We Declared Our Data Source Exhausted. We Had Measured Our Own Query.

We re-ran a sourcing sweep and got 83% the same domains back, with twelve genuinely new ones. We wrote the source up as saturated. One day later, widening the keyword list from 20 terms to 65 — same registry, same script, nothing else changed — returned 5,063 domains at 44% overlap and 2,859 genuinely new. The source was never exhausted. Our query was.

What we were doing

We collect company domains from a public package registry, keep the ones that publish regularly, and check whether an AI assistant names them when asked their category's buying question. The collection step queries by keyword — twenty of them, deliberately paginated deep into the long tail rather than the popular head.

It had worked. Then we re-ran it, and the yield collapsed: 2,285 domains, of which 2,204 we already had. Twelve genuinely new candidates survived filtering. Three had a contact address that a mail server would confirm exists. Three is under one day of work.

The wrong conclusion, and why it was easy

We wrote it up as source exhaustion. The reasoning felt solid, and one part of it still is: a queryable registry is a set, not a stream. Running the same query again mostly re-enumerates what it matched the first time. Yield collapses on the second pass and keeps collapsing. Re-running an unchanged sweep on a schedule is close to worthless, and knowing that is worth something.

The error was attaching that fact to the wrong noun. We had measured the reach of one query design and drawn a conclusion about a registry containing millions of packages. A high overlap number cannot distinguish those two situations — it looks identical either way — and we had run exactly one sweep design.

What the widened sweep returned

Three verifiably live contacts became fifty-seven. One widened sweep produced more genuinely new domains than the entire previous sweep contained.

Adjacent, not synonymous

The widening only works if the new terms reach a different population. Adding synonyms matches the same items and re-proves the saturation you were trying to test — if your seeds are analytics and metrics, you have one slice, not two.

We added observability and delivery, data infrastructure, AI-native tooling, money and trust, identity and security, product surface. The test for an added seed is simple: do you expect it to match things the old seeds could not? If not, it is decoration.

The habit that caught it

The write-up declaring the source exhausted also contained this, in the same document, on the same day:

If widening the seed list to ~60 keywords returns a similar 80%+ overlap, the registry itself is exhausted rather than these seeds being narrow. Read it on overlap percentage, not on raw domain count — raw count is what made this morning's number look healthy.

Both halves earned their keep. The condition made the test worth running instead of leaving a plausible conclusion to harden into a fact. And naming the metric in advance mattered more than we expected: the second sweep returned 5,063 domains against 2,285, and a raw count that much larger is easy to read as an ordinary better day. Overlap is the number that actually distinguishes the two explanations, and we had committed to it before we knew which way it would fall.

We had spent the previous day designing a replacement source. None of that was needed. The fix was forty-five extra words in a list.

Method: npm registry search API, queried by keyword, paginated from result 200 to 1,000 to avoid the popularity head. Sweeps on 25 and 26 August 2026 with 20 and 65 seeds respectively. Overlap is computed on distinct non-repository company domains. "Verifiably live mailbox" means the receiving mail server accepted the address at RCPT while rejecting a random control address at the same domain.