Revenue BenchMethods
Revenue Bench scores every vendor against answers we already knew, taken from real records. Each score has a likely range, and vendors our data cannot separate share a tier. Fedor Kovalev wrote this page.
- Known outcomesBounces, replies, departures
- Test recordsAddresses, people, companies
- Same records to every vendor
- Compare to the known answer
- Score with a likely range
- Group into tiers
Where the right answers come from
They come from a B2B SaaS company's outbound and CRM records, used with its CEO's permission. We do not name the company, publish no individual records, and keep the answer keys hidden so no vendor can tune to them. We built the keys on 30 October 2026 and dropped any record that was not one clear person or one clear outcome.
| Signal | What it shows |
|---|---|
| A real message from a work address in the past 90 days | The address is live |
| A bounce of an address that had written before | The address is dead |
| A reply saying the person left, or a bounce after a real message | A job change |
How each job is scored
| Job | Test set | Headline score |
|---|---|---|
| Email verification | 172 addresses, 82 dead and 90 deliverable | Bounces caught plus good emails kept, minus 1 (Youden's J), out of 100. 0 is a coin flip. |
| Work email finding | 200 people | Right only on the exact known address. An alias-tolerant score sits beside it. |
| Job change detection | 113 people, 13 who left and 100 who stayed | Share of leavers flagged, shown with the overall score. |
| People search | 100 companies with 127 known people | Share of known people returned. |
| AI judges | 702 questions, asked three times each | Share right on the first answer. |
More scoring rules
- Email verification: an unknown verdict counts as neither caught nor kept. Findymail's "no" counts as unknown, not invalid.
- Finding: another address at the same company counts as wrong, because we cannot tell a valid alias from a wrong guess. A call that errors or returns no address counts as none.
- Job change: a source that returns only a company name gets a lower-bound score, because our name matcher misses some same-company names. A record we could not settle reads as unknown.
- People search: non-English learning and development titles count as on-role. A judge model decides whether the other people returned really work there. It confirmed 104 of the 112 known people it could decide on.
- AI judges: a failed call counts as wrong.
- Some vendors ran on fewer records, so their ranges are wider. A finder priced above $0.05 a call ran on a frozen subset of 50 people. A finder that takes only a LinkedIn URL ran on 131 people, or 33 for some. Each vendor's record count is in its data.
Likely ranges and tiers
A few hundred records cannot give an exact score, so every rate comes with a likely range. When a head-to-head test on the same records cannot tell two vendors apart, they share a tier. Heavily overlapping ranges are the usual sign. A wide tier means thin data, not equal vendors.
Illustration, not real data.
The statistics
- Ranges are Wilson 95% intervals, which stay between 0 and 100% and behave well on small samples. The overall score combines two rates, so its range uses the Newcombe method.
- We compare two vendors on the records both answered, with the exact McNemar test, which counts only the records where one vendor is right and the other wrong. We apply a Holm correction across every pair we print, at a threshold of 0.05.
- A vendor is in the top tier when no other vendor beats it and the best vendor does not beat it. We remove the top tier and repeat for the second.
- The Holm correction is conservative, so the top tier can be wide. In this release it holds 7 verifiers, 17 finders, 26 job change sources and 11 people search sources.
- Job change detection has only 13 leavers, too few for a paired test, so its ranges are wide, its board carries a small-sample badge, and two vendors differ only where their overall-score ranges do not overlap.
Accuracy, cost and speed
We do not blend them into one score. Cost is the price per 1,000 calls divided by the share of calls that were right. Speed is the median answer time in seconds, shown only where the store reports it. Prices are list prices on the run date, and Clay credits are priced at Clay's Growth list rate.
Vendors we could not test fairly
We left out 15 vendors we could not test fairly. Three show a risk of buying through a store, whose one shared account can run out or be blocked, and the rest needed their own account, cost too much per company or did not fit our test. The full list with reasons is in the data download, and being on it says nothing about a vendor's quality.
The AI judges report
This report tests 34 models on four questions with known answers. We froze the question wording before any paid call. The first of three answers is the score. The other two measure stability, which is the share of records where the three answers differ.
- Safe hand-off is the share of records a model settles by itself with at most 1 mistake in 100, at its own best confidence cutoff.
- For duplicate and same-person questions a record counts only when the model says yes, because a wrong merge costs the most. For own work email and main employer questions it counts when the model gives a top pick.
- We leave a model with over 20% unsettled records out of the paired tests. Those use the exact McNemar test with a Holm correction across all 40 pairs among the top five models on each job.
- Some rows are free rules we found by looking at the answer key, so their scores run high. They show what the test records give away, not what to deploy.
- The employer job has 102 records, so its likely ranges are about 8 points either way.
What we do not claim
- Results come from one company's records, skewed to learning and development buyers. Yours may differ.
- The verification sample is about half dead by design, so coverage depends on that mix. The overall score and per-class rates do not.
- The keys come from replies and bounces, so we cannot score an address that never replied or bounced.
- The people search key is a partial list of those who wrote to the company in the 90 days to 30 October 2026. Recall is against that list, not the full team.
- Job change detection rests on 13 leavers, and three records marked "there" may be wrong in the key.
- A tier is a statement about our data, not a ranking guarantee.
- We have no commercial tie to any vendor or store, and the results are not advice for any single purchase.
Versions and corrections
Each release is a dated snapshot. Contact Data, Nov 2026 was released on 10 November 2026 with method version 1.0 and test code commit 5406627. It covers 83 vendors through 7 stores, 15,520 calls and $311.69 of spend. The AI judges report ran on 6 November 2026 and cost $8.53 in model calls.
When we change a number, we publish a new version and say what changed. Send corrections to [email protected]. The about page says what happens next.