PeopleSearchBench: how Kretsia's candidate search scores

How does Kretsia score on PeopleSearchBench?

Kretsia scored 84.12 out of 100 Overall on PeopleSearchBench, the mean of three scoring runs on all 30 recruiting queries (standard deviation 1.15). That is 16 to 19 points ahead of the published scores for Lessie (68.23), Juicebox (65.73) and Exa (64.67), and 33.6 points ahead of Claude Code (50.50). On average, 13.3 of the 15 people Kretsia returned for each query met the brief. We ran it on 25 June 2026 with the benchmark's public, unmodified scoring harness.

Kretsia's numbers, run by run

We submitted 444 people for the 30 queries and had them scored three times with the unmodified benchmark harness. All figures are on a scale of 0 to 100, except the qualified count, which is out of 15.

Task completion was 30 of 30 queries in runs 1 and 2, and 29 of 30 in run 3. In each run, Overall is the mean of the three metrics; 84.12 is the mean of the three runs. Effective Coverage varied most between runs (SD 2.66), Information Utility least (SD 0.21).

What PeopleSearchBench measures

PeopleSearchBench (PSB) is a public benchmark for people-search tools, published by LessieAI in the open repository github.com/LessieAI/people-search-bench. We ran its recruiting category: 30 natural-language queries written the way a recruiter would phrase them, in English, Dutch, Portuguese and Spanish. Examples are backend developers in London with microservices experience, secretaries in Dutch government bodies, and strength and conditioning coaches in Portugal.

Each platform returns up to 15 people per query. The benchmark then scores them with what it calls Criteria-Grounded Verification. An LLM judge (google/gemini-3-flash-preview) pulls verifiable criteria out of each query, then checks every returned person against web evidence from Tavily. Three metrics come out, each from 0 to 100, and Overall is their mean.

1. Relevance Precision (84.34). Are the best matches at the top of the list? The benchmark calls it padded nDCG@10. For a recruiter it is the difference between the right person being first and being buried at number twelve.
2. Effective Coverage (87.25). The share of returned people who qualified. At 87.25, about 87 in every 100 people we returned met the brief. Per query, that is 13.3 of 15.
3. Information Utility (80.75). How complete and useful each record is. A result that gives you more to act on scores higher than a bare name with a job title.

Precision@10 is reported alongside. Ours was 89.26, which means roughly 9 of the first 10 people on each list qualified.

How we ran it

1. Date. We ran the benchmark on 25 June 2026.
2. All 30 queries. We submitted 444 people across the 30 queries, up to 15 per query.
3. Three scoring runs. We scored the submission three times with the unmodified harness. Overall came out at 84.85, 84.71 and 82.79. We report all three and the standard deviation (1.15), not just the best run.
4. Other platforms' scores. The scores for Lessie, Juicebox, Exa and Claude Code are the ones LessieAI published in the repository.
5. Candidate records stay private. The 444 returned records contain personal data about real people, so we don't publish them. The queries and the scoring harness are public, so anyone can run the same 30 queries against any platform, including Kretsia, and score the results.

Where Kretsia scored best

Several queries scored a perfect 100 on nDCG@10, which means the ranking was ideal. Among the strongest were:

• Backend developers in London (rec_0001)
• Product managers in Seattle for developer tools (rec_0002)
• Product designers in Amsterdam (rec_0014)
• Full-stack developers in Austin (rec_0019)
• Product marketing managers in San Francisco (rec_0023)
• Senior frontend developers in Stockholm with React and TypeScript (rec_0027)

These are the briefs recruiters write every day: a clear role in a named place, with skills that people list on their own profiles and sites.

What this means if you hire

On the benchmark's 30 recruiting briefs, a Kretsia list averaged about 13 people who met the brief out of every 15, and roughly 9 of the first 10 qualified. You start from a shortlist that already fits, and you decide who to contact.

To see how it does on your own roles, read how Kretsia's candidate sourcing works, or run the same kind of query yourself in a free trial.