The counts, market by market
One research run, dated 1 October 2026, across the four markets QuikSignal researches in the open. "Documents read" is what the agents actually fetched and parsed; "worth printing" is what survived scoring, deduplication and the ninety-day comparison.
- Fintech and payments: 3 companies, 182 documents read, 6 worth printing - 3.3%.
- Cybersecurity: 7 companies, 296 documents read, 16 worth printing - 5.4%.
- B2B SaaS: 4 companies, 214 documents read, 14 worth printing - 6.5%.
- AI infrastructure: 5 companies, 174 documents read, 14 worth printing - 8.0%.
- All four: 19 companies, 866 documents read, 50 worth printing - 5.8%.
Why the ratio is the interesting number
A research tool is usually sold on how much it reads. That is the wrong half of the equation, because reading is the cheap part and getting cheaper every year. The expensive part is deciding what not to tell somebody.
At 5.8%, roughly nineteen out of every twenty documents a careful process reads turn out not to be worth a reader’s attention. That is not a failure of the sources - most of those 866 documents were perfectly good documents. It is what "relevant" actually means when the question is narrow and the reader has ten minutes.
It also sets a floor for anybody doing this by hand. If a person wants the same coverage of nineteen companies, they are reading 866 documents to write fifty lines. Nobody does that weekly, which is why the work usually does not get done rather than getting done badly.
What it says about the "more sources" pitch
The stability of the ratio across four unrelated markets is the part worth arguing about. Fintech ran at 3.3% and AI infrastructure at 8.0% - a factor of two and a half, but both in the same order of magnitude, from different source mixes and different company sizes.
If that holds more widely, then doubling the sources roughly doubles the reading and roughly doubles the output, and the thing that changes a reader’s week is not the source count but the quality of the filter. A tool that reads twice as much and shows you twice as much has not helped.
Method, stated so it can be disputed
Single run, dated 1 October 2026. Nineteen companies across four markets, chosen when the open index was set up rather than for this note. Sources: the companies’ own careers boards (Greenhouse, Ashby), two dated captures of their own pricing pages, SEC EDGAR and the US Federal Register.
A document counts as read when it was fetched and parsed. It counts as worth printing when it survived relevance scoring, near-duplicate collapsing and comparison against the preceding ninety days. The suppression log on each market page lists what was held back and why, so the judgement is reviewable rather than taken on trust.
- It is one week. A single run cannot establish a trend, and this note does not claim one.
- Nineteen companies is a small cohort, chosen for the open index rather than sampled.
- The four markets are technology markets. A regulated manufacturing market with a different source mix could behave differently.
- The ratio is a property of this method and this question. A different relevance threshold would produce a different number, which is why the method is stated above.
How to check it
Open the four market pages in the AI Research Index. Each states its own date, company count, documents read and entries printed in the first lines. Add the four together and you have the table above.
The figures move every week, because the index is re-researched every week. This note is a snapshot of the run dated 1 October 2026, and it says so in its own title line for exactly that reason.