Product update

A Major Upgrade to News Signals

We've made major improvements to how news signals are detected, verified, deduplicated, and delivered, by taking ownership of the whole pipeline end to end. Fewer duplicates, less noise, and more of the events that matter. With a lot more on the way.

Most news signals have the same problem: too many of them, too many repeats, too many wrong. A rep sees the same funding round three days running, a "new client" that's actually the vendor, a lawsuit solicitation dressed up as a business event, and after a week they stop trusting it. A signal you don't trust is worse than no signal at all.

So we invested in the pipeline itself. We now own it end to end, and went deep on the parts that most affect quality. Start with the hardest one at scale: when a major event breaks, it isn't reported once. It's reported thousands of times, across outlets and languages, over days. Your feed should still show you one clean signal, from the source most worth trusting.

One event, illustrated. Our pipeline resolves coverage like this across several thousand fresh articles every day, and it never turns off.

Every company, resolved to the real thing

Articles name companies in glancing, ambiguous ways. We pull out each one that matters and resolve it to its real domain. A few of the calls the pipeline makes:

And when one story involves several companies, we don't flatten it. We emit a distinct, correctly-directed signal for each. One article, two signals:

What else is better

You only see events that matter

An AI relevance layer reads every event and asks what a good rep would ask: is this a real business action worth acting on? Roughly 40% of raw events don't survive. Recalls, sports, politics, obituaries, law-firm solicitations, and events buried in an unrelated article never reach you.

Pointed at the right company

News is full of directional traps: who won the client versus who bought, who raised versus who invested, who has a problem versus who caused it. We get the direction right, so the signal points at the company you actually want to reach.

Correctly classified

A launch mislabeled a partnership breaks every play built on it. In our latest evaluation, events are typed correctly ~93% of the time, and the primary company is correct ~98%.

Fresh and transparent

The feed runs daily, each event dated to when it happened. And every delivery ships with a full record of what was kept, what was dropped, and why.

The quality bar we hold ourselves to

Most providers don't measure this, and almost none publish it. Now that the pipeline is ours, we grade our own output continuously using an independent AI evaluator, event by event, on the dimensions that determine whether a signal is usable. Here is where each one sits now, against where it was before this upgrade:

These aren't marketing numbers. They come from the same evaluation harness that gates every change we ship, which is how we keep improving the feed without quietly breaking it.

How it works

The reason the output is this clean is the architecture. An article passes through more than a dozen stages before it becomes a signal you can trust. Instead of one model doing everything, we assign each job to the model best suited to it, built on the most capable models available today from Google, OpenAI, and Anthropic, and wrap the whole thing in deterministic checks so the AI can't drift.

1. 1Collection. A collector reads company news across roughly 77 sources continuously, all day, every day. It never turns off. Each article's text is cleaned and repaired before anything else touches it.
2. 2Transcription. A frontier language model reads each article and pulls out structured events: what happened, to which companies, when, and a verbatim sentence of evidence. It is constrained to a fixed schema, so it can never invent a field.
3. 3Relevance judgment. A second model acts as an editor, scoring every event for whether it is real, how central it is to the article, and whether it is genuinely B2B-relevant. This is the layer that removes the noise.
4. 4Multi-company resolution. We resolve every company an article names to its real domain, using the article's own text first, then an 82-million-company reference graph, then web search. One story can produce several correctly-directed signals.
5. 5Source selection. When one event is reported thousands of times, we choose the single most authoritative source: earliest confirmed, most complete, closest to the primary.
6. 6Deduplication. Three layers. An exact-identity check guarantees no signal we have ever delivered is delivered again. A deterministic layer catches re-reports by matching facts. A final model compares each event against the prior ten days: same occurrence, or genuinely new?
7. 7Validation and audit. Nothing ships until it passes a strict contract check. After every run, an automated audit re-checks the whole delivery against dozens of quality rules, and anything anomalous raises an alert before it reaches you.

This is just the start

Owning the pipeline means we can move fast, and we're already building the next wave, shipping soon:

• More signal types. A wider set of event categories, so you catch more of the moments that matter to your business.
• Even sharper extraction. Continued gains in accuracy and relevance as we refine the models and our evaluation harness.
• Broader global coverage. More sources and better handling of international and non-English news, so the feed reflects the whole market.