The product, screen by screen

Check the site

Type an address and press Check. The site is fetched the way an AI crawler fetches it and you get one list, in the order worth acting on: blockers where AI cannot read you at all, services AI is not naming you for, being named but not cited, machine legibility, then emerging standards.

Every row says what was found, why it matters, what to do, and who can do it — the engine, a snippet you paste, or a developer. No key, no Google account, no campaign: this works on a brand-new account.

Measure what the engines say

Add an AI key and the engines are asked the questions your buyers ask. Named, cited and recommended are counted separately per engine, each with its 95% range and sample size; live-search answers and answers from training are kept on separate rows, because nothing you publish can move the second kind.

Google’s AI Overviews are read as their own surface from a live results page, from the place you set for the project (the United States where none is set, said on the row) — presence and citation apart, counts across several readings, and the sites each overview drew on instead. Google’s own AI Overview and AI Mode impressions come in from a Search Console export, tied to pages and never to prompts.

See who they cite instead

Every site an AI answer cited for your topics, across all your projects, with links to the exact pages. Your own pages and third-party pages are kept apart, sources the engine declared are kept apart from pages extracted from prose, and you label what kind of site each is — stockist, competitor, directory, publisher, community, reference — so the naming rate per kind means something.

The competitor report shows the searches a rival holds on page one where you are absent, built from rank checks already running, and hands any of them to the Evidence Engine as a topic in one click.

Keep the topic list honest

The topics you confirmed at setup become a living list: what you seeded, what the engine found in recent AI answers, how often each was measured and how often named, its buying-journey stage, and clusters so two phrasings of one question are not measured twice. Topics measured repeatedly and never named are retired; new ones are proposed with the page they were found on.

Let the engine find and close the gaps

For each topic the engine works out what evidence the question needs and what is missing. Each gap names the object it wants — a guide, a comparison, a data table, a deck, a video — and closes only when that object is published, with the link. Before anything is written, two or three of your own AI providers decide the shape of the answer; the writer is given your own pages and videos on the subject and what the strongest outside sources actually say.

On Growth this runs every 6 hours, on Agency every 4, on Enterprise every 3. On Starter you press Run. Review before publishing is on for new accounts, so every page for your site waits for you; switch it off once you are confident and it publishes on its own and reports what it did.

Check the page answers, and keep it aligned

Answerability checks a page against the shape it was promised, quoting the sentence that carries each element, and — if you switch grading on — asks your providers whether it answers a real query from your Search Console, with a verdict and the tally. Pages that miss elements are rebuilt in shape rather than rewritten from scratch.

Re-align rewrites a page already published so what it says about your product agrees with your own description; the sentences dropped are named, and nothing replaces the live page until you approve it. Beside it: Copy HTML, and Send again for a site that cannot take an update.

Publish to your own site first

Connect WordPress, Shopify, Wix, Ghost or any webhook and a project publishes to a destination on its own domain first, with the same facts in Google Docs, Sheets, Slides and Blogger, where Google’s engines read them. Webhook messages are signed the Standard Webhooks way, typed post.created or post.updated, and keyed on an id that never changes, so a resend cannot duplicate a post; a destination that accepts updates receives them in place. New pages carry provenance in schema.org, and are announced through IndexNow once the site serves its own key file.

Prove what happened

Connect the Cloudflare zone that serves your site and every night we read which AI crawlers fetched which of the pages we published there, how often and when last — with “nothing yet” against the pages nothing has come for. A later AI answer that cites a published item is recorded against it with the engine and the date; “never cited” is an honest value.

The outcomes report leads with the work done, then published, crawled, indexed and cited per site. The project overview opens with three sentences written from the latest runs — what changed, what was published and cited, and what needs you — over each engine’s named, cited and recommended counts with their 95% ranges, and a dashboard note lists what is new and what needs a person, each line linking to where it is done.

Read it all from your own AI tools

Mint a scoped key in Settings and point Claude Code, Claude Desktop, the Claude API or your own tooling at the account’s MCP server, or have Claude connect by signing in instead — both work on every plan — and choose one project or every project. Fourteen read tools cover visibility, per-topic measurement, cited sources, gaps, the published-evidence ledger, answerability, what is waiting for your approval and the entity; every rate arrives with its sample size, and nothing can be changed through it unless you tick “Let it make changes” when you make the key or the connection, which is off by default.