Methodology

This site makes one promise: every number traces back to a measurement, a source URL and a rule you can read. Below is how that works, including the parts that do not work well yet.

There is no AI in this product

No model is called anywhere in the running system — not remotely, not locally. That is a constraint we set ourselves, and it forces two useful properties. Identical input always produces identical output, so a number can be reproduced. And every claim can be traced to the rule that produced it, so a wrong claim is a rule you can fix rather than a mystery.

There is no sentiment analysis

We never ask whether an article is positive or negative. Word-counting sentiment is a blunt instrument, and on a subject where half the coverage is rewritten marketing it is worse than useless.

Instead we ask what kind of event an article describes — a release, an outage, a price change, a lawsuit — using explicit rule patterns. Whether an event type counts for or against a tool is written down in a versioned rules file, by a person, in advance. That is the difference between an opinion and a measurement. This layer is not live yet; it arrives with scoring.

How coverage is counted

An article counts for a tool when one of the tool’s registered names appears in its headline or summary. Names that are also ordinary English words carry extra rules: Perplexity is also the standard measure of model quality, Runway is also an airport and a startup metric, Claude is also a first name. Those names only match when supporting words are nearby, and never when a disqualifying phrase is.

Every match stores which rule produced it and the surrounding text, so a wrong match can be found and corrected instead of quietly distorting a count. The tool pages list which of a tool’s names are guarded this way.

“Coverage” on the tools page counts the last 30 days. The sources column counts only publishers independent of the vendor — a company blogging about itself cannot raise its own number.

How articles are grouped

Two articles are treated as the same story when they share at least 35% of their meaningful words, with a minimum of 3 shared words, within a five-day window. Articles from the same feed are never grouped: no publisher reports the same event twice, so overlap within one source is boilerplate rather than a story.

Announcement words are excluded from that comparison. That is not a guess — two unrelated posts, “Introducing OpenAI o1” and “Introducing the SQL Console”, scored a quarter similarity on the single shared word “introducing”.

The grouping is deliberately conservative. A missed grouping counts one event twice, which is visible and fixable. A wrong grouping makes an event disappear, and you never see that again.

Near-identical rewrites

Separately, we measure whether two articles are near-identical in wording, using a SimHash distance of at most 10 over headline and summary. That usually means the same press release rather than separate reporting. It is shown as a count, not as an accusation.

Where this is weak

Text similarity is a stopgap. Most feeds send only a headline, and independent reporting on the same event often shares almost no wording. We measured this: two articles about the same event can sit further apart in SimHash distance than two articles about nothing in common. That is why grouping does not use SimHash at all, and why the real fix is event extraction rather than better text matching.

What we store

Headline, at most 300 characters of the summary the publisher put in their own feed, the date and the link. Never a full article. Every headline links back to the publisher.

Sources

Vendor blogs, independent press and community sources are labelled separately everywhere, because a vendor is not an independent source about itself. Which feeds we read, when each last delivered, and which are switched off and why is public on the sources page — a source that quietly dies lowers the apparent attention around a tool, and you cannot check a number without knowing what fed it.

Scores

There are none yet, on purpose. A score built on three weeks of data would be confident and wrong. When they arrive they will be a weighted sum of measured facts — coverage, shipping cadence, reliability from status pages, transparency checks, adoption counters, and counted incidents — with the formula version and a hash of the inputs stored alongside every score, so an old score stays reproducible after the formula changes.