How a reading is made.
Six steps, run weekly, each one logged. The artifact looks simple (a score, the drifts, the fixes, the verifications) because the work behind it is specific.
- 01
Probing
Every week Butterfly asks a structured question set against ChatGPT, Claude, Gemini, Perplexity, Copilot, Grok, Mistral and Meta AI. The set is shaped to your category, your competitors and your markets, and written the way buyers ask at each stage: which brands to consider, why this one over that one, is the claim real, what does it cost. We ask for interpretation, not retrieval: "what is Marlow" is a question every model answers; "is Marlow worth it over Pilot for a small office" is the one that moves revenue.
- 02
Reading, not counting
For every answer we capture three layers: the mention (were you named), the claims (what was said about you) and the provenance (which pages the model retrieved or appears to have learned from). Claims are extracted into typed statements: location, price, cadence, feature, certification, positioning, comparison.
- 03
Scoring against truth
Each claim is compared with your Brand Truth file and labelled aligned, drifted (true once, or imprecise), wrong, or missing (a differentiator the model never mentions). The answer score is the weighted share of aligned claims across engines and markets, 0 to 100, with missing differentiators penalised so that a short, accurate, forgettable answer does not score as a win.
- 04
Tracing the source
A claim with a known source is a claim you can fix. We trace each drifted or wrong claim to the most likely origin: your own stale page, a third-party listing, a review site, a forum thread, a press release, a Wikipedia edit. Citations are used where engines expose them; otherwise we rank candidate sources by lexical overlap and crawl recency.
- 05
Drafting the fix
The fix is produced in the form the surface needs: a directive in llms.txt, a schema block, a rewritten paragraph with a citable sentence, a partner listing update. Each is validated (schema.org, lint, links) and bundled with evidence. Fixes are opened as pull requests or tickets, never as recommendations in a PDF.
- 06
Verifying
Seven days after a fix lands, the same questions are re-asked on the same engines. Each fix is marked landed, partial or not yet, with the before and after text. Only landed fixes count toward your score change. Verification results are kept so you can audit any claim back to the week it moved.
Three numbers, each with a before and after
Answer score (0 to 100): the accuracy and completeness of what engines say about you, measured against your Brand Truth file.
Share of answer: in questions where any brand is recommended, how often you are one of them, and in what position.
Citation rate: how often your own pages are the cited or inferred source of a claim about you, rather than a third party.
Scores are reported per engine and per market, and as a weighted blend. We publish the weights in your account and never change them mid-quarter without telling you.
What we do not claim
We cannot force a model to say anything. llms.txt states intent; models may ignore it. What we can do is make the truth the easiest thing to retrieve, remove the stale sources that compete with it, and measure honestly whether the answer moved. When it does not, the reading says so.
Change one page. Move every answer.
Run a free scan and see what ChatGPT, Claude, Gemini and Perplexity tell your buyers today. First reading within 48 hours, no card needed.