How to Track Brand Mentions in AI Search

To track brand mentions in AI search, run a fixed set of relevant prompts across the engines and markets that matter, record every eligible answer, and calculate the percentage that contains a verified version of your brand name. Repeat the same measurement over time. Keep recommendation, prominence, sentiment and citations as separate fields so a simple mention is not mistaken for an endorsement.

The difficult part is not copying a question into ChatGPT. It is building a sample that can be compared next week without changing the rules halfway through. The workflow below can be run in a spreadsheet for a small audit or automated when the number of prompts, engines and competitors grows.

What counts as a brand mention?

A brand mention is an eligible answer that contains a verified name from your brand dictionary. That dictionary can include the company name, unambiguous abbreviations, product names and common spelling variants. It should also contain exclusions. A short brand such as “Bolt” or “Monday” can be an ordinary word, so an exact text match alone may produce false positives.

Decide in advance how to treat parent companies, subsidiaries and acquired products. If a customer asks for project-management software and the answer names one of your products but not the corporate brand, you may want both a product-level field and a portfolio-level field. Do not silently switch between them.

A mention is not the same as a link or citation. An answer can name a brand without linking to its website, and it can cite a brand's page without naming the brand in the recommendation. Store these as distinct observations:

Step 1: build a brand and competitor dictionary

Create a controlled list for your brand and each competitor. Include spelling, punctuation and product variants that a reviewer would accept. Add a note explaining ambiguous aliases and the context needed to count them. Review this dictionary whenever a company rebrands or launches a product, but version the change so old and new periods remain comparable.

Competitors need the same care. Otherwise, a carefully normalized numerator for your brand will be compared with incomplete competitor counts. The competitor dictionary also supports a separate AI share-of-voice analysis; it should not be folded into your own mention rate.

Step 2: design a neutral prompt universe

Start from real customer needs: category discovery, problem solving, comparisons, use cases, locations and buying criteria. Most visibility prompts should be unbranded. “Is Acme the best platform?” measures how the engine responds to a named brand, not whether Acme enters the shortlist unaided.

Keep branded reputation prompts in a separate group. They are valuable for checking descriptions, outdated facts and negative narratives, but mixing them with discovery prompts inflates the headline mention rate. A balanced prompt set might contain:

• category prompts, such as “best payroll software for a 20-person company”;
• problem prompts, such as “how to reduce errors in monthly payroll”;
• comparison prompts that do not preselect your brand;
• audience, location or constraint variants that match the market;
• a separate branded group for reputation and factual accuracy.

Freeze an initial version and record which prompts are added or removed. If the prompt set changes, report results for a stable cohort as well as for the new full set. Our guide to building an AI visibility audit covers neutral prompt selection in more detail.

Step 3: define the observation grid

A prompt has no useful meaning without its context. Choose the engines, locale, language, country and collection cadence before the first run. Treat each combination of prompt, engine and run as an answer opportunity. If personalization or logged-in state may affect the output, document that state too.

Do not assume a result from one engine represents another. Report the combined view only after preserving engine-level results. A strong total can hide a complete gap in a commercially important assistant.

Step 4: repeat observations without hiding variability

Generated answers can vary. One answer is useful evidence of what happened once, but it is not a trend. Repeat collection on a consistent schedule and compare equivalent windows. Daily data can be summarized weekly; a smaller manual audit might use several controlled repeats on fixed dates.

Store failures separately from valid answers. A timeout, safety refusal or missing response is not a “no mention”. Exclude it from the eligible denominator and report the failure rate. This simple rule prevents technical collection problems from looking like a visibility decline.

Step 5: store a reviewable record

At minimum, keep the prompt ID and version, engine, market, timestamp, raw answer, response status, brand mention, accepted alias, competitors mentioned, role, prominence, sentiment and source URLs. Preserve enough of the original response to audit classifications later.

Automated entity and sentiment detection should be checked against a human-reviewed sample. Review every negative result and every ambiguous alias until precision is proven acceptable for your use case. Classification rules and model versions belong in the methodology, not only in an internal notebook.

Step 6: calculate metrics with explicit denominators

The basic answer-level mention rate is:

answers mentioning the brand ÷ eligible answers × 100

Prompt coverage answers a different question:

unique prompts with at least one brand mention ÷ unique prompts tested × 100

Imagine an illustrative audit with 20 prompts, three engines and four valid runs per combination. That creates 240 eligible answers. If the brand appears in 72, the answer-level mention rate is 30%. If it appears at least once for 15 of the 20 prompts, prompt coverage is 75%. Those numbers are not contradictory: one measures frequency across answers; the other measures breadth across needs.

Segment both metrics by engine, prompt group, market and period before acting. For a broader KPI framework—including prominence, sentiment and source coverage—use our guide to AI search visibility metrics.

Step 7: turn a mention gap into an action

A missing mention is a diagnosis trigger, not proof of one cause. Review which competitors appear, which sources are registered or cited, how the prompt is interpreted and whether your site contains a credible answer to that need. Then choose the smallest relevant intervention:

• Competitors appear and you do not: compare their evidence, positioning and third-party coverage.
• Your brand appears with weak prominence: clarify category fit, differentiators and evidence on the pages that address that use case.
• Sentiment is repeatedly negative: verify the underlying product issue and the public sources that describe it before changing copy.
• A cited third-party source omits you: decide whether earning legitimate coverage there is useful; do not manufacture mentions.
• Results vary heavily: increase the sample and avoid declaring a win from one favorable answer.

You can pair this work with a structured competitor-tracking process, while keeping the raw brand mention KPI unchanged.

Common measurement mistakes

1. Counting branded prompts as discovery. They answer a different question and usually raise visibility mechanically.
2. Using one screenshot as a baseline. It cannot show repeatability or trend.
3. Treating collection errors as absence. Only valid responses belong in the denominator.
4. Equating a mention with a recommendation. Record role, position and sentiment separately.
5. Changing prompts without versioning. A better score may simply reflect an easier sample.
6. Combining every engine too early. The aggregate can hide the place where customers actually ask.
7. Publishing unsupported causal claims. A visibility change after a content update is an observation unless the design can isolate the cause.

Manual tracking or a monitoring platform?

A spreadsheet works for a one-off audit with a small prompt set. It becomes fragile when answers must be rerun across several engines, markets and competitors, because raw responses, timestamps and classification rules are easy to lose. Automation is useful when it makes the sample repeatable and reviewable—not merely because it produces a dashboard.

Before choosing a platform, ask whether it preserves historical answers, distinguishes failed runs, supports the engines and locales you need, exposes competitor and source detail, and lets you export or audit the observations. The collection method should serve the decision, not define it.

Plan details can change. See the current options on the Bee LLM pricing page.