Reporting

AI answers are non-deterministic, so an AI-visibility number is a sample, not an absolute. To keep that honest, AI-visibility KPI cards — in the app and in reports and shared links alike — show how solid each figure is. Under the metric you'll see "Based on N runs", where N is the number of scan runs behind it (prompts × models × repeats in the period). For Visibility %, the card also shows a margin of error as "± X%". We compute it to be genuinely informative rather than merely reassuring: it is boundary-safe (it never implies an impossible visibility below 0% or above 100%), and it widens when your AI engines disagree with each other or when the sample is thin — so a rate built from many agreeing runs shows a tight ±, while a shaky one shows a wide one. That is why two brands can share the same headline % yet carry very different margins; read the number and its ± together. We take that one step further on changes over time. When Visibility moves versus the previous period, we compare the change with the combined margin of error of both periods. Only a move of at least twice that margin is clearly beyond the margin of error; anything smaller is not, and is flagged "within margin" next to the delta on the KPI cards, and shows as a grey "~" before the change on Home, in the Overview's competitor table and Biggest movers, and, in a report, on the change chips of the Brand table, in the Prompt table's Δ Visibility and in the Brand × Tag, Tag × Month, Brand × Month and Tag tables. It is one rule on every one of them, with the same margin (the "± X%" of the headline), so a move never reads as a gain in one place and as "~" in another. AI summaries follow it too: Crunch is told to describe such a change as unchanged, never as a rise or a fall. A count such as Mentions also follows the number of answers collected, so next to a comparison its card names the answers in both periods ("Based on 1,426 runs · 2,940 in the comparison period"). It is an honest signal that the move may well be noise rather than a real shift, so a small wiggle (say a three-mention drop when citations are highly variable) isn't over-read as a trend. A small coloured dot next to "Based on N runs" summarises how much to trust the figure overall — green for high, amber for medium, grey for low — combining the sample size with the width of the ±. This annotation appears only on AI-visibility metrics, where sampling applies — ad-platform and analytics numbers come straight from the source and aren't sampled. It's the same principle as the free AI-visibility check: show the sample size, never over-claim a single number. The same honesty applies while a scan is running. If individual checks fail — a provider outage, a rate limit — the scan retries them a bounded number of times, then completes without them and tells you how many failed, right in the scan status bar. A failed check is never silently retried forever and never counted as a run, so N only ever includes checks that really happened. After the scan, the Prompts tab shows exactly which prompts had failed checks — an amber badge on the prompt names the engine and the provider's own error message — so a smaller sample size is always attributable, never mysterious. Only whole answers count. An engine reports why it stopped writing: it finished, it reached the length limit the scan set, its web search ran out of rounds, or it declined to answer. ChatGPT, Gemini, Claude, Grok, Perplexity, DeepSeek and Meta AI answers are stored only when they are complete. One that arrives empty or cut off is not a measurement (a brand named after the cut would be invisible), so it counts neither as a mention nor as a miss: the check fails with the reason, a later pass of the scan may ask again, and an engine that keeps answering like this is stopped for the rest of the scan, with the reason in the scan bar. An answer the engine declined to give is stored as its answer, because that is what a real user gets; one its safety filter cut off partway is a fragment and is not. A failure on the provider's side reported inside an answer is temporary and is retried. Google AI Overviews, AI Mode and Microsoft Copilot work as before: an empty result there means no AI answer was shown. N is the number of answers that count. The limits leave room for the longest real answers: 8,192 tokens for Claude and Gemini and 8,000 for ChatGPT, Meta AI and Grok, and Gemini thinks at its low setting because its thinking counts against the same limit. Brand names are read from up to 40,960 characters of an answer. Answers stored before this check existed (October 2026) were not checked this way. Until then Claude answers were limited to 1,024 tokens, so most stopped before the end and some arrived empty, and Gemini answers often ended after a few hundred characters, so Claude and Gemini figures from before the change read low. A project with answers from before gets a note on its Timeline when its first complete answer is stored, and drop and overtake alerts, signals and client digests leave that engine out of any comparison that spans it (with no engine left, they say nothing). In the MCP connector, get_ai_answer shows the reason each stored answer ended (stopReason). Names Google linked in AI Overviews. SerpAPI, which reads Google AI Overviews, AI Mode and Microsoft Copilot for us, sometimes returns a name Google linked without its text: the link's address comes back, the name does not. A list item then reads ": One of the leading full-service agencies …" with no name before the colon. Since 11 October 2026 a scan keeps those links with the answer. Where a list item or paragraph starts without its name (with a colon, a comma or a dash, or empty) and its first link is such a link, the link's domain stands in brackets in the name's place, such as "[example.com]: One of the leading …", so you can see that a linked name stood there; we do not guess the brand's display name. That link, and only that link, counts as a mention when its domain is the client's or a tracked competitor's own domain: the client's website, a domain tagged Owned, the GEO audit's domain or an alias written as a domain, or a competitor domain set in the AI-visibility project's settings. A subdomain or a shared platform (facebook.com, g2.com, amazon.de and the like) stands for no brand. Such a mention counts for visibility and share of voice only, without a position or sentiment, and only when the answer does not name the brand anyway. The Responses view and get_ai_answer in the MCP connector label it: a linked name SerpAPI returned without its text; counted from its link's domain. A link without text in the middle of a sentence, links with text, and the sources listed under an answer never count as mentions. Answers stored before 11 October 2026 were not read this way and can miss a brand Google linked; they are left as they were. A project with earlier answers gets an internal note on its Timeline for each engine, on the day its first answer is read this way, and alerts do not compare across it. A table row whose name SerpAPI dropped together with its link cannot be recovered. One scan per day counts. If a prompt is scanned more than once on the same day — a manual re-run, a retry after a settings change — the later scan replaces the earlier one for that engine and day. It never adds to it. Pooling both would double the sample without any new evidence and make that day a bigger population than every other day. The same rule applies everywhere the runs are read: the headline KPIs, the topic and source cards, report widgets, the API and the MCP connector all measure the same runs, so their numbers agree. "Last 30 days" means 30 whole days. The window is counted in calendar days in UTC and always ends with today, so "last 30 days" on 23 September is 25 August to 23 September inclusive — never 31 days, and never a part-day at either end. That matters because a window measured as "now minus 30 × 24 hours" would cut the oldest day in half and would cover a different span depending on the hour you opened the page. Every surface uses this one definition: the headline KPIs, the charts, the competitive table, the topic and source cards, the per-engine tables, report widgets, the REST API and the MCP connector. So a sample size shown on one card is the same population as the one beside it. The days are UTC days because scans are stored by UTC day. In a report, the dates themselves come from the report's range, which follows your workspace's time zone; AI visibility then counts the scans of those dates by their UTC day.