Methodology
B2BDataBenchmarks publishes empirical, reproducible benchmarks of B2B data providers: email, phone, people, company, live profile, intent, jobs and technographic data. The rules below apply to every suite.
Frozen test sets
Every suite runs against a frozen, versioned test set (queries, companies or people) drawn before any provider is called. All providers answer the same set within the same time window, through their native APIs, mostly behind one gateway that forces a fresh answer on every call, with each query translated to the provider's own taxonomy as faithfully as its API allows. A provider whose price makes the full set too expensive runs on the first entries of the same set; it needs at least 20 attempted entries to be ranked.
Clean, verified inputs by design
Every test set is built from high-quality inputs that were checked before any provider was called: each person's name, title and employer are published by the employer's own website, and each company's facts come from registries and filings. This is deliberate: an equal benchmark needs equal inputs, and inputs we know to be correct. It also means the numbers here are a ceiling: on real-world data, with typos, stale CRM records, wrong domains and ambiguous names, every provider's match rate and accuracy will be lower. We measure the provider, not the noise in your input.
A common core, then suite-specific metrics
Every suite reports a common core of metrics so providers can be compared across data types: coverage (match rate on the frozen set), accuracy (against ground truth), freshness where the data carries a date, latency p50 (measured through the same gateway for every provider, so it includes that hop and compares providers rather than giving a vendor's absolute latency), pricing per useful result on the central pack of each provider's self-serve price ladder, and API quality (a rubric scored from public material: docs, spec, auth, rate limits, error semantics, SDKs, status page, sandbox and changelog), which the Email and Phone suites leave out. Providers that fetch live at request time carry a “live” badge. Each suite then adds only the metrics that are objective and reproducible for that data type: business-email share for email, mobile coverage for phone, headcount accuracy for company, predictive lift for intent, expired-job removal for jobs, and so on. A metric shown as “-” was not measured in that run.
Blind grading and objective truth
Where a judgment call is needed (do two differently worded job titles name the same role?), grading is blind: the judge sees the record, never the provider. Where objective truth exists (does this email verify? does the profile show the person's current employer?), we grade against it directly, with no judge to argue with. Empty results count against coverage, not accuracy; the two are reported separately.
Letter grades
Every measured value carries a letter, A to D. The letter is relative to the field, never to a target we invented: it says how many providers beat this one on that metric, in that run, with the direction taken from the metric itself, so a low latency grades as well as a high match rate. A is the top fifth of the field, B the rest of the upper half, C the lower half, D the bottom fifth. Ties share a grade. Review ratings are the one exception: they come from third-party platforms rather than from a run of ours, so they are graded against fixed bands (A at 4.6 and above, B from 4.2, C from 3.7, D below). A grade is a reading aid; the number beside it is the measurement. Leaderboards are ranked on the overall grade, the weighted mean of a provider's grades on every metric of the suite, so a provider that matches everything but returns poor data does not top the table; ties are broken on the headline metric.
Published aggregates
Raw provider responses hold personal data, so they are not published. The per-provider aggregates behind every number are, with the counts they were computed from, in the repository under runs/, one scores folder per run. Providers that could not run, or ran on fewer than 20 entries, are listed as not measured, never scored as zero.
Review aggregation
Review pages aggregate what real users say across public platforms (G2, Capterra, TrustRadius) and communities (Reddit, Hacker News). Every claim links to its source, summaries are paraphrased rather than copied, ratings belong to their platforms, and pages carry the date they were collected. Providers can request corrections at any time.
Independence
No paid placement, no affiliate links. The dataset is released under CC BY 4.0. Providers win some suites and lose others; publishing the losses is what makes the wins credible.