How we measure what real users see

Mention rate = (runs that mentioned the brand) / (total successful runs). We report Wilson CI95 alongside the point estimate. Wilson is the right interval for binary proportions on small samples — 5 runs is enough to compute it, but you should treat the bounds as wide until you push past 20 runs.

Rank position is integer-valued and skewed; we resample with replacement (default 1000 iterations) to estimate the rank distribution. Bootstrap output is non-deterministic at the 4th decimal — we use the deterministic Wilson interval in production and surface bootstrap only on the deep-mode response.

In deep mode, `providers[name].mentioned` is true when ≥ 50% of runs mentioned the brand. The threshold is a defensible default (matches the natural intuition of 'usually shows up'); flagged for product decision in plan §17.8 #8.

Reported as `variance.rank_stddev`. Excludes runs where the brand wasn't mentioned (rank null). High stddev means a flaky ranking — a tool builder that surfaces this number to their customer is doing the customer a favor.

`brand.toLowerCase()` against `text.toLowerCase()`. Catches every exact mention including nested forms ('Notion AI', 'Notion's').

For each whitespace-separated word in the response within ±25% of the brand length, compute edit distance. ≤2 edits is enough to forgive 'Pereplexity' or 'Notion' → 'Notiom' typos without false-positive matches on unrelated short words.

240 character window around the first match (120 before, 120 after) extracted from the original text so reading the context preserves the model's actual phrasing.

Plan §6.4 mandates deterministic primary brand detection. LLM-based extraction adds 500ms median latency, ~1¢ per call, and varies run-to-run. The deterministic path is sub-millisecond, free, and reproducible.

Our internal §22 test confirmed ChatGPT, Claude, and Gemini UIs only invoke web search on certain queries — and even when they do, the citations they render are inconsistent. Until we can detect/force search-mode reliably, we ship Perplexity-only UI scraping at perplexity_live and call out the rest as roadmap. mode:quick still queries all 4 official APIs in parallel.

When Perplexity gives us a list of source URLs we forward them verbatim. We do NOT validate URL reachability. Spot-checking 5 sample URLs from a recent Perplexity response, 4 returned 200 and 1 returned 404 (a stale Zoho article). If your downstream system needs guaranteed-live links, validate post-hoc.

The deterministic brand detector reports `sentiment: 'neutral'`. If you need positive/negative classification, pass an empty `track_brands` to /v1/check or call /v1/extract_brands directly with the response text.

Perplexity tightens its bot detection unpredictably. The /v1/health endpoint shows live numbers from a 7-day rolling window; partial failures trigger an automatic proportional refund (you only pay for what worked).