How we measure

A tracked prompt is your exact text — what you see on the prompts page is what the engines get asked, word for word, through their official APIs (search-grounded where the provider supports it). We don't rewrite your questions or generate hidden variants. If wording matters to you, track two prompts and compare them.

Every paid plan monitors the same bundle: ChatGPT using GPT-5.4, Claude using Haiku 4.5, Perplexity using Sonar, and Google AI Overviews. Google results are collected through DataForSEO's Standard normal queue at priority 1 and depth 10, with asynchronous AI Overview loading enabled. There are no customer-selectable engines or per-surface schedules.

Lite checks the full bundle every 48 hours. Starter, Pro, Consultant, and Agency check the full bundle daily. One active prompt occupies one shared prompt slot and is asked once per surface on each scheduled check.

AI answers are non-deterministic — the same prompt on the same engine can include you today and skip you tomorrow with nothing changed on your side. So headline numbers pool your recent observed checks: “named in 5 of the last 7 checks” tells you something one answer can't. Lite's 48-hour history remains a 48-hour history; we never interpolate missing dates into fake daily observations. Every number links to the raw answers behind it.

Being named in an answer and being cited as a source are tracked separately — most “appearances” are one without the other, and they call for different work. Citations supplied as structured provider data are normalized deterministically. Provider drilldowns remain separate from equally weighted four-surface aggregate metrics, so you can see both the portfolio view and where systems disagree.

A confirmed change means your recent checks clearly differ from the comparison window and the shift persisted. Ordinary answer wobble renders in the app as “no detected change” and never becomes a claim by itself. Comparisons disclose when the available run counts differ.

The pooling and change gates use standard binomial statistics — Wilson score intervals on pooled windows, and Newcombe difference intervals for window-vs-window change tests (a change must exclude zero and persist). Coverage counts require a pooled lower bound clearing 5%, so a single stray mention can't flip an engine to “names you.” You'll rarely see this notation in the product — it drives what we show, it isn't what we show.

Failed checks and surfaces that returned no AI answer are excluded from denominators and shown separately — a failure is never counted as “absent.” New prompts are labeled until they have enough history; days before a prompt was activated are never treated as zeros.

Raw answers are analyzed by the full-v5 GPT-5.4-mini extraction system. Extraction converts answers into consistent mention, citation, stance, and recommendation evidence without changing the original answer.

A coverage version records the exact provider and model identity behind every observation. When that bundle changes, a new immutable version begins. Older observations remain available, but analytics never silently mix them into the new bundle. Charts and reports treat the boundary as a break in historical comparability.

• No blended cross-engine score — engines are different systems.
• No “rank” metric — position in a list is shown as context, never trended as a headline.
• No claim from a single answer — one check is evidence, a week is a finding.
• No fabricated history — a prompt's data starts the day you activate it.