Measuring marketing when attribution stops working

Several things happened at once, which is why the degradation felt gradual and then sudden.

Browser restrictions on cross-site and long-lived client-side storage removed the identity thread that joined a view in one place to a purchase in another. Apple's App Tracking Transparency prompt cut the signal available from app environments. Consent requirements mean a meaningful share of visitors are never measured at all, and in regions where consent signaling is mandatory for advertisers using Google's ad services, missing configuration shows up as quiet data loss rather than an error. Meanwhile the journey itself got longer and more fragmented across devices, marketplaces, social feeds and dark social links that carry no referrer.

None of that is reversible. The useful response is not a better tracking trick. It is to stop asking a click-path tool a causal question it was never able to answer.

Use all four. The common failure is picking one and treating it as the truth, which is how a team ends up either cutting everything that lacks a last click, or defending everything on the grounds that attribution is unreliable.

Better measurement design will not rescue bad inputs. Work through this before anything else.

1. Own the identifier. Capture click identifiers from each platform at landing, persist them with the order or lead record. This is what lets you reconcile and what lets you send conversions back later.
2. Record the entry context. Landing page and referrer stored on the order itself. Commerce platforms often capture this natively; make sure it is retained and queryable, not just displayed.
3. Send conversions server to server. Platform conversion APIs give a more complete and more durable record than a browser tag alone.
4. Deduplicate properly. Where the same conversion is sent from both browser and server, a shared event identifier must travel with both. Without it you double-count and your targets drift.
5. Hash and send first-party data where you have the rights and consent to do so. It improves match quality on both the measurement and the optimization side.
6. Close the quality loop for lead generation. Push qualified-lead and closed-won status back to the platforms. Optimizing to form submissions buys form submissions.
7. Configure consent correctly for every region you serve, then check what proportion of traffic is consented. That number is now a core data quality metric.
8. Standardise campaign tagging and enforce it. Inconsistent naming fragments channel reporting into near-duplicates that quietly distort every comparison.
9. Pick one system of record for revenue and one for leads. Reconcile everything else to it.

Much of this is integration work rather than marketing work, and it is where workflow automation earns its place: the steps that have to happen on every order are the steps that get skipped when someone is doing them by hand.

Blended metrics have one enormous advantage. They do not care how credit is assigned.

• Marketing efficiency ratio: total revenue over total marketing spend, from your system of record.
• New customer acquisition cost: total acquisition spend over new customers only.
• Contribution margin after marketing: the number that says whether the week paid for itself.
• Baseline check: revenue arriving with no paid click attached, tracked as a trend.

Run them weekly, against the same definitions, with a note on what changed. The discipline is to resist explaining a weekly move. Most weekly moves are noise plus one operational event, and that event is usually a promotion, a stock-out or a release.

Geographic holdouts have become the practical workhorse. The method:

1. Pick the question. One channel, one spend level, one outcome. "Does this upper-funnel video spend produce incremental revenue?"
2. Build matched groups. Split comparable regions by historic revenue pattern and seasonality, not by size alone. More regions per group gives a more stable read than two large ones.
3. Record the pre-period. Several weeks of baseline on both groups, with no changes. Skipping this makes the result arguable.
4. Change one thing. Withhold or add the channel in one group. Leave everything else, including promotions, identical.
5. Run through a full cycle, long enough to cover the lag between exposure and purchase in your category.
6. Compare totals, not tracked conversions. Total revenue, total orders, branded search volume. The point is to escape click tracking, so do not grade on it.
7. Write the conclusion as a decision. Keep, cut, or change the level. Then schedule the re-test, because the answer decays as the market changes.

Where the platforms offer their own lift or conversion-lift tests, they are worth running, with the caveat that the vendor designed and graded the exam.

This is the part that gets skipped and it is the part that makes the system hold.

Write down, before the data arrives: the metric, the threshold, the window, and the action at each outcome. "If blended efficiency holds above the ceiling for three consecutive weeks we raise budget by a step." "If the holdout shows no detectable difference we cut this channel and retest in two quarters."

Without pre-committed rules, a measurement stack becomes a debating resource. Everyone finds the instrument that supports their position, which is easy when you have four of them.

• Summing platform-reported revenue and reporting a number larger than the business took.
• Comparing a platform's figure against analytics and treating the discrepancy as a bug. They use different windows and different rules; they are supposed to disagree. Reconcile against the system of record instead.
• Changing attribution settings mid-quarter, creating a seam in the history that nobody documents.
• Running an experiment during a promotion and reading the result as a channel effect.
• Over-fitting a model to a short history and then defending budget with it.
• Treating consented traffic as the whole picture. Scale up to population level consciously, or compare like with like over time, but do not quietly assume the unconsented behave identically.
• Abandoning platform data entirely. The bidding systems need conversion signal. Starving them to protect measurement purity costs you real performance.

Measurement is the first phase, before any media plan, because everything proposed afterwards is judged by it. We audit the data path, fix the integration gaps, agree the blended steering set and the first experiment, and write the decision rules down. That diagnostic sits inside digital strategy and advisory, with the data plumbing handled as automation and integration work. Agree what better means before the work starts.