How scores on the AI Visibility Index are produced.
Each company is evaluated by generating 5 realistic buyer questions for its category — questions a prospective customer would ask an AI assistant BEFORE knowing which vendors exist, never mentioning the company by name. Those exact 5 questions are then sent as-is to three different AI models: Claude (Anthropic), GPT-4o mini (OpenAI), and Perplexity Sonar Pro (web-grounded search). Perplexity is included specifically because it grounds its answers in live web results, which is the mechanism most likely to reflect real-world current visibility.
For each model, the 5 raw answers are analyzed to determine: whether the company was mentioned at all, what position it appeared in across mentions, and whether its description (if any) was accurate. A deterministic formula — not a model's own self-assessment — converts this into a 0–10 score per model:
| Component | Condition | Points |
|---|---|---|
| Position | 1st mention | 6 |
| Position | 2nd mention | 4 |
| Position | 3rd or later mention | 2 |
| Position | Not mentioned | 0 |
| Accuracy | Description accurate (pass) — only scored if mentioned | 4 |
| Accuracy | Description weak | 2 |
| Accuracy | Description fails / not mentioned | 0 |
model_score = position points + accuracy points (maximum 10, minimum 0). You can recompute any model's score by hand from the position and accuracy shown in that model's row on the company page.
The company's visibility_score is the average of the model_score values across all models with a successful (status = "ok") query, rounded to one decimal place. This is a plain average computed in code — not a separate AI judgment.
Scores can shift slightly on re-scoring, since AI model answers are not perfectly deterministic — the same question asked twice may return slightly different phrasing or ordering. The scoring methodology (best position across 5 questions) is designed to smooth over this variance, but exact re-runs are not guaranteed to produce identical scores.
If a model's API call fails for a company, that model's row is marked "needs manual review" and is excluded entirely from the average — we never substitute a zero, a guess, or an assumed absence for a call that simply didn't complete. A company scored across 2 working models will show its average over those 2 models only, clearly marked as such.
Category is provided by the company being scanned, since automated category detection from a domain name alone is unreliable — a mismatched category will produce a misleading score.
Category wording matters — an ambiguous or broad category description can pull generated buyer questions toward an adjacent but incorrect industry (e.g. "behavioral" read as analytics rather than diagnostics). Companies should describe their category as specifically and unambiguously as possible.
Every company page lists the exact 5 buyer questions used to test it, along with the raw excerpt of what each model actually said. Nothing is summarized away — you can verify the score against the real model output.
The free instant checker at /ai-visibility gives you a quick one-time score. The AI Visibility Index is the permanent, publicly listed, periodically re-scored version — built for companies who want ongoing tracking and a public comparison page.
AI Visibility Pro ($29/mo) unlocks ongoing monitoring, a score history chart, a shareable embeddable badge for your own site, and competitor watch — all built on top of the same scoring pipeline shown above. Below is an example of the badge Pro subscribers can embed:
Unlock AI Visibility Pro — $29/mo
← Back to the AI Visibility Index