Digital marketing's central promise was measurability. Unlike a billboard, you could know exactly which advert produced which sale, and allocate budget accordingly.
That promise depended on tracking infrastructure that has substantially degraded. The dashboards still produce numbers, and the numbers mean considerably less than they used to.
What changed
Several things at once, over a few years.
Third-party cookie restriction. The mechanism that allowed a user to be recognised across different websites has been progressively blocked by browsers. Cross-site tracking, which underpinned most attribution models, no longer works reliably.
Mobile platform changes. Operating system-level restrictions on app tracking, requiring explicit user consent, with the majority of users declining when asked. That removed a large share of the signal available for mobile advertising measurement.
Privacy regulation. Consent requirements mean analytics only capture the subset of users who accepted tracking, and that subset is not representative.
Ad blocking. A meaningful proportion of users, skewed towards particular demographics, are invisible entirely.
Each of these individually would have degraded measurement. Together they've broken the underlying model.
The modelling substitute
Platforms responded by filling gaps with modelled estimates rather than reporting only observed data.
Where a conversion can't be attributed directly, statistical models estimate how many conversions likely occurred and assign them. This is defensible — it's better than reporting zero — and it's not measurement.
The problem is presentation. Modelled figures appear in the same interface, in the same format, as observed ones, frequently without clear distinction. A number that's an estimate produced by the same company selling you advertising is a different kind of number from one that's a count.
The incentive alignment here is worth stating plainly: the party estimating how well their advertising worked is the party selling the advertising.
The attribution model problem, which predates all this
Even when tracking worked perfectly, attribution was conceptually shaky.
A customer sees a brand advert, later searches, clicks a paid search result, doesn't buy, sees a retargeting advert, returns directly a week later and purchases. Which channel gets credit?
Last-click gives it all to the final touchpoint, which systematically overvalues channels near the end of a journey — particularly branded search, which frequently captures people who were already going to buy.
First-click overvalues the opposite. Linear models split evenly, which nobody believes. Time-decay and position-based models make assumptions dressed as analysis.
None of these is right, because attribution is fundamentally trying to allocate a single outcome across multiple causes, which has no correct answer without a counterfactual.
What actually works
The methods that have held up are the ones that create counterfactuals rather than inferring them.
Geographic holdout tests. Turn off a channel in some regions, leave it on in comparable regions, compare outcomes. Straightforward, expensive in foregone activity, and it measures actual incremental effect.
Incrementality experiments. Randomly withhold advertising from a portion of the audience and measure the difference. The gold standard, and platforms have made this progressively harder to run independently.
Marketing mix modelling. Statistical modelling of spend against outcomes at an aggregate level over time. This is an older technique that predates digital tracking and is enjoying a revival precisely because it doesn't depend on user-level data. It's imprecise, needs a lot of historical data, and it doesn't break when cookies do.
The common feature is that these measure incremental effect — what happened because of the spend — rather than correlation between exposure and outcome.
The uncomfortable finding
When companies have run proper incrementality tests, results have frequently been sobering. Channels that appeared highly effective in attribution reporting have turned out to produce much smaller incremental effects.
Branded search is the recurring example: advertising against your own brand name captures people who were already looking for you, and the attributed conversions would substantially have happened anyway.
Retargeting has similar issues. Showing adverts to people who already visited your site and measuring their conversion rate confuses selection with causation.
Several published experiments have found the incremental value of certain digital channels to be a small fraction of what attribution reported. That's not universal and it's common enough to be worth testing before assuming.
What I'd do
Accept that precise attribution isn't available and stop optimising against numbers that don't measure what they claim.
Run holdout tests on your largest spend items. Even one properly designed test per year on your biggest channel will tell you more than a year of dashboard optimisation.
Watch aggregate relationships — total spend against total revenue, over time, with a view to the lags involved. Crude and harder to fool.
And be honest internally about uncertainty. A lot of organisational dysfunction comes from teams defending budgets using numbers everyone privately knows are unreliable, because the alternative is admitting the effect is unclear. It usually is, and saying so is the beginning of measuring it properly.