Article

How to evaluate the match rate of a company enrichment API

Test a company enrichment API on 500 labelled accounts of your own, separate the hit rate from correct matches and read precision, recall and F1 by input type.

9 min read

To evaluate the match rate of a company enrichment API, send it a sample of your own accounts whose correct company has been confirmed by hand, then score every response as a correct match, a wrong match or a miss. A quoted match rate that counts every response returning a company measures how often the API answered. Precision, the share of returned matches that are correct, and recall, the share of your accounts that came back with the right company, measure how often it answered correctly.

This post sets out a five-step test, the measures in UK government guidance on linkage quality, the domain and name cases that inflate a quoted rate and a worked example on 500 illustrative accounts. Sources were read on 15 November 2026.

How do you evaluate the match rate of a company enrichment API?

You evaluate the match rate of a company enrichment API by testing it on a labelled sample: accounts from your own CRM where a person has recorded the right company. The API's answers are compared with those labels, so the test counts right matches, wrong matches and misses separately instead of reporting one blended hit rate.

  1. Draw the sample at random from the accounts you plan to enrich, noting each one's country, size band and record source.
  2. Label every account before the first API call with its website, legal name and country, plus one written rule for groups, such as whether a subsidiary with its own website resolves to itself or to its parent.
  3. Send inputs in the mix production will send, so name-only records appear in the test as often as in your CRM.
  4. Score each response against its label as correct, wrong or missed, and store it with its date.
  5. Compute hit rate, precision, recall and F1 overall and by input type, each with a confidence interval.

A sample of well-known companies chosen by the vendor tests the easy cases, which is why the Fokals guide How to evaluate a company data vendor in ten questions asks for sample data on your own list.

Why do the hit rate and the correct match rate differ?

OutcomeWhat the API didIn the hit rateCorrect
Correct matchReturned the labelled companyYesYes
Wrong matchReturned a different companyYesNo
MissReturned no companyNoNo
  • Hit rate = (correct + wrong) / accounts sent
  • Precision = correct / (correct + wrong)
  • Recall = correct / accounts sent, where every sampled account is a real company with a labelled answer
  • F1 = 2 × precision × recall / (precision + recall)

A wrong match is the costly outcome. A miss leaves an empty field anyone can see, while a wrong match writes another company's headcount and industry into the record with nothing to flag it. In linkage terms it is both a false link and a missed link, so it lowers precision and recall together.

A worked example: precision, recall and F1 from a 500 account test

Assumptions: the figures are illustrative. Acme Robotics, an invented company, tests one enrichment API on 500 active accounts drawn at random from its CRM, each an operating company labelled by hand. A subsidiary with its own website should resolve to itself. It sends 380 accounts with the company's own domain, 20 with a CRM domain belonging to someone else (12 free email domains and 8 shop addresses on a hosted store platform) and 100 with name and country only.

InputSentReturnedCorrectWrongMissed
Own domain3803563451124
Someone else's domain20180182
Name and country10079562321
Total5004534015247

The wrong matches are 11 group parents returned for subsidiaries, 18 email providers or store platforms and, on names, 15 namesakes and 8 other companies in the same group.

  • Hit rate: 453 of 500, or 90.6%, the figure a vendor counting every returned company would quote.
  • Precision: 401 of 453, or 88.5%, with a 95% Wilson interval of 85.3% to 91.1%.
  • Recall: 401 of 500, or 80.2%, with an interval of 76.5% to 83.5%.
  • F1: 2 × 401 / (453 + 500) = 802 / 953, or 84.2%.

Own domains give a precision of 96.9% (345 of 356) and a recall of 90.8% (345 of 380). Names give a precision of 70.9% (56 of 79, interval 60.1% to 79.7%) and a recall of 56.0%. Every hit on someone else's domain was wrong.

The quoted 90.6% sits 10.4 points above the 80.2% of accounts that received the right company, and a rate computed on clean domain inputs alone would read 93.7% (356 of 380). Now suppose a stricter confidence setting had returned nothing for the 18 other-domain inputs and the 11 parent cases. The hit rate falls to 84.8% (424 of 500), precision rises to 94.6% (401 of 424), recall stays at 80.2% and F1 rises to 86.8% (802 / 924). In this illustration the lower match rate comes with the better data.

The intervals use the Wilson method at 95% (z = 1.96) from section 7.2.4.1 of the NIST/SEMATECH e-Handbook of Statistical Methods, which notes that its lower limit cannot be negative. The interval on name inputs spans almost 20 points, too wide to rank two vendors closely on that input type.

How do domain and name matching affect the match rate of firmographic enrichment?

Domain matching and name matching fail in different ways, and both can raise the match rate of a firmographic enrichment test without adding correct data. A domain owned by an email provider or a hosting platform can return that provider. A name sent without a country can return a namesake. Report results by input type for that reason.

Input caseWhat can come backEffect
Free email domain in the domain fieldThe email providerWrong hit
Shop address on a hosted platformThe platform companyWrong hit
example.co.uk cut to co.uk by your own codeNo companyMiss
Subsidiary with its own websiteThe group parentDepends on your rule
Old domain of an acquired companyThe acquirerDepends on your rule
Name without a countryA namesake abroadWrong hit

Before sending, lower-case each domain, remove the scheme, path, port and a leading www, and reduce it to the registrable domain with the Public Suffix List, which began as a Mozilla initiative, is maintained as a community resource and records suffixes such as co.uk. Remove free email domains and send hosted-platform addresses as name inputs. The Fokals guide to matching technographic data to CRM accounts sets out the normalisation in full.

Name inputs need a country, and a city where you have one. Strip legal forms the same way on both sides and treat a name match as a candidate for a person to confirm. The Fokals connection guide says a website address finds a company more surely than a name, and its lookup lists candidates with website and country, surest first.

What match rate methodology should a company data vendor evaluation use?

A company data vendor evaluation should use one labelled sample, sent to every vendor on the same day with the same inputs, and scored by the same rule. Report hit rate, precision and recall by input type, with intervals, and ask each vendor how its own quoted match rate is calculated.

Before comparing a vendor's quoted figure with yours, ask four things:

  • Denominator: every input sent, or only inputs with a usable domain?
  • Sample: the vendor's own list or a buyer's records?
  • Threshold: was a confidence cut-off applied, and what comes back below it?
  • Billing: is a wrong match billed like a correct one, and is a miss billed?

The person scoring responses should not know which vendor produced them. Fokals can be tested the same way. Under its pricing a lookup that finds no company costs nothing, and each company carries one stable Fokals company ID, so a correct match joins to the firmographic, technographic, hiring and intent datasets on the data page.

What do the entity resolution match quality metrics precision, recall and F1 measure?

In entity resolution, precision is the share of links made that are true matches and recall is the share of true matches that were linked. F1 is their harmonic mean, which is high only when both are high. Match quality metrics for a vendor test are these three, read by input type.

The UK guidance Quality assessment in data linkage, from the Office for National Statistics and the Government Analysis Function (updated 16 July 2021), counts true links (a), false links (b) and missed links (c). Recall, or sensitivity, is a/(a+c), precision is a/(a+b) and the F-measure is 2 × precision × recall / (precision + recall).

The guidance does not recommend the F-measure for judging linkage quality, citing Hand and Christen's point that it depends on the match rate of the classifiers compared. Report precision and recall side by side, with F1 as a summary. F1 weighs a false link and a missed link equally, so a wrong match, which is both, counts against it twice and a miss once.

The guidance adds that clerical review, a person ruling on sampled links, is more useful for estimating precision than recall. A labelled sample avoids that limit, because the right company is recorded for every account before the call, so misses are counted too.

How does the Fellegi-Sunter model of record linkage fit company entity resolution?

The Fellegi-Sunter model, formalised in 1969, scores a candidate pair of records by how much more likely each field's agreement or disagreement is among true matches than among non-matches. For company record linkage the fields are domain, name, country, city and identifiers, and a threshold on the summed score decides which pairs become links.

The UK guidance traces the framework to Fellegi and Sunter's 1969 paper A theory for record linkage and notes that it can be seen as naive Bayes, since it assumes identifiers are independent. The Splink documentation defines the terms:

  • The m probability is the chance of an observation, such as two domains agreeing, given that the records match.
  • The u probability is the chance of the same observation given that they do not.
  • The Bayes factor is m / u, and the match weight is its base-2 logarithm, so weights from separate fields add.
  • A total weight M, prior included, converts to a match probability of 2^M / (1 + 2^M), the value used when choosing a threshold.

As an illustration, if the registrable domain agrees for 98% of true pairs and 1 in 10,000 non-matching pairs, the Bayes factor is 9,800, a weight of about 13.3. A name agreeing for 90% of true pairs and 1 in 500 non-matching pairs gives 450, about 8.8. A free email domain is shared by many unrelated accounts, so its u probability is high and agreement on it carries little weight.

Company fields are often correlated, since names and domains share words and cities and postcodes move together, so summed weights can overstate the evidence. Where an API exposes a score or ranked candidates, run the labelled sample at several cut-offs and record precision and recall at each. One workable rule is two cut-offs: accept above the upper one, send the band between them to a person and reject below the lower one.

What can a match rate test not show?

A correct match means the right company was found. Whether its headcount, industry and revenue are right is a separate test against sources you trust.

Recall measured from outside mixes two causes: the vendor may not hold the company, or may hold it and fail to find it. Search the vendor's interface for a sample of misses or ask the vendor to check them, and keep unmatched records in a review queue as the guide to enriching company records in a data product describes.

The result describes your list, and a figure measured on another customer's records does not transfer to yours. Labels can be wrong, so have a second person label 50 accounts and count the disagreements. Results age as companies change domains, merge and close, so rerun the test before a renewal. A 500 account sample will rarely separate two vendors whose precision differs by two or three points, because intervals about six points wide overlap.

Frequently asked questions

How many accounts should a match rate test use?

Size the sample by the smallest group you need to judge. In the illustrative test above, 453 returned matches give a 95% Wilson interval on precision about six points wide, while the 79 returned name matches give one about 20 points wide. If name-only records matter to you, sample more of them so that group alone reaches a few hundred.

Should a vendor see the labels before the test?

Send each vendor the inputs only, in the same form and on the same day. A vendor that sees the labels can tune its answers to them, and the result then describes work done on your file instead of the API you will call in production. Keep the labels back until every response is stored.

Is a vendor's confidence score the same as precision?

A confidence score is the vendor's own estimate for one record, and precision is measured against your labels. Check whether they agree by grouping responses into score bands and computing the share correct in each band. If the share correct rises with the score, the score can help you set a cut-off. If it does not, the score says little about your accounts.

What counts as a correct match for a subsidiary or an acquired company?

The rule depends on how the data will be used, and it should be written down before labelling starts. Fields such as headcount usually belong to the entity the account deals with, so a subsidiary with its own website would resolve to itself. For a group view the parent may be the right answer. Apply one rule to every vendor so the comparison measures the APIs.

How often should the match rate test be repeated?

Rerun it before a renewal, after a vendor announces a change to its matching and when your own input mix changes, for example after a CRM migration. Label a fresh sample each time where you can, because companies change domains, merge and close, and an old label set drifts away from the companies it describes.

Does a high recall show that a vendor has good coverage?

Only in part. Recall measured from outside counts a miss the same way whether the vendor lacks the company or holds it and failed to find it. Search the vendor's own interface for a sample of misses, or ask the vendor to check them, to separate the size of its company index from the quality of its matching.

What this page says about the products it names was checked against their public documentation on 15 November 2026. Product and company names are trademarks of their owners. Fokals is not affiliated with them or endorsed by them.