How we collect data

Everything in your dashboard comes from one repeatable loop: run your prompts against the AI engines, read the answers, detect mentions and sources, and track the change run over run. Here is exactly how that works, including what the data can and can't tell you.

#Running prompts

A project's prompt library holds the questions you may want to track. The prompt set is the versioned selection that actually runs. Free sends the first 30 selected prompts to ChatGPT once. Plus sends up to 80 to ChatGPT and Google AI Overviews weekly. Pro sends up to 200 weekly to ChatGPT, Gemini, Perplexity, Google AI Mode, and Google AI Overviews. We capture and store the full answers from every run.

On Plus and Pro, library and selection edits prepare the next month; they do not mutate the version already running. If you do not prepare a new selection, the current one carries forward. Free keeps its first runnable set fixed while the library remains editable. Only one set runs at a time for a project, and every answer is attributed to the version that produced it.

#Detecting a mention

For each captured answer we determine whether your brand was mentioned. This is harder than a text search: brands have variations, abbreviations, and names that collide with common words. We match against the variations confirmed in your brand profile, which is why keeping that list accurate directly improves detection quality. Repeating a brand several times in one answer still produces one answer-level mention for visibility metrics.

We also extract the other named brands and classify their role. Direct buyer alternatives enter the competitor universe for that weekly run. Marketplaces, retailers, publishers, cited sources, ingredients, suppliers, and adjacent products are excluded from Share of Voice. Name variations and product names are resolved to a consistent brand where the evidence supports it.

Once a mention is found, we read its context to capture placement, sentiment, and any citations and sources attached to the answer.

#Headline metrics and trends

Failed answers are excluded rather than treated as misses. Mention Rate, Citation Rate, and sentiment use the successful answers available to each metric. Share of Voice has an additional answer-level normalization so response length and list-style prompts do not receive extra weight.

Within each eligible non-branded answer, every distinct direct brand receives an equal share. An answer naming your brand and three competitors gives each brand 25%, regardless of how often any name is repeated. The weekly score is the mean of those answer shares. Answers with no qualifying brand are excluded from Share of Voice. If competitor extraction is incomplete for a weekly run, its Share of Voice is withheld instead of treated as zero.

A period score is the equal average of its completed weekly scores. This keeps a week with more engines or answers from dominating the period. The competitor universe is rebuilt for every weekly run, so a brand that appeared last week does not affect this week unless it is named again. The run charts show the completed weekly readings behind the headline.

#Actions stay attributable

Visibility Actions retain the prompt-set version that generated them. Use the prompt-set filter in Actions to review recommendations from the current version or audit an earlier one without mixing the two measurement periods.

#Signal versus noise

Because answers vary, a period headline becomes more stable as completed weeks are added. Use the run chart beneath it to see whether movement persisted across several runs and, on Plus and Pro, across AI engines. Sustained shifts are signal; a one-run blip usually is not.

#What this can and can't tell you

• It can tell you how you appear in AI answers for the prompts you track, how that compares to competitors, and how it's trending.
• It can't tell you about questions you don't track, coverage is only as broad as your prompt set, so invest in prompts that reflect how customers really ask.
• It can't guarantee an AI engine's answer to a given user at a given moment, since answers vary and are personalized. It reflects representative behavior, measured consistently.

For the list of AI engines and why they differ, see AI engines we track.