Build a company research dataset, not a contact dump
Start with company fit, evidence and freshness before deciding which organisations deserve further investigation.
In this article
THE KEY IDEA
A useful company record explains why the organisation fits your criteria. More email addresses do not make that judgement more reliable.
01Define the company-level question
Start with a research question such as whether a company operates a marketplace, sells a specific service or maintains a relevant technology integration. These are observable business characteristics that can be tied to evidence. A vague request for good leads encourages inconsistent labels and unnecessary personal-data collection.
Document the target geography, customer segment and exclusions. An agency offering a service to others is different from a company using that service internally. A company with a blog post about a technology is not necessarily a customer of that technology. Those distinctions belong in the classification rules.
02Keep identity separate from evidence
Use a stable internal company ID and preserve the submitted domain, resolved domain and organisation name. Redirects, acquisitions and regional domains can make a one-domain-one-company assumption unreliable. Queue uncertain identity matches for review before merging records.
Store each fit signal with its own source and date. A careers page may support a hiring signal; a product page may support an offering; a case study may support a relationship at a particular time. Do not collapse these into a single unexplained score.
| Field | Purpose |
|---|---|
| company_id | Join results without depending on a changing name |
| criterion | Explain the question being tested |
| decision | Use a defined label, including uncertainty |
| evidence_url | Let a reviewer verify the decision |
| observed_at | Show when the evidence was collected |
03Collect the minimum information needed
Company research does not automatically require personal phone numbers, individual profiles or email addresses. Start with organisation-level facts and public business channels relevant to the task. Review the applicable privacy and outreach rules before adding personal contact data or using any dataset for communication.
Separate collection from activation. A system that assembles a research queue should not silently send messages, enrich private records or infer sensitive personal characteristics. Each downstream use can introduce new obligations and risks beyond the initial research purpose.
04Score only after you can explain the inputs
If you use a fit score, define its components and retain the underlying observations. An explicit overlapping service may carry more weight than a generic keyword. Missing information should lower certainty, not automatically become a negative fact. Keep uncertainty visible in both the score and any exported label.
Evaluate a sample manually. Inspect false positives such as directories, resellers and content-only sites. Include companies outside the target segment so you can see whether the classifier is merely finding related words rather than answering the actual business question.
05Refresh the decision when the evidence changes
An old product page can remain indexed after a service has been discontinued. Store the last successful observation and the next review date. When a source disappears, mark the evidence stale rather than deleting the company's history or asserting that the business has closed. Deliver a reviewable research dataset with clear limits, not a promise that every row is ready for outreach.
References & further reading
Put it into practice.
Try a public-page workflow, or scope a maintained dataset with our team.