How RecoIndex measures AI recommendations
We measure one thing: when someone asks an AI assistant for a product recommendation, which brands does it name? We run a fixed, public set of buyer-intent questions across several models, repeat them, and report how often each brand appears — with sample sizes and the sources cited. Everything here is reproducible: ask the same questions and you should see broadly the same pattern.
1. The questions
Each category uses the same buyer-intent templates, instantiated with the category's product noun and (for verticals) its industry. We never steer toward a brand. Example set:
2. The models
We sample 5 low-cost models via OpenRouter. We display each vendor below; the exact version is recorded here and in the raw data:
3. Sampling
Language models are non-deterministic — ask twice, get different answers. So we never report a single answer. Each (question × model) is sampled repeatedly at a fixed temperature, and we report mention frequency with the sample size shown (e.g. "named in 18 of 24 answers"). A second low-cost model extracts brand names and cited URLs into structured data at temperature 0.
4. What we report
- Mention rate — share of sampled answers that named the brand.
- Model matrix — mention rate per brand on each model (the heatmap).
- Agreement — how much the models converge on the leader.
- Cited sources — domains the models referenced, ranked by frequency.
5. Limitations
- We query models via API, which can differ from the consumer chat apps (system prompts, browsing, personalization).
- Results aren't personalized — no location, history, or account context.
- Brand-name normalization is automated; near-duplicates are merged heuristically and refined over time.
- Sample depth grows weekly; early weeks have smaller n (shown on each page).
- Models and the web they cite change constantly — this is a moving picture, not a verdict.
6. Raw data
Every category page has a "download CSV" link with the per-model and combined numbers for that week. Updated 2026-W29.