Why we sample real answers (and don't model them)
There's no public API for "what did ChatGPT say last Tuesday". Here's why Citations.io samples real answers instead of modelling visibility scores.
The temptation to model
It's tempting to score AI visibility from inputs you already have: backlink count, domain authority, schema markup, etc. Cheap to compute, no API costs, easy to explain.
The problem: the engines don't use those inputs the way you'd expect. A modelled score predicts should you appear in an answer. A sampled score tells you did you appear.
What sampling actually means
Citations.io runs each tracked prompt against the four major engines (ChatGPT, Perplexity, Gemini, Claude), up to 5 sampled answers per engine per scan. We capture the verbatim answer and tag mentions of your brand and every named competitor.
Everything on the dashboard — visibility %, share of answer, recommendation rate, citation coverage — rolls up from those sampled answers. You can click any number and read the answer it came from.
Confidence labels
A single sampled answer is just an anecdote. We tag every rollup with a confidence label:
- Directional (1–20 answers) — useful for trend direction, not absolute claims.
- Moderate (21–100 answers) — solid signal, worth reporting weekly.
- Strong (101+ answers) — confident enough for board reporting.
This is the difference between "your visibility dropped 12 points" (sometimes a modelled scare) and "your visibility dropped 12 points with strong confidence over 240 answers" (actionable signal).