Recruiting Sourcing Benchmark 2026: Method and Results

The Recruiting Sourcing Benchmark 2026 documents two tests of how well sourcing tools find suitable candidates; Atlas was measured by Sprad itself in both. In our own coverage test (own measurement by Sprad, as of 15 August 2026, method on the sourcing benchmark page), Atlas found 38 % of the candidates that four other systems found; the best of those four found 9 %. In the recruiting area of the open PeopleSearchBench, four systems have published scores (Lessie 68.2, Juicebox 65.7, Exa 64.7, Claude Code 50.5). We ran Atlas through the same tasks ourselves; converted to the published scale that is about 73.5, our own re-measurement and not an entry in the benchmark's table. This page documents both tests: question, method, results, limits and how to repeat them.

Sprad builds Atlas, one of the tested systems. That is a conflict of interest, which is why this page shows how we measured, what we did not measure and where the numbers stop being reliable.

What question does the benchmark answer?

It answers one buying question: if I describe a hard-to-fill role, how many of the right people does a sourcing tool put in front of me? Recruiters often judge tools by how fast results appear or by how many profiles a database holds. Neither says whether the people returned actually fit the role. The two tests below look at fit and at coverage instead.

Test 1: PeopleSearchBench (Recruiting)

PeopleSearchBench is an open benchmark for people search. It contains real search tasks in four areas: recruiting, B2B, expert and influencer. This page only covers the recruiting area, because that is the one that matters for hiring. It has 30 searches, for example a backend developer in London with microservices experience.

Method

• Task: each system gets the search task and returns up to 15 people.
• Rating: a language model breaks each task into checkable criteria. A second model checks every returned person against every criterion with a web search: met counts 1, partly met 0.5, not met 0.
• Score: three values from 0 to 100, relevance (good matches ranked higher count more, measured as nDCG), coverage (how many of the 15 profiles match) and usefulness of the data (completeness and whether you can contact the person). The overall score is their mean.
• Who measured whom: the benchmark was published by Lessie, and the scores of the other four systems are the ones published there. Atlas is not in that table. We ran Atlas through the benchmark's tasks ourselves, with the same prompts and formulas.
• A different judge: the original benchmark rates with Gemini 3 Flash and a web search. We rated Atlas with Claude Haiku and a web search, one search per person. This judge is stricter than the original: in a calibration on the same lists it scored lower. Because of that, the raw Atlas value cannot be placed next to the published values.
• Control run and conversion: we ran Exa with the same judge and the same criteria. In that run (29 October 2026) Atlas measured 80.4 and Exa 73.4 in recruiting. The Atlas value we report, 83.4, dates from 30 October and includes the verified e-mails described under Setup, which Exa's lists did not get. Our Exa came out higher than the published Exa (64.7) although the judge was stricter in the calibration, so the runs differ in more than the judge, and we cannot fully explain that gap. We scaled the Atlas value by the ratio of the two Exa values. That gives about 73.5. It is our own re-measurement, converted to the published scale, not a measured value. Only the comparison with Exa uses the same judge and criteria; every comparison with Lessie, Juicebox and Claude Code rests on the conversion.
• Setup: a local test environment, October 2026. The Atlas lists include a verified business e-mail where the "find e-mail" function returned one for the same person, as a customer would get it by clicking that button. It is not shown in the result list by default.

Result

The Atlas row is not a ranking place. It is an estimate on the published scale, about 5.3 points above the highest published value; with a different judge and a conversion, that gap is too small to call a win on its own. Only the comparison with Exa uses the same judge and criteria (Atlas 80.4 in the same run, 83.4 with verified e-mails; Exa 73.4). On average, 14.2 of the up to 15 profiles per recruiting search counted as matching, in our rating with the stricter judge. Source: PeopleSearchBench (Recruiting), published by Lessie, October 2026. Other systems' scores as published in the benchmark; Atlas measured by Sprad with the benchmark's test procedure and a different judge model, converted as described above. This is not an official entry in the benchmark.

Test 2: our own coverage test of 15 August 2026

The second test asks a different question: of all people that competing tools find for the same role, how many does a system find too? It is our own measurement, run by Sprad as of 15 August 2026, and it is deliberately small.

Method

• Roles: three hard-to-fill roles in three markets, each with the same requirements for every system: a Senior Backend Engineer in Berlin (Go, Kubernetes, five years or more), a Senior Product Manager in London (payments) and a Process Integration Engineer in Eindhoven (semiconductors).
• Systems: Atlas, Juicebox, Metaview, PIN and LinkedIn Sales Navigator. We ran and rated all of them ourselves.
• Input: every system that takes free text got the word-for-word identical prompt. LinkedIn Sales Navigator has no free-text search, so the same criteria were translated into its filters.
• Reference set: for each role, everything the other systems returned, deduplicated. A system is never measured against its own output. Hits were matched on first and last name, accent-free.
• Counting: the documented list contains 218 profiles, each with source and rating. Rating followed one scheme for all systems: strong, usable, weak or miss, based on the must-have criteria of each role.

Result

Read one row like this: of the people the other systems found for role A, Atlas also found 30 percent. Overall, Atlas covers 38 percent and the best of the other four covers 9 percent, about four times as much. Source: own measurement by Sprad, 15 August 2026; method on sprad.io/sourcing-benchmark.

Why are the numbers so low?

Because the tools barely overlap. On role A, 76 people were found in total, and 72 of them by exactly one system. Two of the other systems never share more than a fifth of their hits. Atlas shares up to 40, 47 and 52 percent of its hits with individual systems, so it sits in the middle of the field while the others fish at the edges. On role C, LinkedIn found 18 of its 19 hits at a single employer, while Atlas and Juicebox returned the supplier ecosystem around it. Both answers are defensible, they are simply different slices of the same labour market.

The practical reading: most of the market is found by no single tool. Even the strongest result in our test leaves most people unfound.

What the benchmark cannot say

• Coverage is measured against the field, not the market. The reference set is what the four other systems found. A person no system found does not appear in these numbers.
• Atlas was tuned on this test. The 38 % is the value after several rounds in which we changed Atlas's search against these same three roles and this reference set. It is therefore not a blind result. The other systems were run as they are.
• The counts differ between our documents. The reference sets are 76, 73 and 54 people in the run's report. The consolidated list of all 218 profiles contains 76, 72 and 70 entries per role. We could not fully reconcile the two. The percentages come from the run's reference sets, so read them as approximate.
• Atlas's own finds were not measured here. The test shows how much of the others' finds Atlas also finds, not how good everything else Atlas returns is.
• One run, three roles, one day. Our test ran on 15 August 2026. Search indexes and rankings change, so a rerun will not reproduce the values to the percentage point. Other role families may behave differently.
• Different input for LinkedIn. Sales Navigator has no free-text search. Translating criteria into filters is faithful but not identical.
• We are a vendor. Atlas is our product. In PeopleSearchBench the benchmark's publisher is a competitor, and we measured Atlas ourselves with a different judge. Treat both tests as evidence, not as an independent audit.
• Recruiting area only. We report the recruiting area here because it is the one relevant for hiring.

Can the test be repeated?

Yes, and we want it to be. Everything a repetition needs is on this page or on sprad.io/sourcing-benchmark:

• Same tasks: for the coverage test the three roles and their requirements (see Method), for PeopleSearchBench the 30 recruiting searches of the open benchmark.
• Same input: word-for-word identical prompts for every system that takes free text; criteria translated into filters where it does not.
• Reference set without own finds: collect what the other systems return, remove duplicates, match on first and last name without accents, never measure a system against its own output.
• One rating scheme: strong, usable, weak or miss against the must-have criteria of the role; in PeopleSearchBench met 1, partly 0.5, not met 0 with one web search per person.
• State the judge and the date: a different judge model gives different absolute values, so name it, and calibrate on the same lists before you compare across runs.

What we do not publish are the names of the 218 profiles, because they are personal data. We plan to repeat our coverage test every year with new roles, without tuning Atlas on them beforehand, and to update this page with the new date. Until then, the values above are a snapshot from 2026.

How do I cite this benchmark?

Sprad (2026): Recruiting Sourcing Benchmark 2026: Method and Results. sprad.io/blog/recruiting-sourcing-benchmark-2026, as of November 2026. Underlying tests: PeopleSearchBench (Recruiting), published by Lessie, October 2026; Sprad coverage test, 15 August 2026.

The summary page with the charts is sprad.io/sourcing-benchmark. If you quote a number, please quote the test it comes from and its date.

What does this mean for choosing a sourcing tool?

• Test on your own hardest role. A benchmark is three roles on one day. Your open position is the only test that counts. Active sourcing with Atlas starts free with 100 credits a month, see pricing.
• Do not judge by speed. If results appear instantly, it is usually a keyword search. Checking every profile against your requirements takes longer. Atlas shows first results in about three seconds with Quick Search, the deep search takes a few minutes.
• Expect different tools to show different people. If one search must not miss anyone, combine sources, and keep your own talent pool in the search so that people you already know are found first.
• Know which kind of search you are buying. Boolean strings and requirement search find different people, see Boolean search vs. requirement search.

Frequently asked questions

How does this differ from other sourcing comparisons?

Most published sourcing rankings are written by a vendor about its own product and give no method. Ours is also vendor-made, so we state who measured what, with which judge, what was converted and where Atlas was tuned. If you spot a step that is unclear, you can ask us for it.

Which tool is best for recruiting sourcing?

No single test answers that. In PeopleSearchBench (Recruiting) the published scores run from 50.5 to 68.2 for four systems; our own re-measurement puts Atlas at about 73.5 on that scale, converted and not an official entry. Only the comparison with Exa uses the same judge and criteria, so only that comparison is like for like. Which tool is best for you depends on your roles, markets and channels, so test two or three on a role you know well.

What does 38 % versus 9 % mean?

It is the share of the other systems' finds that each tool also found in our own measurement of 15 August 2026. Atlas found 38 percent, the best of the other four, Juicebox, found 9 percent. It is a coverage figure within three roles, not a statement about the whole market.

Is the benchmark independent?

No, not fully. PeopleSearchBench is published by Lessie, a competitor, and the other systems' scores are as published there. We measured Atlas ourselves, with a different and stricter judge model, and converted the value to the published scale. The coverage test is our own, and Atlas was tuned on it. Both are described above so that you can judge them.

Why does Atlas not cover most of the market?

Because no tool does. The four competing systems barely overlap with each other: on role A, 72 of the 76 people found came from exactly one system. People whom no system found do not appear in the test at all. That is the honest reading of the coverage result.