ChatGPT Brand Tracking: Measure Discovery and Brand Accuracy

ChatGPT brand tracking should answer two different questions: does your brand enter the conversation without being named, and does ChatGPT explain it accurately when a buyer asks about it? Combining these questions produces misleading reports. A strong answer to “What does HarborDesk do?” does not show that HarborDesk appears in an unbranded software shortlist.

This workflow separates discovery from verification, then uses repeated observations to identify work your team can own. It is a platform-specific extension of the LLM brand tracking program.

Define which ChatGPT experience you observed

OpenAI documents that ChatGPT can search the web and that search responses may include citations. Record whether search was explicitly selected, visibly used, or not established from the observed response. Avoid inferring a hidden search process from answer quality alone. See OpenAI's explanation of searching the web with ChatGPT.

Keep consumer interface observations separate from API tests. An API collection can measure the application configuration you run; it cannot establish that a buyer saw the same answer in ChatGPT. Name the surface and collection method in every report, including any model label exposed, mode, language, geography, timestamp, and account or workspace context.

Create fresh conversations without previous brand questions. Record relevant memory or personalization settings and any custom instructions you can control. A new conversation reduces carryover from that conversation but does not establish complete neutrality. Location and account conditions can still matter. Do not call a test “unpersonalized” unless the protocol explains exactly what that label means.

Maintain two prompt panels

The discovery panel asks about a job without naming your brand. Include buyer constraints that actually distinguish your product: team structure, deployment requirements, integration needs, and market. Avoid inserting a list of features selected only because your product has them; that creates a favorable but artificial test.

The verification panel names the brand and asks checkable questions. Choose claims that matter to a purchase, such as implementation requirements, available integrations, intended audience, or known limitations. Keep an approved reference page and an owner for each claim. Do not ask the evaluator to invent a correctness judgment from memory.

Use a fixed competitor list for discovery scoring. Write alias rules so a product name, company name, and ambiguous abbreviation do not become three different entities. If an abbreviation could refer to another business, inspect its context manually before counting it.

Collect answers without steering the baseline

For each core prompt, schedule multiple independent runs under the same conditions. A practical pilot might use three runs per question per collection date; the number is a workflow choice, not a guarantee of statistical precision. Save complete answers before interpreting them.

Do not repair the initial answer with “You forgot our brand” and then count the repaired response as discovery. Follow-up questions can reveal useful diagnostic information, but label them as follow-ups. Similarly, an explicit request to cite your website is a directed retrieval test, not evidence that the website would appear unaided.

Record timeouts, access problems, refusals, and incomplete responses separately. Report the successful answer count beside mention rate. Excluding failures silently can make an unreliable collection look like a clean trend.

Use an evidence ledger

This worksheet is hypothetical. HarborDesk is a fictional product, and the rows describe what an analyst would record rather than actual ChatGPT results.

The ledger should preserve a reference to the original answer. A summary such as “negative sentiment” is insufficient if the real issue is a specific, outdated implementation claim.

Diagnose absence and inaccuracy differently

When your brand is absent from a relevant shortlist, review the answer's selection criteria and visible sources. Is the question appropriate for your product? Do your public pages establish that fit? Does an independent source describe the category differently? An omission does not establish a technical block or a need to rewrite every page.

For an incorrect claim, locate the authoritative evidence first. If your own pages conflict, resolve that inconsistency before seeking external corrections. If a visible citation is outdated, check whether the source actually supports the claim. A citation adjacent to a sentence is not automatically evidence for every part of it.

Search access and model training access are distinct technical questions. OpenAI identifies OAI-SearchBot for ChatGPT search and GPTBot for content that may be used in model training. Review the appropriate control with your technical team using OpenAI's crawler documentation, rather than assuming a training crawler setting governs search visibility. Technical eligibility still does not guarantee a particular answer or citation.

Report a narrow, defensible conclusion

State the panel, collection dates, repetitions, successful answer count, and relevant conditions alongside the finding. Use mention rate for the percentage of successful answers mentioning a brand. If reporting competitive share of voice, count each fixed tracked brand once per answer at most, and divide its count by all tracked brand-answer counts. Keep that metric separate from recommendation rate and accuracy.

Use the citation source audit when visible evidence needs investigation. Citation absence does not prove that no retrieval occurred. After a content correction, retest the same panel and log other changes using measuring AI content updates.

Bring this protocol to any software evaluation. BIU authors this guide from a brand visibility perspective; review the AI visibility offering and validate the actual ChatGPT collection method, evidence access, and coverage in a demo.