Landing page measurement goes wrong when the denominator changes and the label does not

The most damaging landing page software measurement mistakes change what a number means while leaving its label intact. A decision maker then sees a clean trend that may combine different populations, dates, models or privacy conditions.

What to take away

A metric label can stay the same while its denominator changes, so the number no longer means what it did.
Model output distributes credit under rules and does not establish a counterfactual, so call it platform-attributed.
Missing touchpoints from consent, device changes and failed joins should be reported as an observable coverage gap.
Store platform model settings, lookback windows and dates with every extract because definitions are supplier evidence.
Silent corrections hide the old value, so record old and new figures, reason, approver and decisions to revisit.

How the measurement errors were selected

Research date: 6 September 2026.

Scope: internal measurement and reporting for a landing page operated in England, using UK government, regulator and statistical guidance plus a dated first-party platform record.

Inclusions: errors that can alter a metric, hide uncertainty or overstate what the evidence proves.

Exclusions: invented rates, supplier rankings, universal thresholds and mistakes that apply only to one unverified product edition.

Ranking status: non-ranked. Order follows the reporting process.

Conflicts: no supplier payment, affiliate link, product access or account data informed this desk research.

1. Reusing a label after its denominator changes

A form completion rate based on form starts cannot be compared with one based on page sessions. Put event, eligible population, numerator, denominator, time period, timezone and exclusions in the metric card.

The UK Statistics Authority's standards for statistical quality apply to official statistics. Their insistence on suitable data, sound methods and transparent limitations is a useful warning for any report.

2. Treating assigned credit as causal impact

Last-touch, fractional and platform models distribute credit under rules. They do not create a counterfactual. HM Treasury's impact-evaluation quality guidance explains, within government policy evaluation, why observed change alone cannot establish causality. Use "platform-attributed" for model output and reserve "incremental" for evidence from a proportionate causal design.

3. Hiding missing paths

Consent choices, browser settings, device changes, failed joins and offline outcomes can leave touchpoints missing. The ICO's final statistical-purposes exception guidance sets narrow conditions and does not authorise advertising measurement generally. Report the observable coverage gap. Do not replace it with an unsupported uplift.

4. Mixing reporting dates and model settings

An interaction-date series can differ from an outcome-date series, and a new lookback window changes which touches qualify. Google's current Analytics attribution record documents product-specific models, eligible channels and windows. Store those settings with every extract. A platform definition is supplier evidence, not an England benchmark.

5. Copying figures into a dashboard without lineage

Manual transfers can detach a value from its query, input snapshot and correction history. Government Analysis Function reproducible-pipeline guidance is aimed at government analysts, but its version-control, validation, documentation and peer-review practices provide a practical model. If a manual step remains, name it and retain the source record.

6. Correcting a number silently

A quiet overwrite can leave earlier decisions based on a value that no longer exists. Record the old and new value, affected period, reason, approver, time and decisions to revisit. Government guidance on communicating quality, uncertainty and change addresses official statistics and supports making material quality changes visible.

Repair the report

Freeze each metric card and query version. Add explicit missing, unattributed, corrected and not comparable states. Leave decision thresholds blank until the buyer approves the evidence condition and action. A named statistician and measurement reviewer should sign the method; privacy, PECR and finance reviewers retain their separate gates.

Before you act

Keep event, population, numerator, denominator, period and exclusions in the metric card.
Reserve incremental for evidence from a proportionate causal design.
Report the observable coverage gap instead of an unsupported uplift.
Store model settings and windows with every extract.
Record old and new values, reason, approver and time for corrections.
Freeze each metric card and query version before reporting.

Common questions

Why is a form completion rate based on form starts not comparable with one based on page sessions?

The two rates use different denominators, so the same label covers different populations. The article says to keep event, eligible population, numerator, denominator, time period, timezone and exclusions in the metric card, because a clean trend may combine different populations, dates, models or privacy conditions.

What is the difference between platform-attributed and incremental in this article?

Platform-attributed describes output from last-touch, fractional and platform models that distribute credit under rules. Incremental is reserved for evidence from a proportionate causal design. The article notes that observed change alone cannot establish causality, citing HM Treasury impact-evaluation quality guidance.

What should be recorded when a reported number is corrected?

Record the old and new value, affected period, reason, approver, time and decisions to revisit. The article says a quiet overwrite can leave earlier decisions based on a value that no longer exists, and supports making material quality changes visible.