We expected the model makers to dominate the sources. Not one of them appears.
We ran fifteen buying questions through grounded search — the questions a real buyer types, not keywords — and recorded which tools the answers named and which sources they cited. 45 distinct sources, 136 citations. What follows is that data.
What AI models actually name
| Tool | Mentions |
|---|---|
| DeepSeek (also cited as DeepSeek V4, DeepSeek V4 Pro) | 14 |
| Qwen (also cited as Qwen3, Qwen 3.6) | 14 |
| GLM (also cited as GLM-5.2, Z AI / GLM) | 13 |
| Kimi (also cited as Kimi K2.7 Code, Kimi K3) | 11 |
| Gemma (also cited as Gemma 4, Google Gemma) | 10 |
| Mistral (also cited as Mistral Small, Mistral Small 4) | 8 |
| Llama (also cited as Llama 3.2 / 3.3, Llama 3.3) | 6 |
| Nemotron 3 (also cited as Nemotron 3 Super / Ultra) | 5 |
| Meta | 4 |
| MiniMax (also cited as MiniMax M3) | 4 |
| NVIDIA Nemotron | 4 |
| Falcon 3 | 1 |
| gpt-oss-120b | 1 |
| Muse Spark / Muse Glimmer | 1 |
Zero of the most-cited sources belong to a model maker. Not Meta, not DeepSeek, not Mistral, not Qwen. Every page the models cite is an infrastructure vendor: Fireworks, BentoML, Onyx, Vellum, Baseten, Hugging Face.
The companies building the models publish papers and model cards. The companies that serve them publish "best open source LLMs in 2026". Guess which gets retrieved.
This is the clearest structural gap we found in seven categories: the brands being ranked have handed the entire narrative to the layer above them, apparently without noticing. DeepSeek and Qwen tied at the top of the answers — but neither owns a single page that produced those answers.
Where the answers come from
These are the pages the models cited most. The top three carry 29% of all citations in this category, and Reddit threads carry 16% across 11 separate discussions.
Method
Fifteen buying questions, run through grounded search on 3 September 2026:
What are the best Open Source LLMs brands? · Which Open Source LLMs should I buy? · Best Open Source LLMs for beginners · Most trusted Open Source LLMs brands and why · Compare the top Open Source LLMs brands · What Open Source LLMs do experts recommend? · Best value Open Source LLMs · Which Open Source LLMs brands have the best reviews? · Top rated Open Source LLMs 2026 · What should I look for when choosing Open Source LLMs? · Alternatives to the most popular Open Source LLMs · Is there a better option than the leading Open Source LLMs brand? · What do people on Reddit recommend for Open Source LLMs? · Best Open Source LLMs for small businesses · What's the highest rated Open Source LLMs right now?
Every answer carries citations. We recorded all of them and ranked by frequency.
What this does not do. We did not test these tools, run trials, or compare features hands-on. Plenty of roundups claim they did. We measured which tools AI models recommend and which sources those recommendations came from — a different question, and the one that matters if your goal is being recommended rather than choosing.
Reproducibility. Ask the same questions again and substantially the same sources come back. In our testing, two independent runs of a category minutes apart returned the same URLs in the same order.
This is one of seven categories we mapped the same way. Across all seven, Reddit carried 16-27% of citations and the top three sources carried 29-33% — every single time.
This source map was made with Mayin.
Every number above came from running real buying questions through grounded search and recording what the answers cited. You can get the same map for your own category and brand in about 15 minutes, for $49 — one payment, no subscription, delivered in Telegram.