How to track brand mentions in Claude

To track brand mentions in Claude, define the customer questions that matter, run them under controlled conditions, save the complete answers and compare mention quality over time. A useful program records more than whether the name appeared. It captures the model, date, web-search setting, recommendation context, cited sources, factual accuracy and the competitors that appeared beside the brand.

Start with questions that represent demand

Build a prompt library from actual buying tasks rather than repeatedly asking Claude what it knows about you. Include category discovery ("Which payroll platforms support Norwegian companies?"), problem-led prompts, requirement-heavy shortlists, alternatives, comparisons and direct verification questions. Add location, customer type and hard requirements where they affect fit. Direct brand prompts are useful for accuracy, but they cannot measure unprompted discovery. Interview sales, support and customer-success teams for the phrases prospects actually use, then map each prompt to a buying stage and market. This avoids a test set made entirely of marketing language that customers would never enter.

• Discovery prompts reveal whether the business enters a relevant consideration set without being named.
• Comparison prompts show positioning, trade-offs and the competitors Claude considers adjacent.
• Verification prompts expose wrong details about availability, integrations, locations or target customers.
• Citation prompts help identify the pages and third-party sources supporting an answer.

Capture enough context to reproduce the observation

Store the exact prompt and complete response along with a timestamp, model identifier, locale and search mode. If Claude exposes citations, preserve the destination URLs and the claim each citation appears to support. A copied sentence without its surrounding qualifications can turn a cautious mention into a misleading positive result. Screenshots may help review, but structured records make comparison and aggregation easier.

Score mention quality, not just mention count

• Mention rate: the share of eligible prompts in which the brand appears.
• Recommendation rate: the share where Claude proposes the brand as a suitable choice.
• Accuracy: whether testable claims about the business are current and supported.
• Sentiment and framing: positive, neutral, negative, primary choice or conditional alternative.
• Competitive share of voice: brand appearances relative to the relevant alternatives in the same prompt set.
• Citation visibility: whether owned or credible third-party pages are surfaced as evidence.

Run core commercial prompts on a stable schedule and use a separate exploratory set for new questions. Repeat observations because generative output varies. Weekly monitoring may suit a fast-moving category; a slower market may need less frequent checks. Annotate product launches, major content changes and model updates so a later chart has context. Never present one favorable response as proof of sustained visibility. Establish review thresholds before the data arrives: inspect any material factual error immediately, while waiting for repeated evidence before reacting to a modest mention-rate movement. A consistent rule reduces both complacency and overreaction.

Move from a spreadsheet to a monitoring workflow

A spreadsheet can establish the method, but manual tests become inconsistent as prompts, models and markets multiply. ModelSaid helps organize recurring questions, evidence and competitive results in one workflow. Start with a free AI visibility scan, then review pricing and monitoring options when the team needs repeatable measurement. Coverage can expand with the AI answer landscape without changing the core definitions in your scorecard. Assign an owner to each prompt cluster and agree on a response time for high-risk inaccuracies. A dashboard is valuable only when an observable change leads to a review, a decision or a documented reason to take no action.

Act on the finding you can verify

When Claude misses the brand, inspect which competitors and sources it used before changing content. When it states a wrong fact, correct the canonical page and conflicting profiles you control. When it cites an old page, update or redirect that page only if doing so serves users too. Keep a log of the observation, evidence change and retest date. The result is an audit trail linking brand visibility work to real answer behavior rather than guesswork. Report unresolved uncertainty as uncertainty: an uncited answer does not reveal its underlying sources, and a later improvement does not prove that your most recent edit caused it.