A customer can ask an AI assistant about your product and receive an answer that omits you, cites an old page or describes a feature you no longer offer. If you look at only one answer, you cannot tell whether the wording or the session conditions mattered. This guide gives a brand owner or analyst a small, repeatable way to inspect what ChatGPT Search, Perplexity and Gemini actually showed for questions they choose. You will finish with a dated record that a person can check against approved public facts.

The big change

  • What changed: A fixed set of questions and two documented sessions turn a passing look at an assistant answer into a record another person can inspect.
  • Why it matters: Opening the displayed sources and checking claims against dated, approved facts helps an owner identify a specific error or stale page to review.
  • What to watch: Changed account conditions, missing citations and unavailable runs affect what the record can establish. It cannot show how often people see the brand or whether a site edit caused a result.

This is a documentation-based method. BIG CHANGE did not run these questions in any account or observe a brand answer. The tool routes below come from the providers' help pages, checked on October 19, 2026; menus, eligibility and limits can change. The two-session record and review rules are our editorial method, not a measurement endorsed by any provider.

Before you open a chat

Choose one brand you are authorized to assess. Write down its exact name, official domain and the market you care about. Prepare a dated sheet of approved public facts relevant to the questions: for example, the current product description, supported locations, feature availability or publicly listed price. Include a source URL, page date or revision, and the person who approved the facts. If no approved public evidence exists for a claim, you will have to mark that claim unverified. Do not make an assistant's answer your source of truth.

Choose two or three real questions a prospective customer might ask in that market. Include a category or discovery question and a specific product or feature question. Add a comparison only if you have a real, approved basis for checking it. Save each question's exact wording, language, intended market and the fact it should help check before seeing any answers. Do not add the brand name to every question just to force a mention. Keep the wording unchanged across services and both runs.

Use ordinary web interfaces and a permitted account or signed-out session where supported. Check your workplace policy and the account's data settings before submitting the questions. Send only public, approved information. Avoid private customer details, unpublished plans, uploaded files, connected apps and public share links. OpenAI says Search can send rewritten queries and general location to search partners; Perplexity describes the account and search data it collects; Google explains data handling for Gemini Apps. Review the settings for the actual accounts you use.

No paid subscription or API is needed for an eligible manual run: ChatGPT Search includes Free access, Perplexity Standard includes basic searches, and Gemini's web app has a no-plan route with variable limits. Access still depends on region, account, organization controls and available usage. Check those conditions on the day you work; do not upgrade to fill a missing row.

Make the observation sheet

Create one row for every planned combination of service, question and run. With three services, three questions and two runs, that means 18 planned rows; two questions mean 12. Empty planned rows matter because they show what was unavailable instead of quietly shrinking the sample.

For each row, leave space for: brand and domain; approved fact-sheet revision; question ID and exact text; service and account or workspace class; plan and displayed model or mode; run number; date, time and time zone; intended market and location actually known; language and device; visible search and personalization settings; result status; full answer or its permitted saved location; visible citation URLs; opened-source checks; claim decision; reviewer and review date. Record unknown where the interface does not reveal a setting. A new conversation avoids earlier messages in that conversation, but it need not remove saved memory or personalization.

Use these distinct result statuses: answer shown, unavailable (access, region, workspace or limit), and error (a failed search or response). For an answer shown, separately record brand absent, brand named, brand described, brand recommended or brand linked, as the actual words allow. Also record no citation shown when there is no visible source link and broken source when a displayed link cannot be opened. A missing answer is never a missing brand mention.

Run the same questions twice

  1. Log the conditions. Note the date and time with time zone, account or signed-out status, plan, browser or device, language, market, and any model, web mode, location or personalization controls you can actually see. The intended market in your question is different from a location the service may infer. Record both when known, and leave inferred location unknown when it is not visible. Keep settings as consistent as your access allows.
  2. Ask each question in a fresh conversation. At chatgpt.com, open a conversation and choose View all tools > Search, or type / and select Search, before sending the saved question. OpenAI's Search instructions also explain how to open citations and the Sources panel when available. On Perplexity, enter the same question in the homepage Search box as a new Session, using ordinary Standard basic Search rather than Pro Search or a premium model. Perplexity's Session manual uses that current term. At gemini.google.com, choose New chat and send the same question; see Google's web-app instructions. Repeat for every frozen question. If a control has changed, consult the current manual and record the route you actually used.
  3. Save before judging. Preserve the exact answer as permitted text, screenshot or export, plus its time, displayed mode and each visible source URL in order. A private conversation link can help you find the answer later, but keep a local permitted copy: Perplexity says signed-out Sessions expire after 14 days. Label a link's position honestly. In Gemini, a Sources button or inline source may appear for a response; a related link is not automatically a citation for a particular brand claim. Do not turn a missing citation into a cited answer by searching for a source afterward.
  4. Open and check the sources. For each link relevant to a material brand statement, record the page title, publisher and domain, any page date shown, whether it opened, and the passage that supports or contradicts the statement. Compare the answer with your dated approved fact sheet. Mark each material claim supported, wrong, stale or not verified, and write a brief reason. A working link can still fail to support the claim beside it. If no source is shown, check the claim against your approved facts and record the absence of a visible citation separately.
  5. Make the second run. In a second documented session, repeat every frozen question in fresh conversations on the three services, using the same available settings. Record any changed plan, model, mode, location, memory or timing conditions. Save and check the second answers as new rows, even when they look identical. Do not fill a failed row from memory or a previous run.

Review what the rows actually show

Have a person authorized to speak for the brand inspect the answers, source passages and claim decisions, then sign and date the sheet. The closing note should name the exact questions and observation dates. It can say, for example, that a particular row omitted the brand, lacked a citation, used an outdated page or contained a claim contradicted by the approved record. It should identify the factual page or product documentation that deserves review. Any public correction needs the normal approval process; this exercise cannot promise that an edit will change future answers.

If Search is missing in ChatGPT, OpenAI suggests checking a new conversation, browser or app version, intended account, usage limit and workspace permission. If basic Perplexity Search is blocked, note the displayed plan and exact failure. If Gemini supplies neither inline source links nor a Sources button, record no citation shown for that response and check material facts independently. Google says links do not appear for every response. A regional block, organization restriction or exhausted limit leaves unavailable rows and an incomplete three-service sample. Resume later only with a new timestamp and recorded conditions.

This sheet is a small observation of chosen questions in documented conditions. It gives you specific statements and pages to investigate, not a representative visibility rate, ranking, share of voice or causal SEO finding. For a different question about a verified site's appearances in Google Search features, see our Google Search Console AI search report guide. Its property impressions are a different record from the individual assistant answers collected here.