Brand Perception

Being mentioned is table stakes. Brand Perception measures the sentence after your name — how consistently engines recognize you, who they put you beside, which sources shape it, and what they actually said.

This screen matters because AI answers are read as summaries, not opinions. A wrong claim in an assistant's answer is absorbed as fact by a buyer who will never see your correction.

What is measured today

The deterministic half of perception is live, read from your collected answers over a rolling 30-day window:

Always read Awareness together with Answers analysed. An awareness figure over a handful of answers carries an interval wide enough to be meaningless, which is exactly why the interval is shown rather than hidden.

Voice of AI

Verbatim snippets from the answers themselves, with the engine, the prompt that produced them, and when they were captured.

This is the most useful panel on the screen and the least abstract. Reading ten snippets tells you how the engines frame you in a way no aggregate can — including the phrases they reuse, which are usually lifted from a source you can find in Sources.

Strengths, weaknesses and misconceptions

The panel that answers the actual question: what do the engines think, and which parts of it are a problem.

A pass reads each collected answer that names you and extracts the claims it makes, split three ways:

• Strengths — what you are credited with.
• Weaknesses — what is held against you.
• Misconceptions — what is simply wrong, each carrying the correction.

Claims are folded into themes, and each theme reports how many distinct answers made the claim — the number that carries weight — its share of analysed answers, which engines it appeared on, and whether it is about your brand or one of your products.

Every claim carries its evidence

Each claim comes with a verbatim span from the answer it was taken from, and that is the design rather than a nicety.

A model summarizing model output is a machine for producing a plausible, tidy, unfalsifiable list. Requiring the quote makes the claim checkable: any claim whose evidence does not literally appear in the source answer is dropped before it reaches your screen. Answers are read one at a time rather than batched, so claims cannot bleed between them and be attributed to the wrong source.

The practical effect is that a theme on this screen can always be verified by reading the quote next to it.

Coverage comes first

Above the themes sits a coverage line: how many of the answers naming you have been read, out of how many are worth reading, and how many of those said anything at all.

Read it before the themes. Every number under it is only as good as that fraction, and a gap between analysed and said something is not a failure — an answer that lists you without asserting anything about you genuinely contains no claim, and inventing one to fill the gap is exactly the failure mode the design guards against.

Perceived competitors

The brands engines co-mention with you, ranked by how often you appear together.

Often not your slide-deck competitor set. When engines pair you with a segment you do not compete in, that is a positioning problem visible before it reaches your pipeline — and when a name recurs that you do not track, add it in Competitors so it starts counting toward Share of Voice.

Sources behind perception

The domains cited in the answers that mention you, with their citation counts and the URLs involved. When a description of your product is wrong, this is where the wrong description came from — and correcting the source is more effective than publishing a rebuttal on your own site.

What is not measured yet

Per-mention sentiment scoring is still pending, so the sentiment figures on this screen and on the Overview come from the sample layer rather than your answers.

Panels that depend on unavailable data are hidden rather than estimated, and unmeasured values render as a dash. You will not see a number here that was inferred from nothing. See Product status.

Working the screen

1. Start with misconceptions. They are cheap to fix and expensive to leave — usually one outdated third-party page and one correction request. Each carries its evidence and the correction.
2. Read the weaknesses honestly. Engines are summarizing sources; a weakness that recurs across engines is a real market position, not a bad review.
3. Read Voice of AI weekly. Five minutes, and it catches misdescriptions early.
4. Check the co-mention list against your intended competitive set.
5. Follow every problem to its source. Find the page rather than arguing with the output.
6. Watch coverage and Answers analysed. If either is thin, the fix is more prompts on perception intent — see Prompts — not a bolder reading of the same small sample.

Perception moves more slowly than visibility. Expect weeks, not days.