How Ansedd measures AI visibility
Every number on your dashboard links back to this page. Here is exactly how we sample the assistants, what each metric means, how accurate our reading is, and where the limits are.
Sampling design
We run every tracked prompt once per engine, once per week. That is one sample per prompt-provider pair, taken on the same weekly cadence for every account below the Studio tier. Studio runs twice a week.
Why one weekly sample and not many per day? AI answers are noisy. A single prompt asked twice an hour apart can name different products. Sampling hard and often would multiply cost without buying you a truer picture, because the noise is in the model, not in our counting. Instead we sample steadily and read the trend across weeks. This keeps the product cheap enough to price for one-person software businesses, and keeps the signal legible.
Each engine is asked with web search or grounding enabled, so the answers reflect what a real user would see today. We never fall back from one engine to another inside a run: if Gemini fails, that week has no Gemini data, and we say so, rather than substituting another engine and corrupting the comparison.
Metric definitions
These definitions are quoted verbatim from the code that computes them. A confidence gate applies to every mention-based metric: a mention our extractor is less than 70 percent confident about is stored for auditing but excluded from every number below. When a metric has nothing to divide by, the result is 0, never a blank and never an error.
Visibility Score
The percentage of your tracked prompts where your brand was either mentioned or cited by the AI answer, on a 0-100 scale. A prompt counts as a hit if the brand appears in ANY of that prompt's answers within the scope (a single provider, or all providers combined). Mentions below the confidence threshold do not count; citations to your own domain do. Returns 0 when there are no prompts.
Share of Voice
Your brand's mentions as a percentage of ALL entities' mentions (your brand plus every tracked competitor) across the scope, on a 0-100 scale. The denominator is the total mention count for brand and competitors combined; mentions below the confidence threshold are excluded from both the numerator and the denominator. Returns 0 when no entity was mentioned at all.
Citation Share
Citations pointing to your own domain as a percentage of ALL citations across the scope, on a 0-100 scale. A citation is counted for your domain when its native source URL is on your domain or a subdomain of it (same-site match). The denominator is every citation returned by the AI answers in scope. Returns 0 when there are no citations.
Recommendation Rate
The percentage of your tracked prompts where the AI answer actively recommended your brand (not merely mentioned it), on a 0-100 scale. A prompt counts if the brand was recommended in ANY of that prompt's answers within the scope. Recommendations below the confidence threshold do not count. Returns 0 when there are no prompts.
Mention Count
The total number of times your brand was mentioned across the scope, above the confidence threshold. Unlike Visibility Score this counts every occurrence, not distinct prompts. Returns 0 when the brand was never mentioned.
Extraction accuracy
Reading whether an answer names your product is harder than it sounds: brand names like Linear, Notion or Arc are also ordinary words. We check our reading against a hand-labeled golden set of real answers, weighted toward these generic-name traps.
The extraction precision gate is enforced at 0.95 on the golden set in CI. Any release that touches extraction must clear that bar, or it does not ship. The live number is published here once the first gated eval runs.
Variance and stability bands
A single week's number can move for reasons that have nothing to do with you. So we never show a weekly value alone. We show it inside a stability band: the minimum and maximum of that metric over its trailing weekly values, up to the last four weeks, rendered as a shaded range on the chart.
Read against its band, a number tells you whether this week is inside the normal range or genuinely out of it. A dip that stays within the band is noise. A move that breaks out of the band is worth your attention. The band is computed at read time from the weekly rollups, not stored, so it always reflects your most recent four weeks.
Position, reported coarsely
We report where your product appears in an answer as one of three buckets: top, middle, or absent. We do not report a precise ordinal rank, because the underlying answers do not support that precision. Claiming you are "number 3" when the model would happily reorder its list on the next call would be a fake number, and we do not ship fake numbers.
Limitations
What Ansedd does not claim to do:
- We measure the assistants we query directly (ChatGPT, Claude, Perplexity, Gemini). We do not measure Google AI Overviews or in-product assistants we cannot call.
- We do not predict outcomes. We report what the answers said, never a forecast of what they will say.
- One weekly sample per engine captures the trend, not every possible phrasing. Rare or highly variable answers may not all be seen.
- Citations come only from an engine's native citation objects. If an engine returns no citations, Citation Share for that engine is empty, not zero-by-assumption.
- Extraction is very accurate but not perfect. Low-confidence readings are excluded from metrics rather than guessed.
Questions about how a specific number is computed? Run the free grader and see it applied to your own product.