Why AI recommends the same tools every time, and which slots you can actually win

Last week I tore down what four AI models recommend for databases: same buyer questions, 20 runs each, count who gets named. Neon, Upstash and Turso already own the generic slots. The specialists, Tigris and Tinybird, are close to invisible.

The post did fine. The comments did better. A handful of sharp people pushed on the why, and the answer is more useful than the findings were.

Two competitions, and you are probably in the wrong one

The models are not judging which product is best. They surface whatever their training data already corroborated for that exact phrasing. A worse tool with denser, clearer writing tied to a question beats a better tool nobody wrote about that way.

So there are two competitions running at once:

• Best product. What you actually build.
• Best-corroborated answer. What the sources a buyer's question pulls from already named.

Most founders pour everything into the first and assume the second follows. It does not. The model only scores the second.

Some slots are welded shut. Stop fighting them.

"Object storage" has meant Amazon S3 for fifteen years. That mental model is set in concrete across millions of pages. No comparison article, no migration guide, no amount of content moves "best object storage" off S3 on any timeline that matters to a startup. If your growth plan depends on winning a query like that, the plan is the problem.

The tell for a welded-shut slot: ask the four models the same generic question a few times and they all agree, every run. Agreement across models and across runs means consensus has formed. You are not getting in.

Some slots are wide open. That is where the work pays.

Now ask a narrow sub-job instead. "Vector database for X." "Analytics for Y." Watch what happens: the models hedge, name different tools, and the first pick shuffles run to run. That disagreement is the signal. Consensus has not formed yet, which means the slot is still being decided, which means you can be the one it decides on.

The challengers that won did exactly this. Neon, Upstash and Turso did not beat Postgres at "best database." They became the corroborated answer for "serverless Postgres / Redis / SQLite" while those mental models were still forming, and rode them as they widened.

So the move in a locked category is not to attack the incumbent's query. It is to find the sub-job nobody owns, become the best-corroborated answer for it in the buyer's own words, and let the category grow around you. Manufacture a young category you can actually win.

You can see which is which

This is the part founders miss: you do not have to guess whether a slot is welded or winnable. You can observe it. Run the buyer questions across the models, more than once, and look at the agreement. Tight agreement across models and runs is a settled slot. Disagreement is an open one. That turns "get recommended by AI" from a vibe into a map of where to spend.

That map is what Bersyn builds: who each model names, who gets recommended first, where they disagree, and the verbatim answers behind every number. If you want it for your own category, that is the whole product.

Method: category-representative buyer questions across ChatGPT, Claude, Gemini and Perplexity, multiple runs, reported with model versions and scan dates. No claim that any tool is good or bad, only what the models answered.