The Ecomm Analyst

For a few years now every attribution conversation has started with the same eulogy. The pixel got worse, iOS made opt-out the default, third-party cookies finally went, and the clean deterministic tracking everyone built their reporting on stopped being clean or deterministic. That part is true. What gets lost is that the eulogy is for one specific method, user-level click tracking, not for measurement itself. Plenty still works in 2026. It just works differently, and the operators who adapted are making better decisions than the ones still refreshing a Meta dashboard hoping the numbers come back.

Start with what actually degraded, because the panic is usually broader than the damage. The thing that broke is the platforms’ ability to follow an individual person from ad click to purchase across apps and devices. When a large share of users never share the identifiers that made that possible, the platforms fill the gap with modeling, which is a polite word for educated guessing. So platform-reported conversions are now part measured and part estimated, and the estimated part is a black box you cannot audit. That is the real change. Your reports did not get less confident-looking. They got less true while looking exactly as confident as before.

What still works, and works better than it gets credit for, is blended measurement. Total revenue from your store divided by total ad spend does not depend on tracking any individual at all. No identifier, no cookie, no opt-in required. It cannot tell you which channel deserves credit, but it tells you the truth about whether your marketing in aggregate is paying off, and signal loss does nothing to it. For most operators the blended ratio has quietly become the most reliable number they own, precisely because it was never built on the plumbing that broke.

Incrementality testing is the other thing that came through intact, and it is the closest thing to a real answer the discipline has. Instead of asking a platform to claim credit, you turn a channel off in a region or for an audience and watch what happens to total sales. If you pause Meta in three states for two weeks and total revenue in those states drops, that drop is incremental and it is real, because you measured the world with and without the ad rather than asking the ad to grade itself. Geo holdout tests take effort to run and they are not continuous, but they answer the one question platform attribution cannot, which is what would have happened anyway.

Post-purchase surveys came back into fashion for the same reason. A single question at checkout, how did you hear about us, collects first-party intent data that no privacy change can touch, because the customer is volunteering it. The data is noisy and people misremember, but across thousands of responses the pattern is stable, and it catches channels your click tracking will never see, like podcasts, word of mouth, and the influencer whose link nobody tagged. Used as a directional cross-check against your other numbers, it is cheap and surprisingly hard to fool.

Server-side tracking and the conversions API recover some of what the browser pixel lost, and they are worth setting up, but it is worth being honest about what they do. They improve the quality of the signal you are allowed to collect. They do not bring back the users who opted out. Treat them as patching a leak, not refilling the tank, and you will size your expectations correctly.

Media mix modeling used to be a thing only brands with a data science team could run, and a lighter version of it has become reachable for mid-sized operators. At its core it looks at spend and outcomes over time across channels and estimates each channel’s contribution statistically, without tracking anyone. It needs a decent history of data and it answers slowly, in weeks rather than clicks, but it is built on aggregate numbers, so signal loss leaves it largely alone. For a store spending across several channels it is increasingly the backbone that platform numbers get checked against rather than the other way around.

The shift underneath all of this is a shift in what you trust. The old model trusted one precise-looking number per channel and ran the business on it. The model that works now triangulates. Blended efficiency tells you if the whole machine is profitable, incrementality tells you which channels actually move sales, surveys catch what tracking misses, and platform numbers get demoted to directional signals you read for trend rather than truth. No single one of those is as satisfying as the clean dashboard used to feel. Together they are far harder to fool, which matters more.

So the honest answer in 2026 is that attribution did not die, it just stopped being free and automatic. The measurement that survived requires you to do something deliberate, run a test, ask a question, do the blended math, rather than read a number a platform hands you. That is more work. It is also the reason the operators who made the switch are quietly more confident in their spending decisions than they were back when the tracking looked perfect and was already starting to lie.