How we handle your documents

AI provider

Claude, made by Anthropic, used through Anthropic’s commercial API. No other AI company receives anything from your documents or your website.

Training

Under Anthropic’s commercial terms, what we send through its API is not used to train its models. We do not use your documents to train anything either. How long Anthropic keeps what it receives is set by those same terms.

Turning AI off

Anyone signed in to your organization’s account can turn AI off on the Account page, or ask us to. With it off, nothing from your documents or your website is sent to any AI company.

Fixed PDFs

Available with their originals for a week after fixing finishes, then no longer available, and removed from our database within about an hour.

The free PDF page

Calls no AI model and never saves your file.

These run on our own servers, whatever your AI setting, and none of them sends your document anywhere:

• adding a text layer to a PDF that has no text at all, by reading its scanned pages;
• setting the title and the language;
• naming form fields from the words printed beside or above each box;
• building the tags, where the copy can carry them;
• checking the reading order of those tags against where the text sits on the page;
• converting a document to a web page with our own layout reader, where your account has one;
• the guided review, where a person on your team answers what software cannot.

Two of these use machine-learning software: text recognition, which reads the words on a scanned page, and the layout reader. Both run on our servers. Neither sends your document anywhere, and neither writes words of its own: they read what is printed on the page. If your policy counts that as AI, tell us before you start.

A description of each picture

Where the fixed copy has tags and picture descriptions are on for your account, and in a web page version where picture descriptions are on for that work: that one picture, the document’s file name (for a web page version, the document’s title instead: the one our layout reader read from it, or the one AI wrote for the version), the picture’s caption if it has one, and any words already read inside it. Never the rest of the page. In a fixed PDF the description is written into the copy you download, and labelled as written by AI beside its picture in the report and the review; the PDF itself does not mark it as AI’s, so read each one before you publish. In a web page version, which has no review, it is written in as the picture’s text alternative without a label, and a picture AI calls decoration is given none, so read the version before you publish it. The report on each fixed copy says whether its pictures were described.

A web page version AI writes

Where our own layout reader cannot convert a document, or the document is longer than it takes, or the account has no reader: the document’s text, page by page, and its file name. Its pictures go separately, as above. With AI off, such a document is reported as not converted instead, and costs nothing.

A worked example on a website finding

Only when somebody presses “Show an example fix”: the public HTML of that one element of your website and the rule it failed. No document content.

Nothing is sent during a website scan, and nothing is sent unless somebody on your account asked for a fix, a conversion or a worked example.

For everything: anyone signed in to your organization’s account turns AI off on the Account page, and can turn it back on there. We can also set it for you. Turning it off applies from the next document: work already running finishes the documents it is on and sends nothing more to AI after them. What was already sent cannot be recalled.

For one set of files: choose “Don’t use AI on these files” when you send them, and leave AI on for everything else. Fixing the set again, or making a web page of one of its PDFs, keeps that choice.

With AI off your PDFs are still fixed. Pictures come back without a description, and where the fixed copy has tags each picture gets its own card in the review for a person on your team to describe. Turning AI off does not change what a page costs.

It fixes up to 5 pages of one PDF, calls no AI model, and never saves your file. Where your PDF has no text, text recognition reads it on our own servers. The fixed copy waits in our server’s memory for up to fifteen minutes until you download it, and is gone once you do.

If you give us your email address to download the copy, that address is kept as a contact, and we email you at most once to ask how it went. Our privacy page says how long it is kept and how to have it removed.

It is one of our public pages, so the advertising and visitor-measurement tags the privacy page lists run on it. The panel that shows your file’s name and results is hidden from the session recording.

• Every PDF you upload: held in its upload set for three days after you add it.
• A PDF you fix: also kept, beside its fixed copy, for a week after fixing finishes, so you can compare and download them. The record that it was fixed, meaning its name, the date and what happened, stays.
• A PDF converted to a web page: the copy sent for conversion is removed in the same step that saves the result, and the PDF you uploaded stays in its upload set for three days after you add it. The web page version can be downloaded for three days after it is made.
• Assessments, findings and document lists: kept for as long as you have an account, and deleted when you ask.

Once a window closes, those files are no longer available, and they are removed from our database within about an hour.

Our database provider keeps backups for 7 days. A file removed from our database leaves those backups when they expire.

Only people signed in to your organization’s account can see your documents in the product. Our application runs on Render in Oregon, United States, and our database is hosted by Supabase in Oregon, United States; every company that handles any of your data is listed on our privacy page.

We answer security and AI questionnaires directly. Send yours through the contact page, and see security for how customers are kept apart and how sign-in works.

Back to the list of sections