Six social commerce measurement mistakes, from young orders to worker league tables

These six mistakes can change a commercial decision without any underlying change in orders. Current UK authority guidance and a first-party TikTok reporting record inform the corrections. The list is not ranked, and it claims no prevalence or performance effect.

Invented examples of fraud, uplift or customer harm are excluded. Each entry names a reproducible correction and a source that supports the underlying evidence boundary.

What to take away

1. Reporting a rate without its denominator

A conversion rate is meaningless until visits, people, clicks or another eligible population is defined. Publish numerator, denominator, exclusions, period and source. The ONS quality definition in Quality in Official Statistics treats clarity, coherence and comparability as dimensions of fitness for purpose. Retail metrics are not official statistics, but incompatible bases still mislead.

Illustrative worked example: numerator 120 orders, denominator 4,000 eligible visits, exclusions 300 automated sessions and 50 staff visits, period 1 to 31 March 2026, source platform orders export and analytics export.

2. Mixing young and mature orders

Placed orders can later be cancelled, returned, refunded or disputed. GOV.UK's returns record for Accepting returns and giving refunds explains treatment for covered online purchases, with exceptions. Define the qualifying state and cohort closure. Show provisional results separately rather than comparing them with settled history.

3. Calling attribution causal lift

Platform or last-touch rules allocate observed outcomes. The 2026 Magenta Book, Central government guidance on evaluation, explains that experimental and quasi-experimental impact methods compare against an unaffected group or period. Without a defensible counterfactual, describe attribution and avoid an incrementality claim.

4. Forcing unlike reports to agree

TikTok's reporting-discrepancy guidance identifies differences in attribution method, lookback window, timezone and reporting date. Preserve each rule, align what can be aligned and show residual exceptions. Do not allocate an unexplained gap across channels.

5. Collecting every identifier available

More customer-level data can increase exposure without improving the decision. The ICO sets out in its data-minimisation record in the Guide to the data protection principles that collected personal information must be sufficient and relevant for the stated purpose, without exceeding what is needed. Start from the report's decision, then justify each field, access group and retention period.

6. Turning an operations dashboard into a worker league table

Individual response or sales figures may reflect shift, stock, allocation or case difficulty. The ICO's worker-monitoring guidance in Monitoring workers discusses data analytics used to infer performance and says monitoring should be considered in design. Obtain current privacy and employment review before identifiable use.

Correction rule

When a mistake is found, preserve the old definition, identify affected periods and rerun only where source records support it. Mark a series break rather than silently rewriting history. Record who approved the correction and which decision must be revisited.

Diagnose the failure at source, transformation and presentation layers. A missing platform event, a broken join and a filtered chart can produce the same visible zero but require different repairs. Add a controlled test record and expected output, then confirm it is excluded from commercial totals.

Keep an evidence pack with the metric definition, source extracts, transformation version, active filters, exception table, old output, corrected output and reviewer decision. Record the earliest affected date and whether downstream budgets, creator reports or customer actions used the faulty result.

Assign the repair to the layer that failed. A source owner addresses missing events, an analyst fixes a join, and a report owner corrects presentation. The final checker confirms that test records behave as expected and stay outside commercial totals. Close the incident only after related dashboards and scheduled exports use the approved definition.

Reissue a decision only after the corrected population and order maturity are stated. Notify readers who received the old result where the difference was material. Keep both report versions and an explanation; an erased mistake cannot support learning or audit.

A report can remain incomplete. Missing evidence is preferable to a precise dashboard that combines different populations, immature outcomes or unjustified personal data.

Before you act

Common questions

Why is a conversion rate meaningless without a denominator?

Define visits, people, clicks or another eligible population before reporting a rate.

What should you do when a measurement mistake is found?

Preserve the old definition, mark a series break and record who approved the correction.

How should you handle reporting discrepancies between platforms?

Preserve each platform's rule, align what can be aligned and show residual exceptions.