Ask three data vendors what industry a company is in and you will routinely get three different answers - and all three can be wrong at once. Industry fields are the least reliable columns in B2B data, and campaigns targeted on them alone quietly leak money on every send.
Why the field is wrong so often
Industry labels are usually self-reported once, mapped to a taxonomy that forces one category onto companies that straddle several, and then left to age. A machining company that pivoted to medical devices four years ago is still "Industrial Machinery" in half the databases that list it. The label is not malicious - it is just old, coarse and unowned.
The company's website is the primary source
The company itself tells you what it sells, to whom, in its own words, on a page it maintains because customers read it. That is the evidence we qualify on. Our operating rule is that a vendor category can route a company into review, but only company-level evidence decides whether it enters a campaign:
- If the website is dead, parked, or contradicts the targeting thesis, the company does not enter the list - whatever the database says.
- When the website and the vendor field disagree, the vendor field loses. The record is reclassified or dropped, never sent on the strength of a label.
- Titles are matched to how that market actually names the role. The person who owns the budget is called something different in manufacturing than in SaaS.
When to use AI, and when to use rules
Reading a company description and deciding "does this match the ICP" is language work, and it is a job AI classification is genuinely good at. But hard gates should stay deterministic: a dead website is a dead website, a missing decision-maker is a missing decision-maker. We use models to interpret evidence and plain rules to enforce boundaries - and we never let the model invent evidence that is not on the page.
Failures are not answers
The quietest way bad data enters a campaign is through a failed lookup that nobody noticed: an API timed out, the record came back empty, and the pipeline shrugged and moved on. Our rule is that a failed enrichment is requeued for another pass - an error must never silently become a targeting decision. If you build lists at any scale, this one rule will save you more embarrassment than any copy improvement.
This is the unglamorous half of lead generation: not finding more companies, but refusing the wrong ones. Campaigns feel the difference immediately - fewer confused replies, fewer instant deletes, and a list where the copy actually has a chance.