Enrichment Desk

Explainer / Judging data

Match rate, fill rate and freshness: how to judge an enrichment API

Understand match rate, fill rate and freshness, then compare enrichment APIs by cost per usable company record for your agent.

Enrichment Desk · · Updated · 6 min readShare

Ask which inputs the vendor counted before comparing enrichment match rates. For your own decision, keep the list and acceptance rules fixed, then compare cost per enriched record that is correct, complete enough and recent enough for the agent's job.

What does a match rate actually count?

Match rate is matched input records divided by input records, under a stated rule. A percentage calculated from a screened list answers a different question from one calculated from everything submitted. Ask whether rejected inputs, unknown companies and unsuccessful requests were included before treating two percentages as comparable.

For your own comparison, we recommend keeping every unique submitted company in the denominator. Give rejects, unmatched records and timeouts separate counts, but leave them visible in the total. Otherwise, removing difficult inputs can improve the apparent rate without finding another company.

A returned candidate still needs an identity check. People Data Labs' company enrichment parameters let the caller change the minimum confidence needed for a match. People Data Labs (PDL) describes that choice as a trade-off between precision and recall: a stricter setting can return fewer, stronger candidates.

Its required-field setting also changes which responses qualify as matches and charges. PDL says a field supplied as an input cannot also be enforced through that setting. Record the requested fields and matching settings alongside the denominator; do not assume different vendors' confidence scores mean the same thing.

What does a filled field tell you?

Fill rate describes how often a particular field has a value across a named population. For an office-location lookup, that population might be every company submitted or only the companies matched. The latter fraction can look better while leaving the same companies without a location.

A field can be populated and still be wrong. The UK's Government Data Quality Framework distinguishes completeness from accuracy: having values is different from having values that reflect reality. A workable acceptance rule rejects empty strings and placeholders, then checks the values against the company's reference information.

Keep per-field fill separate from the share of records containing all the fields the job requires. A generous response full of optional fields cannot compensate for a missing field the agent needs to make its decision.

How fresh does the data need to be?

Freshness is the recency of the evidence relevant to a value, judged against its intended use. The framework defines timeliness in relation to that use, including the delay between collecting information and making it available. Choose the window for the decision, rather than accepting one blanket definition of fresh.

A recent response is not proof of a recent observation. PDL's company schema defines dataset_version as the data release identifier. Its office-location last_seen field records the latest observation linking a location to a company; that date differs from the release version.

Ask what each date measures and which field it covers. Treat last_seen as evidence of that recorded association, not verification that the company currently operates there. Keep undated values marked unknown, and report how much of the required population has dated evidence; excluding unknowns can inflate the fresh share without improving the data.

What should your company panel contain?

Use a fixed panel of public companies you already know. Our recommendation is to include the sizes, countries and sectors your workload encounters, with difficult cases alongside familiar brands. This gives you a diagnostic comparison for your job, rather than a market-wide ranking.

Before asking providers for data, freeze the company list, duplicate handling, available inputs, requested fields and acceptance rules. Set the freshness window at the same time. Give each provider the information the agent would actually have, rather than richer identifiers added just to rescue a disappointing result.

Save a reference for each company: its known domain and the required values from its own public pages, with the date you checked them. Decide whether the intended identity is a branch, subsidiary or parent, so a plausible name cannot pass for the wrong business.

When you later run the panel, check the returned identity and values against those references. Preserve the original inputs, matching settings, dated outputs and actual charges. Report counts by relevant segment as well as the overall total, with unknowns and failed requests still visible.

How does the denominator change the buying decision?

The same panel can produce several impressive percentages without yielding enough usable records. Keep the underlying counts beside each fraction so the buying decision survives a change in how it is presented.

Hypothetical arithmetic, not vendor results or a benchmark. Assume all returned identities are confirmed and office location is the only required field. An accepted record also needs observation evidence inside the chosen freshness window.

Measure Numerator Denominator Hypothetical result
Match rate 40 matched companies 50 unique submitted companies 80%
Office-location fill rate 30 usable locations 50 submitted companies 60%
Documented freshness share 18 locations observed within 30 days 30 populated locations: 2 older, 10 undated 60%
Cost per accepted enriched record $10 total spend 18 qualifying companies $0.56 per record, rounded

Counting fill only among matches would raise that fraction. Counting freshness only among dated values would raise another. Neither change supplies a missing location or dates an unknown one. Under this example's acceptance rule, an undated location does not qualify; that is a purchasing rule, not proof that the value is stale.

Can an agent pay at the request instead?

ZeroClick, which publishes this site, builds machine-readable agent storefronts over sellers' existing services. An agent can pay for an individual request by answering a priced HTTP 402 challenge and repeating the request with proof. Each call carries an agent credential the agent registers for itself in one free call, with no human step. Advertised options can include USDC through x402 or MPP, a card the agent is authorised to use via an MPP shared payment token, or human approval through Stripe-hosted checkout with a card, Link, Apple Pay or a bank account.

What should go into cost per enriched record?

Divide total spending attributable to the work by unique company records that meet the identity, required-field and recency rules. Count each company once, even when several attempts are needed. With no qualifying records, the cost per accepted record is undefined, not zero.

Include billed failures, retries, extra field charges and the relevant committed spending, including unused allowances. Compare lookup charges, paid matches, credits, seats and plans over the same workload and billing period. When weighing a seat or plan, check that it grants the API access and usage your agent needs; a lower lookup price is only useful if the records qualify.

Agents can also purchase plans and prepaid credits against a durable credential. Card minimums can make these more practical for tiny calls. The payment route changes how the agent buys, not whether the resulting company data passes your rules.

Before buying, freeze the panel, denominator and acceptance criteria. Then compare the bill for qualifying company records, rather than the price of asking for them.

Frequently asked questions

What is a good match rate for an enrichment API?

Judge a match rate against the inputs and acceptance rules for your workload, rather than a universal target. Returned candidates still need to be checked for correct identity.

What is the difference between match rate and fill rate?

Match rate counts matched input records over a stated input population. Field fill rate counts usable values for a particular field over a stated population, which may include all inputs or only matches.

Does a new dataset version mean the data is fresh?

A release identifier does not establish when each field was last observed. Check the relevant observation dates and keep unknown dates visible in the comparison.

How do you calculate cost per enriched record?

Divide spending attributable to the work, including retries and relevant commitments, by unique records that meet your identity, field and recency rules. A billed match that fails those rules does not belong in the denominator.

Read this article as Markdown Open in ChatGPT Open in Claude