How accurate is AI Peekaboo?

You are about to report a number to a client, act on a missing mention, or decide whether to trust a sentiment label. This page sets out how accurate AI Peekaboo's measurement is today, with the sample behind each live figure, and how we keep results trustworthy.

At a glance

• Sentiment agreement: 96%, on 160 tracked mentions across 160 production runs (September 2026). On high-confidence overrides, agreement with the blind review was 100% (122 of 122).
• Fan-out query type: 93% matched or was judged defensible by an independent blind review, on 300 search queries from 164 brands (September 2026).
• Page type: 86%, on 200 cited pages from 192 sites (September 2026), compared with 34% from guessing by URL path alone.
• Name-variation linking: 47 of 47 newly linked canonical-name pairs right in blind review across a 7-day production replay, and 10 of 10 high-confidence competitor merges right with 0 wrong merges on 80 split candidates (September 2026).

Three parts of the pipeline do not carry a single standalone percentage today:

• Brand mentions are extracted from each full AI answer and linked through deterministic canonical-name rules. A separate full-answer mention sweep is currently running in shadow testing and does not change customer numbers yet.
• Site types classify each cited domain into one of 21 site types using domain rules plus a high-confidence classifier; a separate blind-review percentage has not been published for the 21-type set yet.
• Citation detection records the exact source URLs each AI model returned with its answer, so it has no separate classifier accuracy figure.

What these numbers mean

• Agreement: how often AI Peekaboo's label matched an independent blind review that saw the same input and the same definition, without knowing which result came from AI Peekaboo. For fan-out queries, where a search can reasonably sit on the border between two intents, agreement counts labels the blind reviewer either picked first or judged defensible.
• Name-variation precision: when two spellings or forms of a name are linked as the same company, how often an independent reviewer checking each company's website confirmed they really are the same entity.

Each figure comes from its own test on its own production sample, so they are not blended into one overall score.

How we keep results accurate

• Strict canonical matching first. Every extracted name is checked by deterministic rules covering case, spacing, punctuation, legal suffixes (such as Inc or LLC), your saved website domain, and the recognized brand names you enter in Settings. Ambiguous short names shared by two companies are left unlinked on purpose so one company never absorbs another's mentions.
• A second, independent check on tone and competitor variants. On paid runs, a separate verification model reviews the sentiment of every tracked mention (for your brand and your competitors) and checks whether newly discovered names are variations of a tracked competitor. It never merges names into your own brand automatically, so your headline visibility score is never inflated by a model guess.
• Low confidence is not trusted on its own. A sentiment override or competitor name merge only takes effect when the verification step is at least 90% confident (or confirmed consistently with zero conflicts across repeated runs). Anything less keeps the original result.
• Regular blind re-testing. Classifiers and matching rules are tested against reviewed production samples and independent blind reviews, and every benchmark on this page carries its sample size and month.

For how tone is judged, see Sentiment, explained. For page and site labels, see Source types and ownership. For search intents, see Fan-out queries.

Check a result yourself

Every number in the dashboard traces back to a real answer you can read.

1. Open Recent Chats from your dashboard, or open the prompt and pick a run.
2. Read the full AI answer, word for word.
3. Compare it with the mentions, positions and sentiment recorded for that run.

See Dashboard overview for where Recent Chats lives.

Report a miss

If an answer names your brand and the run does not count it, or a competitor is merged or split wrongly, tell us. Write to team@aipeekaboo.com with the brand, the prompt, the AI model, the date of the run, and what you expected to see. See Contact support.

If AI models write your own brand under a name we do not recognize, add it as a recognized brand name in Settings so deterministic matching picks it up on future runs.