How We Track AI Releases
Updated
Every entry on an /updates page is a release a company announced on one of its own channels, dated and linked to that page. This is how the logs are built and what the numbers on each page count.
Sources
For each company we keep a list of official sources: changelogs, release notes, GitHub releases, product blogs and news pages. Third-party coverage is never used as the source of an entry. When a release is announced in several places, we link the most specific official page.
How often we check
GitHub releases and RSS or Atom feeds are polled every 15 minutes. Other official pages are watched through a stream of web page changes, processed every minute, so a release usually appears here soon after the company publishes it.
How an entry is written
A language model reads the official page and drafts the entry: a title, a one or two sentence description, the release date, the product it belongs to, and one of eight categories (Model, Product, Developer Tools, API, Infrastructure, Research, Enterprise & Security, Ecosystem). The draft is checked against a fixed schema before it is published, and entries that fail are retried or held back.
The release date is the date the company published, not the date we found it. Month summaries at the top of company pages are generated from the listed releases and labelled as summaries.
Duplicates and corrections
The same release is often posted to a changelog, a blog and a social account, sometimes a day apart. We match new entries against recent ones from the same company and keep one entry per release. When we find a duplicate, a wrong date or a wrong link after publishing, we archive the bad entry rather than silently editing history.
What the numbers count
A release is one announced change, so a large launch and a small fix each count as one. Counts compare a company with itself over time more fairly than with other companies, because some companies post every small change and others bundle them into monthly notes.
When a page is indexed
A company page is open to search engines once the company has at least 12 releases in the last 12 months and at least one in the last 90 days. Pages below that bar still work and are linked, but ask search engines not to index them until there is enough to read.