# Why AI Recommends Differently for Pet Food Than for Dumbbells

Yesterday I published a controlled experiment showing that simply turning web search on changed 77% of AI product recommendations.

Same model.

Same 50 buying prompts.

One variable.

If you missed it, you can read the full study here:

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

But that 77% is just the average.

The real story appears when you break the data down by category.

Recommendation Changes by Category

A Clear Pattern Emerged

The pattern was surprisingly consistent.

The more fragmented the market, the more browsing overrides memory.

After reading my previous research, data engineer Rami Yahyaoui suggested a hypothesis that fits these results remarkably well.

In fragmented markets, the model has a much weaker training signal.

There are simply too many brands, too little historical exposure, and not enough consistent information for the model to build strong internal knowledge.

As a result, retrieval becomes much more influential.

In categories dominated by a handful of household names, the model already "knows" the likely answer. Live search mostly confirms what it remembers.

In long-tail markets full of smaller brands, memory has much less to rely on, so whatever the retrieval layer surfaces at query time has a much greater impact on the final recommendation.

Why Pets Keeps Standing Out

Pets has been the most extreme category in every study I've run.

Not just in this experiment.

It also showed the strongest anomalies in my previous research comparing store quality with AI recommendation frequency.

Something about this market makes AI recommendations unusually unstable.

At the moment, my best explanation is exactly what the data suggests:

A fragmented market combined with a weak training signal makes retrieval dominate memory.

It's still a hypothesis, but one that consistently matches the data.

What This Means for Ecommerce

One conclusion is becoming increasingly clear.

There is no universal strategy for getting recommended by AI.

In categories like Pet Supplies, live retrieval appears to drive most recommendations. Being the brand that AI can discover and verify right now becomes critical.

In categories like Fitness, the model's memory still carries much more weight, making long-term brand prominence significantly more valuable.

Same AI.

Same buying question.

Different category.

Different recommendation dynamics.

Final Thoughts

One of the biggest lessons from this research is that we shouldn't treat AI recommendations as a single system.

Different categories appear to operate under different recommendation dynamics.

Understanding those differences may become just as important as understanding the AI model itself.

Full study and methodology:

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