How to track brand mentions in Perplexity

To track brand mentions in Perplexity, run a stable set of customer questions on a schedule, save the complete answers and citations, and score how the brand is presented. A mention count by itself is weak evidence. The useful questions are whether Perplexity discovers the company without being prompted, recommends it for an appropriate buyer, describes it accurately and links to sources that support the answer.

Build a prompt set around customer decisions

Collect the questions prospects ask before they know your brand: category discovery, problem solving, shortlisting, comparison and fact checking. Add constraints that determine real fit, such as country, company size, budget, integration or delivery date. Keep direct brand prompts in a separate accuracy panel. Asking "Is Acme good?" measures recall about Acme, not whether it would be discovered for an unbranded need. Assign every prompt a market, audience and buying stage so aggregate results remain interpretable.

• Discovery: "What tools help independent retailers forecast inventory?"
• Shortlist: "Which options support Shopify and operate in Norway?"
• Comparison: "Compare the strongest options for setup, support and reporting."
• Verification: "Does this provider offer the integration and service region I need?"

Capture enough context to reproduce an observation

Store the exact prompt, full response, date, locale, account state, visible product or mode, follow-up context and every surfaced URL. Perplexity answers can vary with wording, available web evidence and product changes. Repeat commercially important prompts rather than treating one run as permanent. When a follow-up changes the shortlist, preserve the preceding turns; otherwise a reviewer cannot tell which constraint caused the brand to appear. Separate desktop and mobile observations if the experience or sources differ materially.

Score mention quality, not just frequency

• Eligible mention rate across prompts where the business could reasonably qualify.
• Unprompted recommendation rate, excluding direct brand questions.
• Accuracy of audience, capabilities, availability, pricing approach and limitations.
• Framing as a primary choice, conditional option, alternative, source or poor fit.
• Citation visibility for owned pages and credible independent evidence.
• Competitor share of voice within the identical prompt and market set.

Report raw counts beside rates. Three mentions out of ten prompts mean something different from three out of one hundred. Flag material factual errors immediately, but require repeated evidence before reacting to a small shift in wording or order. Annotate product releases, website migrations and important Perplexity changes in the timeline without claiming they caused an observed movement. Generated answers are variable, and monitoring is observation rather than proof of ranking logic.

Turn monitoring into a weekly operating loop

ModelSaid helps teams organize recurring questions, answer evidence, citations and competitor appearances instead of relying on scattered screenshots. Begin with an AI visibility scan, then use the AI readiness checker to inspect whether important facts are clear and accessible on the site. Product coverage can expand as new answer engines emerge while your prompt taxonomy, scoring rules and historical evidence remain useful.

Review the highest-value prompt clusters on a cadence matched to the market. Give each issue an owner, a source hypothesis and a retest date. If Perplexity omits the brand, inspect which eligible companies and sources appear. If it states a false capability, correct the canonical page and any profiles you control. If the company does not meet the prompt, accept the result rather than changing eligibility rules. A good report ends with verified actions: fix a fact, improve decision-ready evidence, investigate a source conflict, monitor another cycle or take no action. Keep a fixed core panel for trends and an exploratory panel for new buyer language. Add a monthly quality review in which someone who did not collect the answers checks a sample against the rubric. This catches generous scoring, missed caveats and name-matching errors. Preserve retired prompts with version notes, and show unresolved inaccuracies beside any headline score. Send product, support and communications teams only the findings they can own, with the evidence attached. Before every review, confirm that prompt eligibility and locale still match the market, so normal business changes do not masquerade as visibility losses. That discipline makes the program resilient when interfaces, models and answer products change.