나는 2026년에 어떤 데이터베이스를 사용할 것인지 네 가지 인공지능 모델을 물었다. 네온이 이미 이겼습니다. 네 명의 도전자는 보이지 않습니다.

Every week I take one buyer category, ask ChatGPT, Claude, Gemini and Perplexity the five questions a real buyer would type, and count who gets named and who gets recommended first. Same questions for every company, so it is a fair board and not a vibe.

This week: databases and storage. I expected the incumbent reflex. I got the opposite, then a twist.

The serverless newcomers already won

To the models, the challengers are already the answer:

• Neon (serverless Postgres): recommended first in 14 of 20 conversations.
• Upstash (serverless Redis): first in 11 of 20.
• Turso (edge SQLite): first in 9 of 20.

That is real proof the door is not locked. A company younger than the incumbent it replaced can become AI’s default pick. Neon did it.

Then I asked about the specialized jobs

Same category, different sub-job, and the model snaps back to the incumbent every time:

• Vector search: the models pick Pinecone, Milvus and Weaviate. Qdrant is named a lot but recommended first only six times.
• Object storage: the answer is Amazon S3, Cloudflare R2 and Backblaze. Tigris is invisible, named zero times on ChatGPT, Claude and Gemini.
• Real-time analytics: ClickHouse, not Tinybird.
• The ORM: Prisma, not Drizzle.

Why

The newcomers that won did not win on features. By the time these models trained, enough independent writing already named them as the answer. They are in the inputs the model reads.

The invisible ones shipped great products, and it does not register, because the model is not evaluating products. It repeats what its inputs said. You cannot test your way into a recommendation you were never part of, and you cannot out-feature your way in either. The buyer who types “best vector database” and takes the first answer never sees Qdrant, no matter how good it is, until the inputs change.

See it yourself

The full board, with every model’s answer and the verbatim text, is here: What AI recommends for Databases and Storage. Counts only, no score, every number links to the actual answer.

If you want the same teardown for your own category, that is what Bersyn does.