29,633 Reasons. 26,812 Were Unique.

I am building Atom Foundry, where we study how AI systems discover, evaluate and recommend e-commerce brands.

One of the hardest parts of this work is understanding why a model chooses one brand over another. We can observe the recommendation itself, but the reason behind it is much harder to measure.

At first, the obvious approach seemed simple. If the model recommends a brand, just ask it why.

So we did that.

We collected 29,633 reasons given by an AI model for its recommendations and then checked how often the model repeated the same explanation.

The result surprised me.

26,812 of the 29,633 reasons were completely unique text. That is about 90% of the entire dataset.

The most repeated reason appeared only 12 times.

Out of almost 30,000 recommendations.

Why I started looking at the explanations

This experiment came after several attempts to find measurable signals behind AI recommendations.

We had already tested things such as store quality, public fame and domain recognition. Some signals had a measurable effect, but they did not explain enough of the recommendation behavior to give us a reliable model.

That left another obvious source of information.

The model itself was already giving us an explanation for every recommendation.

If those explanations reflected a stable decision process, we should have been able to find some recurring patterns.

For example, imagine an AI system repeatedly choosing brands because they have stronger customer reviews. We would expect explanations related to reviews to appear frequently across similar recommendations.

The wording would not have to be identical. Language models can express the same idea in many ways. But we would expect to see a relatively small group of recurring explanations.

Instead, we found almost the opposite.

We checked every reason

The experiment itself was simple.

We took the complete reason field from our recommendation dataset and checked all 29,633 entries for exact text repetition. We wanted to see whether the model was repeatedly using the same explanations when making recommendations.

The results were:

29,633 reasons checked

26,812 distinct reasons

90% unique text

12 occurrences for the most repeated reason

The test used exact text matching. This is important because two explanations can communicate a similar idea while using completely different words. Those explanations would still count as different in our analysis.

So this study does not say that 90% of the underlying ideas were different. It says that 90% of the actual text strings were unique.

That makes the finding quite narrow, but also very clear.

Nine out of ten explanations were unique

The main result was difficult to ignore.

Nine out of ten explanations were unique.

The most common explanation in the entire dataset appeared only 12 times, which represents about 0.04% of all the reasons we collected.

If there were a small set of stable explanations sitting underneath the recommendation behavior, we expected those explanations to appear much more often.

They did not.

Instead, the model kept generating new explanations.

This is where the difference between a recommendation and an explanation becomes important.

The explanation may come after the decision

The result suggests that we should be careful about treating the model's explanation as a direct description of the mechanism behind its recommendation.

The model produces a recommendation, and then it can produce a plausible explanation that fits that recommendation. Because the model is very good at generating natural language, the explanation can sound specific, logical and confident.

That does not mean the explanation is false.

It means something more subtle.

The explanation may not tell us what actually caused the decision.

It may be a description generated after the decision has already happened.

That is what we mean by confabulation in this study. The model is capable of producing a convincing reason for a choice without that reason necessarily being a reliable representation of the process that produced the choice.

This distinction matters because a convincing explanation can easily be mistaken for evidence.

Why this can be a problem for ecommerce

This creates a very practical problem for anyone trying to optimize a store for AI recommendations.

Imagine that an AI recommends a competitor and you ask it why. The model might tell you that the competitor has better reviews. You could then invest time and money into improving your own review profile.

The next time you ask, the model might talk about product quality. You could respond by rewriting your product pages.

Later, it might mention price, selection, availability or brand reputation.

Every answer sounds useful because each answer gives you something that you can act on.

But our experiment shows why you should not automatically treat these explanations as evidence of what actually caused the recommendation.

You could end up optimizing your website around the story the model tells you instead of the signal that actually changes its behavior.

That is a very different problem.

What we should measure instead

For us, this changed the way we think about recommendation research.

If we want to understand what actually moves an AI recommendation, asking the model to explain itself is not enough. We need to measure the behavior directly.

The stronger approach is to change one variable while keeping the rest of the experiment as stable as possible, run the recommendation again and measure whether the selected brand changes.

That gives us evidence about whether a particular signal can actually influence the outcome.

The model's explanation gives us a story about the outcome.

The experiment gives us evidence about the outcome.

Those two things should not be treated as the same.

This is also why we are building our research around controlled comparisons rather than simply collecting what models say about themselves.

An important limitation

There is an important limitation to this study.

We used exact text matching, so two reasons that mean roughly the same thing can still be counted as different.

For example, better customer feedback and stronger reviews from buyers would be treated as two different strings.

We did not run a semantic similarity analysis in this test, so we cannot say how many of the 26,812 unique strings represented genuinely different concepts.

We can say something more specific.

The model did not repeatedly use the same textual explanations for its recommendations.

That is enough to show that the reason field is not a stable text based signal that we can use to reverse engineer the recommendation process.

We are also not claiming that this experiment reconstructs the model's internal computation. It does not.

The finding is narrower than that, and that is intentional.

What I took away from the experiment

I started this experiment thinking that the model's own explanations might give us a useful shortcut into its recommendation behavior.

Instead, the data made me much more cautious about using those explanations as evidence.

There is a meaningful difference between observing what a model chooses and reading what it says about that choice afterward.

The recommendation is behavior that we can measure.

The explanation is generated language that we can also measure, but it should not automatically be treated as the cause.

That distinction has become important in our research at Atom Foundry.

We are less interested in what an AI model says it values and more interested in what actually changes its decisions.

That means running controlled experiments, changing individual signals and measuring the result across repeated recommendations.

It is slower than simply asking the model why.

It is also much harder.

But if the goal is to understand AI commerce, I think that is the more useful path.

29,633 reasons gave us a simple lesson.

A convincing explanation is not necessarily the cause of a decision.

Sometimes it is simply a very good explanation.

Research

The full study, methodology and findings are available here:

Atom Foundry Research: 29,633 Reasons. 26,812 Unique. The Model Confabulates.