AI Commerce Score™, suggesting that store readiness alone does not explain recommendation behavior.

The Surprising Result

Across all five categories, recommendation frequency showed little to no meaningful correlation with AI Commerce Score™.

Some highly recommended brands scored relatively poorly. Some highly optimized brands were rarely recommended.

The relationship was far weaker than expected.

This finding appeared repeatedly across Beauty, Supplements, Coffee, Pets, and Home & Living.

The implication is important.

Store readiness alone does not appear sufficient to explain why AI recommends certain brands more frequently than others.

Something else is happening.

The Familiarity Hypothesis

As we analyzed recommendation patterns, another explanation began to emerge.

Many of the most frequently recommended brands shared one characteristic:

They were already familiar. They had existing brand recognition. Existing awareness. Existing market presence. Existing references across the web.

In other words, recommendation behavior often appeared to resemble memory more than evaluation.

This led us to a framework we call: ## Recommendation by Memory™

Under this model, AI systems frequently recommend brands they have encountered repeatedly during training and exposure.

Not necessarily because those brands are objectively better.

But because they are more familiar.

A Potential Future Shift

However, we do not believe recommendation behavior will remain static.

As AI systems gain access to richer retrieval systems, real-time information, structured commerce data, and increasingly sophisticated evaluation capabilities, recommendation behavior may evolve.

We call this future state: ## Recommendation by Understanding™

Figure 2. The Evolution of AI Recommendation Systems

Recommendation behavior is evolving from what AI remembers to what AI understands. This transition may define the next phase of AI commerce and product discovery.

Under this model, recommendations become less dependent on historical familiarity and more dependent on:

Store quality
Trust signals
Product fit
Verifiable information
Real-time relevance
Semantic understanding

In other words: The best understood businesses may eventually outperform the most familiar businesses.

Why This Matters

If recommendation behavior is driven primarily by memory today, businesses face a difficult challenge.

Optimization alone may not immediately increase recommendation frequency.

But if recommendation systems gradually move toward understanding, the businesses investing in AI readiness today may be building an advantage for tomorrow.

The future of AI commerce may not belong to the brands AI remembers.

It may belong to the brands AI understands best.

The Emerging Question

Most businesses are still asking: Can AI find us?

A more important question may be emerging: Why does AI choose one business instead of another?

That question sits at the center of what we call Recommendation Intelligence™.

And as AI increasingly influences product discovery, recommendation behavior may become one of the most important areas of research in commerce.

About Atom Foundry

Atom Foundry is building the field of AI Commerce Intelligence™.

Our research explores how AI search engines, LLMs, Apple Intelligence, and shopping agents discover, understand, evaluate, trust, recommend, and route customers to businesses.

Current research initiatives include:

AI Commerce Intelligence™
Recommendation Intelligence™
AI Commerce Graph™
AI Readability™
AI Understanding™
AI Trust™
Decision Confidence™

Learn more:

🌐 https://atomfoundry.dev

📊 https://github.com/Atom-Foundry

Research by Atom Foundry. Based on 20,000 AI recommendations captured across Beauty, Supplements, Coffee, Pets, and Home & Living categories.

The State of AI Recommendations in Home & Living

Daniel Pokorný — Thu, 18 Jun 2026 16:01:40 +0000

We Captured 4,000 AI Recommendations in Home & Living. The Results Were Unexpected.

AI is supposed to reward better websites.

At least that's what many brands assume.

If a store is easier for AI systems to understand, trust, and interpret, it should receive more recommendations.

Our data suggests otherwise.

We analyzed 4,000 AI-generated recommendations in Home & Living and compared recommendation behavior against AI Commerce Score™.

The correlation came back at:

r = 0.108

Effectively zero.

This is the fifth study in our Recommendation Intelligence Research™ series.

And after more than 20,000 captured recommendations across five industries, the pattern remains unchanged.

AI does not appear to recommend the most AI-ready brands.

It recommends the brands it already knows.

The Research

This study follows the exact same methodology used in our previous research across Beauty, Supplements, Coffee, and Pets.

Everything below is based on captured responses.

Nothing is estimated.

Nothing is projected.

Category: Home & Living

Model: GPT-4o-mini

Shopping Intents: 20

Runs Per Intent: 20

Prompt Runs: 400

Recommendations Captured: 4,000

Distinct Brands: 271

We measured three core metrics.

Recommendation Share™

The percentage of all captured recommendations belonging to a brand.

Recommendation Frequency™

The percentage of prompt runs in which a brand appeared at least once.

Recommendation Position™

The average position a brand occupied when recommended.

Lower numbers indicate stronger placement.

As with every study in this series, this represents one model, one category, and one point in time.

The Brands AI Recommended Most

The leaderboard immediately revealed the same pattern we observed in previous categories.

IKEA appeared in 48.3% of all prompt runs.

West Elm appeared in 47.0%.

Pottery Barn appeared in 41.5%.

These brands dominated recommendations despite only moderate store readiness.

IKEA scored 64 on AI Commerce Score™.

West Elm scored 60.

Meanwhile, some of the strongest stores in the category received very little attention.

Article scored 80.

Burrow scored 78.

Uplift Desk scored 77.

GhostBed scored 76.

Yet each appeared in fewer than 7% of prompt runs.

The strongest stores were not winning recommendations.

The most familiar brands were.

The Central Finding

The correlation between Recommendation Frequency™ and AI Commerce Score™ came in at:

r = 0.108

Across 71 measured brands, recommendation frequency and store readiness were effectively unrelated.

In practical terms, knowing how often AI recommends a home brand tells you almost nothing about how AI-ready its store actually is.

The result mirrors what we observed in Beauty, Supplements, Coffee, and Pets.

Once again, recommendation behavior appears disconnected from store quality.

Perhaps the most surprising finding is the average readiness of the brands AI recommends most often.

The average AI Commerce Score™ across all on-index recommended brands was just 50.7.

That sits directly on the border between Low Readiness and AI Invisible.

In other words:

AI is frequently recommending stores that are barely readable to the very systems doing the recommending.

Five Categories. One Result.

Home & Living was the final test.

At this point, the pattern is difficult to dismiss as noise.

Across more than 20,000 captured recommendations, recommendation frequency has never demonstrated a meaningful positive relationship with store readiness.

Five categories.

More than 20,000 recommendations.

Not one meaningful positive relationship.

Three categories produced essentially zero correlation.

One produced a negative relationship.

Home & Living produced essentially zero again.

The conclusion is no longer a single finding.

It is a replicated result.

AI recommendation frequency is not positively related to store readiness.

The Most Marketplace Driven Category Yet

Home & Living also produced another notable finding.

Approximately 21% of all recommendations went to retailers and marketplaces.

That is the highest retailer share we have measured across the entire research program.

Amazon. Wayfair. Home Depot. Macy's. Lowe's.

Retail remains deeply embedded in how AI understands this category.

At the same time, Home & Living produced the highest on-index coverage of any category in our research.

Roughly 60% of recommendations mapped directly to a single-brand store already included in our database.

This gave us more visibility into the relationship between recommendation behavior and store quality than any previous study.

The result remained unchanged.

Recommendation frequency still failed to track with readiness.

What It Means

Visibility is not recommendation.

Recommendation is not readiness.

And readiness is not currently being rewarded.

Today's AI systems still recommend heavily from memory.

They reach for the brands they encountered repeatedly during training.

Brands with decades of awareness.

Brands with large retail footprints.

Brands that appear constantly across reviews, articles, forums, and public discussion.

That helps explain why IKEA, West Elm, and Pottery Barn dominate recommendations despite only modest levels of AI readiness.

But memory is unlikely to remain the dominant signal forever.

The future of AI commerce will not be powered entirely by what models remember.

It will increasingly depend on what agents can verify.

As shopping agents gain the ability to browse websites, compare products, evaluate trust signals, interpret structured information, and complete purchases autonomously, recommendation behavior may gradually shift from memory-based recommendation toward evidence-based recommendation.

When that happens, readable stores become more important than famous names.

The brands benefiting most from familiarity today may be the brands with the most to lose tomorrow.

The brands investing in AI Readability™, AI Understanding™, and AI Trust™ today may become the brands best positioned for the next generation of AI commerce.

That transition is exactly what the Recommendation Intelligence Framework™ was designed to measure.

Recommendation Intelligence™

After five studies and more than 20,000 captured recommendations, we believe the industry may be looking at the wrong layer.

Most AI search conversations focus on visibility.

Can AI find you? Can AI crawl you? Can AI cite you?

Those questions matter.

But they do not explain recommendation behavior.

Recommendation behavior appears to be a separate system.

A separate problem. A separate competitive advantage.

That layer is what we call: # Recommendation Intelligence™

Not whether AI knows a brand.

But whether AI chooses it.

What's Next

Home & Living closes the first phase of the Recommendation Intelligence Research™ program.

Across five industries and more than 20,000 captured recommendations, we found no meaningful relationship between store readiness and recommendation frequency.

The next phase asks a different question.

If store quality does not explain recommendations, what does?

The upcoming State of AI Recommendations Across Commerce 2026 report will combine findings from all five industries into a single cross-category analysis.

After that comes a new hypothesis.

The State of Fame in AI Recommendations

Perhaps recommendation behavior is not driven primarily by store quality.

Perhaps it is driven by fame.

If recommendation frequency consistently ignores store readiness, the next logical question becomes whether AI systems are simply recommending the brands they remember best.

After 20,000 recommendations, that possibility is becoming increasingly difficult to ignore.

Explore The Research

Get your free AI Commerce Score™

https://atomfoundry.dev

Explore the full Home & Living dataset

https://atomfoundry.dev/research/state-of-ai-recommendations-home-living

Explore the complete Recommendation Intelligence Research™ series

https://atomfoundry.dev/research

Because the future of AI commerce will not be decided by visibility alone.

It will be decided by recommendation.

The State of AI Recommendations in Pets

Daniel Pokorný — Wed, 17 Jun 2026 09:34:45 +0000

Recommendation Intelligence Research™ · Atom Foundry · June 2026

We asked one AI model 20 high intent pet shopping questions, 20 times each.

Then we checked the same thing we checked in beauty, supplements, and coffee.

Does being recommended by AI have anything to do with how AI ready a store actually is?

Pets produced the strongest answer yet.

Across another 4,000 recommendations, recommendation frequency was not positively related to store readiness.

In fact, it moved in the opposite direction.

Methodology

Same method as the first three reports, so all four categories are directly comparable.

Everything below is computed from real captured responses.

Nothing is estimated or projected.

Category: Pets

Model: one model (gpt-4o-mini)

Intents: 20 high intent shopping prompts

Runs per intent: 20

Prompt runs total: 400

Recommendations captured: 4,000

Distinct brands: 405

Three metrics carry the analysis.

Recommendation Share™ is a brand's share of all recommendations captured. The field sums to 100 percent.

Recommendation Frequency™ is the percent of the 400 prompt runs in which a brand appeared at least once.

Recommendation Position™ is the average rank in the answer when the brand appeared. Lower is better.

Marketplaces are not brands.

Retailers and marketplaces such as Chewy, Petco, Amazon, PetSmart, and Frisco were separated out and excluded from the brand level analysis.

The contest measured here is between single brands and their own stores.

A note on pets.

Several of the largest pet food brands including Purina Pro Plan, Royal Canin, Hill's Science Diet, and Taste of the Wild are not in our store index, so they appear as off index and are excluded from the correlation analysis.

One brand, Wellness, was removed from the on index set after we found its name had auto matched to an unrelated domain. Rather than attach an incorrect store, we classified it as off index.

Other limitations remain the same.

One model.

One category.

One point in time.

Brands map only to stores we actually measure.

Anything else stays off index.

The Pet Recommendation Leaderboard

Top brands by share of voice across all 4,000 recommendations, with retailers excluded.

The AI Commerce Score™ column reflects each brand's actual score from our index where a measured store exists.

Pet category results showing Recommendation Share™, Recommendation Frequency™, Recommendation Position™, and AI Commerce Score™ for leading AI recommended brands.

The Central Finding

If AI recommended the stores that are easiest for AI systems to read, recommendation frequency and AI Commerce Score™ would move together.

They do not.

In pets, they move slightly in the opposite direction.

Across the on index single brand stores, the correlation between Recommendation Frequency™ and AI Commerce Score™ is r = -0.366 (n = 39).

That is the first category where the relationship is meaningfully negative.

The brands that appear most often in AI recommendations tend to have weaker stores.

Blue Buffalo appears in 27.8 percent of prompts while scoring only 30.

Merrick appears in 24.8 percent of prompts while scoring 42.

Purina Pro Plan appears in 26.8 percent of prompts despite not even being represented in our store index.

Meanwhile some of the strongest stores in the category barely appear at all.

**Benebone scores 79.

Casper scores 73.

Pawstruck scores 72.**

Yet none of them approach the recommendation frequency of the legacy pet food brands.

This is the clearest inversion we have measured so far.

The brands AI remembers are not the brands with the most AI ready stores.

The average AI Commerce Score™ across all on index recommended brands is just 52.7.

We have now measured four categories.

Beauty: r = 0.17
Supplements: r = -0.015
Coffee: r = 0.019
Pets: r = -0.366

Across more than 16,000 recommendations, recommendation frequency is never positively associated with store readiness.

In three categories the relationship is effectively zero.

In pets it becomes negative.

Brands, Not Retailers

Pets is also the most marketplace driven category we have measured.

Retailers and marketplaces such as Chewy, Petco, and Amazon accounted for 7.1 percent of all recommendations.

That is significantly higher than beauty, supplements, or coffee.

Only 19.8 percent of recommendations map to a single brand store that we actually measure.

That is the lowest share we have seen.

Two factors drive that result.

First, pet commerce is heavily concentrated around large retailers and marketplaces.

Second, some of the most recommended brands in the category are not represented in our store universe.

Purina
Royal Canin
Hill's Science Diet
Taste of the Wild

These brands are recommended because of brand familiarity, not because an AI system evaluated their stores.

That is itself an important finding.

What It Means

Visibility is not recommendation.

Recommendation is not readiness.

Today AI still recommends largely from memory.

It reaches for names that appeared most often in training data and public discussion.

That is why the dominant pet food brands continue to win recommendations even when their stores are weak, absent, or difficult for AI systems to evaluate.

But that advantage is temporary.

As AI shopping evolves toward retrieval, browsing agents, comparison engines, and autonomous purchasing, recommendation decisions will increasingly depend on what agents can actually read, understand, verify, and trust.

The brands relying on historical recognition have the most to lose.

The brands building readable, trustworthy, machine legible stores today are the ones most likely to keep the recommendation when memory is no longer enough.

That is exactly the gap the Recommendation Intelligence Framework™ was built to measure.

AI Readability™
AI Understanding™
AI Trust™
Recommendation Intelligence™
Decision Confidence™

The future of AI commerce belongs to brands that are understandable by machines, not just familiar to humans.

Want to Know Where Your Store Stands?

Get a free AI Commerce Score™

https://atomfoundry.dev

Full pets dataset and interactive results

https://atomfoundry.dev/research/state-of-ai-recommendations-pets

Decision Confidence™ Is The Missing Layer Between Recommendation And Revenue

Daniel Pokorný — Tue, 16 Jun 2026 12:39:51 +0000

Decision Confidence™ Is The Missing Layer Between Recommendation And Revenue

Most discussions about AI commerce focus on visibility.

Can AI find your business? Can AI understand your products? Can AI recommend your brand?

These are important questions.

But they are not the final question.

Because recommendation is not the purchase.

A customer can discover your business. A customer can trust an AI recommendation. A customer can visit your website.

And still leave without buying.

Why?

Because recommendation creates consideration.

Purchase requires confidence.

This is where Decision Confidence™ begins.

Decision Confidence™ measures whether customers have enough certainty to move from consideration to purchase. The framework evaluates trust signals, clarity, proof, risk reduction, and decision friction to understand how confidently a customer can make a buying decision.

The Hidden Gap Between Recommendation And Purchase

Most commerce frameworks stop at recommendation.

They assume that if a customer arrives on the website, the hard part is over.

In reality, recommendation only creates an opportunity.

The customer still needs to answer several questions:

Is this business trustworthy?
Is this product right for me?
Am I making the right decision?
What happens if something goes wrong?
Is there a better alternative?

Every unanswered question creates uncertainty.

And uncertainty reduces the probability of purchase.

The sequence looks like this:

AI Readability™

↓

AI Understanding™

↓

AI Trust™

↓

Recommendation Intelligence™

↓

Decision Confidence™

↓

Purchase

↓

Revenue

Recommendation creates attention.

Decision Confidence creates action.

What Is Decision Confidence™?

Decision Confidence™ measures the level of certainty a customer has before making a purchase decision.

It evaluates whether a business provides enough information, proof, reassurance, and trust signals to reduce hesitation and uncertainty.

In simple terms:

Recommendation answers:

Should I consider this business?

Decision Confidence answers:

Am I comfortable enough to buy?

These are not the same thing.

Many businesses successfully generate consideration while failing to generate confidence.

The result is abandoned sessions, delayed decisions, and lost revenue.

Why Confidence Matters

Every purchase involves risk.

Customers constantly evaluate that risk.

They ask themselves:

Will this product meet expectations?
Can I trust this company?
Is the price justified?
What if I need a refund?
What if I choose incorrectly?

The greater the uncertainty, the lower the likelihood of purchase.

Confidence reduces uncertainty.

And reducing uncertainty is often more important than increasing traffic.

The Five Dimensions Of Decision Confidence™

Decision Confidence™ is measured through five core signals.

1. Trust Signals

Does the business provide evidence that it is legitimate and trustworthy?

Examples include:

Reviews
Ratings
Testimonials
Security indicators
Business information
Guarantees

Trust reduces perceived risk.

Without trust, confidence collapses.

2. Clarity

Can customers immediately understand what is being offered?

Confusion creates hesitation.

Customers should instantly understand:

What the product does
Who it is for
Why it is different
What they receive

Clarity reduces cognitive effort.

3. Proof

Can the business demonstrate that its claims are true?

Proof includes:

Case studies
Customer outcomes
Before-and-after examples
Expert endorsements
Independent validation

Claims create interest.

Proof creates belief.

4. Risk Reduction

Does the business actively remove fear from the decision?

Examples include:

Refund policies
Free returns
Free trials
Warranties
Satisfaction guarantees

The lower the perceived downside, the easier the decision becomes.

5. Decision Friction

How difficult is it for a customer to move forward?

Friction includes:

Complicated checkout flows
Excessive form fields
Hidden costs
Missing information
Poor user experience

Every obstacle increases hesitation.

Every hesitation reduces confidence.

Recommendation Without Confidence Creates Revenue Leakage

Many businesses focus heavily on visibility.

Others focus on AI recommendations.

But recommendation alone does not create revenue.

A business can be:

Visible
Readable
Understandable
Trusted
Recommended

And still lose sales.

Because the customer never reaches sufficient confidence to act.

Recommendation generates visitors.

Confidence generates buyers.

This distinction becomes increasingly important as AI systems take on larger roles in product discovery and evaluation.

The Future Of AI Commerce

As AI systems become responsible for more discovery, comparison, and recommendation activity, confidence becomes one of the most important commercial variables.

AI may recommend. AI may compare. AI may shortlist.

But many purchasing decisions still require human approval.

The businesses that win will not simply be the most visible.

They will not simply be the most recommended.

They will be the businesses that create the highest level of decision confidence.

Because confidence converts consideration into action.

And action creates revenue.

The AI Commerce Intelligence Framework™ explains how AI systems move from discovering and understanding businesses to influencing recommendations, purchase decisions, and ultimately revenue. Decision Confidence™ represents the critical layer between recommendation and purchase.

Decision Confidence™ In The AI Commerce Intelligence Framework™

Decision Confidence™ is the fifth layer of the AI Commerce Intelligence Framework™.

AI Readability™ enables access.

AI Understanding™ enables interpretation.

AI Trust™ enables credibility.

Recommendation Intelligence™ enables selection.

Decision Confidence™ enables action.

Without confidence, recommendation has limited commercial value.

Without confidence, purchase does not happen.

Without purchase, revenue does not exist.

Decision Confidence™ measures the final layer before commerce becomes economic outcome.

Explore The Framework

AI Commerce Intelligence Framework™

https://atomfoundry.dev/framework

Decision Confidence™

https://atomfoundry.dev/framework/decision-confidence

Decision Confidence™ is part of the AI Commerce Intelligence Framework™ developed by Atom Foundry.

Building the AI Commerce Intelligence Layer™.

Recommendation Intelligence Is Becoming The Next Layer Of AI Commerce

Daniel Pokorný — Mon, 15 Jun 2026 11:52:52 +0000

Recommendation Intelligence Is Becoming The Next Layer Of AI Commerce

Most businesses are focused on AI visibility.

Some are starting to focus on AI understanding.

A few are beginning to think about AI trust.

Very few are asking an even more important question.

Does AI actually recommend your business?

Because visibility and recommendation are not the same thing.

Understanding and recommendation are not the same thing.

Trust and recommendation are not the same thing.

A business can be visible. A business can be understood. A business can even appear trustworthy.

And still never become the recommendation.

As AI systems become increasingly involved in product discovery, evaluation, and decision-making, recommendation is becoming a critical layer of AI commerce.

The Recommendation Intelligence Framework™ measures how often, how prominently, and for which buyer intents AI systems recommend businesses. Recommendation Frequency™, Recommendation Position™, Recommendation Share™, Competitor Comparison, and Intent Match help quantify recommendation behavior across AI systems.

The Difference Between Visibility And Recommendation

AI Visibility asks a simple question.

Can AI see me?

Recommendation Intelligence asks a different question.

Does AI choose me?

Those are not the same thing.

A business may appear inside AI-generated answers.

It may be cited. It may be mentioned.

Yet never become the actual recommendation.

Visibility creates possibility. Recommendation creates influence.

The Recommendation Problem

Imagine a customer asks:

Best protein powder
Best CRM for startups
Best moisturizer for sensitive skin

An AI system may know dozens of valid options.

The challenge is no longer discovery. The challenge is selection.

Only a handful of businesses will actually be recommended.

Most will remain alternatives.

The difference between being known and being chosen is where Recommendation Intelligence begins.

Why Recommendation Matters

Recommendation systems operate differently than search systems.

Search systems retrieve information.

Recommendation systems prioritize options.

When multiple businesses appear relevant, AI systems must determine which businesses deserve attention.

This is where recommendation becomes important.

Recommendation reduces complexity. Recommendation narrows choice. Recommendation influences decisions.

The businesses that consistently receive recommendations gain a significant advantage over businesses that merely appear.

What Recommendation Intelligence Measures

Recommendation Intelligence™ measures how AI systems recommend businesses, products, and brands.

Several factors influence this outcome.

Recommendation Frequency™

How often does a business appear across recommendations?

Recommendation Position™

Where does the business appear within recommendation lists?

Recommendation Share™

How much recommendation visibility does a business capture compared to competitors?

Competitor Comparison

Which businesses consistently outrank others?

Intent Match

For which buying intents does a business appear?

Together these metrics help quantify recommendation behavior across AI systems.

Recommendation Intelligence™ introduces a measurable framework for understanding recommendation behavior. Recommendation Frequency™, Recommendation Position™, Recommendation Share™, Competitor Comparison, and Intent Match help quantify how AI systems choose businesses during product discovery and evaluation.

Recommendation Is Not Readiness

This is where many businesses become confused.

Readability answers one question.

Can AI access and interpret my business?

Recommendation answers a different question.

Will AI choose my business?

Those are not the same thing.

Recent Atom Foundry research analyzed:

20,000 AI-generated recommendations
1,490 brands
5 ecommerce categories
100 shopping intents

Across every category, Recommendation Frequency™ showed little to no measurable relationship with AI Commerce Score™.

Many highly recommended brands had weak stores.

Many highly optimized stores received little recommendation visibility.

Recommendation appears to operate according to different rules.

The Hidden Cost Of Low Recommendation Visibility

Most businesses never realize when recommendation becomes a limitation.

There is no dashboard. No recommendation ranking report. No recommendation analytics platform.

The business simply appears less often.

Competitors receive more recommendations.

AI systems route more attention elsewhere.

The business is not invisible.

It is simply not being chosen.

Why Recommendation Matters More Than Visibility

Many businesses focus on becoming visible to AI systems.

Others focus on becoming understandable.

Some are beginning to focus on trust.

The next challenge is becoming recommendable.

Visibility creates discovery.

Understanding creates interpretation.

Trust creates confidence.

Recommendation creates selection.

The sequence looks like this.

AI Readability™

↓

AI Understanding™

↓

AI Trust™

↓

Recommendation Intelligence™

↓

Decision Confidence™

↓

Purchase

↓

Revenue

Recommendation is the bridge between trust and action.

Without recommendation, commercial influence remains limited.

The AI Commerce Intelligence Framework™ maps the stages AI systems use to discover, understand, evaluate, recommend, and route customers to businesses. The framework connects AI Readability™, AI Understanding™, AI Trust™, Recommendation Intelligence™, and Decision Confidence™ to the commercial outcomes that ultimately drive revenue.

Learn more about the AI Commerce Intelligence Framework™

https://atomfoundry.dev/framework

Final Thought

The first generation of AI optimization focused on visibility.

The second generation focused on understanding.

The third generation focuses on trust.

The fourth generation will focus on recommendation.

Because recommendation systems do not simply retrieve information.

They choose.

And the businesses that AI consistently choose will increasingly shape the future of commerce.

Recommendation Intelligence™ is the fourth layer of the AI Commerce Intelligence Framework™.

It is where confidence becomes selection.

And where influence begins to become revenue.

Related Frameworks

The State of AI Recommendations in Coffee

Daniel Pokorný — Sun, 14 Jun 2026 20:27:35 +0000

Recommendation Intelligence Research™ · Atom Foundry · June 2026

We asked one AI model 20 high intent coffee shopping questions, 20 times each. Then we checked the same thing we checked in beauty and supplements. Does being recommended by AI have anything to do with how AI ready your store actually is?

The answer, across another 4,000 recommendations, is again almost nothing.

Three categories now. The same result.

Methodology

Same method as the beauty and supplements reports, so all three are directly comparable. Everything below is computed from real captured responses. Nothing is estimated or projected.

Category: Coffee

Model: one model (gpt-4o-mini)

Intents: 20 high intent shopping prompts

Runs per intent: 20

Prompt runs total: 400

Recommendations captured: 4,000

Distinct brands: 228

Three metrics carry the analysis.

Recommendation Share™ is a brand's share of all recommendations captured. The field sums to 100 percent.

Recommendation Frequency™ is the percent of the 400 prompt runs in which a brand appeared at least once.

Recommendation Position™ is the average rank in the answer when the brand appeared. Lower is better.

Marketplaces are not brands. Retailers and grocery marketplaces such as Amazon, Costco, Trader Joe's, and Trade Coffee were separated out and excluded from the brand level analysis. The contest measured here is between single brands and their own stores.

Other limits we are honest about. One model, one category, at one point in time. AI recommendations vary run to run, which is why each intent ran 20 times. Brands map to the real stores we measure where a match exists. Brands with no measured store are reported as off index, never invented.

The coffee recommendation leaderboard

Top brands by share of voice across all 4,000 recommendations, with retailers excluded. The AI Commerce Score™ column is each brand's real score from our index, where the store is something we measure.

Coffee category results showing Recommendation Share™, Frequency™, Position™, and AI Commerce Score™ for leading AI recommended brands.

The central finding

If AI recommended the stores that are easiest for AI to read, recommendation frequency and AI Commerce Score™ would move together. They do not. Across the on index single brand stores, the correlation is just r = 0.019 (n = 18), indistinguishable from zero and far below the threshold for significance. This is not a weak link. It is no measurable relationship at all.

The clearest examples are Peet's Coffee and Stumptown Coffee Roasters. Peet's is recommended in 96.5 percent of prompts and Stumptown in 78.3 percent, yet their stores score just 56 and 58. Blue Bottle Coffee appears in 75.3 percent of prompts while scoring only 36, deep in AI Invisible Risk.

The mirror image is the best built stores in the set. Bean Box at 74, Coffee Circle at 70, and Laird Superfood at 69 barely appear at all. AI recommends the names it knows, not the stores it can read.

The most recommended brand overall, Peet's Coffee, sits at a middling 56. The average AI Commerce Score™ across all on index recommended brands is only 55.7, and the ten most recommended brands average essentially the same score as the field. Being recommended buys you nothing on readiness.

We saw the same thing in beauty and supplements. Beauty came in at r = 0.17. Supplements came in at r = -0.015.

Three independent categories. More than 12,000 recommendations. The same result.

Recommendation frequency has no measurable link to store readiness.

Brands, not retailers

When buyers ask AI for the best coffee, AI answers overwhelmingly with brands, not shops. Retailers and marketplaces like Amazon, Costco, and Trader Joe's accounted for just 72 of 4,000 recommendations, under 2 percent.

The contest is overwhelmingly between brands and their own stores.

49.8 percent of recommendations map to a single brand store we actually measure. Coffee maps to more direct stores than supplements because many leading coffee brands operate strong direct to consumer storefronts alongside wholesale and retail distribution.

A striking share of the most recommended names, including Starbucks, Lavazza, Dunkin', and Illy, are not in our index at all. These brands are recommended despite not being part of the store universe we measure.

That is itself a finding.

AI is often recommending brand familiarity rather than store readiness.

What it means

Visibility is not recommendation.

Recommendation is not readiness.

Today AI recommends from memory. It reaches for the names it saw most during training, which is why a brand can be named in almost every answer while its store scores only in the middle of the pack.

That bias is baked in.

And it is temporary.

As AI shopping moves from recalling names to live retrieval and agents that browse, compare, and check out, the advantage shifts to the stores an agent can actually read, trust, and act on.

The brands coasting on fame today have the most to lose when the model changes.

The brands building readable, trustworthy, machine legible stores now are the ones that keep the recommendation when memory stops being enough.

Want to know where your store stands?

Get a free AI Commerce Score™

https://atomfoundry.dev

Full coffee dataset and interactive results

https://atomfoundry.dev/research/state-of-ai-recommendations-coffee

# The State of AI Recommendations in Supplements

Daniel Pokorný — Sat, 13 Jun 2026 19:13:45 +0000

Recommendation Intelligence Research™ · Atom Foundry · June 2026

We captured 4,000 AI supplement recommendations and compared them against real e-commerce stores.

The question was simple:

Does being recommended by AI have anything to do with how AI-ready a store actually is?

Across another 4,000 recommendations, the answer is again: almost not at all.

This is the second category we have analyzed. Beauty showed the same pattern. Supplements replicate it.

Key Finding

Recommendation Frequency™ and AI Commerce Score™ show no measurable relationship.

Correlation: r = -0.015 (n = 39)

In other words:

Stores that AI recommends more are not more AI-ready.
Stores that score higher on AI readiness are not recommended more often.
Recommendation behavior appears largely disconnected from store quality.

Methodology

Everything below comes from captured model outputs.

Dataset

Category: Supplements
Model: GPT-4o-mini
Shopping intents: 20
Runs per intent: 20
Total prompt-runs: 400
Recommendations captured: 4,000
Distinct brands: 371

Metrics

Recommendation Share™

A brand's share of all recommendations captured.

Recommendation Frequency™

The percentage of prompt-runs where a brand appeared at least once.

Recommendation Position™

Average ranking position when a brand appeared.

Lower is better.

Important Note About Retailers

Retailers and marketplaces were excluded from brand analysis.

Sites such as:

Amazon
iHerb
GNC

sell hundreds of brands and would distort the results.

This study measures competition between individual brands and their own stores.

Supplement Recommendation Leaderboard

Brand	Share™	Freq™	Pos™	AI Commerce Score™	Readiness
Garden of Life	6.7%	66.8%	3.6	52	Low
Optimum Nutrition	4.3%	42.8%	1.4	81	Moderately
Cellucor	3.2%	31.5%	5.8	67	Low
NOW Foods	3.0%	30.3%	5.4	14	AI Invisible
Nature Made	2.5%	25.3%	3.3	n/a	Off-index
Thorne Research	2.3%	22.8%	6.2	n/a	Off-index
BSN	2.1%	20.8%	3.3	n/a	Off-index
MyProtein	1.9%	19.3%	6.0	n/a	Off-index
Kirkland Signature	1.8%	17.8%	7.4	n/a	Off-index
Dymatize	1.7%	16.8%	4.5	n/a	Off-index
Vital Proteins	1.7%	16.8%	5.0	61	Low
MegaFood	1.6%	16.0%	5.0	n/a	Off-index
New Chapter	1.6%	15.8%	4.6	n/a	Off-index
Kaged Muscle	1.6%	15.5%	5.2	n/a	Off-index
MuscleMilk	1.5%	14.5%	4.0	n/a	Off-index
Nature's Way	1.4%	13.5%	6.3	14	AI Invisible
Vega	1.3%	12.8%	6.0	52	Low

Recommendation Frequency Does Not Follow Readiness

If AI recommended the stores that are easiest for AI to read, recommendation frequency and AI Commerce Score™ would move together.

They do not.

Across all measured single-brand stores:

r = -0.015

Statistically, that is indistinguishable from zero.

Not weak. Not small.

Simply no measurable relationship.

Examples

Highly Recommended, Poorly Built

NOW Foods

Recommendation Frequency™: 30.3%
AI Commerce Score™: 14

Nature's Way

Recommendation Frequency™: 13.5%
AI Commerce Score™: 14

Both stores sit deep inside what we classify as AI Invisible Risk™.

Yet AI recommends them frequently.

Highly Built, Rarely Recommended

Some of the strongest stores in the category barely appear at all.

Examples include:

Liquid I.V. (90)
Moon Juice (81)
DripDrop (75)

Excellent store readiness.

Minimal recommendation visibility.

Beauty Showed The Same Pattern

This is not the first category.

Our beauty research found:

r = 0.17

Again, statistically insignificant.

Combined:

Beauty: 4,000 recommendations
Supplements: 4,000 recommendations

8,000 recommendations total.

The same conclusion appears twice.

Store readiness does not predict recommendation frequency.

Brands, Not Retailers

Supplement recommendations are overwhelmingly brand-driven.

Retailers and marketplaces represented only:

195 of 4,000 recommendations

Less than 5%.

Unlike beauty, however, supplements show a stronger marketplace dependency.

Only:

33.5% of recommendations

mapped to direct-to-consumer stores we measure.

Many supplement brands rely heavily on:

Amazon
iHerb
GNC

rather than their own storefronts.

That fragmentation is itself a signal.

Recommendation by Memory™

The pattern points in one direction.

AI appears to recommend brands it already knows.

Not stores it can necessarily read.

We call this:

Recommendation by Memory™

The model reaches into parametric memory and returns familiar names it has seen repeatedly during training.

Examples:

Garden of Life
NOW Foods
Optimum Nutrition
Nature Made

These brands win despite having average, weak, or unknown store readiness.

Why This Matters

Most AI search discussions focus on visibility.

Questions like: Was I mentioned? Was I cited? Did I appear?

Those questions matter.

But recommendation is a different layer.

A brand can appear in the candidate set and still never make the shortlist.

Our data suggests:

Visibility is not recommendation.

And:

Recommendation is not readiness.

At least not today.

What Happens Next?

Today commerce appears heavily memory-driven.

Tomorrow may be different.

As AI systems evolve toward:

live retrieval
browsing agents
autonomous purchasing agents

the advantage shifts from brands AI remembers to stores AI can actually:

read
trust
understand
transact with

The brands winning today because of fame may be the most exposed when that transition happens.

Conclusion

Across 4,000 supplement recommendations:

Recommendation Frequency™ did not correlate with AI Commerce Score™
Famous brands dominated recommendations
Poorly optimized stores frequently outperformed better-built stores
Recommendation behavior appears driven more by memory than readiness

The question is no longer: Can AI see your brand?

The harder question is: Why did AI choose your brand instead of someone else's?

That is the question Recommendation Intelligence Research™ is designed to answer.

Get a free AI Commerce Score™: https://atomfoundry.dev

Research Hub: https://atomfoundry.dev/research

# AI Trust Is Becoming The Next Layer Of AI Commerce

Daniel Pokorný — Fri, 12 Jun 2026 07:59:33 +0000

Most businesses are focused on AI visibility.

Some are starting to focus on AI understanding.

Very few are asking a more important question.

Can AI trust your business enough to recommend it?

Because understanding and trust are not the same thing.

An AI system can fully understand what you sell and still choose another option.

As AI systems become increasingly involved in product discovery, evaluation, and recommendation, trust is becoming a critical layer of AI commerce.

The AI Trust Framework™ measures the signals that influence recommendation confidence. Reviews, authority signals, reputation, consistency, and brand mentions help AI systems evaluate whether a business appears trustworthy enough to recommend.

The Difference Between Understanding And Trust

AI Understanding asks a simple question.

Can information be interpreted correctly?

AI Trust asks a different question.

Can that information be trusted?

A business may clearly communicate what it sells.

It may clearly define its products.

It may clearly describe its category.

Everything may be perfectly understandable.

Yet AI systems may still hesitate to recommend it.

Not because they do not understand the business.

Because they lack confidence in the signals surrounding it.

Understanding creates meaning.

Trust creates confidence.

The Recommendation Problem

Imagine a customer asks:

Best protein powder
Best CRM for startups
Best moisturizer for sensitive skin

An AI system may understand dozens of valid options.

The challenge is not understanding. The challenge is deciding which option deserves the recommendation.

That decision requires confidence.

And confidence is often built on trust.

Why Trust Matters

Recommendation systems operate differently than search systems.

Search systems primarily retrieve information.

Recommendation systems evaluate options.

When multiple businesses appear relevant, additional signals help determine which option should be surfaced.

This is where trust becomes important.

Trust reduces uncertainty.

The stronger the trust signals, the easier it becomes for recommendation systems to develop confidence in a particular business.

What AI Trust Measures

AI Trust™ measures the signals that help AI systems evaluate credibility, authority, and reliability.

Several factors influence this outcome.

Reviews

Customer reviews provide external validation.

They help confirm that products and services perform as expected.

Brand Mentions

Independent mentions across websites, publications, communities, and media help establish legitimacy.

Authority Signals

Research, expert references, industry recognition, and authoritative citations strengthen perceived expertise.

Reputation

Long term positive signals contribute to overall credibility.

Consistency

Businesses that present a consistent identity across channels create stronger confidence signals.

Visibility Does Not Guarantee Recommendation

This is where many businesses become confused.

Visibility answers one question.

Can AI see me?

Trust answers a different question.

Would AI risk recommending me?

Those are not the same thing.

A business can be visible. A business can be understood. A business can even be relevant.

Yet still fail to become the recommendation.

The Hidden Cost Of Low Trust

Most businesses never realize when trust becomes a limitation.

There is no notification. No warning. No ranking report.

The business simply receives fewer recommendations. The products appear less frequently. The confidence behind recommendations becomes weaker. The business is not invisible.

It is not fully trusted.

Why Trust Matters More Than Visibility

Many businesses focus on becoming visible to AI systems.

Others focus on becoming understandable.

The next challenge is becoming trustworthy.

Visibility creates discovery.

Understanding creates interpretation.

Trust creates recommendation confidence.

The sequence looks like this.

AI Readability™

↓ ↓

AI Understanding™

↓ ↓

AI Trust™

↓ ↓

Recommendation Intelligence™

↓ ↓

Decision Confidence™

↓ ↓

Purchase

↓ ↓

Revenue

Trust is the bridge between understanding and recommendation.

Without trust, recommendation becomes uncertain.

Learn more about the AI Commerce Intelligence Framework™

https://atomfoundry.dev/framework

Final Thought

The first generation of AI optimization focused on visibility.

The second generation focuses on understanding.

The third generation will likely focus on trust.

Because recommendation systems do not simply retrieve information.

They make decisions.

And every decision requires confidence.

AI Trust™ is the third layer of the AI Commerce Intelligence Framework™.

It is where understanding becomes confidence.

And where recommendation begins to emerge.

Related Frameworks

AI Readability™

https://atomfoundry.dev/framework/ai-readability

AI Understanding™

https://atomfoundry.dev/framework/ai-understanding

AI Trust™

https://atomfoundry.dev/framework/ai-trust

AI Commerce Intelligence Framework™

https://atomfoundry.dev/framework

AI Trust™ is part of the AI Commerce Intelligence Framework™ developed by Atom Foundry.

# I Analyzed 4,000 AI Product Recommendations And Found Almost No Relationship Between Recommendations And Store Quality

Daniel Pokorný — Thu, 11 Jun 2026 10:24:23 +0000

Most discussions around AI commerce focus on visibility.

Can ChatGPT find my brand? Can Gemini mention my products? Can Perplexity recommend my store?

Those questions matter.

But I became curious about a different question.

Does being recommended by AI actually have anything to do with how readable and machine friendly a store is?

To find out, I ran a small experiment.

The Setup

I used a single AI model and asked 20 high intent beauty shopping questions.

Examples included:

Best vitamin C serum
Best moisturizer for oily skin
Best anti aging skincare

Each prompt was executed 20 times.

The final dataset contained:

400 prompt runs
4,000 recommendations
238 unique brands

I then matched every recommended brand against its measured AI Commerce Score.

The Result

Across measured brands, the correlation between recommendation frequency and AI Commerce Score was:

r = 0.17

In practical terms, recommendation frequency and store quality appeared largely unrelated.

The Surprising Part

Some of the most frequently recommended brands had some of the weakest stores.

Brand	Recommendation Frequency	AI Commerce Score
Clinique	38.5%	14
SkinCeuticals	31.8%	14
Kiehl's	25.5%	14

All three stores scored only 14 out of 100.

Deep in what I classify as AI Invisible Risk.

Yet they continued to appear repeatedly in recommendations.

Meanwhile Drunk Elephant achieved the highest score in the dataset at 87 out of 100 but was not the most recommended brand.

What Seems To Be Happening

My interpretation is that today's AI systems still recommend heavily from memory.

Large brands have accumulated decades of awareness.

They appear in training data.

They appear in articles.

They appear in reviews.

They appear across the web.

As a result, recommendation systems already know them.

That knowledge appears to outweigh store quality in many recommendation scenarios.

At least for now.

Brands Beat Retailers

One finding surprised me even more.

Retailers barely appeared.

Amazon.

Sephora.

Ulta.

Combined, they accounted for only 14 recommendations out of 4,000.

AI overwhelmingly recommended brands rather than stores.

That suggests recommendation systems currently think more like category advisors than shopping directories.

Why This Matters

Today AI recommendations appear to be driven largely by familiarity.

Tomorrow may look different.

As AI systems increasingly browse websites directly, compare products, verify information, and eventually complete transactions on behalf of users, store quality may become significantly more important.

A brand can survive on recognition.

An AI agent cannot act on recognition alone.

It needs information. It needs structure. It needs trust signals. It needs something it can actually understand.

The Bigger Question

If recommendation frequency and store quality are largely disconnected today, what happens when AI systems move from remembering brands to evaluating businesses in real time?

Full Research

Full dataset, methodology, and leaderboard:

https://atomfoundry.dev/research/state-of-ai-recommendations-beauty

About Atom Foundry

Atom Foundry researches how AI systems discover, understand, trust, recommend, and route customers to businesses.

Current research areas include:

AI Readability™
AI Understanding™
AI Trust™
Recommendation Intelligence™
AI Commerce Intelligence™

Would you expect AI to recommend the most famous brands or the best built stores?