Method

How it grows

Every company in the Glyph Companies registry that makes AI, data, cloud or developer software is in scope. Each Monday the check enrolls new ones as candidates, then looks for legal pages (terms, privacy, subprocessors, data processing) on a few candidates' and vendors' own sites. A page a vendor has no watched page of that kind for is watched from then on. Sentences that look like answers are kept as leads, unverified, until a person or a research pass confirms them against the page; only then does a vendor get answers here. Research passes read vendors' pages by hand and every quoted clause is checked against the live page before it goes in.

The vendor's own words

Every answer rests on a clause quoted word for word from the vendor's own public page: terms of service, privacy policy, data handling documentation, help center article or subprocessor list. The short answer and its colour are a reading of that clause; the clause and its link are there so you can judge it. When no watched page answers a question the cell says so rather than guessing.

Colours

Green means the clause protects your data (no training by default, deletion within a stated period, a self-serve zero retention setting, a published list). Grey means it depends on a plan, an approval or a setting. The warm colour means the vendor uses or keeps the data by default. Each is also written out in words.

Checked weekly, cheaply

At 06:40 UTC every watched page is read once and reduced to its policy text (menus, link cards, scripts and styling removed, spacing and quote marks normalised), and the text is hashed. An unchanged hash costs one database write. Only a changed page is parsed further: the new text is stored, compared paragraph by paragraph with the previous version, and recorded as a change with the added and removed paragraphs. Page noise is not a change: event banners and promotions, cookie bars, page widgets, related-article links, footer years, spacing, and text the vendor’s other watched pages share (its menus and banners). A paragraph that reads like policy, or that an answer quotes, is never treated as noise. Every answer quoting that page is then looked for in the new text; a clause that is gone is flagged on the answer until a person re-reads the page. Nothing is deleted.

Pages that block robots

Some vendors (OpenAI, xAI and Perplexity among them) refuse automated reads, and some pages only render in a browser. Those are read from their newest Internet Archive capture, so a change shows up once the archive has captured it. Each page says whether it was read live or from the archive.

History

Before the weekly check began, each page's history was rebuilt from Internet Archive captures since January 2024 (up to ten per page, evenly spread, plus the newest). A change found this way is dated by the first sampled capture that shows it, so it happened on or before that date, and links to the archived copy.

About this data

Each answer quotes a clause word for word from the vendor's own terms, privacy and documentation pages. The pages are checked every week and every change is dated, with older history from the Internet Archive. A weekly run also finds new vendors and pages; what it finds stays an unverified lead until the clause is confirmed on the page. Not legal advice: your contract and the current page decide. Logos via /api/logo; trademarks belong to their owners. Anonymous usage stats via Google Analytics. Public pages are a summary: your contract, order form or data processing addendum can say more and takes precedence. Use Suggest a correction on any vendor.

Across Glyph Ibis

Certifications and attestations (SOC 2, ISO 27001, FedRAMP, HIPAA) are on Compliance. Vendor names and slugs follow the shared registry at companies.fru.dev.