Brand perception in AI answers: what AI says about you

Visibility says whether you appear

Visibility answers a yes or no question: did this answer mention your brand. Perception answers the harder three: how does the answer describe you, what does it argue against you, and is any of it true.

Visibility is the easier number to live with. The share of answers that mention your brand moves on a trend line, and you either appear or you do not. A brand can sit comfortably above 50% Visibility and still lose deals, because the answers that mention it call it expensive, enterprise-only, or freshly shut down. Appearing and being described well are separate problems that need separate measurements.

Your tracked prompts were chosen to catch mentions, so they are the wrong instrument for description. Asking “best help desk software” tells you whether you made the list. It does not tell you what the list said about you after you left the room.

The three questions perception answers

Which attributes get tied to your brand, which objections recur against it, and whether the claims in answers match the facts you have registered. In advisible those are the three tabs: Market, Objections and Fact-checking, each with a different owner and a different next move.

• Market. Two independent scores per attribute, Association and Market prominence, each 0 to 100 with a 95% confidence interval. Positive attributes only: anything negative goes to Objections rather than dragging a score down.
• Objections. Arguments against your brand that recur across answers, each with a category (product, pricing, trust, support, comms), a trend, and the sources cited behind the hits.
• Fact-checking. Every atomic claim in an answer that touches one of your registered brand facts, judged Contradicted, Supported, Inconclusive or Not covered.

Read in that order and the tabs tell a story: what you are known for, what is said against you, and which of it is checkable at all.

How discovery works when you supply no topics

Perception runs its own question sets instead of your tracked prompts, in a weekly cycle of two waves. Wave one asks about your brand and your category; the answers get broken into claims, embedded and clustered, and the groups that recur become attributes and objections. Wave two then measures how those attributes play out across the category.

The clustering step is what makes discovery possible without you naming topics. Ten answers that say roughly the same thing in ten different wordings collapse into one objection with a frequency count, while one-off claims that never recur get dropped. You never typed “onboarding is slow”; the answers produced it, and the count tells you whether it is a theme or a bad day.

What to do with each output

Each tab points at a different owner: Market goes to whoever owns positioning, Objections goes to the team that can fix the complaint, Fact-checking goes to whoever maintains the page the answer cited.

From Market, pick the attribute where your Association runs far ahead of Market prominence (or the reverse) and decide whether it is worth owning. The full logic of that comparison is in association vs market prominence.

From Objections, route by category rather than debating them: pricing objections go to whoever sets packaging, comms objections go to PR, support objections go to the support org. The cited sources behind each hit tell you which review article or thread is doing the repeating, which is usually a faster fix than arguing with the model.

From Fact-checking, Contradicted verdicts become page work: find the source, update it, re-run. The workflow is written out in fact-checking what AI says about your brand.

Why scores stay hidden until there are enough answers

A score built on four answers is noise wearing a percentage, so scores stay hidden until the sample clears a minimum, and every published score carries the range the true value most likely falls in.

The interval does real work. A gap between two attributes only gets surfaced when the intervals no longer overlap, and a week-over-week change only gets claimed when both weeks clear the minimum and the move is larger than the combined uncertainty. That is why the dashboard sometimes shows no movement at all while the raw answer count keeps climbing: nothing has been proven yet, and a number you cannot defend in a meeting is worse than no number.

While a score waits, the raw material is not hidden. Every original answer sits in the answer archive, so you can read what is being summarised, and each metric states its sample size as “based on N answers” instead of a bare percentage.