To check whether ChatGPT recommends your brand, test a stable set of buyer questions, save each answer and its displayed sources, and record mentions, citations, and recommendations separately. Then verify material statements against current authoritative evidence. A favorable mention can still misstate a capability or price. Treat your record as a sampled investigation log, not market-wide reputation, and give every material error a named reviewer and an appropriate correction path.
Separate a mention, source, and recommendation
A communications team needs to know whether an answer represents the offer accurately. A growth team needs evidence of buyer action. Neither question is answered by counting every favorable appearance as a win. Use separate fields rather than a combined “ChatGPT recommendation” score.
| Signal | What to record | What it does not establish |
|---|---|---|
| Brand mention | The answer names your brand | Suitability, endorsement, or preference |
| Linked citation | A displayed source reference with a URL | That the page supports the associated claim |
| Recommendation | Explicit wording presenting the brand as suitable for the stated need | A recommendation for other buyers or contexts |
| Referral | A recorded website visit with an identifiable referring source | The full answer seen or why someone visited |
| Inquiry | A contact request received by your business | Commercial fit or an AI-caused decision |
| Qualified opportunity | A prospect meeting your declared sales qualification criteria | Revenue or attribution to an observed answer |
A source listed in a broader Sources panel is not automatically a citation supporting a particular sentence. Preserve that distinction when coding the answer. OpenAI cautions that Search results and citations can be incomplete, outdated, or incorrect. Inspect the page rather than treating the link as approval.
Check the factual claim
Classify each material statement as supported, contradicted, uncertain, or unverifiable. Supported means current authoritative evidence matches the statement's scope. Contradicted means reliable evidence conflicts with it. Uncertain means evidence is ambiguous; unverifiable means you cannot establish the claim from accessible evidence. Do not silently count either as accurate. If you code sentiment, describe it as a reviewer's assessment of the answer's framing, not population opinion.
Keep test conditions visible
Start with buyer concerns from support conversations, sales objections, and branded paid-search terms. These inputs help choose questions; paid campaign outcomes do not prove organic citations or buyer preference. Agree which concerns matter before running the test, so a favorable answer cannot redefine success afterward.
Record the exact prompt, engine and product surface, model label shown, Search selection or visible use, market, language, account or workspace, personalization settings, timestamp, complete answer, displayed URLs, and reviewer. Save a capture before follow-ups alter the conversation. Keep inaccessible sources and missing interface labels as explicit limitations.
OpenAI documents query rewriting and possible use of location and saved memories in Search. Identical prompts therefore do not establish identical retrieval conditions. Where available, choose an Unpersonalized Temporary Chat before the first message. The current Temporary Chat guidance says this choice does not use memory, custom instructions, or plugins. Temporary Chat alone is not sufficient: Personalized is a separate option, and account restrictions and limited safety context can still apply.
Neutral discovery versus brand-named questions
Use neutral questions to observe discovery, such as “Which providers should a procurement team consider for [use case] in [market]?” Then use brand-named questions to inspect evaluation, such as “Does [brand] offer [capability], and what evidence supports that?” Label the second question as prompted brand evaluation, not unprompted discovery.
Run each question in a fresh conversation and repeat on a declared cadence. Record interface or setting changes rather than quietly merging unlike observations. For the panel setup and tools-comparison boundaries, use the existing citation-panel operating guide. This article adds claim investigation; it does not replace that owner's operating sheet.
Investigate an inaccurate answer
Preserve the exact wording before requesting another answer. Copy the relevant sentence, retain the full capture, and record any attached link. A later correct answer does not erase the earlier observation or establish that a correction worked.
Inspect sources
Use this investigation checklist before choosing a response:
- Open the displayed source and locate the passage that supposedly supports the claim.
- Check publication or update date, market, product version, and any conditions or exclusions.
- Compare the statement with current approved product documentation, terms, or other authoritative evidence.
- Separate an error on your own page from a third-party error or wording unsupported by the displayed sources.
- Save counter-evidence and mark anything you cannot verify instead of filling the gap with an assumption.
If the source discusses a different product tier, do not generalize its capability to the whole offer. If an inaccessible page prevents review, retain that gap. An answer's confident tone is not corroboration: OpenAI's accuracy guidance identifies possible incorrect facts and fabricated citations and recommends verifying important information.
Assign correction ownership
Route an owned-page error to its content or product owner for approval and correction. Route a third-party error to that publisher with the exact passage, proposed correction, and supporting evidence. For an unsupported output, use the platform's available feedback or reporting route when appropriate. These are proposed response practices, not provider-guaranteed remedies.
If source maintenance or technical access needs investigation, consider SEO, AEO, and GEO support. Do not assume crawler changes will correct a specific answer. Material changes to approved claims belong in the marketing claim ledger; that article governs evidence and approval, not this test's measurement.
Copyable correction log
This blank log implements a proposed Observe → Verify → Route → Recheck workflow. It records decisions, not measured performance.
| Field | Entry to complete |
|---|---|
| Test ID and exact claim | Identifier; verbatim disputed sentence |
| Capture and displayed source | Full answer capture; URL as displayed or none |
| Counter-evidence | Authoritative passage, date, scope, and URL |
| Claim status and severity rationale | Supported, contradicted, uncertain, or unverifiable; buyer consequence |
| Owner and approval | Named investigator; required qualified reviewer |
| Correction request and status | Owned-source, third-party, or unsupported-output route; action taken |
| Recheck date and result | Planned repeat; observed wording; remaining limitation |
For example, a claim about an unavailable capability or an outdated price would need product or commercial review. These are hypothetical issues, not findings about a client. Source updates, publisher requests, and platform feedback do not guarantee propagation into future answers.
Tools organize evidence, not certify reputation
Start manually with a panel your team can review. Consider monitoring software when additional markets, repeat frequency, exports, or reviewer workload make the manual process impractical. Ask for an evidence-preserving export, not just a headline score.
Coverage, exports, cadence, and review
Check which engines and product experiences are actually observed, which settings are captured, how often tests run, and whether you can inspect answers and source URLs. Require a way to retain exceptions and reviewer decisions. Do not assume API output reproduces a consumer interface or that a vendor covers every buyer's experience.
Keep feature, price, and vendor comparisons with the existing tools owner. For share-of-voice definitions and denominators, use the AI visibility benchmark scorecard. A local prompt-panel measure is not market share or a reputation certificate. This guide's decision is narrower: which observed claim needs investigation, approval, or rechecking?
Report exceptions and escalation
Lead the review with material unresolved claims, their buyer consequence, and the next action. A favorable mention with an incorrect eligibility condition may deserve more attention than an absent mention. Record severity reasons in words rather than assigning unsupported precision to risk scores.
Buyer impact and next review
Escalate statements about safety, legal status, security, regulated eligibility, or contractual commitments to qualified reviewers. Keep sensitive evidence in access-controlled records. Use AI governance support when the business needs a defined review and escalation process, not a promise to control external answers.
Recheck the documented question after an approved action. Report whether the wording changed, remained wrong, or could not be compared. Do not infer causation from a later improvement. Keep referral, inquiry, and opportunity records separate and preserve attribution gaps.
Separate checklist for other engines
Add other experiences only when buyer evidence justifies the workload:
- Identify the engine and exact surface; Google AI Mode is not ChatGPT.
- Save each surface's conditions, full output, displayed links, and reviewer separately.
- Keep first-party reports distinct from manual claim-verification records.
Google's Generative AI performance documentation describes combined AI Overviews and AI Mode link impressions, not an AI-Overviews-only click, CTR, or query report. It does not reveal answer accuracy. Use the separate Google tracking guide for reporting detail; do not relabel its impressions as ChatGPT evidence.
Use the log before commissioning more monitoring
Modi's editorial position is simple: a mention is not a citation, recommendation, or qualified demand. Ask whether the answer accurately represents the offer and who owns the correction. Use the blank log first, then request a Free AI Visibility Score for a high-level website-readiness baseline. The score is not live reputation monitoring, an answer-correction service, or a guarantee of future recommendations.
