Every list of companies arrives with a reason attached. The reason is usually a sentence — “expanding into EMEA”, “likely evaluating vendors”, “strong fit, recently funded”. The question nobody asks in the moment, because the list has already been paid for, is what the sentence is standing on.
Most of the time the answer is: another sentence. A model summarised a page, a person summarised the model, and by the time the claim reached the rep it had been paraphrased twice and dated never. It reads like evidence. It behaves like a rumour.
The three tests
A claim is usable when it survives all three of these. Not two. The combinations that fail are more instructive than the ones that pass.
1. It has a date
Not the date the list was generated — the date the underlying thing happened, or failing that, the date the source page was published or last changed. “Hiring three backend engineers” means something different if the roles went up last week versus last March. In the second case the roles are probably filled, the budget is spent, and the trigger you are calling about closed a quarter ago.
A claim without a date cannot decay, which means it will sit in the system forever, being wrong quietly.
2. The source can be opened
Not cited — opened. A URL that 404s, sits behind a login, or points at a data provider’s own record of the claim is not a source. It is a receipt for a purchase.
This is where most enrichment fails, and it fails invisibly, because nobody clicks. The test is cheap: take twenty rows, open twenty links, count how many render a page that contains the claim. Teams who run this for the first time are usually surprised, and the surprise is rarely pleasant.
A citation nobody can open is decoration. It exists to make the row look researched.
3. The link between them is stated
The weakest part of most signals is not the source and not the date — it is the unstated leap between what the source says and what the row claims. A page says the company opened a Dublin office. The row says “expanding into EMEA, likely to need multi-region support”. The first half is reported. The second half is invented, and it is the half the rep will repeat on the call.
Stating the link forces the leap into the open, where it can be judged. Sometimes it is a good leap. The point is that it should be visible rather than baked into a sentence that reads as fact.
What passes and what does not
| Claim | Verdict | Why |
|---|---|---|
| Posted a Head of RevOps role on 12 June, link to the posting | Passes | Dated, openable, and the claim is what the page literally says. |
| Raised a Series B (source: provider record) | Fails test 2 | The provider is reporting the round, not sourcing it. Find the announcement or the filing. |
| Growing quickly | Fails all three | No date, no source, and no stated relationship between anything and anything. |
| Migrating off a legacy stack (inferred from three job posts) | Passes as an inference | Only if the three posts are linked and the word 'inferred' survives to the reader. |
| CTO said on a podcast in Q1 that observability is a 2026 priority | Passes, but decaying | Real and dated, but a stated priority is weaker than a funded one and should age faster. |
Why this changes the list, not just the footnotes
Applying these tests does not produce the same list with better citations. It produces a shorter list. Companies that looked strong because three vague claims stacked up drop out when the three claims turn out to be one claim, undated, repeated.
That is the point. A list of 40 companies where every row can be opened is worth more than a list of 400 where the rep discovers the problem on the call — which is the worst possible place to discover it, because that is the moment the prospect learns what your research is actually made of.
- Pick twenty rows at random from your current list.
- Open every source link. Count how many load a page you can read.
- For each that loads, check the page actually contains the claim.
- Note the date on the page, not the date on the row.
- Anything that fails is not a lead you have — it is a lead you were told about.
Where the tests come from
None of this is novel. It is how a desk researcher, a due-diligence analyst or a journalist has always worked: claim, source, date, and an explicit note where the reasoning jumps. What is new is that generating a plausible sentence became free, so the volume of unsourced claims went up by orders of magnitude while the cost of checking them stayed the same.
The response is not to stop using models. It is to make the model show its working and to accept a shorter list as the price. That is what Leads does — it researches, and it publishes what it rejected and why alongside what it kept.
If you are building the filter that sits in front of this, the companion piece is how to write an ICP that returns companies worth calling. Evidence tells you whether a claim is true. The profile tells you whether a true claim is relevant.