How a report gets made

What happens after you pay

Four steps, and they are the same four the progress page counts off as it runs.

Finding your pages

Every address your site will admit to.

Reading your site

Up to 100 of your pages, best first, as they stand today.

Writing your report

A brief per page, plus the easy fixes. This is the long one: nearly all of the wait is here.

Printing your PDF

The file you keep, typeset from the report itself.

The link, and the email

The report goes up at a private link and the same link is emailed to you.

Reports so far have taken between five and fifteen minutes. A large site takes longer, and the link is emailed to you the moment it is ready, so there is nothing to sit and wait for.

What we open, and what we never touch

We read your site in an ordinary browser, because what the report is about is what a browser and a search engine can see. That means the pages that are publicly available, which are the same pages a search engine reads, and nothing else.

• We read your robots.txt and obey it: a page it disallows is not requested. Three requests come before it or beside it, and the note below names them.
• We never attempt to reach anything behind a login, and we never submit a form.
• We open at most 100 pages, best first. On a larger site the rest are counted rather than read, and the report says so rather than implying the whole site was seen.
• We space our requests out, and we identify ourselves in a header on every one of them.

The crawler has a page of its own, written for whoever runs the site, with the exact header and how to slow it down.

What is measured before any prose exists

This is the part worth knowing if you have been sent an AI-written audit before. The measuring code calls no model, fetches nothing and rolls no dice: it is a function of what the crawl brought back. All of it is computed and stored before anything is asked of a model, so the findings a report is graded on exist whether or not the writing succeeds.

Held against your own answers, it works out which of the things you say you sell appear anywhere on the site, whether a price, a contact and a real postal address can be found, what a machine can state about the business from its structured data, and whether the site agrees with itself about those facts. Then, on the pages themselves:

• Pages sharing one title
• Pages sharing one summary
• Pages with no title
• Pages with no summary
• Titles cut off, or too short to say anything
• Summaries cut off, or too short
• Pages a machine cannot read as a business
• Pages that point at another page as the real one
• Pages that do not say which address is the real one
• Pages telling search engines to ignore them
• Pages the server tells search engines to ignore
• Language versions that do not point back at each other
• Pages with no main heading
• Pages with more than one main heading
• Photographs with no words behind them
• Pages too short to answer anything
• Pages that do not adapt to a phone
• Pages that only answer after two or more forwards
• Pages asking for a file that is not there
• Pages where a script stopped part way
• Addresses of yours that should be sent to one of the others
• Pages of yours your robots.txt hides

Everything printed around the rating comes out of those measurements: the four dimensions that say what is holding the figure down, and how much of your site we managed to read. The figure itself is the model's, proposed when it writes the report and clamped by the code to the scale it was asked for. We would rather say that than let you assume otherwise: it is a judgement, the things printed beside it are not, and the report is laid out so you can hold the one against the other.

Where the model fits, and where it does not

Plainly: a language model writes the prose. The text of your site and your answers are sent to Anthropic, and that is the one place your information leaves our servers as content. It is unavoidable, because it is how the report gets written, and the privacy notice says so too.

What is worth knowing is how narrow that job is. It is used twice:

Choosing what not to download

On a large site, deciding which shapes of address are not worth opening, so the crawl spends its pages on the ones a buyer lands on.

Writing the prose

Given everything measured above, plus the text of the site, it returns the verdict, the briefs, the buyer questions and the order of work.

That is also why the report is worth arguing with. Every finding is supposed to be traceable to a page we actually read, so if one states something about your site as a matter of fact that is not true, it is a fault of ours and we will correct it or refund you.

What we keep afterwards

The rendered HTML of every page we read, the text of your site as one document, and the report itself with its PDF. Those are the evidence behind the findings and what makes one checkable months later. The privacy notice says how long, and what happens when you ask us to delete it.

There is no account, no cookie, no advertising pixel and no mailing list. The email address you give us is used to send you the report and for nothing else.