Measurement and Attribution Under US Privacy Constraints

Every advertising platform reports on its own performance, using its own attribution, and each one claims the conversions it can see. Add the claims together and most companies find they sold considerably more than they actually sold. That arithmetic failure sits underneath nearly every measurement problem in US marketing, and nineteen or more state privacy regimes degrading the underlying signal has made it worse rather than better. The response is not a better attribution model. It is measuring things the platforms do not control.

A capability page from Digital, Ecommerce & Performance Marketing in the United States. This page ties together the measurement threads running through the other capability pages. Last reviewed August 2026.

1. The double-counting problem

A customer sees a social ad, later clicks a search ad, and buys. Both platforms record a conversion. Neither is lying — each reports what it observed within its own attribution window — but the business had one sale.

Illustrative of overlapping platform attribution claims. The magnitude varies considerably by channel mix and attribution window settings; the mechanism is universal.

Your order count is the only number in the entire reporting stack that nobody is incentivised to inflate. Start there and work backwards.

2. Attribution is not measurement

These two words get used interchangeably and mean different things. Attribution assigns credit for a conversion that happened. Measurement establishes whether the conversion would have happened anyway.

Distinction between attribution and incrementality measurement. Operational framing.

The last row is the important one. Attribution arrives free from a party with a commercial interest in the answer. Measurement costs you time and volume, and it is the only version that survives scrutiny.

3. The three questions worth answering

Most measurement programmes fail because they try to answer everything. Three questions cover almost all real decisions.

Prioritised measurement questions. Operational judgement — the third is the most neglected and often the most useful for budget decisions.

The third question deserves attention. Average return tells you nothing about what the next dollar does. A channel returning well on average may be saturated, with incremental spend producing far less than the average implies.

4. Geo holdout testing

Geographic holdouts are the most defensible measurement available to most advertisers, because they are platform-independent and measure total business outcome rather than attributed conversions.

Design requirements for geographic holdout testing. Operational guidance; statistical design should be validated for your volume before relying on the result.

The most common failure is the second row. A control group chosen after the campaign has run is not a control group, and no amount of analysis fixes that.

5. Incrementality on branded terms

Branded search is the single highest-value incrementality test available, and the one most companies avoid because the result is often uncomfortable.

People searching your brand name have already decided to find you. Paying to appear above your own organic result may be capturing demand you already owned. The test is simple: pause branded paid search for a defined period and watch total conversions rather than paid conversions.

If paid brand conversions fall 80% and total conversions fall 4%, you have found a budget line that was mostly buying traffic you already had.

There are legitimate reasons to keep bidding — competitors bidding on your name, controlling the message, protecting against misleading ads. But those are strategic reasons, and they should be chosen knowingly rather than justified by a conversion count that was never incremental.

6. Marketing mix modelling, honestly

Marketing mix modelling has returned to fashion because it needs no user-level tracking, which makes it resilient to privacy fragmentation. That is a genuine advantage and it comes with genuine limits.

Assessment of marketing mix modelling in current conditions. Operational judgement; model quality depends heavily on data history and specification.

The useful position is that modelling and experimentation answer different questions and validate each other. A model that disagrees with a clean geo holdout should be re-examined, not defended.

7. The measurement stack that works

Layered measurement approach. The third row carries the key instruction: platform reporting is useful for optimising within a platform and should never be aggregated across platforms as a revenue claim.

8. What to report to a board

Marketing reporting frequently fails upward: it presents platform metrics to people who need business metrics, which erodes credibility every time the two diverge.

Reporting guidance. Operational judgement based on the divergence between platform-reported and business-reported outcomes.

9. What this page does not cover

Scope statement. Test design in particular benefits from statistical review — an underpowered experiment produces a confident-looking result that means nothing.

10. The 90-day measurement rebuild

Indicative sequencing. Reconciliation comes first because it is free, immediate, and usually reveals the size of the problem better than any argument about methodology.

11. Mistakes to avoid

Recurring errors in marketing measurement; illustrative.

12. What changes in 2027

Modelling replaces tracking further. As state privacy regimes multiply without a federal standard, user-level measurement continues shrinking and modelled or experimental approaches become the default rather than the supplement.

Retail media keeps offering the cleanest read. Closed-loop environments where the retailer observes purchase remain the strongest attribution available — bounded, as ever, by that retailer’s footprint.

Experimentation becomes routine. As platform-reported numbers become less defensible, running structured holdouts moves from advanced practice to something mid-market advertisers do as a matter of course.

Key Takeaways

• Never sum conversions across platforms. Each claims what it saw in its own window; added together they exceed your actual sales.
• Your order count is the only number nobody is incentivised to inflate. Start there and reconcile backwards.
• Attribution assigns credit; measurement establishes causality. The first arrives free from a party selling you media.
• Geo holdouts are the most defensible method available because they are platform-independent and measure total business outcome.
• A control group chosen after the campaign is not a control group. Design the test before spending.
• The branded search holdout is the highest-value test and the most avoided, because the result is frequently uncomfortable.
• Average return says nothing about the next dollar. Test at the margin before scaling a channel.

Frequently Asked Questions

Conclusion

Marketing measurement in the US has arrived somewhere genuinely awkward: the tracking that made attribution feel scientific has degraded, and the platforms still producing confident numbers are the ones selling the media. Meanwhile budgets are cooling and costs are rising, which makes the quality of these decisions matter more than it did when everything was cheap.

The way through is not a better model. It is accepting that the most trustworthy number in the business is the order count, that the only way to learn what a channel truly contributes is to turn it off somewhere and watch, and that a biased self-reported field on a form frequently reveals more than a dashboard built on partial data. That is a less satisfying answer than a unified attribution platform. It is also the one that holds up when someone from finance asks a hard question.