The evidence layer behind Compass.
See what Compass reads, how articles become measurements, how fresh the corpus is, and where the limits are. The methodology is public because the output is meant to be cited.
A selected, source-linked technology press corpus.
Compass measures a chosen, dated corpus of the technology press — and that choice is the instrument. A curated corpus, held consistent and compared to itself over time, means something an unbounded pile can't. So this month's number and last year's number mean the same thing.
The sources. We read more than 140 of them — the major technology newsrooms, the specialist press, and the labs and platforms in their own words, always labeled as such. Together they cover the full sweep of the field: AI and machine learning, chips and hardware, cybersecurity, cloud, software, consumer tech, robotics, space, AR and VR, green tech — and the startups, deals, and policy moving underneath it all. A source earns its place by doing real reporting over a long stretch, not by landing one big story. We change the panel by deliberate review, never to chase a story. And you never have to take our word for it: the measurements open to the dated articles underneath them.
Two kinds of source make up that frame, and they never blur. Most of it is independent press — and that's what shapes how Compass reads the framing of a story. A smaller, clearly-labelled set of primary sources, like the AI labs' and platform vendors' own posts, grounds first-hand launches and results, but stays out of that framing read. Wherever a signal rests on one, you'll see it in the source trail.
The method. Compass reads its panel of sources on the same steady schedule, month after month. The discipline of a lab, applied to the news. Read that way, the numbers gain a property news rarely has: they can be set against one another over time, instead of one headline at a time. When a signal moves, that movement is a change in the selected coverage record — it may reflect real events, a shift in what the press chose to cover, which sources were reachable that month, or how stories were classified. The dated articles behind it are there so you can judge which.
Who changes it, and why. Compass is maintained as an instrument, so changes to the corpus or to how it is processed are deliberate, reviewed, and versioned — never made to favour a company, a topic, a viewpoint, or a number someone wanted. When processing changes in a way that would move a published figure, the change is dated and the affected measurements are rebuilt rather than quietly left to drift. Continuity is the point: a measurement is only worth citing if the same question, asked next quarter, is still being answered the same way.
From articles to signals, in four passes
Compass measures what the press actually published. The four passes below turn the article record into structured measurements.
- 1
The record, placed. Each article is tidied up and tagged to its topic threads and its timeframe — the scaffolding everything else hangs on.
- 2
The players, identified. The topics and actors in play — companies, technologies, products, and the people behind them — are surfaced and tied back to the articles that name them.
- 3
The names, resolved. Aliases are merged — so "Amazon Web Services" and "AWS" count as one company — and the citations in each article are linked back to what they cite.
- 4
Signals. From that structure Compass computes the measurements you see — monthly presence and velocity, persistence, emergence, and the connectedness score, which reflects how far a topic's coverage reaches across the field. The scores are arithmetic over that structured record — no editorial opinion enters at the scoring step.
Current, and always dated.
New coverage flows in continuously, and the measurements are rebuilt on a regular schedule to take it in. So the picture you get is a recent one — and the dashboards carry the moment they reflect, so you can always see how current your view is.
Every signal opens to its evidence.
Compass shows you what a number rests on. Open a measurement and you get the dated articles behind it; the reports also grade how much coverage a reading stands on, so a broad, well-supported signal reads differently from an early one. That is what makes the output citable — you can point a colleague at the evidence, not just at the number.
And Compass is careful about what a connection means: it marks two topics that turn up together far more often than chance would predict, and leaves the why for you to draw.
What the coverage percentages mean
You will see a few different percentages around the product. They are measured over the same set of articles but count different things, so they do not have to agree:
- Extraction coverage — of the articles read in a run, the share Compass identified at least one company or technology in. This is the ceiling on how much of the corpus a given run's signals could draw on. It is the figure shown on the Signal Map.
- Citation coverage — of the same articles, the share carrying at least one parsed source citation. Citation parsing is best-effort and older material is thinner, so this moves independently of extraction coverage.
- Evidence behind a figure — when we say every figure carries its source, we mean the drill-down above: open a measurement and you get the dated articles it was computed from. That is a property of the interface, not a percentage of the corpus.
That discipline is the whole point. Compass measures a real, dated, source-linked body of evidence — and stops at the measurement. The conclusion is yours to draw.