/ BENCHMARKS
What AI engines recommend, by category.
Each table below is a recorded run: a set of the questions buyers in that category actually ask, put to answer engines, with every mention counted and every source kept. The numbers carry 95% Wilson score intervals, so where a sample cannot separate two brands the table says so instead of ranking them anyway.
We publish the runs whole, including the results that are awkward for the brands in them and for us. The method is written up separately, together with the run we did against our own brand, in which we were named in none of the answers.
Vector database
Recorded 2026-09-17 · Recall v0
6 buyer questions × 4 runs = 24 probes. 100% produced a usable answer (86–100%).
| Brand | Named in | 95% interval | |
|---|---|---|---|
| Pinecone | 88% | 69–96% | |
| Qdrant | 88% | 69–96% | |
| Weaviate | 83% | 64–93% | |
| Milvus | 83% | 64–93% | |
| Elasticsearch | 21% | 9–40% | |
| Chroma | 17% | 7–36% | |
| pgvector | 0% | 0–14% | |
| Redis | 0% | 0–14% |
Every bar is a Wilson score interval at n=24. The top four overlap, so this sample cannot rank Pinecone, Qdrant, Weaviate and Milvus against one another, and reporting that it can would be the error. What it does establish is the drop below them: 83% and 21% do not overlap, so that gap is real.
pgvector and Redis are named in none of the 24 answers, and pgvector ships inside Postgres. That is not low visibility, it is the wrong question: these ask for companies, vendors and platforms, and an extension is none of those. Different cause, different fix — and a ranking alone would have shown it as a zero with no explanation.
What the engines cited
Five of the ten most-cited sources are the measured vendors' own sites, which is the category a team wholly controls. Three belong to nobody in the table — semantic.io, aiworkflowlab.dev and inductivee.com — and each was cited five times or more.
| Source | Citations | Tier |
|---|---|---|
| qdrant.tech | 26 | own site |
| pinecone.io | 19 | own site |
| docs.weaviate.io | 16 | own site |
| weaviate.io | 12 | own site |
| semantic.io | 8 | — |
| aiworkflowlab.dev | 7 | — |
| learn.microsoft.com | 7 | — |
| github.com | 6 | — |
| milvus.io | 6 | own site |
| inductivee.com | 5 | — |
Decentralised exchange
Recorded 2026-07-27 · Recall v0
14 buyer questions × 3 runs = 42 probes. 100% produced a usable answer (92–100%). Engine: ChatGPT (gpt-5-mini, web search enabled).
| Brand | Named in | 95% interval | |
|---|---|---|---|
| Uniswap | 67% | 52–79% | |
| 1inch | 52% | 38–67% | |
| Jupiter | 36% | 23–51% | |
| Curve | 29% | 17–44% | |
| Orca | 24% | 13–39% | |
| Raydium | 21% | 12–36% | |
| Aerodrome | 14% | 7–28% | |
| PancakeSwap | 12% | 5–25% |
Every bar is a Wilson score interval at n=42. Uniswap and 1inch overlap, so this sample cannot establish that one beats the other, and saying so is more useful than a confident ranking the data does not support.
The most cited domain of all belongs to a wallet, not an exchange, answering questions about a category it does not compete in. A share-of-voice number alone would never have surfaced that.
The leaders absorb 61% of all naming (51–69%) across 107 mentions.
What the engines cited
Six of the ten most-cited sources are product documentation. That is the largest citation category for crypto questions, and the one a team wholly controls.
| Source | Citations | Tier |
|---|---|---|
| help.phantom.com | 15 | docs |
| coinbureau.com | 13 | press |
| defillama.com | 12 | analytics |
| developers.uniswap.org | 11 | docs |
| docs.jup.ag | 11 | docs |
| eco.com | 10 | — |
| docs.orca.so | 10 | docs |
| app.uniswap.org | 9 | — |
| docs.raydium.io | 7 | docs |
| help.1inch.com | 7 | docs |
Your category is not here yet
These are the categories we have run so far. A run for yours starts from the questions your buyers ask rather than the ones we guessed, and the result belongs to you whether or not it flatters you.