A loop, not a dashboard.
Butterfly is built around one rule: a finding only counts when the model says it differently. Everything below exists to make that happen without a meeting.
- 01 Listen
Ask what your buyers ask
Every week Butterfly asks hundreds of real buyer questions across ChatGPT, Claude, Gemini, Perplexity and more, in every market you sell in.
- 02 Trace
Find the source of every claim
Each claim is extracted, scored against your Brand Truth file, and traced to the page, listing or thread that taught the model to say it.
- 03 Fix
Ship the correction, not a slide
Butterfly drafts llms.txt, JSON-LD, page rewrites and partner copy, then opens the pull request or ticket with proof attached.
- 04 Prove
Re-ask until the answer moves
Nothing is done until the model says it differently. The next reading verifies every fix and shows the before and after.
A question set shaped like your market
You give Butterfly a website and a category. It builds a question matrix the way a buyer would ask: consideration ("which brands should I look at"), comparison ("X or Y for a small team"), verification ("is the sustainability claim real"), and transaction ("cheapest plan with SSO"). Questions are localized per market and refreshed from Search Console and real assistant logs where you connect them.
questions:
market: en-CA
stages: [consider, compare, verify, buy]
count: 600
engines: [chatgpt, claude, gemini, perplexity, copilot, grok, mistral, meta]
cadence: weekly # Monday 06:00 localEvery claim, scored and sourced
Each answer is split into claims about you: location, price, features, positioning, who you are compared with. A claim is scored against your Brand Truth file as aligned, drifted, wrong or missing, then traced to the page, listing or thread the model most likely learned it from. A claim with a known source is a claim you can fix.
claim: "ships every two weeks"
status: drift
truth: "ships weekly"
source:
url: retailer.example/marlow
type: third-party-listing
owner: partnerships
seen_in: [chatgpt, gemini]Finished artifacts, opened where work happens
Butterfly drafts the correction in the form the surface needs: an llms.txt directive, a JSON-LD block, a rewritten paragraph with a citable sentence, a partner listing update. It validates the output, attaches the evidence (the drifted answer, the source, the truth) and opens a pull request or a ticket. You review in GitHub, Linear, Jira or Slack. Nothing waits in a slide.
PR #88 fix(about): roast location and shipping cadence
+ src/content/about.md (2 claims)
+ public/llms.txt (3 directives)
+ public/schema/org.jsonld (validated)
evidence: 4 answers, 2 sources, 1 truth diffNothing is done until the answer moves
Seven days after a fix ships, the same questions are asked again on the same engines. Each fix is marked landed, partial or not yet, with the before and after. Your Monday reading reports answer score, share of answer and citation rate with deltas, and the one decision that still needs a person.
verification · PR #88
chatgpt landed "roasted in Vancouver, ships weekly"
claude landed
gemini landed
perplexity partial price still $32 → cache, re-ask Thu
score 61 → 88 share 32% → 41%One file every fix is drafted from.
Your Brand Truth file is the versioned, machine-readable statement of what your brand is: positioning, products, prices, locations, proof points, claims you can substantiate, phrasings you reject. Butterfly drafts it from your site in the first hour, you correct it once, and every score and fix is measured against it from then on. Export it any time as Markdown, JSON-LD and llms.txt.
Positioning
Category, frame of reference, who you are for, who you are not for.
Claims and proof
Each claim paired with the evidence that substantiates it. Citable by design.
Facts
Prices, locations, cadences, certifications, integrations. The things models get wrong first.
Voice
Preferred phrasings, prohibited phrasings, register per audience.
Competitors
Who you accept being compared with and how the comparison should read.
Versioning
Every change tracked: who, when, why. Diffable in your repository.
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.