This is the end-to-end pipeline behind every report. For grading specifics see Your visibility grade; for a summary version see the public methodology page.
1. Question selection
The scan selects up to 30 buyer-question templates for your industry and location — natural-language questions phrased the way buyers actually ask (details). On monitoring dashboards, your customized prompt set replaces the defaults.
2. Sampling across channels
Each question is asked on 12 channels:
- Eleven direct model channels — GPT-5.6 (OpenAI), Claude Sonnet 4.6, Gemini 3 Flash, Claude Opus 4.6, Gemini 3.1 Pro, DeepSeek V3, Kimi K3, GLM-5.2, Llama 3.3 70B, Qwen3 30B and Mistral Small 3.1 answer from their own knowledge via their model APIs, showing what they have internalized about your category.
- GPT-5.6 + Brave Search (grounded) — live Brave web results are retrieved for the question and fed to the model, approximating retrieval-augmented assistants and producing citations.
Up to 30 × 12 — hundreds of answers per scan, each stored verbatim.
3. Parsing and attribution
Each answer is parsed for:
- Brand mentions — your name and close variants, attributed at sentence level, so the dashboard can show the exact sentence in which you (or a competitor) appeared.
- Competitor names — every recommended brand, normalized and counted (details).
- Sentiment — whether each brand mention is positive, neutral or negative in context.
- Citations — every source URL the answer relied on, grouped by domain (Citation Gap).
4. Website diagnosis
In parallel, your own site gets the 9-point AI-readiness audit — at most three fetches (robots.txt, llms.txt, homepage), each capped and time-limited.
5. Scoring and storage
Mention rate and the A–F grade are computed, and everything is written to the report — permanently addressable at /r/<id>. On monitoring dashboards the scan becomes a data point in your trend history, feeding SOV, sentiment and trend charts.
Honest limits
- We sample the channels listed above, nothing more — see platform coverage.
- Model outputs vary run to run; single-scan numbers carry noise. Weekly re-sampling is how signal emerges.
- A few hundred samples per scan is a deliberate cost/robustness trade-off, and the identical protocol every week is what makes your trend line valid.