Perception

Visibility tells you whether a model mentioned your brand. Perception tells you what it said about you. Open a client's AI Visibility project and pick Perception from the Brand group in the left-hand nav, and you get the words the models reach for when they describe each tracked brand — "responsive", "expensive", "white-label", "enterprise-focused" — with how often each one came up. Every tracked brand gets its own card: your client's brand first, then the competitors, ordered by how much evidence there is behind them. Switch the range at the top between the last 7, 30 or 90 days to see whether the language around a brand is shifting or holding steady. It's the view that answers the question a client actually asks after "are we visible?" — namely, "and how are we being described?"

Nothing new is being computed here and no second AI pass is run to produce it. When a scan reads an answer and records a brand mention, the extractor also pulls one to eight short descriptive terms for that brand, taken verbatim from the answer text, under an explicit instruction never to invent a term that isn't there. That has been happening on every scan for a long time — Perception is simply the page that finally shows it, which is why a project with scan history has something to read on day one rather than starting from zero. To be precise about what this is not: it isn't entity extraction, and the terms aren't resolved against any external database or knowledge base. They're descriptive words the model used, attached to a brand you already told us to track. If a model didn't describe a brand in a given answer, that mention simply carries no terms — nothing is filled in for it. Near-identical wordings are folded together before you see them: "affordable" and "affordability", or the plural "deals" and "deal", count as one attribute, so a single quality isn't split across three thin rows or three radar axes. Only clear variants of the same word are merged — genuine synonyms like "cheap" and "affordable" stay separate, because merging those would be a judgement the data doesn't support. The engine filter at the top of the page slices all of this to a single AI engine — how ChatGPT alone describes the brands versus Gemini alone — and every view (terms, radar, heat map, sources) recomputes from just that engine's answers. The picker lists only engines that actually produced described mentions in the current window, with each one's evidence count, and a single engine is naturally a thinner sample: expect more brands to fall below the five-mention floor there. That's the floor doing its job, not data going missing.

Each card lists that brand's top 20 terms, most frequent first, with a count and a share. Read the two sample sizes in the card header before the terms themselves: it shows how many mentions carried terms alongside the brand's total mentions, because those two numbers are usually different. The share is calculated against the mentions that actually carried terms — never against all mentions — since a mention with no descriptive terms could never have produced one, and using the bigger number would quietly overstate the evidence. The bars are scaled within each brand, against that brand's own top term. That's deliberate: a brand is read against itself, not against a competitor who happens to have five times the sample. Comparing raw bar widths across two cards would tell you more about mention volume than about language. And a brand with fewer than five attributed mentions is labelled "too few mentions to read as a pattern" — it's still shown, because hiding data is worse than qualifying it, but the label is there so nobody builds a strategy on three data points.

A term marked with a fingerprint icon is distinctive: across the whole project, in the range you're looking at, it was used for that brand and for no other tracked brand. That's where positioning becomes visible. "Fast" applied to everyone in the category is table stakes and tells you very little; "white-label" applied only to you is a position you own in the models' language — and a competitor's distinctive terms are the ground they own that you don't. Distinctiveness is relative to the brands in that project and to the selected range, so it will change as you add or remove competitors or move between 7, 30 and 90 days. Treat a distinctive term as a prompt for a question — is this how we want to be described, and is the competitor's distinctive term something we should be competing for? — rather than as a score.

Above the per-brand cards, three views compare brands at a glance. Four summary tiles name the headline facts — the attribute AI most associates with your client, a term distinctive to them, the biggest gap (an attribute a competitor is described with far more than your client is), and the most-described rival. The "Brand shape" radar plots your client against its two most-described competitors across the top shared attributes by default — and the chips above it let you choose exactly which brands to compare, up to your client plus three competitors at once (more shapes than that stop being readable). The axes re-derive from the brands on the chart, so the comparison always shows the attributes those brands actually share; brands below the evidence threshold appear greyed rather than hidden. The "Brand comparison by attribute" heat map does the same across every tracked brand at once — a grid where each cell is a brand's share for that attribute, darker meaning a larger share, and "—" meaning the attribute never came up for that brand. Every one of these is share-based and gated to the same five-attributed-mention floor as the cards, for the same reason: a brand is compared on the proportion of its own descriptions that carried an attribute, never on raw counts, so a small brand and a large one sit on the same scale. Brands without enough evidence don't appear in the comparison at all.

Under the comparison views, "Attributes & sources" pairs your client's top attributes with the pages AI cited in the answers that used each one — the concrete places to influence if you want to change how the models describe you on that quality. The number next to a page is how many of those answers cited it. Read it as an association, not a proof. Citations are recorded per answer, not per word, so these are the sources present in the answers that carried the attribute — not evidence that one page caused that exact term. Your own domain is included when it was cited, and a page cited several times within one answer counts once, the same "one answer's worth of evidence" rule the rest of the view uses.

Each card shows a sentiment score out of 100 next to the sample sizes. That score is the brand's overall tone across its mentions in the range. It is not attached to any individual term, and Perception will never show you a sentiment figure beside a term — because we don't measure that. The reason is worth stating plainly rather than burying: the descriptive terms and the sentiment score are recorded separately on each mention, with no link between them. So the honest reading is "the models described you as 'slow support', and your overall tone score is 45" — two facts, side by side. The reading we can't support is "your sentiment on support is 45", because nothing in the data connects a specific term to a specific score. If you want to know why a tone score is what it is, the terms are strong evidence to read alongside it; they are not a per-aspect breakdown of it.

The same data goes into reports. Add a Table widget, choose the AI Visibility source, and set its dimension to "Perception". The table has five columns — Brand, Term, Mentions, Share and "Only this brand" — so a client sees the language, its frequency, its share of that brand's attributed mentions, and whether it's used for them alone. As on the page, the client's brand sorts first, and there's deliberately no sentiment column. One difference from the in-app page: in a report, a brand with fewer than five attributed mentions is dropped from the table rather than labelled. On screen you can read a caveat next to a thin sample; in a client report a footnote nobody reads is worse than simply not showing a row that can't carry its own weight.

You can also have this land in your inbox. In an AI-Visibility project's Settings, turn on the "AI perception email" and once a month the workspace owner gets a short, forwardable summary for that brand: the top terms AI models used, which ones rose or faded since last month, and one attribute a competitor owns that you don't yet — the exact gap worth earning. It's the same live-computed data as the Perception page, so the email and the page never disagree. Two deliberate rules keep it honest. It only sends when there are at least five attributed mentions in the window — below that the terms are too thin to read as a pattern, so no email goes out rather than a misleading one. And on white-label plans it's sent under your agency's brand and logo, with no CrunchJunkie footer, so you can forward it straight to a client's brand or PR team. It costs nothing to run (no AI call — just a read of the mentions already captured on each scan).

Perception tells you what AI says. The messaging-gap scorecard tells you whether that matches what you WANT it to say. In an AI-Visibility project's Settings, list your target attributes — the things you're investing in being known for, like "affordable", "fast delivery" or "sustainable". Then add a Table widget to a report, choose the AI Visibility source, and set its dimension to "Messaging gaps". Each target attribute gets a row, scored against what the models actually said this period: Picked up (AI regularly uses it — 15%+ of your described mentions), Partial (it shows up, but thinly), or Not yet (AI hasn't associated it with you at all). For anything short of Picked up, the last column names the competitor who already owns that attribute — so a gap arrives with the name of who to catch. Rows are ordered gaps-first, because that's the to-do list. Like the rest of Perception it needs at least five attributed mentions to score honestly; below that the widget stays empty rather than labelling every pillar a gap on thin data. Matching is forgiving — "fast delivery" counts a mention of "fast" or "delivery" — so you can write pillars in natural language.

The per-brand terms tell you the words models reach for; Objections tells you the case against you they build. Open Perception and click Analyze on the Objections card, and Crunch reads the scan answers that mention your brand and pulls out the recurring concerns, hesitations and downsides a buyer would read alongside the recommendation — Every objection carries its receipts: expand it to read the exact answers it came from (engine and date shown, attribution validated against the analysed text — the count is derived from those receipts, never estimated) and, aggregated across them, the sources those answers drew on. Treat the source list as leads for changing the narrative — the pages those answers cited, not proven causes of the specific claim; an answer can also carry a fact from the model's training memory. The same receipts attach to supported and contradicted fact-check verdicts. "limited integrations", "pricing unclear", "smaller than the incumbents" — with how many of those answers raised each one. It's a separate AI pass over answers you've already scanned, run on demand and cached, so it doesn't slow the page or recompute on every load — click Re-analyze after new scans to refresh it. Same discipline as the rest of Perception: it surfaces only concerns actually stated or clearly implied about your brand, never generic category caveats or a competitor's problems, and an answer set with no real objections returns nothing rather than an invented list. It needs at least three answers that mention the brand before it will run.

Models sometimes state things about you that are out of date or simply wrong. Fact-checking turns that into something you can act on. In the project's Settings, under "Brand facts to check", list the factual claims you make about the brand — one per line, like "Founded in 2019", "Based in Berlin" or "Free 14-day trial". Then open Perception and click Check. Crunch reads each fact against the scan answers that mention you and marks it Supported (an answer says something consistent), Contradicted (an answer conflicts with it) or Not mentioned (the answers don't address it), with a one-line note paraphrasing the relevant answer text. A Contradicted verdict is the one to chase — it's a factual error the models are repeating to buyers, and the note tells you where. As with Objections it's an on-demand, cached AI pass that needs at least three brand-mentioning answers, and it judges strictly from the answer text: if the answers never touch a fact, it's Not mentioned, never a guess.

If Perception shows nothing, it usually isn't a fault. The most common cause is a project imported from Peec AI: Peec-imported mentions are stored without descriptive terms, so an imported project has nothing for this page to show, no matter how much visibility data it carries. That's a limitation of what the import supplies, not a bug — and it resolves for that client the moment you add native prompts and run CrunchJunkie's own scans, which populate terms from then on. Two other honest reasons for a thin or empty page: a brand-new project simply hasn't scanned enough yet, and a shorter range may fall outside your scan history — try 90 days before concluding there's nothing there. And within a range, some mentions genuinely carry no terms because the model named the brand without describing it, which is exactly why the card reports attributed mentions separately from total mentions.