Public company signals are evidence, not purchase intent

01Do not turn correlation into a fact

A company posting data-engineering roles may be investing in data infrastructure. It does not prove that the company needs your service, has a budget or plans to buy soon. A technology mention in a job description can describe an existing stack, a future preference or a broad hiring requirement.

Use public signals to rank research work, not to manufacture certainty. Name the observation precisely: a role was published, a location was announced or a product page changed. Then record the hypothesis it suggests and what evidence would confirm or contradict it.

02Represent the signal as a dated event

Store the organisation identity, event type, source URL, source date where available and observation time. These dates are different: you may discover a six-month-old announcement today. Without that distinction, an old expansion story can repeatedly look like a new opportunity.

{
  "event": "data_engineering_role_observed",
  "company_id": "example-001",
  "observed_at": "2026-10-11T09:00:00Z",
  "hypothesis": "may_be_expanding_data_team",
  "purchase_intent": "not_established"
}

03Define freshness and repetition rules

A reposted vacancy should not necessarily count as another hiring event. A syndicated press release should not become multiple independent signals. Deduplicate by the underlying event, retain the source history and define when an updated page represents a materially new observation.

Set expiry rules appropriate to the signal. A current open role may deserve frequent rechecking, while a historical acquisition remains relevant for a different reason. Keep expired signals in history where permitted, but remove them from a current-priority score unless the scoring rule explicitly needs them.

04Make scoring reviewable

Use a small set of understandable criteria before building a complex score. For example, direct product overlap, a relevant dated announcement and a current operational need can be separate dimensions. Do not add unsupported precision by assigning a numerical probability without calibration data.

Review a sample of high-scoring and low-scoring companies. Look for directory pages, reposted news and generic job descriptions that inflate results. Track whether reviewers agree with the signal interpretation, not merely whether the model can reproduce its own previous answer.

05Keep downstream actions proportionate

A research signal can create an internal review task or update a company note. Outreach needs its own eligibility and privacy checks. Avoid collecting personal details unrelated to the stated purpose, and do not infer sensitive traits about employees from company-level observations.

The final output should let a colleague answer three questions: what happened, where was it observed and why might it matter? If those are clear, the dataset can support useful judgement without pretending to read a company's private plans.