How it works

One URL, a free card, a confirm step, a fenced queue, fourteen checks, two review passes and a cost ledger — as a diagram and as four rules.

The pipeline

Step	What happens there
A URL	One address is the whole input.
Instant card	That one page, read directly. Free, no account, no provider spend.
Confirm step	Is this the right business? A model adjudicates the match — and may only rule one out.
The queue	A fenced, exactly-once lease. A pull cannot be run twice or half-charged.
Fourteen checks	Website first, because five others read the pages it captured. The rest run alongside it.
Voice analysis	A model call over the pages already captured. Claims that cannot quote your copy are dropped.
Quality review	Every section read back against its own evidence. It may flag a claim and quote why. It may not change one.
Cost ledger	Every call's own price, written as it happens, summed at the end.
The report	Each section carries its evidence, or says it has none.

A report is produced once, from calls made once, and it is finished when you read it. Nothing here needs ninety days of history before it can say something — where a trend line is worth having, we buy the history rather than waiting to accumulate it.

Nothing derived speaks unless its evidence arrived

Each of the fourteen checks either produced evidence or it did not. A check that failed says so, by name, with the reason. A section with no data renders as absent rather than as a zero, because those are different facts and a reader deciding something needs to tell them apart.

Every number has a receipt

Figures carry where they came from. A claim about your tone of voice has to quote your own copy or it is dropped before you see it. A follower count that a platform did not return is reported as not returned, not estimated.

The run meters its own cost

Every external call a pull makes writes what it cost into that pull's own ledger, and the ledger is summed at the bottom of the report. The prices published on the pricing page are read out of real runs' ledgers, not modeled — and when the two disagree, the ledger wins.

Where a language model touches your report — all of it

Three of them are described below in transport detail — which model, which schema, which cap. The complete list is on the models page, and it is longer than three: the quality review and the confirm-step suggestions joined it. Nothing else in the pipeline asks a model anything; the remaining checks are API reads and page captures, and their output reaches you unmediated.

Call	When	Model	What it does	What it may NOT do
The brand adjudicator	On the confirm step, before a pull is charged.	`gpt-5-mini, strict JSON schema`	Reads the candidate business we matched and gives a second opinion on whether it is really you.	It may only DEMOTE a match, never promote one. A model that could talk a wrong business into your report would turn a hallucination into a charge; one that can only rule candidates out turns its worst case into a question you get asked.
The voice analysis	Once per paid pull, in the voice leg.	`gpt-5-mini, strict JSON schema, token-capped`	Reads the copy captured from your own site and proposes claims about how you sound.	Every claim must quote the text it came from. A claim whose quote is not found in the captured copy is dropped before the report renders — so the section can come back with fewer claims than the model produced, or with none, and it says so when it does.
The AI-visibility grid	On the deeper settings, in the AI-visibility leg.	`six families — GPT, Claude, Gemini, Grok, GLM and Kimi — three runs each`	Asks each one the same questions about you and your category, three times, and shows you every answer beside the others with how many of the three named you.	It cannot look anything up. Every one of these calls is a brand-new logged-out session with no browsing, no memory and no account, so it reports what the models absorbed in training and not what a searching customer would be shown — the panel says 'from training, no search' for exactly that reason, and the ChatGPT-app panel beside it is where the searching customer's answer lives. On a free-tier Gateway key it also cannot burst all six at once without most of them 429ing — it paces itself instead (burst until throttled, then slow that one model down), and if the pull's own time budget runs out first it says exactly how many of the six answered rather than presenting a partial grid as a full one.

All four, with what each may and may not do: Where a language model touches your report.

What the website capture gates

The website check runs first because 5 others read the pages it captured: Screenshots, Logos, Colors, Typography, Voice. This is not a preference about ordering — a site that cannot be captured takes all of them with it, and each says so on its own page rather than rendering empty.

The cross-section dependency map

The pull is not fourteen independent checks — it is a chain of custody for identity. Everything downstream depends on knowing WHO the brand is, and that knowledge is built in layers, each one narrowing what the next is allowed to claim.

Layer 0 — what you told us

Feeds: everything

The confirm-step interview seeds the entity anchor before any money moves: brand name, search phrase, industry, location, confirmed social profiles, services, aspirations. A confirmed field outranks every inference below it, and wears a 'you confirmed this' receipt wherever it is used.

Layer 1 — the website capture

Feeds: looks, voice, search seeds, social ownership, structure, the anchor itself

The one source the brand fully controls. It hands the looks check its logos, colors, fonts, favicons and screenshots; the voice check its quotable sentences; the search checks their site vocabulary and seed phrases; social discovery its ownership proof (a profile your own site links to is yours by evidence, not by name); and the site structure its map. If this capture is poisoned, everything downstream inherits it — which is why palette picks are pixel-verified against the run's own screenshots.

Layer 2 — the entity anchor

Feeds: every check that spends money on a name

Assembled from layers 0 and 1 plus the reviews check's domain-verified city. It is the pull's answer to 'who is this brand, provably' — and every identity-sensitive check consults it before spending: search seeding disambiguates a generic name, the AI mentions check refuses to spend when the name cannot be attributed, social candidates classify owned or lead, reputation hits classify verified, lead or off-entity, and the industry directories know which shelves to check and in which city.

Layer 3 — the market context loop

Feeds: keyword universe, market shelves, competitors, content gaps

The search checks buy the brand's ranked keywords and its market's vocabulary; the keyword universe clusters them into topics; the topics' category terms drive the recurring-domain sweep (which directories and listicles own the brand's searches); and the competitor set — entity-gated so a film character cannot be a rival — feeds overlap, gap, and the positioning map. This is the layer that searches what the brand IS, not just what it is called.

Layer 4 — the cross-checks

Feeds: voice, reach, AI answers

Sections check each other where their evidence overlaps: voice claims are compared against real social captions; the AI mention corpus's own entity categories disprove wrong-entity mentions; the app-surface answer is diffed against the API answer; and follower counts, dates and media all carry their provenance or their honest absence.

Layer 5 — the quality review

Feeds: the reader

Last, and deliberately so: a model re-reads every section's derived claims against that section's own evidence, with a digest of the sibling sections and the anchor in front of it — so 'the palette disagrees with the screenshots' or 'these keyword themes do not match what the site sells' is caught after everything has landed and before anyone reads a wrong fact. It can flag and demote; it can never silently fix. Its verdicts ride the report as caution chips with quoted receipts.

Why the order is the order

This is why the pipeline's order is what it is: website before the wave (the anchor needs layer 1), reviews feeding the anchor before search and AI spend (the city and the domain-verified listing narrow both), the universe before the shelves (category terms drive the sweep), and quality last (an auditor that runs before the evidence lands has nothing to audit). A check that ignored this order would not be cheaper — it would be confidently wrong about a different brand.

Two narrower, machine-derived answers complement this map. The website-gate diagram above is drawn from each check's own dependency sentence, not from a list kept beside them. And the quality review's reading permissions — what it may look at when judging one claim — are enumerated on the models page.