Auditing a Purchased Data Sample Field by Field

You bought a batch of contact records. Before you renew, renegotiate, or walk, you need to know how accurate it actually is, field by field, on your own data rather than on the vendor's homepage number. This guide carries one accuracy audit all the way through: the sample math, the per-field verification, the counts, and the decision at the end. It is written for recruiting and revenue operations, and for anyone who has to defend how the numbers were gathered.

I will be honest about one thing up front. The professional-profile index I usually build these guides around was unavailable during this session, so there are no Refolk-exclusive figures here. Every number below is externally sourced and cited to the method it came from. That is a limit, and I would rather state it than dress it up.

What a field-level data accuracy audit actually measures

A field-level data accuracy audit measures the proportion of records where a given field is correct against ground truth, computed separately for each field, with a margin of error attached. It is not one accuracy score. It is one score per field, because the fields fail in different ways and clear different thresholds.

The reason this matters is that a single blended "accuracy" number averages away the field that decides your renewal. In one same-sample head-to-head, two providers were 6 points apart on email but 26 points apart on mobile match. If you had reported one blended figure, you would have hidden the entire story. Independent phone testing across B2B providers has ranged from 63 to 91 percent accuracy with coverage from 26 to 92 percent, so phone is almost always where the real spread lives.

Throughout this guide I work one example: a purchased list of roughly 100,000 B2B contacts, audited across five fields (business email, direct phone, job title, company HQ, employee count), stratified by three regions and two freshness bands. I will show the counts as I go so you can follow along on your own file.

How to size the sample and set the margin of error

Fix your confidence and target margin first, then compute the sample size; do not start from a round number. For a proportion at 95 percent confidence with a plus-or-minus 5 percent margin, using the worst-case planning value of 0.5, you need at least 384 records. A sample of 381 reaches roughly plus-or-minus 5 percent, and for a population of 100,000 the requirement is about 383 respondents, so the population size barely moves the number once it is large.

The formula is n = p* x q* x (z/E)^2, where p* = q* = 0.5 maximizes the required sample and therefore guarantees your margin will not exceed E. Using 0.5 is the honest default when you have no prior estimate. The margin of error for a binomial proportion is largest exactly at 0.5, so planning there means every field you measure comes in at or below your target margin.

Here is the trade-off in one table. Read it as: pick the margin you can defend, then read off the n you owe.

The margins are computed from the standard proportion formula, E = 1.96 x sqrt(0.25/n). The row at 384 lines up with the 381 and 383 anchors from the sample-size references.

Now the trap that sinks most provider tests. A 500-record or 1,000-record sample looks robust, and vendor-comparison guides routinely use those round numbers. But the moment you split across strata, the per-stratum n collapses. Five hundred records across five ICP segments is about 100 each, which is a plus-or-minus 9.8 percent band. Any per-region accuracy claim from that is nearly meaningless.

For my worked example I want per-region reads, so I size 384 per region across three regions, which is 1,152 records. A published NCI example lands in the same territory: an overall sample of 513 gives 95 percent confidence at plus-or-minus 4 percent, stratified by regional subgroups.

How to draw a stratified sample that does not lie

Stratified sampling divides the population into strata based on characteristics relevant to the audit, selects independently from each stratum, and guarantees adequate representation of each segment. For a data audit, the strata that matter are region, segment, channel, and lifecycle stage, plus record freshness.

Freshness is the one people skip, and it is the one that decides whether your title numbers are honest. Title changes drive 65.8 percent of annual data decay, and B2B data decays roughly 2 to 3 percent per month, or 22 to 30 percent annually. Some sources put it higher, above 5 percent per month. If your sample over-weights records verified last week, your title accuracy will look far better than the database deserves, because the median record is months old.

Two rules keep the sample honest. First, sample your own live export, because the vendor's demo file is drawn from the curated, recently refreshed part of the database. Second, freeze the sample with a stable unique ID per record before verification begins, so you can de-dupe, track each record's outcome, and reproduce the audit later.

In my example the population splits 55 percent North America, 30 percent EMEA, 15 percent APAC. I confirm the drawn sample matches those proportions inside each freshness band before I compute anything. If it does not, I redraw. A convenience sample that drifted toward fresh domestic records would overstate accuracy across the board.

How to verify each field against ground truth

Each field has its own authoritative source and its own failure mode, so verification is field-specific work, not one script. Name the ground-truth source per field before you start, and specify the match logic: exact, fuzzy, geocode, or checksum.

Email

SMTP verification checks deliverability by opening a real-time connection to the recipient's mail server and asking whether the mailbox exists, without sending anything. The server response classifies the address as valid, invalid, temporarily unavailable, greylisted, or catch-all. The known limit is catch-all domains: most such servers return a generic 250 accept code for every address during the handshake, including fake or mistyped ones. Between 30 and 40 percent of B2B email domains are configured this way.

This sets the honest ceiling of the whole audit. You cannot confirm mailboxes on catch-all domains, so you report email accuracy over resolvable records only and quarantine catch-all as its own bucket. Counting catch-all as valid is precisely how vendors reach numbers in the 90s while real-world deliverability sits lower.

Phone

Run a controlled connect test and track four rates by source and recency: deliverability rate, bounce rate, connect rate (connected divided by total dials), and answer rate (human answers divided by connected). Keep connect and answer separate. A low connect rate is often call-window or routing friction, voicemail and gatekeepers, not a wrong number, so test time-of-day before you condemn the data.

Title and company

Pull the sampled contacts and cross-check the exported job title against the current LinkedIn profile or employer site. Use three buckets, not a binary: current and accurate, outdated role, wrong company. For firmographics, anchor to public records. Firmographic data comes from SEC filings, state business registries, and company pages, so verify against those primary records rather than against another aggregator that may share the same upstream source.

Finding the right people to run the phone and title checks against, at the exact seniority and region your strata require, is its own sourcing problem. This is where Refolk earns a place in the workflow: you describe the segment in plain English and get a verified, reachable list to sample against, which is faster than filtering a general database by hand.

The audit procedure, end to end

Run these eight steps in order. The whole provider test takes roughly one week and can cost nothing beyond free trial access. Each step names its owner and its "done" condition.

How to compute accuracy and read it against thresholds

Accuracy for a field is correct values divided by total tested, with the margin of error for that sample size attached; a point estimate without an interval is not a finding. Reporting "84 percent accurate" from 100 records hides a band of roughly plus-or-minus 8 to 10 percent that can straddle your acceptance line.

Here is a documented same-sample comparison to calibrate what a real gap looks like. Two providers were run against the same 1,000 B2B leads, with a 20 percent subset manually verified against LinkedIn profiles updated within the prior 90 days.

The gap row is a plain subtraction. Notice again how small the email spread is against the mobile chasm. If your renewal hangs on phone data, a blended number would have told you almost nothing.

Now compare each field's interval to its published acceptance threshold. These are practitioner and regulatory lines, not vendor marketing.

On phone specifically, 90 percent or higher is excellent and anything below 75 percent should be a major concern. The reading rule is the same for every field: if the accuracy interval sits entirely above the threshold, the field passes. If the interval straddles the threshold, you do not have a verdict, you have an underpowered sample, and you enlarge n.

A point estimate without an interval is a claim, not a finding, and it will straddle the very line you needed it to answer.

The most dangerous thresholds are regulatory rather than aspirational. Most ESPs flag accounts when hard bounce rates exceed 2 percent consistently, review above 5 percent, and risk suspension above 10 percent. Google and Yahoo treat a 0.3 percent spam-complaint rate as a critical problem, with best practice below 0.1 percent. A source you renew that pushes you past that line costs weeks of remediation, so weight compliance-linked fields heaviest in the final call.

How this audit goes wrong

Most bad audits fail in one of eight predictable ways, and each one either inflates or deflates a field in a direction you can name. This is the section to keep open while you work.

• Catch-all counted as valid. SMTP returns 250 for every address on 30 to 40 percent of B2B domains. Scoring these as valid inflates email accuracy toward the vendor's marketing number. Check: quarantine catch-all as its own bucket and report it separately, never in the numerator.
• Vendor's curated segment, not your pull. Marketing quotes accuracy from curated or recently refreshed segments, while the record you pull may sit in a part of the database untouched in months. Check: sample your own live export.
• Connect rate confused with data accuracy. A low connect rate is often call-window or routing friction, voicemail and gatekeepers, not wrong numbers. False positive: condemning good numbers dialed at the wrong hour. Check: separate connect rate from answer rate and test time-of-day.
• Unstratified sample skewed to easy records. A convenience sample over-weights fresh and domestic records and overstates accuracy. Check: confirm stratum proportions match the population before computing anything.
• Point estimate with no interval. Reporting a bare percentage from a small sample hides a band that can straddle the threshold. Check: always attach the margin; if the interval crosses the line, enlarge n.
• A functionally equivalent title marked wrong. A promoted-but-same-function title triggers a false negative. Check: use current, outdated, and wrong-company buckets rather than a binary match.
• Firmographics checked against another aggregator. Two providers can share an upstream source, agree with each other, and both be wrong. Check: anchor firmographics to registry or EDGAR primary records where possible.
• Spam-trap contamination on purchased lists. Purchased lists are routinely seeded with spam-trap addresses, and hitting even one can trigger a Spamhaus listing. Check: never live-send the raw sample to measure bounce; verify through SMTP first.

The firmographic ground-truth point deserves one caution. The US has no single national company registry; each of the 50 states plus DC runs its own Secretary of State portal, and SEC EDGAR, the federal filing system since 1994, covers publicly listed companies only. So EDGAR is authoritative for public-company HQ and legal name, but for private firms you fall back to state registries, and match logic there is fuzzier.

Turning the numbers into a renew, renegotiate, or replace call

Map each field's interval against its threshold, then let the pattern of passes and misses pick the decision. The judgement is not "what is the average accuracy" but "which fields miss, and do they miss on a compliance-linked line or a nice-to-have one."

Read the grid this way. If misses are cosmetic and few, renew and note the re-audit date. If a low-stakes field misses in isolation, renegotiate price or demand a refresh SLA. If several fields miss but the failures are functional rather than compliance-linked, renegotiate scope and buy only the passing fields. If compliance-linked fields, email validity and bounce-adjacent quality, miss broadly, replace, because the cost of a blocklisting dwarfs the subscription.

Prices in one 10-provider test on the same 500 contacts ranged from 0.01 to 2.50 dollars per record while accuracy ranged from 62 to 94 percent, so cost and accuracy are only loosely linked. A cheap source that clears your thresholds is a better renewal than an expensive one that does not, and your audit is what tells them apart.

Before you sign off, run this checklist.

Keeping the audit honest over time

An accuracy audit is a snapshot of a decaying asset, so schedule the re-audit before you file the report. B2B data decays roughly 2 to 3 percent per month, which means a source that cleared your thresholds today can drift below them within two or three quarters without anyone touching it.

Set a fixed re-audit cadence, quarterly for compliance-linked fields and semi-annually for the rest, and reuse the frozen sample design so results are comparable across runs. Keep the stable record IDs from the first sample; tracking the same records over time separates genuine source degradation from sampling noise. And when the index I usually publish exclusive figures from is back online, I will rerun this against a live professional-profile index and update the spine with numbers you will not find on any vendor's homepage. Until then, the method above stands on its own, and it runs on your data rather than on anyone's marketing claim.