To check what AI says about your brand, ask a small set of real buyer questions across the answer products your customers use, save each complete response and its visible sources, then compare every important claim with a current official reference. The result is a dated sample, not a census of all AI answers or a permanent accuracy score.
This procedure is designed for a first manual review. It complements a repeatable audit: use the manual run to discover important questions, then version the questions and conditions if you plan to compare later observations.
Set the scope before you start
Choose answer products and surfaces relevant to your buyers. ChatGPT, Gemini, Perplexity, DeepSeek, and Grok are examples, not a ranking of audience share or a requirement to test all five. Interfaces and search options change. Record the exact product and visible mode you used, and do not compare a web-search answer with a non-search answer as if they were the same condition.
Open at least one fresh conversation for each planned question, unless the test specifically measures follow-up context. Keep personalization and location settings consistent where possible. If you cannot control or identify a setting, mark it as unknown rather than silently assuming it is off. OpenAI: searching the web with ChatGPT.
Ask five questions tied to buyer decisions
Replace bracketed text with facts that matter to your customers. Ask the same version on each selected surface, then record any follow-up separately. The prompts below are templates to adapt, not a validated benchmark or a promise that a platform will disclose its reasoning.
| Decision | Prompt template | What to verify |
|---|---|---|
| Fit | What does [brand] do, and which kinds of teams is it for? | Product description, audience, exclusions, and unsupported superlatives |
| Price | What does [product or plan] cost, and what affects the price? | Plan, amount, currency, billing period, region, and terms |
| Capability | Does [product] support [specific feature or workflow]? | Feature availability, plan limits, and current documentation |
| Comparison | How does [brand] compare with [competitor] for [specific use case]? | Criteria, omissions, dated sources, and claimed differentiators |
| Availability | Is [product, location, or service] available in [place or condition]? | Current coverage, hours, eligibility, and effective dates |
Save enough context to repeat the check
Copy the complete answer and record the exact prompt, product, interface or mode, timestamp, locale, and visible source links. If a response includes citations, inspect the linked passages and mark whether they support the whole claim. OpenAI notes that web-search citations and results can be incomplete or outdated, so a link should be checked rather than treated as proof. OpenAI: searching the web with ChatGPT.
| Question | Observed claim | Official truth and source | Verdict | Context |
|---|---|---|---|---|
| Price template | Exact quoted price and terms | Current pricing URL and effective date | Supported / contradicted / insufficient / not comparable | Product, mode, date, locale, citations |
| Feature template | Exact capability statement | Current product documentation | Supported / contradicted / insufficient / not comparable | Product, mode, date, locale, citations |
Use four verdicts, not a forced true-or-false score
Mark supported when the cited or observed statement matches the approved source and applicable conditions. Mark contradicted when the statement conflicts with reliable current evidence. Mark insufficient evidence when the answer or its source does not establish the fact. Mark not comparable when the product, plan, region, or time period differs. Keep failed runs in a separate collection-status column; do not drop them from the denominator without explanation.
A manual review is a snapshot of the questions and surfaces you chose. It does not measure every customer exposure, guarantee a model update, or establish that your brand is broadly understood. Use repeated, versioned observations if you need to assess change over time.
Turn the audit into a source correction
Group findings by the source a team can act on: an unclear owned page, a stale third-party listing, a missing answer, or an unsupported statement. Assign an owner and a concrete acceptance check, such as the exact current price and billing condition appearing together. Then repeat the saved question and compare the response with the original record.
If the answer is still wrong, preserve that result. A changed website is not the same thing as a changed answer, and a changed answer is not proof that a source edit caused the change. Keep the correction, rerun, and conclusion as separate entries.
See how a claim and its evidence appear in the illustrative Tailscale sample report. When you are ready to compare options, review Pevatron audit plans and pricing
Questions answered
Which AI products should I check?
Choose products and surfaces your buyers use. You can include ChatGPT, Gemini, Perplexity, DeepSeek, or Grok, but record the exact interface and mode because experiences differ.
What should I ask an AI about my business?
Ask about product fit, price and terms, a specific capability, comparison for one use case, and availability or eligibility. Use the same wording when comparing products.
Do citations mean the answer is correct?
No. Open each source and compare its passage with the full claim and relevant conditions. A citation may be incomplete or fail to support the sentence.
Can a manual check estimate how many buyers see a wrong answer?
No. It covers only the questions, products, and conditions in your sample. It is a diagnostic snapshot, not an exposure estimate.
How often should I rerun the audit?
Rerun after a material source change and on a cadence appropriate to the risk. Keep question versions and dates so changes in the test are visible.
Sources and research notes
Primary sources checked on 14 November 2026. Our checklists and examples are recommendations, not measured customer results. Recheck changing provider policies before acting in 2027.
- OpenAI: searching the web with ChatGPT — OpenAI Help Center