Playbook · methodology
Protect brand facts with Fact Check
AI engines state prices, dosages, specifications and terms with confidence — and sometimes wrongly. This playbook uses Fact Check to compare AI claims with your approved documents and turns contradictions into fixes.
- Goal
- Fewer contradicted, outdated and unsupported claims
- For
- Pharma, finance, health and complex B2B products
- Baseline
- 7–14 days of daily tracking
- Re-measure
- After every run; review monthly
Goal
Reduce the share of AI statements about your products that contradict your approved reference documents, go beyond them (off-label), are not supported by them or reflect outdated facts. Correct statements should be confirmed as matches.
Who it's for
- Pharma and medical device companies with approved labels such as an SmPC
- Banks, insurers and fintechs with conditions, fees and terms
- Manufacturers with technical specifications
- Any brand whose prices, policies or product facts change often
Prompt set
Twelve templates that make engines state concrete facts. Fact Check examines tracked answers that mention an asset, so these prompts name the product.
Import it in AutoSEO under AI Visibility → Tracker → Import CSV (paste the text or upload the file).
Replace these placeholders before you import — the importer takes every line literally:
[product]— the product as named in Fact Check (add spelling variants there as aliases)[condition]— an indication, use case or customer situation[country]— a market you check, e.g. “Germany”[brand]— your brand
prompt,tags"What is [product] used for?","Facts|Indication""How do you use [product] correctly?","Facts|Usage""What are the side effects or risks of [product]?","Facts|Risk""Who should not use [product]?","Facts|Risk""How much does [product] cost?","Facts|Price""What are the terms and conditions of [product]?","Facts|Terms""Is [product] suitable for [condition]?","Facts|Suitability""What are the technical specifications of [product]?","Facts|Specs""How is [product] different from its previous version?","Facts|Versions""Is [product] approved in [country]?","Facts|Regulatory""What do experts say about [product]?","Facts|Reputation""What does [brand] offer for [condition]?","Facts|Portfolio"Engines to track
Check the engines your customers — and in regulated industries, patients and advisors — actually use.
| Engine | Why it matters for fact checking |
|---|---|
| ChatGPT | The assistant people use most for health, finance and product questions. |
| Google AI Overviews and AI Mode | Answer factual questions directly on the search results page. |
| Perplexity | Its cited sources make it easy to trace a wrong claim back to the page behind it. |
| Gemini, Claude, Microsoft Copilot | Add them for complete coverage: wrong facts often differ from engine to engine. |
Baseline measurement in AutoSEO
- 1
Create the assets
In Fact Check, create one asset per product with its name, spelling variants and the markets to check, each with its regulator, e.g. DE · EMA or US · FDA.
- 2
Upload the reference documents
Upload the approved label, specification sheet or terms as PDF, pasted text or URL. AutoSEO splits them into sections for the comparison.
- 3
Import and track the prompts
Import the prompt set. After every tracking run, answers that mention an asset are checked against its reference documents.
- 4
Record the verdicts
Note the counts per verdict — matched, needs review, off-label, contradicted, unsupported, outdated — per engine and market on the Accuracy page.
Actions
- 01
Triage the findings
Start with contradicted and off-label findings, then outdated ones. Each finding shows the claim and its verdict.
- 02
Trace the source of each error
Open the answer and its cited sources. Wrong claims usually come from outdated pages of your own, old press releases, third-party databases or forums.
- 03
Fix your own pages first
Update or redirect outdated pages, state current facts as plain, dated text, and take old PDFs that still circulate offline.
- 04
Ask third parties for corrections
Ask cited publishers, directories and databases to correct wrong facts, with a link to the current source document.
- 05
Route findings through review
Give medical, legal or compliance reviewers access to the project with a suitable role. Findings marked “needs review” are for people to decide.
- 06
Keep reference documents current
When a label, price or term changes, upload the new version right away — otherwise correct new answers look like errors.
How to measure change
- Verdict counts per engine and market on the Accuracy page, compared with the baseline
- Share of contradicted and off-label claims among all checked claims
- Findings closed versus new findings per month
- Whether the sources you corrected still appear behind wrong claims
Document for compliance
Keep records of findings and the actions you took. Fact Check shows what AI said; your records show how you responded.
Expected variance
Factual errors are sticky and random at the same time. An engine may repeat an outdated price for months because an old page ranks well — and state a different wrong price the next day. Judge progress by the rate of problematic verdicts across many answers, not by single answers.
Limitations
- Fact Check can only judge claims that your uploaded reference documents cover.
- It checks answers to tracked prompts, not every conversation people have with AI.
- You can't force an engine to correct a claim; you can only fix and strengthen the sources it relies on.
- This is not a pharmacovigilance or regulatory reporting system. If an AI answer could harm people, follow your internal escalation process.
Questions about this playbook
Which documents can I use as a reference?
Anything that states the approved facts: a product label or SmPC, a specification sheet, terms and conditions, a price list. Upload them as PDF, as pasted text or by URL.
What do the verdicts mean?
Matched means the claim agrees with your reference. Contradicted means it states the opposite. Off-label means it goes beyond what the reference covers, for example an unapproved use. Unsupported means the reference does not back it up. Outdated means it reflects facts that are no longer current. Needs review means a person has to decide.
Can I check several countries?
Yes. Each asset lists the markets it is checked for, with the regulator, for example DE · EMA and US · FDA. Prompts carry a country, so you can track the answers for each market separately.
Does this replace our medical or legal review?
No. Fact Check helps you find the claims worth reviewing, quickly and consistently. Decisions about what is correct, and what to do about it, stay with qualified people in your organization.
Where are our reference documents processed?
They are stored in your AutoSEO instance. For the comparison, the relevant text is sent to the AI provider configured on the instance, or processed through your local Claude Code or Codex agent. Self-host AutoSEO if documents may leave your infrastructure only toward an AI provider you choose.
Run this playbook in AutoSEO
Every feature is included in the free self-hosted edition and in AutoSEO Cloud.