Lookalike API Benchmark:
10 Providers Compared on 48 Companies

Quick summary

The October 16 Openbenchmarks update covers 48 seed companies in 12 categories. It compares the relevance of the first ten results and, for endpoints that support it, the first hundred. CompanyEnrich leads both published rankings. PredictLeads is second at ten results (89.8%); Seltz is second at a hundred (63.4%).

The practical distinction is list depth. A product showing a handful of similar companies can use Precision@10 as its starting point. A workflow consuming a longer account list also needs to consider Precision@100. Neither score measures email quality, conversion rates or how many relevant companies a provider misses.

Relevance at two depths

Source: Openbenchmarks, October 16, 2026 update. Scores are means over each provider’s judged seeds (47 or 48). Bars use the same 0–100% scale.

Full provider comparison

Rows below are ordered by Precision@10. Percentages follow the source’s displayed leaderboard; the saved JSON includes more decimal places. “Not ranked” means an endpoint cap prevents a comparable top-100 result, not a score of zero.

All providers were attempted on 48 seeds. Ocean.io, Parallel, Discolike and CUFinder have 47 judged seeds in the source API. Averages are over judged seeds, so their denominators differ from providers with 48. A cap of 100 describes the requested depth in this test, not a provider’s maximum possible product limit.

How the benchmark measures lookalikes

Openbenchmarks defines relevance through two checks: the candidate must operate the same kind of business for the same kind of buyer, and it must share at least one product or service capability with the seed. A shared industry label alone is insufficient.

1. Fix the cohort. Use 48 named seed companies, each with a domain and business context, across 12 categories.
2. Request ranked results. Ask each provider for up to 100 similar companies using the input its interface accepts. Deduplicate before scoring.
3. Judge relevance. Apply a common rubric with a three-model majority-vote panel. The publisher identifies the panel as claude-opus-5, gpt-5.6-terra and kimi-k3.
4. Score each cutoff. Divide the relevant candidates within the first K positions by K, then average across that provider’s judged seeds.

Precision@K = relevant companies in the first K positions ÷ K

Example: 8 relevant results in the first 10 positions means 80% Precision@10. Returning only 8 results, all relevant, still means 80%.

An underfilled response therefore loses points. Result caps below 100 are handled separately: in the tested setups, PredictLeads and ZoomInfo returned at most 25 results per seed and CUFinder at most 10, so Openbenchmarks omits their Precision@100 ranking. Current documentation allows more per seed: up to 50 for PredictLeads and up to 100 for ZoomInfo.

Provider inputs are documented, but differ

Domain-based providers receive the seed domain. Exa and Parallel receive an authored natural-language query; Seltz uses its submitted query format. ZoomInfo is tested through its GTM CLI using the seed name. Its returned records have less descriptive context for the judge. These interface differences are part of the test.

For CompanyEnrich, the publisher reports POST /v2/companies/search in lookalike mode with pageSize: 100, lookalike.weight: 0 and lookalike.minScore: 0, excluding the seed with filter.domain.notIn. The minimum-score override allows the benchmark to request a deeper list. These results apply to that configuration.

See the publisher’s methodology for the complete scoring and query rules.

Where these results apply

• A defined cohort. Forty-eight seeds across 12 categories cannot represent every market, geography or niche. Check your own account segment.
• Model judgments. The scores reflect a shared relevance rubric and model panel, not a fully human-labeled ground truth. Ambiguous companies and limited response descriptions can affect labels.
• Similarity from one seed. The study does not evaluate multi-seed audiences, constrained firmographic search, enrichment accuracy, contact data or downstream sales outcomes.
• Different run dates. CompanyEnrich was added on October 16. The source changelog says existing providers were not rerun for that update; this is not a simultaneous October test of every provider.
• Incomplete judgment coverage. Four providers have 47 judged seeds. The published scores do not all average the same number of cells.
• No universal winner. Aggregate precision does not establish a win on every seed, statistical significance, or the best performance for every integration.

Sources and saved snapshot

The charts and tables above adapt the Openbenchmarks lookalike report and its public results API, checked October 28, 2026. The API reports a last update of October 16, 2026. Openbenchmarks licenses its benchmark data under CC BY 4.0. This page adds CompanyEnrich’s presentation and interpretation.

Download the dated results snapshot (JSON). It contains a relevance-only extract of the source’s provider and seed-level results, with attribution and source notes. It contains no new provider measurements.

What can be reproduced from the available files?

The publisher’s repository contains benchmark runners and judge code. At our source check, its file tree did not contain the data/ directory or manifest.json described in its README, and the README leaderboard still listed eight providers. We use the newer webpage and API for the ten-provider comparison.

The saved API snapshot lets readers inspect the aggregate and seed-level figures. It does not include candidate-level labels or raw provider and judge calls, so it cannot independently establish whether each relevance decision was correct. Reproducing the original judgments requires those artifacts; a fresh run requires API access and may yield different results.

Frequently asked questions

Which lookalike API ranked highest?

In the October 16, 2026 Openbenchmarks update, CompanyEnrich ranked first at Precision@10 (91.9%) and Precision@100 (73.5%). That is a result for this cohort and evaluation method, not a guarantee for every company or workflow.

How was relevance decided?

Three AI models judged each returned company against the seed, and the majority vote decided. A company counted as relevant only if it runs the same kind of business for the same kind of buyer and shares at least one product or service capability with the seed. Being in the same industry wasn’t enough on its own. Every provider was judged by the same panel and the same rubric. See how the benchmark measures lookalikes for the models used and the scoring formula.

Why are three providers missing from the top-100 ranking?

In the tested setups, their endpoints returned at most 10 or 25 results per seed. They are compared at Precision@10, where they can fill the cutoff, rather than assigned a misleading top-100 score. PredictLeads now documents up to 50 results per seed and ZoomInfo up to 100.

Are Exa, Parallel and Seltz lookalike APIs?

Not in the same sense. Exa, Parallel and Seltz are web search APIs with company search modes. Those modes take a written description of the companies you want rather than a seed domain, so in this benchmark each one received a query written for the seed. The query becomes part of the result: describe the same company differently and the list can change. Exa’s older findSimilar endpoint accepts a URL, but Exa has deprecated it in favor of descriptive queries.

In a pipeline or an AI agent, that adds a step before every lookup. Someone, or a model, has to research the seed company and write a good query first. A domain-based lookalike API takes the domain your CRM or agent already has and returns a ranked list in one call, with no query to write and no web pages to read through.

Who ran the benchmark?

Openbenchmarks Labs publishes the original measurements and method. CompanyEnrich is one of the evaluated providers and authored this analysis. The source data and our commentary are linked separately above.

Test lookalikes on your own accounts

Start with companies your workflow already understands. Compare the first ten results and the full list you plan to consume, using the same relevance criteria across providers.

Explore the Similar Companies API and run the same queries against your own target accounts.