Open menu
Data Enrichment

How to Measure Data Enrichment Accuracy Before You Buy (Match Rate, Precision and Decay)

How to Measure Data Enrichment Accuracy Before You Buy (Match Rate, Precision and Decay)

You measure data enrichment accuracy by running the same stratified sample of 500 to 1,000 real records through each vendor in the same week. Then you score four numbers separately: match rate, fill rate per field, precision on a hand-checked holdout, and verified rate for emails and phones. Finally, you run it again 60 to 90 days later to measure decay. Below is the exact protocol, the formulas, and the first-party numbers that explain why a single “accuracy” figure is never enough.

I’m going to be honest with you.

In 2021, in Hamburg, I bought a data provider on a demo that showed a 94% match rate. It looked wonderful. Then our first campaign bounced 11%, and the bounces got one of our sending domains throttled for two weeks.

The match rate was real. It just wasn’t accuracy. Nobody had told me those were different numbers, and I hadn’t asked.

I get it. Vendors lead with the biggest figure they have, and “accuracy” has no agreed definition. So the burden lands on you, the buyer, to measure it yourself before the contract.

So let’s fix it. Here’s how I measure data enrichment accuracy now, step by step. Let’s get into it.

TL;DR: The Four Numbers and the Five-Line Protocol

Short on time? Data enrichment accuracy is four numbers, not one.

  • Match rate: how many submitted records the vendor recognised at all.
  • Fill rate per field: how many records came back with a value in each column.
  • Precision: how many returned values are actually correct, checked by hand on a holdout.
  • Verified rate: how many returned emails and phones you can confirm deliverable or connected.

And the protocol fits in five lines.

→ Build a 500 to 1,000-row sample that looks like your real data → send the same file to every vendor in the same week → compute match and fill per field → hand-check 100 holdout rows for precision → verify emails and phones separately, then re-run in 60 to 90 days for decay.

Most vendor demos measure the first number and call it accuracy. So the rest of this guide is about the other three, plus the clock. And it all starts with ONE file that every vendor gets, unchanged.

What Does Data Enrichment Accuracy Actually Mean?

Data enrichment accuracy means the share of returned values that are correct and current, which is a different number from the share of records a vendor manages to match. The confusion between those two is where most bad purchases start.

So let’s define each term with a formula, because formulas end arguments:

  • Match rate = records matched Ć· records submitted. “We found this person” is a match, even if every field is stale.
  • Fill rate (per field) = records with that field returned Ć· records submitted. Email fill and job-title fill are different numbers.
  • Precision (per field) = correct values Ć· values returned, measured on a holdout you’ve checked by hand.
  • Verified rate = emails or phones confirmed deliverable or connected Ć· emails or phones returned.
  • Decay = share of values that were correct in test one and wrong in test two Ć· months elapsed.

Notice that precision is the only one with “correct” in it. Match rate and fill rate measure volume. Only precision and verified rate measure truth.

šŸ“Œ Example: A vendor matches 940 of 1,000 records (94% match), returns an email for 420 (42% email fill), and 290 of those emails pass verification (69% verified rate). Hand-checking 100 returned job titles finds 83 correct (83% precision). Four numbers, one "94% accurate" sales deck.

Fill rate versus verifiable fill, with real numbers

Here’s how big the gap between “returned” and “verifiable” can be. The numbers below come from a 1,088,655-profile random sample of the 270.5 million contact profiles in the CUFinder database, taken on October 29, 2026.

MeasureProfilesShare of sample
Profiles in sample1,088,655100%
Any work email found (fill rate)211,12219.4%
Found on a standard domain110,44910.1%
Found on an accept-all domain100,6739.2%
Phone number found51,5784.7%

Fill rate vs verifiable fill: the number a vendor quotes vs the number you can send to

Work email fill rate split by domain type in a 1,088,655-profile random sample of the CUFinder contact database Emphasis bar chart from a 1,088,655 profile random sample of the CUFinder contact database, October 2026: a work email is found for 19.4 percent of profiles, but only 10.1 percent of profiles have an email on a standard domain where verification can confirm the mailbox, while 9.2 percent sit on accept-all domains, and a phone number is found for 4.7 percent, which is why fill rate and verified accuracy must be measured as two different numbers. Fill rate (any work email found) 19.4% Found on a standard domain (verifiable) 10.1% Found on an accept-all domain (unverifiable) 9.2% Phone number found 4.7%

Source: CUFinder contact database, 1,088,655-profile random sample of 270.5M contacts, October 2026. Cache aig2_how-to-measure-data-enrichment-accuracy_fill_verified.

A work email exists for 19.4% of profiles in that sample. But only 10.1% of profiles carry an email on a standard domain, where a verifier can confirm the mailbox. The other 9.2% sit on accept-all domains, which say yes to any address and so can’t be verified without a real send.

So the honest “fill rate” is 19.4%. And the honest “verifiable fill” is 10.1%. If a vendor quotes you the first and you budget on the second, you’ve already halved your expectations before paying anything. Good.

That’s what the words mean. Now, why the biggest number misleads.

Why Match Rate Alone Misleads You

Match rate misleads because a vendor can match a record loosely, return stale values, or hand you an email that no verifier can confirm. All three still count as a match. My 2021 campaign hit all three at once.

Here’s what went wrong, in order:

  • Loose matching. “M. Fischer at Fischer GmbH” matched a different Fischer at a different Fischer GmbH. Still a match.
  • Stale values. Dozens of contacts had changed jobs. The vendor returned their old title and old company email. Still a match.
  • Unverifiable emails. Many returned emails sat on accept-all domains. The vendor’s “verified” badge meant the server didn’t reject the address, not that a mailbox existed. Still a match.

Ninety-four percent of records matched. Eleven percent of a 6,000-send campaign bounced. Both numbers were true at the same time!

Was the vendor lying? No. The vendor measured the thing it could measure cheaply, and I didn’t measure the things I needed. That’s on me, and it’s why the next seven steps exist.

šŸ” Did You Know? In our 1,088,655-profile sample, 47.7% of all found work emails sit on accept-all domains. A verifier returns "accept-all" or "unknown" for those, not "valid." So any vendor's "verified email" claim needs a definition before it means anything.

Bounces were the symptom. Here’s the measurement that would have caught them.

Step 1: Build a Test Sample That Looks Like Your Real Data

Build a test sample of 500 to 1,000 real records from your own CRM, stratified to match the mix you actually enrich, with 20 to 30 percent deliberately messy rows. The sample decides what the test can see, so this step matters more than any vendor’s demo.

In 2022, for a Frankfurt bank prospect list, I let a vendor pick the sample. They took our cleanest 200 rows: full names, domains, LinkedIn URLs. The pilot looked perfect.

Then the real 4,000-row export, full of name-plus-company rows from a conference, matched at a fraction of the rate. Same vendor, same month.

How to build the sample

So build the sample yourself, from YOUR data, like this:

  1. Pull 500 to 1,000 rows from the real source. CRM export, form fills, badge scans. Not a vendor-provided list.
  2. Stratify by company size. Mix 1-to-10-person companies with enterprises in the proportion your pipeline actually has.
  3. Stratify by seniority and region. Founders, managers, C-level; home market and abroad.
  4. Include 20 to 30 percent problem rows. Name plus company only. Misspelled names. Personal email domains. The rows that make vendors nervous.
  5. Set aside a 100-row holdout you already know the truth for. Current customers, colleagues’ contacts, people you spoke to last week. This is your precision check later.
  6. De-duplicate and snapshot the date. The date matters for Step 6.

Why the mix matters

Why does company size matter so much? Because fill rates move with it, for every vendor. In the same 1,088,655-profile CUFinder sample, a work email exists for 10.7% of people at 1-to-10-person companies and 34.1% at companies with 10,001 or more.

So a sample skewed to big companies flatters everyone, and a sample skewed to small ones punishes everyone.

The same goes for seniority. Senior individual contributors and managers carry work emails far more often than C-level executives do in our data, and phones are rare at every level. A sample built from a “VP and above” filter will under-report email coverage for every vendor you test.

And region matters too. A vendor strong in North America can look weak on a DACH list, and the reverse. If 40% of your pipeline is in Germany, 40% of the sample should be too.

If your CRM is too messy to sample cleanly, run a quick cleanup first. I’ve written a step-by-step on how to audit your data quality before enrichment, and it takes an afternoon.

šŸ’” Pro Tip: Name the sample file with the date, like enrichment_test_2026-10-29.csv, and never edit it. Every vendor gets a copy of the identical file. The moment you "fix a few rows" for one vendor, the comparison is dead.

That’s the sample. Now send it.

Step 2: Send the Same File to Every Vendor in the Same Week

Send the identical file to every vendor within the same week, with the same columns, and ask for raw output including confidence scores and last-updated fields. Anything else and you’re comparing different tests.

Here’s the checklist I use:

  • Same columns for everyone. If one vendor gets LinkedIn URLs and another gets names only, the match rates aren’t comparable.
  • Normalise inputs first. Trim spaces, one company name format, lower-case domains. Give every vendor the same chance.
  • Log the request date per vendor. Data changes weekly. A two-week gap between vendors is a confound.
  • Ask for raw output. Every field, plus confidence scores and any last-updated or last-verified timestamp. If a vendor won’t share a last-updated field, note it. That’s a signal.
  • Never let the vendor pick or clean the sample. See Frankfurt.
  • Keep the credits honest. Note whether misses were charged. It changes the cost math in Step 7.

One more thing. Tell every vendor they’re in a bake-off. Good ones behave the same either way, and the ones who suddenly offer a “special enrichment pass” have told you something useful.

Test real-time enrichment separately

If you’ll also enrich inbound leads in real time, through a form or an API call, test that path on its own. Batch enrichment and real-time enrichment often draw on different caches, and a vendor that’s excellent on a 1,000-row file can return thinner records in 300 milliseconds.

So take 100 rows from your sample and push them one at a time through the API or the form integration. Log latency, fields returned, and whether the record matches what the batch file returned for the same person. Any gap between the two is a gap your routing rules will live with.

That’s the send. Next, the counting.

Step 3: Measure Match Rate and Fill Rate per Field

Measure match rate once per vendor, then fill rate once per field, because the per-field numbers are where vendors differ most. A Google Sheet does this in five minutes.

Put each vendor’s output in its own tab, one row per submitted record. Then add these formulas.

  • Match rate → =COUNTIF(matched_column, “yes”) / ROWS(matched_column)
  • Fill rate for a field → =COUNTA(field_column) / ROWS(field_column)
  • Fill rate among matched records only → =COUNTA(field_column) / COUNTIF(matched_column, “yes”)

Then build one table: fields down the side, vendors across the top, fill rate in each cell. Here’s the shape, with illustrative placeholders you’ll replace with your own results:

FieldVendor A fillVendor B fillVendor C fill
Matched at all%%%
Job title%%%
Company domain%%%
Employee range%%%
Work email%%%
Direct phone%%%
LinkedIn URL%%%

Why per field? Because fill varies wildly by field, for physical reasons. Across our October 2026 samples, a job title is present on roughly nine profiles in ten. In the 1,088,655-profile sample, a work email is present on 19.4% and a phone number on 4.7%. So a vendor claiming “90% fill” is almost certainly talking about titles and company names, not emails and phones.

And compare fill among matched records too. A vendor with a lower match rate but a higher email fill among matches may be the one that’s honest about what it doesn’t know.

🧠 Fun Fact: The cheapest field to fill is the one you already have. Some vendors count "company name returned" as a filled field when you submitted the company name. Exclude input fields from your fill-rate table, or every vendor scores a free 100%.

Volume counted. Now the number that actually matters.

Step 4: Measure Precision on the Holdout

Measure precision by hand-checking the 100 holdout rows you set aside in Step 1, field by field, against a source you trust. It’s the only step that measures whether the data is true, and it takes one evening per vendor.

In 2024 I did exactly this for a 500-row test. One evening, 100 job titles, each checked against the person’s current LinkedIn profile. Seventeen were stale: the person had moved roles or companies, and the vendor had returned the old title.

The match rate for those 100 rows was 100%. Precision on job title was 83%.

The hand-check method

Here’s the method:

  1. Pick the fields that drive your decisions. Usually job title, current company, location, and employee range. Not all 60 fields.
  2. Check each returned value against a challenger source. The person’s LinkedIn profile, the company’s own site, a recent conversation. Mark correct, wrong, or stale.
  3. Compute precision per field → correct values Ć· values returned. Stale counts as wrong.
  4. Weight the fields by what they cost you when wrong. A wrong employee range mis-routes a lead. A wrong title embarrasses a rep. A wrong company wastes the whole touch.
  5. Record the vendor’s confidence score next to each result. If high-confidence rows aren’t more often correct, the score is decoration.

This is where data matching quality shows itself. A vendor that resolves “M. Fischer, Fischer GmbH” to the right Fischer has done hard work. One that resolves it to any Fischer has inflated its match rate at the expense of your precision.

Weighting the fields, and why 100 rows

If you want the textbook version of precision and its sibling recall, Wikipedia’s precision and recall entry is the cleanest definition I’ve found. In enrichment terms: match rate is a recall-flavoured number, precision is precision.

Does every field deserve the same weight? No. A correct direct dial for a VP in your ICP is worth far more than a correct employee range for a company you’d never call.

So score precision per field, then weight the fields by the decision they feed. Routing fields (size, industry, country) carry one weight. Outreach fields (title, email, phone) carry another, usually heavier.

Why only 100 rows? Because you’ll actually do 100. A 100-row hand check gives you a rough but honest precision figure. A 1,000-row hand check never gets finished.

How long does it take? About one evening per vendor.

šŸ“Œ Example: Two vendors, same holdout. Vendor A: 96% match, 78% title precision. Vendor B: 81% match, 93% title precision. Vendor B found fewer people and was right about almost all of them. For outbound, B wins. For a TAM count, A might. Know which job you're buying for.

Precision done. Now the two fields that bounce and ring.

Step 5: Verify Emails and Phones Separately

Verify emails and phones as their own metric, because a returned contact field is a claim, not a fact, and the only honest test is whether it delivers or connects. Here’s how I do each.

Emails: verifier status first, then a small real send

Start by running every returned email through a verifier and recording the status. Count as “verified” only the addresses with a clear valid status. Accept-all, unknown, and risky are not verified, whatever the vendor’s badge says. My roundup of email verification tools covers which verifiers report accept-all honestly.

Then do what verifiers can’t: a small real send. Take 100 to 200 “valid” emails per vendor, send a plain, honest message from a warmed secondary domain, and measure hard bounces. Keep the volume low and the content legitimate, and follow Google’s email sender guidelines on authentication and complaint rates, because this test is not worth a blocked domain.

This step is data verification in the plainest sense: checking a claimed value against reality.

The accept-all problem gets worse with company size

Here’s the part that surprised me when I pulled the numbers.

Employer sizeWork emails foundOn accept-all domainsAccept-all share
1 to 10 employees7,8763,64646.3%
11 to 5015,2205,92838.9%
51 to 20020,6938,11339.2%
201 to 50016,5476,89841.7%
501 to 1,00014,3356,22143.4%
1,001 to 5,00036,06717,15547.6%
5,001 to 10,00017,0678,55150.1%
10,001 or more78,23941,06752.5%

The verification gap grows with company size: accept-all share of found work emails

Share of found work emails on accept-all domains by employer size in a 1,088,655-profile random sample of the CUFinder contact database Ordinal bar chart of the share of found work emails that sit on accept-all domains, by employer size, from a 1,088,655 profile random sample of the CUFinder contact database, October 2026: 46.3 percent at companies with 1 to 10 employees, 38.9 percent at 11 to 50, 39.2 percent at 51 to 200, 41.7 percent at 201 to 500, 43.4 percent at 501 to 1,000, 47.6 percent at 1,001 to 5,000, 50.1 percent at 5,001 to 10,000 and 52.5 percent at 10,001 or more, so an enterprise-heavy test sample needs a real send or a bounce test, not just a verifier status, to measure accuracy. 1 to 10 employees 46.3% 11 to 50 38.9% 51 to 200 39.2% 201 to 500 41.7% 501 to 1,000 43.4% 1,001 to 5,000 47.6% 5,001 to 10,000 50.1% 10,001 or more 52.5%

Source: CUFinder contact database, 1,088,655-profile random sample, October 2026. Cache aig2_how-to-measure-data-enrichment-accuracy_aa_by_size.

Among found work emails, the accept-all share rises from 38.9% at 11-to-50-person companies to 52.5% at companies with 10,001 or more employees. Large companies run mail servers that accept everything and filter later. So an enterprise-heavy test sample needs the real-send step more than an SMB sample does, not less.

Phones: a 30-call connect sample

For phones, pick 30 returned numbers per vendor and call them. Record four outcomes: right person, right company but wrong person, wrong number, dead line. Your connect rate is “right person” divided by 30.

Thirty calls takes an hour. It will tell you more about a vendor’s phone data than any accuracy page ever will.

Is 30 enough? For a buying decision, yes. If two vendors land within a few calls of each other, call 30 more for each before you pick.

And log the time of day you called, because connect rates swing between 9 a.m. and 4 p.m. in ways that have nothing to do with the data.

šŸ’” Pro Tip: Log the verifier status next to every email in your sheet, then compute precision for "valid" rows and "accept-all" rows separately. If the vendor's accept-all emails bounce at three times the rate of its valid ones, you've just learned how to filter its output before every campaign.

Found, verified, connected. Now the clock.

Step 6: Measure Decay With a Second Run

Measure decay by re-running the same sample file through the same vendor 60 to 90 days after the first test and counting how many previously correct values are now wrong. Accuracy has a half-life, and a vendor’s refresh cadence decides it.

The reason is simple. People change jobs constantly, and every job change turns a correct title, company and email into a wrong one. Here’s the scale, from CUFinder’s own job-change tracker:

Month (2026)Job-change events logged
June 17 to 30 (partial)14,623,884
July100,192,408
August87,009,908
September36,137,333
October 1 to 29192,991,894
Total, June 17 to October 29430,955,427

Why accuracy decays: LinkedIn job-change events logged per month, June 17 to October 29, 2026

LinkedIn job-change events logged by CUFinder tracking per month, June 17 to October 29, 2026 Bar chart of LinkedIn job change events logged by CUFinder tracking each month: 14.6 million in the partial window of June 17 to 30, 100.2 million in July, 87.0 million in August, 36.1 million in September and 193.0 million from October 1 to 29, 2026, a total of 431.0 million events in about 19 weeks against a database of 270.5 million contact profiles, which is why an enrichment accuracy test measured once is already decaying by the time the contract is signed. Full month Partial window June 17 to 30 (partial) 14.6M July 100.2M August 87.0M September 36.1M October 1 to 29 193.0M

Source: CUFinder LinkedIn job-change tracker, rolling window June 17 to October 29, 2026 (not a year). Partial month in gray. Caches aig2_how-to-measure-data-enrichment-accuracy_jc_2026*.

Between June 17 and October 29, 2026, the tracker logged 430,955,427 LinkedIn job-change events. That’s a rolling window of about 19 weeks, not a year, and the monthly figures swing with how LinkedIn surfaces changes. Against a database of 270.5 million contact profiles, the message is blunt: the contact data you tested in January is not the contact data you’re using in April.

In Q1 2026 I proved it to myself. I re-ran a January test file in April, same 300 rows, same vendor. Fill rate looked identical.

But precision on job title had dropped by a visible margin, because the people had moved and the vendor hadn’t caught up yet. The first test had been right. It just had a date on it.

Computing decay

So compute decay like this.

→ Decay per month = (values correct in test 1 and wrong in test 2) Ć· (values correct in test 1) Ć· months between tests.

And ask every vendor two questions in writing: how often is this field re-verified, and does the output carry a last-verified timestamp? A vendor that tracks job changes and timestamps fields is selling you freshness. One that doesn’t is selling you a snapshot.

For the bigger picture on how fast B2B records go stale, our job change statistics page tracks the same feed month by month. It’s the cleanest argument I know for budgeting re-enrichment rather than a one-time purchase.

šŸ” Did You Know? 430,955,427 LinkedIn job-change events were logged by CUFinder's tracker between June 17 and October 29, 2026. That's more events in 19 weeks than there are contact profiles in the 270.5-million-record database they're matched against, which is why data decay is a schedule, not a risk.

Four numbers plus a clock. Time to put them in one sheet.

Step 7: Score Vendors With a Weighted Sheet

Score vendors with a weighted scorecard that puts precision and verified rate ahead of match rate, then compute cost per correct record. Because the cheapest vendor per record is rarely the cheapest per correct record.

Here are the weights I start from. Change them to fit your job, but change them BEFORE you see the results, not after.

MetricWeightHow you measured itVendor AVendor B
Precision on key fields30%100-row holdout hand check (Step 4)%%
Verified rate (emails and phones)25%Verifier status + real send + 30 calls (Step 5)%%
Fill rate on the fields you need20%Per-field table (Step 3)%%
Match rate10%Matched Ć· submitted (Step 3)%%
Freshness and decay10%Second run + last-verified fields (Step 6)%%
Cost per correct record5%Formula below$$

And here is the cost formula, the one I wish I’d had in 2021.

→ Cost per correct record = total spend Ć· (records returned Ɨ precision Ɨ verified rate).

In 2025 I ran this sheet for two vendors. Vendor one had the higher match rate, the lower price, and the nicer dashboard. Vendor two had fewer matches and a phone connect rate almost twice as high. On cost per correct record, vendor two won, and the SDR team’s connect rate said the sheet was right.

Two optional rows are worth adding.

  • Net-new lift. How many contacts inside your target accounts did the vendor find that you didn’t have at all? For outbound, this can outweigh everything else.
  • Coverage by segment. If you sell to German manufacturers, score precision and fill on the German manufacturing rows only. Global averages hide local holes.

If you want a broader decision framework around the scorecard, covering integrations, compliance and contract terms, I’ve written separately on how to choose between data enrichment providers.

That’s the score. Here’s the sheet that holds it all.

A Scoring Sheet You Can Copy

One spreadsheet holds the whole test: a Rows tab per vendor, a Holdout tab, and a Score tab that reads from both. Here’s the layout, so you don’t build it from scratch at 11 p.m. the night before a renewal.

Rows tab (one per vendor), one row per submitted record:

  • Columns A to F: your inputs (name, company, domain, LinkedIn URL, country, size band).
  • Column G: matched (yes or no).
  • Columns H onward: one column per returned field (title, company, email, phone, employee range, and so on).
  • Column S: email verifier status (valid, accept-all, unknown, invalid).
  • Column T: credits charged for this row.

Holdout tab, 100 rows, one per holdout record and field:

  • Columns: record id, field, returned value, true value, verdict (correct, wrong, stale), vendor confidence score.
  • Precision per field → =COUNTIFS(field_col, “title”, verdict_col, “correct”) / COUNTIF(field_col, “title”)

Score tab, one row per metric:

  • Match rate → =COUNTIF(Rows!G:G, “yes”) / COUNTA(Rows!A:A)
  • Email fill → =COUNTA(Rows!email_col) / COUNTA(Rows!A:A)
  • Verified rate → =COUNTIF(Rows!S:S, “valid”) / COUNTA(Rows!email_col)
  • Credits per correct email → =SUM(Rows!T:T) / (COUNTA(Rows!email_col) Ɨ precision Ɨ verified rate)
  • Weighted score → =SUMPRODUCT(weights_range, scores_range)

Set the weights BEFORE you paste a single vendor result. Then paste, then read. Doing it in that order is the whole difference between measuring and rationalising.

šŸ’” Pro Tip: Freeze the weights row and colour it grey. When a vendor's rep asks why they lost, you can show them the weights were set on a date before their file arrived. Nobody argues with a timestamp.

Before you sign anything, here’s what to watch for when a vendor runs the test for you.

Red Flags in a Vendor-Run Accuracy Test

A vendor-run accuracy test is a sales asset, not a measurement. Six patterns tell you the number won’t survive contact with your data.

  • The vendor picks the sample. You’ll get the rows that match. Insist on your file, including the ugly rows.
  • “Accuracy” with no denominator. Accuracy of what, divided by what? If the deck can’t say, the number is marketing.
  • No last-updated or last-verified field in the output. You can’t measure freshness, and they’d rather you didn’t.
  • “Verified” that includes accept-all. Ask for the verifier status breakdown. If accept-all is counted as verified, the verified rate is inflated.
  • No disclosure on misses. If you can’t learn whether unmatched records are charged, you can’t compute cost per correct record.
  • Refusing raw output. A vendor that only shows you a summary dashboard is hiding the rows that would lower the summary.

None of these mean the data is bad. They mean you haven’t measured it yet.

Public benchmarks can help you calibrate what “good” looks like for a category. Dropcontact publishes its own email finder benchmark across 20,000 contacts, and sites like apibenchmarks.com compare enrichment APIs on match rate and latency. Read them as other people’s tests, on other people’s data, and then run yours.

b2bbench’s employee count test goes field by field, which is the shape I want. It covers 282 companies (release 2026.11.1). CUFinder had the highest band accuracy at 90.4%. Explorium and People Data Labs sit in the same top group, so read the three as level.

And one quiet red flag: the vendor who says data quality isn’t a problem any more. Harvard Business Review reported years ago that only 3% of companies’ data met basic quality standards. I haven’t met the 3%. Neither has any vendor.

Common Mistakes When Measuring Data Enrichment Accuracy

The vendor isn’t the only one who can skew a test. Here are the seven mistakes I’ve made myself, so you can skip them:

  • Testing your cleanest data. Clean rows match everywhere. The test has to include the rows that don’t, because those are the rows you’re paying to fix.
  • Treating one test as permanent. A single bake-off tells you about one week. Schedule the second run before you sign, and a quarterly re-test after.
  • Weighting every field equally. A wrong employee range and a wrong email are not the same mistake. Weight by consequence.
  • Comparing vendors across different weeks. Data moves. Two weeks apart is two different tests.
  • Ignoring what misses cost. If unmatched rows are charged, your cost per correct record can double on a messy list. Put credits charged in the sheet.
  • Counting accept-all as verified. It feels like a win in the fill column and becomes a bounce in the campaign. Keep the status column honest.
  • Skipping the write-up. If the result lives in your head, the next renewal starts from zero. Save the sheet, the sample file and the date together.

The thread through all seven is the same. Measure the thing you’ll actually use, on the data you actually have, at the time you’ll actually use it.

How I Run This on CUFinder’s Own API

Full disclosure: I work at CUFinder, and I run this same test on our own data, because our billing model makes fill rate flatter us. CUFinder charges credits only for records found, and “Not Found” rows are free, so a fill-rate-only comparison favours any pay-on-found vendor, including us.

So I don’t stop at fill. I run the 100-row holdout, the verifier pass, and a small real send every quarter, and I publish what comes out. The email finder accuracy test is the most recent one, 2,000 lookups with the method laid out so you can repeat it on any vendor, us included.

The company-side counterpart is a public frame. In October 2026 we ran the public Openbenchmarks company-enrichment frame on our own API. 282 domains, the board’s own judge prompts.

The result: 90.1% correct-field yield, the highest on the frame, 1.1 points ahead of People Data Labs. 93.0% accuracy on returned fields, also the highest.

It’s a self-run, not an official board entry. So check the full run before you decide.

A third-party board agrees on accuracy. b2bbench’s company enrichment test ran the same 282 companies through ten APIs. CUFinder posted the highest correct-field yield at 93.1%, with a blank counted as wrong. People Data Labs came next at 88.8%.

Bar chart of b2bbench company enrichment scores on 282 companies, CUFinder highest at 93.1%
Correct-field yield across five firmographic fields, n = 282. Source: b2bbench.com, release 2026.11.1

The free plan gives 50 credits a month, which is enough for a 50-row smoke test of the ugliest rows in your CRM, not for the full 500-row protocol. For that, you’ll spend a few dollars on each vendor you test. It’s the cheapest few dollars in the whole procurement.

Frequently Asked Questions

What is a good match rate for B2B data enrichment?

A good match rate depends on your inputs, so judge it against your own sample rather than a universal number. Lists with domains or LinkedIn URLs match far higher than name-plus-company lists. More importantly, match rate without precision is a vanity metric. A vendor that matches 80% of your rows and is right 93% of the time beats one that matches 95% and is right 75% of the time.

What is the difference between match rate and fill rate?

Match rate is the share of submitted records the vendor recognised at all, while fill rate is the share of records that came back with a value in a specific field. A vendor can match 90% of records and fill a work email for only 20% of them. Always measure fill per field, and exclude the fields you supplied as input.

How big should my test sample be?

Use 500 to 1,000 rows for match and fill rates, plus a 100-row holdout you already know the truth for to measure precision. Stratify the sample by company size, seniority and region, and include 20 to 30 percent messy rows. Fifty rows is enough for a free-tier smoke test but not for a buying decision.

How do you make sure enriched data is accurate?

You check returned values against a source you trust, verify emails and phones separately, and re-test on a schedule. Concretely: hand-check 100 holdout rows against LinkedIn or company sites, then run a verifier and a small real send on emails. Call a 30-number phone sample, and repeat the test 60 to 90 days later to measure decay.

How do I verify emails without sending to them?

Run them through an email verifier and count only clear “valid” results as verified; treat accept-all, unknown and risky as unverified. In our 1,088,655-profile sample, 47.7% of found work emails sit on accept-all domains, which a verifier cannot confirm. For those, a small real send from a secondary domain is the only honest test.

How do I check phone numbers without calling everyone?

Call a 30-number sample per vendor and record right person, wrong person, wrong number, or dead line. Thirty calls takes about an hour and gives you a usable connect rate. Line-type checks and carrier lookups can flag disconnected numbers in bulk, but only a conversation confirms the person.

How often should I re-test an enrichment vendor?

Re-run the same sample every 60 to 90 days, and treat any drop in precision as decay you’ll pay for. CUFinder’s tracker logged 430,955,427 LinkedIn job-change events between June 17 and October 29, 2026, so a quarter is long enough for a measurable share of titles and emails to go stale.

Why does more data increase accuracy?

It doesn’t by itself; more fields and more records increase coverage, not correctness. Accuracy improves when a vendor cross-checks values across sources, re-verifies on a schedule, and timestamps each field. A bigger database with no refresh cadence just gives you more stale values to be wrong about.

What are the pillars of data quality?

The commonly cited pillars are accuracy, completeness, consistency, timeliness, validity and uniqueness. In enrichment testing, precision measures accuracy, fill rate measures completeness, the second run measures timeliness, verification measures validity, and de-duplication before the test handles uniqueness. For a sense of how the market scores on these today, our B2B data quality statistics page collects the published numbers.

Should I test several vendors at the same time?

Yes, test two to four vendors on the identical file in the same week, because data changes fast enough that a month’s gap makes results incomparable. Running them together also lets you compute net-new lift and see where a waterfall of two vendors beats either one alone.

It’s Time to Run Your Own Test

You don’t need a data science team for this. You need one honest sample file, one evening with LinkedIn open, and a sheet with six rows.

Here’s the whole method in one line.

→ 500 to 1,000 real rows → same file, same week, every vendor → match and fill per field → 100-row precision check → verify emails and phones separately → re-run in 60 to 90 days → score and compute cost per correct record.

That’s it. That’s how I’d have caught the 2021 bounce disaster before the first send, and it costs an afternoon plus a few dollars of credits.

So, which number did a vendor sell you on, and which one turned out to matter? Tell me in the comments. The stories are how this guide gets better.

And if you want to see the protocol applied end to end, with the bounce test included, the email finder accuracy test above is the worked example. Run it on us. Run it on everyone. Then buy the one your sheet picks!

Share this post with your friends!

References

  • CUFinder contact database, 1,088,655-profile random sample of 270.5M contacts, queried October 29, 2026 (cache tags aig2_how-to-measure-data-enrichment-accuracy_fill_verified and _aa_by_size)
  • CUFinder LinkedIn job-change tracker, rolling window June 17 to October 29, 2026, 430,955,427 events (cache tags aig2_how-to-measure-data-enrichment-accuracy_jc_2026*)
  • Wikipedia, Precision and recall: https://en.wikipedia.org/wiki/Precision_and_recall
  • Google Workspace Admin Help, Email sender guidelines: https://support.google.com/a/answer/81126
  • Dropcontact, Email Finder Benchmark across 20,000 contacts: https://www.dropcontact.com/email-finder-benchmark
  • apibenchmarks.com, independent enrichment API comparisons: https://apibenchmarks.com/
  • Harvard Business Review, “Only 3% of Companies’ Data Meets Basic Quality Standards” (2017): https://hbr.org/2017/09/only-3-of-companies-data-meets-basic-quality-standards
  • Clearbit, announcement of the April 30, 2025 retirement of its free tools, an example of why vendor facts need dates: clearbit.com/blog
How would you rate this article?
Bad
Okay
Good
Amazing
Comments (0)
Related Posts

Keep on Reading

How to Find Decision Makers in Your Customer List (2026 Sales Guide)
Data Enrichment

How to Find Decision Makers in Your Customer List (2026 Sales Guide)

How to Enrich Companies With Fundraising Data
Data Enrichment

How to Enrich Companies With Fundraising Data

How to Do a Reverse Email Lookup (and Find Who Owns Any Email)
Data Enrichment

How to Do a Reverse Email Lookup (and Find Who Owns Any Email)

How Enriched Data Improves Customer Segmentation (2026 Practitioner Guide)
Data Enrichment

How Enriched Data Improves Customer Segmentation (2026 Practitioner Guide)

Comments (0)
98% accuracy, GDPR & CCPA ready

Find and enrich your next customers

Prospect 85M+ companies and 1B+ contacts, enrich your CRM, and act on buying signals — powered by AI and Big Data.

  • Free plan available
  • 50 credits / month
  • No credit card required