Four vector databases were named in every single answer: Milvus, pgvector, Qdrant, and Weaviate each landed 6 of 6 times when we asked ChatGPT and Gemini to recommend the best vector database software on 6 August 2026. No other product came close. The next tier, Pinecone and Redis, appeared in only 3 of 6 answers each.
We put six buyer questions to ChatGPT and Gemini, gave each one live web search, and recorded every product named and every source domain behind the answer. Eighteen product names showed up at least once across the six answers, but only 4 of them were universal. The gap between being named in 6 of 6 answers and being named in 1 of 6 is the whole story when a buyer asks an assistant to choose for them.
Pinecone, probably the most recognised commercial name in this category, was named in just 3 of 6 answers. It showed up for "small teams," "what should I use," and "recommendations," but not once in the bare "best," "best free," or "most affordable" answers. Brand recognition did not carry it: half of these answers left Pinecone out entirely.
How the shortlist changed from question to question
The four constants held across every phrasing, but the rest of each shortlist moved with the question, and this is the part worth watching. The bare "best vector database software" answer named 8 products, and two of them, Vespa and Zilliz, appeared in that one answer and nowhere else in the six.
Asking for "best free" reshuffled the supporting cast. That answer also named 8 products and brought in Chroma, OpenSearch, and Faiss, three names tied to open-source and self-hosting, while dropping pricier managed names like Vespa, Zilliz, and Turbopuffer. Redis was the one that survived the shift, named in the free answer as one of the 8.
"Most affordable" produced a 7-product shortlist led by Turbopuffer and ChromaDB, neither of which appeared in the "best free" answer. Turbopuffer was named in 2 of 6 answers total, and both were price-framed questions, the bare "best" and "most affordable," never the free-tool question. So "cheap" and "free" produced visibly different lists from the same two assistants.
Neither the "free" nor the "affordable" answer introduced a product unique to itself: every name in those two shortlists also appeared in at least one other question. The seven products named in only 1 of 6 answers clustered in the other four questions instead. MongoDB Vector Search appeared only in the "small teams" answer. LanceDB and Milvus Lite appeared only in "what should I use." Vald and MongoDB Atlas Vector Search appeared only in "recommendations." Vespa and Zilliz appeared only in the bare "best" answer.
Two products were named under two different labels, which is why the raw list looks longer than the real field. Chroma (named in 2 of 6) and ChromaDB (named in 2 of 6) are the same tool, split by how the assistants spelled it. MongoDB Vector Search (1 of 6) and MongoDB Atlas Vector Search (1 of 6) are the same product too. We are keeping both labels in the table below rather than folding them into a combined count the measurement never actually recorded.
The source map: who fed the answers
Reddit fed every answer. reddit.com was the most-cited domain, sitting behind 6 of 6 answers, tied with zenml.io, a roundup publisher, also at 6 of 6. Community threads and roundup articles, not vendor marketing, carried this category.
Instaclustr's roundup (instaclustr.com) was behind 5 of 6 answers. The OpenAI community forum (community.openai.com), the developer blog mastra.ai, and Medium each fed 4 of 6 answers. Below them sat firecrawl.dev, liveblocks.io, and redis.io at 2 of 6 apiece, then a long tail of single-answer sources including a Facebook thread and a YouTube video, each cited once.
The key insight is in two vendor-owned domains. redis.io, Redis's own site, was cited in 2 of 6 answers, and Redis the product was named in 3 of 6. milvus.io, Milvus's own site, was cited in 1 answer, and Milvus was named in all 6. Publishing your own explainer that the assistant then reads and repeats is one of the few levers a vendor controls directly, and in this measurement it plainly worked.
Review directories barely registered. g2.com appeared as a source in exactly 1 of 6 answers, and Capterra did not appear at all. In a category buyers actually research on forums, the classic directories account for 1 of 6 answers between them, against Reddit's 6 of 6.
What a vendor here would do to get named
If you sell a vector database and want to be in these answers, the source list is the target, not your market share. The four products named in 6 of 6 answers are the four woven through the recurring sources: the Reddit threads, the zenml.io and instaclustr.com roundups, the OpenAI forum. Pinecone's 3 of 6 is the counter-example, a strong brand that the free-and-affordable sources simply did not carry.
In practice that means being present where these answers were assembled: the Reddit discussions cited in 6 of 6 answers, the two roundups cited in 6 of 6 and 5 of 6, and the community forum and dev blogs cited in 4 of 6. It also means keeping your own explainer indexed and genuinely useful, the way redis.io and milvus.io fed answers that named Redis and Milvus. The assistants repeated pages that already existed; they invented nothing.
The full list, counted
Products named in at least 2 of the 6 answers. Chroma and ChromaDB are the same tool recorded under two labels, and we have kept both rows exactly as measured.
| Product | Named in | Share |
|---|---|---|
| Milvus | 6 of 6 | 100% |
| pgvector | 6 of 6 | 100% |
| Qdrant | 6 of 6 | 100% |
| Weaviate | 6 of 6 | 100% |
| Pinecone | 3 of 6 | 50% |
| Redis | 3 of 6 | 50% |
| Chroma | 2 of 6 | 33% |
| ChromaDB | 2 of 6 | 33% |
| Faiss | 2 of 6 | 33% |
| OpenSearch | 2 of 6 | 33% |
| Turbopuffer | 2 of 6 | 33% |
Seven more products were named in 1 of 6 answers each: LanceDB, Milvus Lite, MongoDB Atlas Vector Search, MongoDB Vector Search, Vald, Vespa, and Zilliz.
What this means for your category
Glotier does not sell a vector database, so we are correctly absent from all 6 of these answers. That is the honest result, and we are not going to dress it up as anything else.
What you just read is exactly what Glotier runs for a customer's own category: the same kind of six buyer questions, the same two assistants, the same record of who got named and which sources decided it. If you sell in this space, your name is either in those 6-of-6 answers or it is not, and that is a measurable fact rather than a guess. The check is free, needs no account, and takes about a minute.