Mining Buying Signals from Public Exhaust — Jobs, News, and Import Records

Intent Data You Don't Have to Buy

The intent-data industry sells inferences: this company visited these websites, so maybe they're shopping. Meanwhile, companies publish direct evidence of what they're investing in, continuously, for free. A manufacturer hiring three maintenance technicians and a plant manager is expanding production capacity — that's not an inference, that's a budget line with a job req attached. A company importing production equipment cleared customs with it; the manifest is public record. An acquisition announcement is a press release before it's ever a data product.

The catch is scale. Reading one company's careers page is trivial; watching hiring, news, and imports across 19,000 companies in a market — continuously, without an army of interns — is a pipeline problem. This article covers the signal architecture I built into the market intelligence platform, signal by signal, plus the one design rule that keeps the whole thing trustworthy.

Signal 1: Job Postings — and the Coverage Lesson

Hiring is the highest-value signal because job postings name the exact function receiving investment. The mechanics: for each tracked company, find the careers page, detect which applicant-tracking system it runs, and pull postings through that ATS's public endpoints — Greenhouse, Lever, Ashby, SmartRecruiters, Recruitee, Workday, UltiPro, ADP, Paylocity, BambooHR, Workable, Breezy. Postings are filtered against target-role keywords — for an industrial equipment seller, that's plant managers, production supervisors, maintenance techs, line operators — and matches write a hiring flag onto the company record with the posting attached as evidence.

The instructive part is what happened on the first live sweep: roughly 1,300 companies checked, and about half a percent showed open target-role postings. My first read was "hiring is a rare signal." The correct read was "my coverage is bad." Mid-market manufacturers don't cluster on the two or three trendy ATS platforms that tech companies use — they're on payroll-suite ATSes like ADP and Paylocity. Every adapter added found postings the previous sweep had marked as silence. The transferable rule: with detection pipelines, a low hit rate is a claim about your detector until proven otherwise. Measure per-adapter yield before you conclude anything about the world.

Signal 2: News, Classified by an LLM That Reads So Reps Don't

News monitoring uses the free tier of everything: a Google News RSS query per company, swept on a schedule. The raw feed is noisy — obituaries, local sports sponsorships, name collisions — so each company's results get one LLM call that does two jobs: is this actually about this company? and is it a growth signal or a risk signal? Expansions, acquisitions, product launches, new facilities on one side; recalls, closures, layoffs on the other. Classified headlines attach to the company record, so a rep opens a call already knowing the news. The store currently holds about 3,000 classified news items across 400-plus companies.

Two cost controls make it sustainable: only qualified-tier companies get swept — the same score-gating logic the enrichment waterfall uses — and it's one classification call per company per sweep, not per headline.

Signal 3: Import Records — the Signal Companies Can't Not Send

Customs manifests are public in the US, and they're the closest thing to ground truth about capital investment: a company that imported a production line has a production line. Aggregators like ImportYeti make shipment histories searchable by company — I've built an entire data pipeline on that source — and for signal purposes the query is simple: which tracked companies have recent equipment shipments, from whom, at what cadence? Unlike hiring and news, imports can't be un-published or spun. The company may not answer your email, but the manifest already told you they're tooling up.

Signal 4: The Negative Signal — Dead Domains

The quietest signal in the set: a company website that stops resolving. A monthly probe checks every tracked domain at the DNS, connection, and certificate level. A dead domain usually isn't a dead company — it's the fingerprint of a rebrand or an acquisition, which are buying-moment events. Failures get flagged for a repair job that hunts the successor domain and recovers the contacts that would otherwise rot. Data hygiene and signal detection turn out to be the same job wearing different hats.

The Design Rule: Signals Rank the Queue, Never the Score

Here's the architectural decision that matters more than any individual signal: signals never touch the lead score. The score measures durable fit — what the company makes, its size, whether your product belongs in its plants. Signals measure timing — is something happening right now. The two decay at completely different rates: fit changes over years, a hiring spike is stale in a quarter.

Mix them into one number and you get the worst of both: scores that thrash weekly as news cycles pass, no way to explain to a rep why a company dropped fifteen points, and a slow erosion of trust in the whole system. Keep them separate and each stays explainable — the score says whether to care about a company, the signals say when. In the platform, signals feed a daily "who do we call today, and why" queue where news-growth and open-hiring flags rank qualified companies — and every row shows its reason: the headline, the posting, the tier. The score sits beside it, unmoved.

That separation is also what keeps thin signals useful. A signal covering only a few hundred companies would be a terrible scoring input — most companies would be scored on absence of data. As a queue-ranker it's perfect: the companies it does fire on go to the top of today's list, and everyone else keeps their honest score.

Build Order, If You're Starting

Domain health first — it's a day of work and it protects the dataset everything else depends on. News second: RSS plus one LLM classification call is the cheapest real signal per unit of effort. Hiring third, and budget for many more ATS adapters than you think you need. Imports last — highest evidentiary value, but only worth wiring up once the companies you'd match against are deduplicated and canonical, or you'll attach ground truth to the wrong duplicate.

The whole approach — map the market, keep it fresh, read its signals — is what the AI Deep Search solutions page describes end to end, and the case study shows running in production.