How to Track Brand Mentions and Citations in AI Answers

Define the questions before collecting results

Choose prompts tied to buyer decisions, such as a task, budget, compatibility requirement, or comparison. Group them by intent and market. Save the exact wording rather than paraphrasing after seeing a result. If a platform personalizes an answer or changes its interface, note the conditions you can observe and avoid claiming a perfectly controlled experiment.

Build the panel from sales inquiries, product objections, retailer questions, or interviews when available. Mark each prompt as unbranded discovery, comparison, objection, or branded navigation. For a consumer-technology product, “best mini PC for quiet office work under a budget” tests an actual use case; “Tell me about [Brand]” tests name recognition. Keep branded questions in a separate segment. If there are no real inquiry records, say that the initial panel is researcher-selected and validate it with the client before drawing purchase-intent conclusions.

A useful log has one row per observed answer. Preserve the raw answer or a durable reference alongside the extracted fields. A later repeat belongs in a new row with its own date, even when the question is identical. Keep the platform, language, target market and use of live search visible. Do not mix a United States shopping result with a German text-only answer in one unexplained rate.

• Question ID, exact wording, intent, target country, platform, and date.
• Named brand and exact product model, with a note on the role given to each.
• Each cited domain and full URL, linked claim, and whether the page supports it.
• Approved product fact checked, issue severity, and reviewer note.

Three outcomes that should never be merged

A brand mention means the name or model appears in the answer. A citation means a link to a particular page appears as evidence. A shopping result or product card is another surface with merchant data and a destination. One observation can have any combination of these outcomes, so a single visibility percentage hides the real problem.

For example, a page may be cited while the brand is never named; a brand may be named but linked to a retailer with an outdated variant; or a shopping card may carry the right price but the answer discuss a different use case. The remedy differs in each situation. Mark a recommendation only when the assistant actually proposes the product as an option for the buyer, rather than naming it in a warning or a background sentence.

Define every denominator before reporting a rate

The mention rate is answer rows that name the brand divided by eligible answer rows in the frozen panel. The recommendation rate uses rows where the product is positively offered as a candidate for the buyer, divided by the same eligible rows; it need not equal the mention rate. An owned-domain citation rate can use answer rows that link at least one approved owned URL divided by all eligible rows. If an answer contains no citations, that row still counts in the answer-level denominator; do not silently discard it.

Citation share answers a different question: owned-domain citation instances divided by all citation instances in that specified answer set. The denominator excludes answers with zero links because they contribute no citation instances. A competitor share-of-voice figure must define whether it uses rows mentioning each brand, all brand appearances, or exclusive first choices. Multiple brands can occur in the same row; therefore separate brand mention rates need not add to 100%. Never label a panel result as “market share.”

Illustration only: twenty preapproved questions on two platforms with two repeats create 80 answer rows. If the brand is named in eight rows, recommended in five, and its site cited in two, the three answer-level rates are 8/80, 5/80 and 2/80. If there are 20 citation instances across those answers and two point to an owned page, citation share is 2/20. These invented counts demonstrate arithmetic, not a BrandInAnswers customer outcome or a benchmark.

• Mention: brand named anywhere in the answer.
• Recommendation: product offered as a relevant choice, with position recorded only within that answer.
• Citation: full URL present, domain owner identified, claim supported or unsupported.
• Product fit: correct variant, availability, compatible features and limitations.

Review source gaps before prescribing content

For each cited URL, classify the page as brand site, retailer, independent review, comparison, industry directory or another source. Record the specific claim it supports and its publication or update date when available. A competitor review may be absent from your brand plan for a good reason: perhaps your model was not sold in that country or the reviewer tested it under different conditions. Identify the reason before commissioning a response article.

Map each gap to the person who can correct it. A mismatched model identifier belongs with the product owner. A stale retailer listing needs the merchant or distributor. A missing explanation of compatibility belongs on the approved product or support page. An independent publication decides whether it has reason to cover a product; outreach is an action, not a guaranteed citation. Treat a competitor citation as a research lead rather than a request to imitate that source.

Compare a baseline with a follow-up carefully

Record the number of prompts and observations behind any reported rate. If five of twenty observations named the brand, state “5 of 20 observations in this panel” and include the dates and question set. Do not call that a universal probability or a stable ranking. Repeat where variation is material and report the range rather than selecting the most favorable answer.

When pages change, record the approval and publication date in a separate change log. At follow-up, compare mention, citation, accuracy, and source patterns, while allowing for other causes: new competitors, product launches, search index changes and assistant updates. Report per-platform and per-market counts before any roll-up. Referral visits and inquiries can be included when client analytics are available, but they should not be silently attributed to one content edit.

Turn measurement into a decision

For each gap, link to the underlying observations and name the next owner: product team for incorrect specifications, editorial team for an unanswered buyer question, retailer manager for stale listings, or analyst for a recurring platform discrepancy. A dashboard is useful only if its underlying rows and interpretation are reviewable.

Our sample audit report shows how the prompt panel, answer log, source review, and action list fit together. Use it as a starting structure and adapt the questions to your catalog. Our own new-site discovery audit demonstrates why a live URL test, a search index status and an AI provider answer should never be presented as one measure.