answer-accuracy.md
84 lines · skills/ai-seo/references/answer-accuracy.md at 88c79f6
Most marks come from matching words in the file, so treat them as pointers.
Answer Accuracy: Auditing What AI Says About You
Visibility counts whether you appear. Accuracy checks what the answer says once you do. A brand can show up in every answer and still be described with last year's pricing, a missing integration, or a limitation it fixed months ago.
Much of an answer is commentary the user never asked for. Profound's analysis of 50,000 responses (Profound, June 2026) found about 47% of response content was unsolicited editorial material: comparisons, rankings, caveats, and recommendations. That editorial layer is where wrong claims about you tend to live.
Profound sells AI visibility tracking, so treat its figures as vendor research. Other trackers that capture answer text for this kind of audit include Ahrefs Brand Radar, Semrush's AI toolkit, Peec AI, Otterly, and ZipTie.
Raw lines 1 to 8When to run it
- An assistant states a wrong price, plan limit, feature, or compatibility claim
- Sales hears prospects repeat something false that "ChatGPT said"
- After a repositioning, pricing change, rebrand, or product retirement
- Quarterly for any brand with a regulated, safety, or pricing-sensitive product
The claim ledger
Capture answers first, then log one row per claim. Record the engine, model or surface, locale, prompt, and capture date for every answer, since a claim you can't reproduce is hard to fix.
| Field | What goes in it |
|---|---|
| Engine and date | ChatGPT (search on), Gemini, AI Overviews, Perplexity, Copilot, Claude; capture date |
| Prompt | Exact wording, plus which prompt family it belongs to (see prompt-portfolio.md) |
| Claim | The sentence as the model wrote it |
| Claim type | Price, availability, compatibility, feature, limitation, comparison, ranking, recommendation, caveat, sentiment |
| Verdict | One of the seven below |
| Evidence | The current source that confirms or contradicts it |
| Likely source | The page the claim probably came from: your page, docs, a feed, a profile, a third-party article, a forum thread, or unknown |
| Severity | Critical, high, medium, or low |
| Owner | Who fixes the source |
Extract every claim that could change a buying decision, including the ones the user didn't ask about. A question about pricing that comes back with an unprompted "but it lacks SSO" is two claims.
Raw lines 16 to 33Seven verdicts
- Accurate and current
- Accurate but incomplete: true, but missing the fact that matters (the free plan exists, but the answer omits its limits)
- Ambiguous: could be read as true or false
- Stale: was true, isn't now
- False
- Subjective framing: an opinion stated as fact ("best for small teams," "hard to set up")
- Unverifiable: no current source either way
Check first-party facts against your own current product material, and claims about the outside world against reliable independent sources. Your preferred marketing wording is not evidence.
Raw lines 34 to 45Severity
| Level | Examples |
|---|---|
| Critical | Safety, legal, eligibility, price, availability, or compatibility errors that could cause a bad purchase or real harm |
| High | False comparisons, invented limitations, stale product facts in high-value prompt families |
| Medium | Incomplete or slanted framing that could plausibly change a decision |
| Low | Tone and wording that don't change the decision |
Weight severity by exposure too: a medium error that appears in every run of your top commercial prompt outranks a high error in one obscure prompt.
Raw lines 46 to 56Fix the source first
The answer is downstream of what the model retrieved or learned. Rewriting your homepage copy rarely fixes a claim that comes from a stale third-party review.
- Trace it. Search the claim's wording, check which sources the engine cited for that answer, and look for the page where the wrong fact lives.
- Fix your own sources. State the correct fact plainly on the page that should own it, with a date: pricing page, docs, changelog, comparison page, help center. Remove or redirect the outdated page. Update feeds and structured data that carry the same fact.
- Fix your profiles. G2, Capterra, app marketplaces, Crunchbase, LinkedIn, and partner directories often hold the stale version.
- Ask third parties for factual corrections only. Send the publisher the correct fact and a source. Don't ask them to remove legitimate criticism.
- Retest the same prompts after a set observation window (4–8 weeks is a reasonable starting point), with a few controlled variants. Report movement over time; one clean run proves nothing.
Never promise that a source fix will change the answer. Models retrain and re-retrieve on their own schedules.
Off limits: fabricated corrections, editing Wikipedia about yourself, astroturfed forum posts, and pressuring publishers to pull fair criticism. These break platform rules, and models are getting better at discounting them.
Raw lines 57 to 70Report
- Accuracy rate across the cohort, the riskiest narrative, and the main source problem
- The claim ledger
- Owned-source fixes, each with the exact fact and the page that should state it
- Third-party correction requests, factual and sourced
- The retest plan and date
- Unknowns
This audit pairs with the positioning work in positioning-and-consensus.md: accuracy fixes wrong facts, positioning fixes the wrong story. The AI-accuracy audit loop in marketing-loops can schedule it.
Framework lineage: claim ledger, verdict set, and severity scale adapted (re-expressed and restructured) from Profound's MIT-licensed ai-answer-accuracy-auditor skill and its "Parrot Problem" research (June 2026), credited.
Raw lines 71 to 84