Creator discovery platforms measurement: quality controls

Measurement for creator discovery platforms should begin with a state register, not reach and engagement charts. For an England buyer, the useful question is whether an authorised record moved through a defined workflow on adequate evidence.

A profile being found is not approval to contact its subject. A published item is not proof of an earned response, retained value or causal effect.

This guide supplies the method and the control register. It does not contain an England benchmark, campaign dataset or supplier assessment, and it applies the same controls across named creator discovery platforms rather than scoring them. A named UK statistician or evaluation specialist must review the calculations; privacy and PECR, safeguarding, creator authority and rights, advertising, security, accessibility and commercial reviewers must clear their own gates before publication.

Start with the decision and identity unit

Write one decision at the top of the register. An example is: should the buyer continue limited discovery for the stated England campaign brief? Do not quietly broaden this into whether a creator is authentic, suitable or willing. The workflow cannot establish those characteristics from a profile or score.

Choose the identity unit before counting: a source record, an evidenced creator account, a represented creator dossier, a proposed contact instruction or a contracted creator. It might also be an approved asset or a published item.

Keep the source identifier beside the buyer's internal identifier. Never merge accounts just because names or images match, and log later merges and splits as versioned decisions.

The UK Statistics Authority's standards for official statistics require suitable sources, sound methods and openness about limitations. They govern official statistics, but the principles are a useful quality test for this private register. They do not certify its contents.

Use states that cannot collapse into one another

Each record has one current operational state and a retained history. The allowed sequence is deliberately restrictive:

There is no automatic edge from discovery to outreach. The ICO says that an open social-media page does not make its personal information freely reusable for direct marketing in its lead-generation guidance. UK GDPR and the Data Protection Act 2018 set the lawful basis, transparency and rights for the processing itself, while PECR governs the contact route. A privacy and PECR specialist must decide the actual processing and contact route.

Give every measure a complete card

A metric name alone is not a definition.

The three minimum discovery measures are evidence admission, state reconciliation and correction propagation. Admission admits records, state reconciliation decides whether the register view can be trusted, and correction propagation releases blocked use. All three must pass before the next authorised step.

Shortlist size, follower counts and engagement labels cannot answer those control questions. Outcome measures such as shortlisting or contacting need separate eligibility and authority before any combined funnel.

Per measure, record the decision informed, eligible population, identity unit, source and provenance, numerator and denominator or other unit.

Record period and timezone, inclusion and exclusion rules, duplicate, missing and late-data treatment, and claimed or verified per field. Record privacy and PECR authority; query and register version; uncertainty; owner; correction or recheck trigger.

Leave it blank until the buyer has evidence and risk authority to set it.

The Government Analysis Function advises authors to communicate quality, uncertainty and change so readers do not draw conclusions the data cannot support. This guidance addresses official statistics, yet its insistence on transparent limitations is directly useful here.

Evidence admission completeness

Decision: may submitted records enter limited discovery review? Population: records proposed during declared reporting period. Identity: buyer record ID paired with source ID. Numerator: eligible records with source, access basis, field definition, timestamp and claimed/verified status; also correction route and review date. Denominator: all eligible proposals. Period and timezone: [ ]. Exclude documented synthetic records and withdrawn submissions.

Duplicates stay visible; only the designated survivor counts in the numerator. Missing fields fail; late records enter the next stated cut-off. Source: register query [ ], version [ ]. Privacy authority and owner: [ ]. Uncertainty arises from unresolved identity and source-access questions. Threshold: [ ]. Recheck on source change, challenge, expiry or a changed purpose.

Unresolved duplicates stay in the denominator.

Illustrative example: 82 proposals are eligible and 61 carry provenance, access basis, field definition, source timestamp, status, correction route and expiry. The admission ratio is 61 divided by 82, or 74.4 per cent.

Gather one full period of baseline proposals before fixing the threshold, confirm the register version and record the privacy authority. A pilot floor might sit between 85 and 95 per cent complete records, then be adjusted after review.

Platform exports can include a YouTube Studio CSV or a Meta Business Suite export. Company details can be checked against Companies House where relevant.

State reconciliation exceptions

Decision: can the buyer rely on the current workflow view?

Population: records expected in master register at cut-off, with an immutable internal event ID. Unit: count of unmatched, conflicting, impossible or unversioned state events. Period and timezone: [ ]. Include late events received before the published close; retain later arrivals as adjustments. Do not discard duplicates until lineage is resolved.

Source: master register, supplier export, contract ledger, each with version [ ]; claim status: observed reconciliation result, not creator quality. Privacy authority, owner, threshold: [ ]; recheck after import, schema revision, correction, suppression, export.

Illustrative example: of 500 expected events, 18 unmatched IDs, 6 impossible transitions, 9 conflicting states and 7 unversioned events give an invalid count of 40, or 8 per cent. The four failure types are counted before the register can be trusted for a workflow view. Missing supplier or internal events remain open exceptions.

Correction propagation time

Decision: should affected use remain blocked? Population: accepted correction or withdrawal instructions requiring downstream action; Identity: case ID; Unit: elapsed time from approved instruction to verified completion by named recipients. Clock, timezone, permitted pauses: [ ]; exclude rejected requests only when reason and reviewer are recorded.

Missing recipient evidence leaves case open; duplicate notices link to first valid case; late confirmation changes next report and preserves prior release. Sources: case log, recipient evidence, publication check, versions [ ]; recheck on challenge, failed propagation, new recipient discovery. Owner and threshold: [ ]; uncertainty covers unobservable third-party copies.

Correction propagation stays open until every named recipient confirms verified completion.

The ICO's accuracy guidance supports documenting information sources and challenges. The page is marked under review following the Data (Use and Access) Act, so it needs a publication-day check.

Apply the controls to named creator discovery platforms

The register is platform-agnostic, but the evidence is not. A buyer may run it against Modash, HypeAuditor, CreatorIQ, Traackr, Upfluence, Aspire, GRIN or Julius, and each delivers a different export with different definitions behind the columns.

Ask for the field dictionary and the metric definition: how a follower count was sourced, how engagement rate is calculated, which attribution window applies to a conversion, and whether a profile is a raw platform record or a merged creator dossier. Ask which API or export version produced the file, when the snapshot was cut, and what the export leaves out — deactivated accounts, private profiles, suppressed records and deleted campaigns are common gaps.

Then re-run admission, state reconciliation and correction propagation on that file before any activity figure reaches a dashboard. A platform that cannot answer those questions fails the control, however polished its charts look.

Put reconciliation and exceptions before activity

The first dashboard panel should answer whether the data can be trusted. Show unmatched event IDs, impossible transitions, expired evidence, unresolved duplicates, missing privacy authority, overdue rechecks and open correction cases. The second panel can show evidence admission and synthetic-test status. Only after those panels clear should activity counts appear.

Store a data contract beside the display: field name, definition, type, allowed state, system of record; add owner, update rule, retention rule. Record extraction cut-off, timezone, query hash or version, source snapshots; log row counts, suppression decisions, corrections since the previous issue.

The AQuA Book distinguishes verification from validation and calls for proportionate assurance across an analysis's life. This is government analysis guidance, not approval of a buyer's dashboard.

Security evidence matters because a polished chart can hide missing logs. NCSC's cloud audit guidance suggests questions about which events customers can see, log format and retention. Apply those questions to the configured service and preserve access controls around personal and commercial data.

Separate observation, credit and effect

Reporting needs distinct evidence columns:

No method wins universally. The decision, feasible design, data quality and risk determine what can be said. HM Treasury's updated Magenta Book separates process, impact and value-for-money evaluation. Its companion quality guidance for impact evaluation explains why monitoring change does not establish that an intervention caused it.

Paid placement, an earned editorial response and an audience reaction require separate evidence. So do reconciled revenue, retained commercial value and incremental contribution. Platform credit cannot bridge those gaps by itself, and several systems may credit the same event. If the counterfactual is absent or damaged, label the result observational or attributed and stop short of causal language.

Treat advertising and control evidence as reportable states

Campaign reporting must not obscure whether commercial content was recognisable and supported. The CMA's current brand guidance on social-media endorsements says businesses should check content and act when labelling is wrong, and the CAP Code rules on the recognition of marketing communications, enforced by the Advertising Standards Authority, apply to paid and gifted creator content. Record the agreed incentive, disclosure instruction, observed rendering, claim approval, monitor result and correction state. A disclosure observation is not a measure of campaign effectiveness.

Likewise, accessibility is an acceptance gate, not a decorative score. GOV.UK's assistive-technology testing guidance recommends documenting the browser, operating system, assistive technology and task. Its scope is government services; an England buyer can adapt the reproducibility principle while arranging competent review for the actual creator and audience tasks.

Safeguarding, creator authority, rights, privacy, advertising, security and accessibility cannot compensate for one another. A commercial outcome does not release a failed gate. Record who stopped the workflow, what evidence is required, who may restart it and when the evidence expires.

Decide whether an external benchmark is admissible

There is no verified England benchmark in the sources used for this guide for creator-platform admission, contact, response, publication, reach or return. Do not manufacture one from a global supplier report or a different campaign population.

Before comparing any external rate, record publisher, research date, geography, platform/service configuration and creator/audience population. Record selection frame, sample, campaign/product category, channel/format and paid/earned status.

Record identity rule, numerator, denominator, period and missing-data treatment. Record corrections, privacy coverage, uncertainty and funding. If one field prevents like-for-like interpretation, publish "no admissible benchmark"; an external rate remains context, never the buyer's target.

Issue a versioned decision report

Each reporting release should retain its cut-off, state schema, data contract, queries, source snapshots, adjustments, unresolved exceptions, uncertainty and approvals. Do not overwrite an earlier number after a late record or correction. Show the old issue, the revised issue and the reason.

The report ends with a decision, not a celebratory summary: continue limited discovery, hold contact, pause publication, correct and recheck, verify export, or close. Name the owner and the next evidence date. Publication remains on hold until the named statistical, causal, privacy and PECR, safeguarding, rights, advertising, security, accessibility and commercial reviewers confirm that their separate evidence is current.