Scope: what this audit covers and what it does not
The question this audit answers is whether the business's budget decisions rest on measurement that fits them, from credit to cause, and which gap in the evidence matters most. It covers how marketing is judged rather than how tags are installed: the KPI tree, one agreed count of conversions, new and returning customers, platform and analytics gaps, data quality, the AttributionAssigning conversion credit to touchpoints. approach, self-reported answers, offline media, IncrementalityConversions caused by marketing. tests, mix models, the change log and reporting.
It does not score the collection of data. Tags, consent behaviour and the configuration of the analytics property are scored by the analytics implementation audit, and that audit should come first when nobody trusts the raw numbers. Keeping measurement trustworthy through releases, outages and staff changes is run by the measurement governance playbook, and choosing an approach with a staged roadmap belongs to the guide to choosing an attribution approach. This scorecard rates the result of that work; it does not repeat it.
A single paid search account, including how its conversion actions are set up, is scored by the Google Ads account audit. The attribution models framework summarises the models the credit rows refer to. Every scorecard sits in the audit library.
Before you score: access and evidence
Ask for read access to the analytics property, each ad platform, the order system or CRM that holds the business's own record of sales or leads, and the folder where reports and any test or model results are kept. Without the order system or CRM you cannot judge the source of truth, so the gate stays provisional until you have it. Fix an evidence windowThe fixed period of data an audit scores, agreed before scoring, long enough to cover a full cycle for the business and kept the same at every rescore. that spans at least one full budget cycle, and keep the same window for every rescore.
Collect:
- the KPI tree or the written list of measures, and the business outcome each serves;
- the source of truth for orders or leads and how platforms and analytics compare with it;
- the attribution settings in each platform and analytics, and the model used in reporting;
- any incrementality tests, lift studies or mix models run, with their plans and results;
- the measurement change log, if one exists;
- the last few performance reports and the decisions recorded from them.
Before reading any results, list the budget decisions the business has taken in the window and the evidence each one cited. That list is what the audit is really testing: whether the evidence behind each decision could have supported it.
Criteria and scoring
Rate each row missing (absent or broken), partial (present but incomplete, inconsistent or out of date) or solid (matches the "what good looks like" column). The scale is your judgement against that column, not a number. Rows marked gate decide the band on their own when they score missing, because without them no comparison between channels holds.
What is measured
| Criterion | What good looks like | Score guidance |
|---|---|---|
| KPI tree (gate) | Channel measures trace to the business outcome through a written KPI tree that leadership recognises | Missing: each channel reports its own favourite measure / Partial: a tree exists but leadership does not use it, or the arithmetic between levels is unstated / Solid: written, recognised and used to read movements |
| New and returning customers | Acquisition is judged on new customers, with returning customers reported separately | Missing: analytics new users treated as new customers / Partial: status recorded but not reported by channel / Solid: status taken from the business's records at conversion, reported separately |
The KPI trees guide explains how to build a tree that ties each KPIA metric tied directly to an objective. to the outcome, with the arithmetic between levels written down, and the KPI tree framework gives the general method. The second row is scored against the guide to new and returning customers, which explains why a new user in analytics is not a new customer.
Source of truth
| Criterion | What good looks like | Score guidance |
|---|---|---|
| One count of conversions (gate) | One agreed count of orders or leads, from the business's own records, is the yardstick every platform is judged against | Missing: each platform's claims are added up or compared with each other / Partial: a business count exists but reports still lead with platform figures / Solid: one named system of record for each outcome, used in every report |
| Platform gaps explained | Differences between platforms, analytics and the business's records are measured and their causes known | Missing: gaps unknown or argued about / Partial: measured once, causes guessed / Solid: a standing reconciliation with each gap explained and stable |
| Data quality checks | Breaks, gaps and spikes in the data are caught by routine checks before they reach a report | Missing: problems found by the reader of a report / Partial: checks run after something looks wrong / Solid: a routine check, with causes fixed rather than reports patched |
The gate follows the governance section of the guide to choosing an attribution approach, which names one system of record per outcome. The guide to why platforms and analytics disagree draws on Google Analytics Help and Google Ads Help to explain why platform totals summed together exceed any single count of orders, and sets out the reconciliation method. Data quality is scored against the marketing data quality guide.
Credit
| Criterion | What good looks like | Score guidance |
|---|---|---|
| Attribution approach fit | The attribution approach matches the decision, the data volume and the team's maturity, and its limits are written down | Missing: whatever each tool defaults to, unlabelled / Partial: one reporting model, but used for budget questions it cannot answer / Solid: approach chosen for the decision, figures labelled with source and model, limits recorded |
| Self-reported answers | How-did-you-hear answers are collected and read beside click-based credit | Missing: not asked / Partial: asked but uncoded, or treated as proof of cause / Solid: asked well, coded and compared with the tracked source on the same record |
| Offline media measured | Television, radio, print and podcast spend has a measurement plan rather than none | Missing: offline spend judged on clicks, or not at all / Partial: response codes only / Solid: evidence chosen for the money at stake, planned before the campaign (or no offline spend) |
The credit rows are scored against the guide to choosing an attribution approach, the self-reported attribution guide and the guide to measuring TV, radio, podcasts and print. A business with no offline spend scores the third row solid and moves on.
Cause
| Criterion | What good looks like | Score guidance |
|---|---|---|
| Incrementality tests | The largest spend lines have holdout or lift evidence, planned before the results | Missing: budgets set on attributed figures alone / Partial: tests run but read after the fact, or on small lines only / Solid: a planned test with a Holdout GroupA randomly chosen part of an audience kept from exposure to measure what exposure caused. or lift design behind the largest lines, with dated results |
| Mix models used where they fit | A mix model is used only with enough history and varied spend, and is checked before it is trusted | Missing: a model's budget split acted on unreviewed, or a model forced onto thin data / Partial: reviewed, but not calibrated against experiments / Solid: used where the data supports it, reviewed and calibrated, or deliberately not used |
Score the first row against the incrementality testing guide, which explains why attribution cannot say what would have happened anyway. The second row is scored against the guide to evaluating a marketing mix model, and the marketing mix modelling basics explain what an MMMStatistical model of channel impact. needs before it can be trusted. Neither guide's sizing or fit advice is repeated here.
Change and reporting
| Criterion | What good looks like | Score guidance |
|---|---|---|
| Measurement change log | Changes to tracking, definitions and settings are logged with dates, so trends can be read across them | Missing: no record / Partial: some changes noted in tickets or emails / Solid: one log, read before anyone explains a movement |
| Reports that end in decisions | Reports compare like with like and end in decisions with owners | Missing: platform exports with no comparison / Partial: plan against actual but no recorded decisions / Solid: like-for-like comparison, reconciled figures and a decision log |
The tutorial on keeping a measurement change log explains why a measurement change can look like a business change, and the performance reporting guide sets out the report structure and the decision log.
How to score and prioritise
Score the two gates first. If there is no KPI tree, channel figures cannot be tied to the outcome and the credit and cause rows describe evidence for an unstated goal. If there is no single count of conversions, every channel is judged on its own numbers, so the platform gap, attribution and test rows are provisional. Mark them so.
Then rank the remaining gaps in this order:
- The largest spend lines judged on credit alone, because a holdout or lift test there changes the most money.
- The explanations around the numbers: platform gaps, data quality and the change log, so a movement can be trusted before anyone acts on it.
- Self-reported answers, offline media and mix models next.
- Report layout last, unless it hides a problem the other rows need to see.
Within each group, rank by the budget the gap touches. A missing test on the line that takes most of the spend outranks a perfect self-reported survey on a small channel. Ranking gaps by the spend they touch applies the Pareto principle to marketing budgets, so the few lines carrying most of the money are fixed first.
Reading the result: what to do with each band
Red means a gate scored missing. Hold budget moves that rest on platform figures, agree the KPI tree with leadership and name one system of record for each outcome, and restate the last few reports against it before drawing conclusions. Rescore after the first reporting cycle on the agreed count.
Amber means the gates hold at least partially but several other rows are partial or missing. This is the usual outcome. Plan the first test on the largest spend line, start the change log and the reconciliation table, and work through the rest in order so each new layer of evidence is in place before the next budget cycle.
Green means the gates are solid and most rows are too. Move effort into a steady testing rhythm and into calibrating any model with the test results, and rescore each quarter or after any large tracking change.
From findings to a prioritised fix list
Write each partial or missing row as a finding in plain words: what you saw, the evidence, and why it matters. Then turn the findings into one fix list with a line per fix: the finding, the fix, the owner, the guide that explains it and the rescore date. Order the list by the priorities above, not by how easy each fix is.
Measurement fixes need owners on both sides of the business. The KPI tree belongs with leadership and finance as much as marketing; the source of truth with whoever owns the order system or CRM; reconciliation, data quality and the change log with the analytics lead; tests and models with the analyst and the budget holder together. Present the band, the gate findings and the first test to the budget holder, and keep the full scorecard as the working copy for the rescore. The academy capstone on a measurement governance plan for a used camera dealer practises turning an audit like this one into a fix list inside a year's plan for events, consent, outages and corrections.
Illustrative example
A mattress retailer audits its evidence before a budget review
Joxtera Mattresses, an invented mattress retailer spending about £90,000 a month across paid search, paid social and display, runs the audit before its annual budget review. The KPI tree scores partial: revenue is the stated goal, but channel teams report their own measures. One count of conversions scores missing: the platforms together report 3,100 orders for the quarter against 1,900 in the order system, and the board pack adds them up. Incrementality scores missing: nobody has tested what branded search or social add, though they take most of the budget. The change log scores missing, and the audit finds that a tracking change last spring, which began counting a second purchase event, explains a jump everyone had credited to new creative. Self-reported answers score partial, collected at checkout but never coded. The band is red. The fix list puts one count of orders from the order system first, a holdout test on branded search second and a change log, backfilled to cover the tracking change, third, and holds the planned shift of budget into social until the test reads out.
Common mistakes
- Starting with the model. Arguing about attribution models before agreeing a KPI tree and one count of conversions polishes numbers that cannot be compared. Score the gates first.
- Adding up platform claims. Each platform credits the conversions it touched, so the sum overstates the total. Score the source of truth on the business's own records.
- Treating attribution as proof of cause. Credit shows which touches were present, not what would have happened without them. Score the cause rows on tests, not on attributed figures.
- Trusting a mix model because it produced an answer. A model always produces contributions. Score it on review and calibration, not on whether it exists.
- Reading a trend across a tracking change. A jump that coincides with a new tag is a measurement change until proven otherwise. Check the log before scoring any movement as real.
- Auditing the reports instead of the decisions. A beautiful report that changes nothing scores no better than a plain one. Score the reporting row on recorded decisions.
