Measure Coverage Separately from Accuracy

Decide what one observation represents

Start by writing the unit of analysis in a sentence. For a company enrichment project, one observation might be a distinct operating company identified by a supplied row ID. For contact verification, one observation might be a named person and business email pair. These units cannot be combined casually. Ten contacts at one company do not establish coverage across ten companies, and five domains belonging to one organization do not automatically represent five prospects.

Create separate counters for input rows, distinct companies, distinct websites, contact records, and required field opportunities. The last counter is the number of places where a field should legitimately be researched. A manufacturing capability field may apply only to manufacturers, while a company website field applies to the entire accepted list. Record these rules before examining results so that a difficult row does not quietly become ineligible merely because it is difficult.

Also distinguish the delivered file from the accepted research population. If the buyer supplies 120 rows and 10 are duplicates, a company-level report may legitimately use 110 unique companies, but the reconciliation should still account for all 120 original rows. A clean denominator is a documented choice, not a smaller number discovered after the work is done.

Give coverage and accuracy different jobs

Use field coverage as a statement about availability: populated eligible cells divided by eligible cells. Use verified accuracy as a statement about evaluated values: supported values divided by values that received a determinate judgment. Publish the unknown or unresolved share alongside that second measure. Otherwise, an apparently high accuracy result can hide a large number of values that nobody could check.

The Government Data Quality Framework distinguishes completeness from accuracy and treats quality as fitness for purpose. Its useful contribution here is the separation of dimensions. Our audit goes further operationally by keeping a small ledger for each field: eligible, populated, reviewed, supported, contradicted, and unresolved. That ledger is a proposed buying and review method, not a statistic or an official certification.

Do not label a list simply '95% accurate' unless a reader can identify what was tested. Was it company names, active domains, email syntax, current job roles, or every claim in the dossier? A field may pass a format check while being wrong in substance. A syntactically valid URL can resolve to an unrelated business, so validity belongs in the report without replacing accuracy.

References: Government Data Quality Framework ↗

Set essential and optional fields before the audit

Turn the campaign into a field priority map. A seller of warehouse automation may require a confirmed warehousing service, the operating company's website, and a relevant business contact. A public phone number might be useful but optional. An ecommerce agency may need evidence of an active store and a supported commerce platform, while employee count is merely contextual. The priority map prevents optional completeness from compensating for a missing mandatory condition.

For each essential field, state what qualifies as populated. 'Manufacturer' may be too broad when the offer requires injection molding. A blank founder name is missing; 'team' is also missing if a named person was required. A range is valid only when the specification permits ranges. Treat 'unknown' as a status with explanatory value, not as a filled business attribute.

Agree which missing fields block acceptance, which trigger review, and which remain harmless blanks. This is a business decision rather than a universal score threshold. If a contact role is essential to routing but a direct number is not, your acceptance rule should reflect that difference even when the optional number is easier for the researcher to obtain.

Freeze a reproducible audit copy

Save the delivery being evaluated before anyone cleans it. Keep the original row IDs, file version, receipt date, and supplied field names. A reviewer should be able to find the exact value that was judged, even after corrections. If a spreadsheet application reformats phone numbers, changes date interpretation, or converts an identifier into scientific notation, the audit otherwise risks measuring the import process instead of the research.

Build an audit sheet alongside the frozen copy. Suggested columns are input_id, company_id, field_name, delivered_value, eligibility, review_status, reviewer_value, evidence_url, evidence_date, reviewer, and note. One row per field judgment is often easier to analyze than a company row with many overloaded comment cells. The format also lets you count several claims about one company without pretending they are independent companies.

Prevent correction work from overwriting the initial outcome. Add a corrected_value and correction_status, or keep a second version with the same identifiers. An audit should be able to say both that the initial delivery contained an error and that the error has since been resolved. Combining those states into a single final value erases the evidence needed to improve the process.

Sample across the risks that matter

If the delivery is small, reviewing every essential field may be practical. For a larger list, choose a sample that covers the meaningful subgroups: name-only matches, redirects, parent companies, ambiguous domains, different research methods, and records with many unknowns. Include a random component so that the sample is not merely a collection of obvious failures or unusually easy successes.

Keep purposive risk checks separate from a representative audit summary. If you intentionally select the twenty most suspicious records, the resulting error rate describes those records, not the full file. You can still learn a great deal from that review: whether a matching rule is unsafe, whether historical pages are being treated as current, or whether a particular source repeatedly describes the wrong entity.

Write down sample selection before inspecting the underlying evidence. Record sample size and which segments it represents. Do not imply statistical precision that the design does not support. For a commercial acceptance decision, a transparent small review may be more useful than a polished percentage with an unexplained sample. The practical objective is to discover whether the observed failure modes threaten the campaign.

Worked example: a field-level quality report

Consider an illustrative delivery of 100 distinct accepted companies. Every company has a website, 80 have a named business contact, and 60 have a platform value. The buyer reviews all websites and a defined sample of 20 populated contact records and 20 populated platform values. These different review denominators must remain visible; a single average would erase the distinction between file-wide availability and sampled correctness.

The website review finds 94 correct company matches, four unrelated sites, and two unresolved ownership cases. Website coverage is 100/100, or 100%. Determinate website accuracy is 94/98, approximately 95.9%, with two unresolved cases out of 100 reviewed. Supported website coverage across the entire population is 94/100. Each metric answers a different question, and the unresolved count prevents the determinate rate from looking like universal confirmation.

The contact sample has 18 supported person-company relationships, one contradicted relationship, and one unresolved relationship. Report 18/19 determinate accuracy and 1/20 unresolved, together with the 80/100 file coverage. Do not report 18/100 as overall contact accuracy, and do not project 18/19 across the file as a guaranteed result. The sample is evidence for an acceptance discussion, not a rewrite of unreviewed records.

Test the claim against the source

Open the evidence and ask whether it supports the exact delivered value. A company page saying it serves retailers does not make it a retailer. A customer logo does not establish an active current contract. A named leadership page may support a role, but it does not automatically support the person's business email. Break a compound field into the separate claims it contains before deciding that the entire field is correct.

Record entity, scope, and time. The source may describe a subsidiary while the row represents the parent, or a case study may discuss an implementation completed years earlier. A supported historical fact can be inaccurate when exported as an unqualified current state. The reviewer should identify the mismatch instead of marking the source itself as bad.

Use an unresolved category when reasonable evidence remains insufficient. An unavailable page is not proof that the delivered claim is false; an inaccessible page is also not confirmation. If two credible current sources disagree, preserve both and request adjudication. Assigning every disputed case to 'wrong' is as misleading as assigning it to 'correct,' especially when the purpose of the report is to diagnose research quality.

Worked example: the better-filled file loses

Imagine two illustrative proposals for an agency seeking active ecommerce businesses. Delivery Alder contains platform labels for all 100 companies. A complete review finds 70 active merchants on the required platform, 20 companies with only old platform traces, and 10 entities that are not merchants. Delivery Birch contains 85 platform labels and 15 explicit unknowns; review confirms 80 suitable merchants and contradicts five labels.

Alder has higher raw platform coverage: 100% versus 85%. Birch has more confirmed usable companies for this campaign: 80 versus 70. Neither percentage alone establishes value, because price, contact quality, the supplied specification, and the treatment of unknowns also matter. But the example shows why maximizing filled cells can push a team toward the wrong dataset.

An appropriate correction request differs between the files. For Alder, ask for a method change that distinguishes historical traces from active store evidence and checks business eligibility. For Birch, ask whether the unknowns can be resolved economically and whether five incorrect labels share a cause. Requesting 'higher fill rate' from both providers misses the actual weaknesses and can reward fabrication in the second file.

Explain missingness before aggregating it

A missing field may mean the researcher did not look, the source was inaccessible, the company does not disclose the fact, the input was ambiguous, or the field does not apply. These are different process states. Record a small controlled vocabulary so that the buyer can decide whether to retry, clarify the input, accept the uncertainty, or change the requested attribute.

Never improve coverage by substituting the string 'N/A' everywhere. If 'N/A' sometimes means not applicable and sometimes means not found, the denominator becomes impossible to defend. Use a separate status column and leave the attribute value empty when no supported attribute exists. For display, a readable label is fine; for calculations, it must remain distinguishable from a researched value.

Inspect whether missingness clusters in important segments. A high overall contact fill rate can conceal almost no coverage among the larger accounts the buyer cares most about. Cross-tab missing fields by relevant company type or input quality rather than assuming all blanks are interchangeable. The resulting pattern often suggests a better scope or source strategy more directly than the overall coverage figure.

Turn findings into acceptance decisions

Write acceptance rules in the language of the intended workflow. For example: release a row only when company identity, business eligibility, and required platform evidence are supported; route unresolved contact roles to manual review; permit phone numbers to remain blank when they were optional. This is more actionable than a rule requiring a single blended quality score.

Separate record correction from systemic correction. One transposed URL may need a local fix. A recurring parent-subsidiary mismatch may require rechecking an entire segment. A status-mapping bug that turns unknowns into positives can affect every downstream score. Describe the failure mechanism and affected population before asking for replacements, because replacing a few examples may leave the same error in the rest of the list.

Agree how corrected delivery will be evaluated. Keep the initial audit results, identify replaced or changed values, and recheck the error class as well as the repaired examples. Do not require unlimited rework for fields outside the accepted scope. The quality report should support a proportionate resolution based on the specification, the evidence, and the practical cost of the error.

Troubleshoot misleading quality dashboards

If accuracy improves while usable records decline, inspect the denominator first. The team may be excluding hard cases or counting only easy fields. If coverage rises abruptly after a schema change, check whether default values are now classified as populated. If every provider appears nearly perfect, inspect whether reviewers are checking truth or merely confirming that each row contains a URL.

If reviewers disagree often, compare their definitions before averaging their scores. One reviewer may accept a company biography from any date, while another requires a current management page. Another may treat a corporate group as a single account while a colleague audits legal entities separately. A short adjudication meeting with two or three disputed rows can expose this mismatch faster than adding more reviewers.

If a chart combines contacts and companies, split it. If optional fields dominate the composite score, show the essential field results separately. If the dashboard cannot show unresolved cases, improve the model before using it for procurement. A reporting system that insists on a binary answer may be less capable of representing research quality than the original annotated spreadsheet.

Run a repeatable quality review

Assign an owner for the audit definition, someone responsible for evidence review, and someone authorized to accept the delivery. These may be the same person on a small team, but the decisions still need to be distinguishable. Keep a short change log when field definitions, sample rules, or acceptance criteria change so that month-to-month comparisons remain interpretable.

At the next delivery, reuse the structure without blindly reusing the old thresholds. A different campaign may require stronger company classification and tolerate lower optional contact coverage. A higher-stakes import may justify a larger review. The durable asset is the measurement discipline: explicit units, stable identifiers, evidence-based judgments, and an honest account of what remains unknown.

• Confirm the accepted input population, unique-company count, and duplicate treatment before counting fields.
• List required fields and define what qualifies as a supported value for each one.
• Freeze the delivery and preserve the original identifiers and unmodified values.
• Choose and record the sample method, including separate risk-based reviews.
• Calculate coverage, reviewed accuracy, unresolved share, and supported coverage where the data permits.
• Read disputed sources, adjudicate differences, and retain the reason for each decision.
• Release, hold, correct, or exclude records under the agreed campaign rules.
• Archive the audit and corrected delivery together so the next review can identify recurring causes.