Six paid-social measurement failures that quietly corrupt decisions

Most reporting damage starts before the chart: an event is vague, a denominator shifts or an attribution rule is mistaken for causality. Fix the specification before polishing the dashboard. The six failures below come in the order a report usually breaks, not in a ranking.

The failures apply to any paid-social account, including Meta, TikTok, LinkedIn, Snapchat and Pinterest, and to the analytics tool beside it, for example Google Analytics 4. The repair is the same on every platform: write the definition down and keep it with the number.

What to take away

1. Treating a column name as a metric definition

A column called "results" in one platform export and "conversions" in another can count different events, whether they come from a pixel, a server-side feed or an uploaded offline file. Windows and modelling change the count too.

The repair is a written specification: store the event, eligible population, numerator, denominator or unit, source account, query version, period, timezone, inclusions, exclusions, owner and model status. The key metrics article shows the fields for five common measures, and the measurement and reporting guide shows how to publish them.

2. Changing the denominator unnoticed

A click-through rate is clicks over impressions, cost per acquisition is spend over acquisitions and return on ad spend is revenue over spend. Each needs a compatible population. Mixing eligible impressions with all impressions, or England-only orders with UK-wide spend, gives a different measure under the same name.

The Government Analysis Function guidance on quality statistics defines comparability as the degree to which statistics can be compared over time, region or another domain, and coherence as the use of the same concepts and harmonised methods across outputs. Campaign figures are not official statistics, but the test transfers. Repair: version the denominator and label geography beside the result.

3. Calling attributed credit incremental impact

Attribution allocates credit among ad interactions inside click-through and view-through windows, such as a 7-day click window against a 1-day click window. That answers a different question from the counterfactual: the outcome in the absence of the campaign, against which any observed change is judged.

Causal confidence usually needs a strong comparison group or baseline set before launch. Typical designs are a conversion lift test, a geographic holdout, a randomised A/B test or media-mix modelling. Repair: label platform attribution and causal estimates separately, with a qualified evaluator for the latter. The attribution methods comparison and the return on investment article explain why ROI stays blank until the counterfactual is defined.

4. Reporting before lag and revisions are visible

Conversions arrive after the interaction that prompted them, so a young period always looks weaker than it will. Statistical practice handles this with scheduled revisions: planned amendments that incorporate data unavailable at first publication, and a reliability measure comparing early estimates with later ones.

Borrow the habit. Repair: publish extraction time, maturity status and revision schedule, and show the earlier value alongside any restatement.

5. Hiding observed and modelled events in one claim

An estimate is judged by its closeness to the unknown true value, which is accuracy. A modelled conversion is an estimate; an observed one is a record. Adding them into a single figure hides which part carries the uncertainty.

Repair: expose model status and methodology, and state when an export does not offer a usable split. Never rename modelled platform events as verified internal orders. The service standards guide shows how to write a metric card that carries this flag.

6. Letting measurement override release controls

A desired report does not create permission to use pixels, tags or audience data. Regulation 6 of PECR bars storing or accessing information in a user's terminal equipment, which is what a platform pixel or analytics tag does.

The ICO guidance on storage and access technologies was finalised on 29 April 2026.

Repair: keep privacy and PECR, claim review, accessibility, regulated-product and finance decisions as separate gates. The data protection article covers who acts when a rights request or incident arrives.

Apply the repair test

For each reported number, ask whether another authorised analyst can recreate it from frozen sources and the named query. If not, mark it unavailable. Record the owner, uncertainty and action trigger. A precise chart without those fields is presentation, not reproducible evidence.

Before you act

Common questions

Why is a column name not a metric definition?

The same column name can count different events in different accounts or exports, because settings, windows and modelling differ. The repair is to store the event, eligible population, numerator, denominator and source account, so the number can be recreated and compared.

How should attributed credit be reported?

Attribution credit answers a different question from what would have happened without the campaign. Label platform attribution and causal estimates separately, and use a qualified evaluator for the causal estimate, since it needs a counterfactual that attribution alone cannot supply.

What should accompany any reported number?

Ask whether another authorised analyst can recreate it from frozen sources and the named query. If not, mark it unavailable. Record the owner, uncertainty and action trigger. A precise chart without those fields is presentation, not reproducible evidence.