Did the ad even work? The attribution problem

A line usually attributed to a nineteenth-century retailer captures it: half the money spent on advertising is wasted, and the trouble is not knowing which half. Digital advertising was supposed to end that, because online, unlike a billboard, you can see clicks and sales.

But the core question is harder than counting clicks. It is a question about causation: of the people who bought, how many bought because of the ad, and how many would have bought anyway?

That second group is the whole difficulty. Someone who already wanted your product, searches your brand, clicks your ad, and buys, generates a perfect-looking record: impression, click, conversion. But the ad caused nothing. They were going to buy regardless. Counting that sale as an advertising win overstates the ad's effect, sometimes enormously.

So attribution is not really a tracking problem. It is a causation problem wearing tracking's clothes, and causation cannot simply be read off a log of who did what.

The default measurement for years was last-click attribution: whichever ad someone clicked immediately before buying gets full credit. It is simple, and it systematically rewards the wrong things.

Consider a real path. A video ad first makes someone aware of a product. Days later they read a review. Finally they type the brand into a search engine, click the brand's own search ad, and buy. Last-click hands 100 percent of the credit to that final search ad, and zero to the video that created the demand.

So the channels that harvest existing intent, mainly branded search and retargeting, look spectacular, because they are always last in line. The channels that created the intent look worthless. Budget then flows toward harvesting and away from creating, until there is less demand to harvest, and nobody can see why from the dashboard.

This is the central trap of attribution: the easiest thing to measure is the last touch, and the last touch is the least likely to be the cause.

The obvious response is multi-touch attribution: spread credit across every step in the path rather than giving it all to the last. Simple rules split it evenly or weight the first and last touch; data-driven versions train a model to assign credit from patterns across many journeys.

It is a real improvement, and it hits two hard ceilings.

First, coverage. Multi-touch needs to stitch a single person's touchpoints together across sites, apps and devices, which requires exactly the cross-site identifiers that privacy changes are removing. As those identifiers vanish, the model sees a shrinking, non-random fragment of each journey.

Second, and deeper, correlation is not causation. Even a perfectly observed path only tells you the order in which someone encountered ads, not which ones changed their mind. A model trained on who-converted learns the shape of typical journeys, and typical journeys are dominated by people who would have converted anyway. It can describe the path beautifully and still not know what caused the destination.

That second ceiling is why the field turned to a fundamentally different method.

The only clean way to answer did the ad cause the sale is the method used to test a drug: a controlled experiment.

Split the audience at random. One group is eligible to see the ad, the other is held out and sees nothing. Both groups then buy at some rate. The difference between them, the lift, is the sales the ad actually caused. This is incrementality, and randomisation is what makes it trustworthy: because the two groups differ only by chance, the people who would have bought anyway are spread evenly across both, and cancel out.

exposed group:   8.0% bought
held-out group:  6.5% bought
incremental lift: 1.5 points  <- what the ad actually caused

The result is often humbling. Channels that looked heroic under last-click, especially retargeting and branded search, frequently show small incremental lift, because they were mostly reaching people who would have converted regardless. Incrementality is unpopular for exactly that reason: it strips the flattering numbers away and shows the real, smaller causal effect.

It is also the same idea as an A/B test, applied to whether a whole channel is worth its budget.

While the industry debated attribution, the ground moved. The identifiers all this measurement relied on were deliberately restricted by the companies that control browsers and phones.

Apple's App Tracking Transparency (2021) required apps to ask permission before tracking users across other companies' apps and sites. Most people declined. Industry analytics reported opt-in rates starting low, in the mid-teens of percent, and rising only partway over time. Overnight, a large share of mobile ad measurement went dark.

Browsers moved too. Safari and Firefox block third-party cookies by default. Chrome spent years planning to remove them, then in 2025 changed course, keeping third-party cookies under user choice rather than deleting them, while also winding down much of its Privacy Sandbox replacement effort.

The net effect for measurement is blunt: cross-site and cross-app tracking is now partial, inconsistent between platforms, and shrinking. User-level attribution, which assumed you could follow one person everywhere, lost the ground it stood on. Reported figures vary by source, but the direction is not in doubt, and it forced a rebuild.

With per-person tracking degraded, measurement reassembled from three complementary tools, each answering a different question. No single one is the source of truth any more.

• Media-mix modeling (MMM). A top-down statistical model relating overall sales to overall spend per channel over time, plus outside factors like season and price. It needs no individual identifiers, because it works on aggregates, which is precisely why it came back. It answers the strategic question: how should I split budget across channels?
• Incrementality experiments. The causal check from earlier. It answers: is this specific channel actually working?
• Privacy-preserving and first-party reporting. Platforms report aggregate conversions rather than per-person trails; advertisers lean on their own first-party data and server-side signals. It answers the tactical question: what happened, roughly, in-platform?

The mental shift is the real content here. Measurement moved from a single deterministic ledger, this click caused this sale, to a portfolio of estimates that are triangulated: a broad model for allocation, experiments for causal truth, and aggregate reports for tactics. Less precise per event, and considerably more honest.

Pull the thread together and a mature view of advertising measurement looks different from the dashboard fantasy.

The uncomfortable conclusion the best practitioners have reached: the old retailer was right, and pretending otherwise was the error. You cannot know exactly which half of the spend is wasted. You can estimate it, bound it with experiments, and improve the allocation over time, but the precise per-sale attribution that dashboards promised was always partly an illusion built on tracking that is now going away.

Good measurement is not a ledger of certainties. It is a disciplined way of being usefully uncertain.

Last-click gave a single confident-but-wrong number. The modern approach answers three different questions with three tools, and cross-checks them, because no single method survived the loss of per-person tracking.