The cookie deadline kept moving and then stopped moving at all. Attribution carried on degrading regardless, because the cookie was never the only thing breaking.

In short

• Third-party cookies were not removed from Chrome, and it made less difference than expected. Other sources of signal loss continued.

• No single measurement method is trustworthy on its own any more. Serious teams triangulate across three.

• Platform reporting tells you what happened inside one platform. Modelling tells you what your spend did overall. Experiments tell you what was actually caused.

• The first fix is almost never a new tool. It is agreeing what a conversion means and sending it reliably.

  
  

For five years, the industry had a deadline to plan around. Third-party cookies were going away, the date moved repeatedly, and a great deal of budget was spent preparing for a specific event.

Then in April 2025 the position changed. Google confirmed it would keep its existing approach to third-party cookie choice in Chrome rather than introducing a new prompt, and the Privacy Sandbox programme was steered in a different direction. The deadline everyone had planned around stopped existing.

What did not change is the thing that actually mattered. Attribution kept degrading. It just did so for reasons that never had a countdown attached.

Where signal actually goes missing

The cookie was one leak in a system with several. Most teams underestimate how many of these apply to them simultaneously.

• Mobile app tracking permissions. A large share of users decline tracking when asked directly, which removes a substantial slice of app and cross-app measurement regardless of what browsers do.

• Browser restrictions on first-party storage. Several major browsers cap how long client-side cookies survive, which quietly shortens attribution windows and makes long consideration cycles look shorter than they are.

• Consent. In regulated markets a meaningful proportion of visitors decline analytics consent. The visit still happens, the conversion still happens, and neither is recorded.

• Ad blocking and network-level filtering, which removes both the pageview and the conversion event.

• Closed environments. Discovery increasingly happens inside platforms, communities, summaries and messaging apps that pass no referrer data at all.

Every one of these predates any cookie deprecation timetable and none of them was solved by the timetable being cancelled.

The practical position is simple. You will never again see the full path. You can still make good spend decisions, but not with one method.

Three methods, used together

Mature measurement is not a single source of truth. It is triangulation, where three methods with different blind spots are compared and the disagreements are treated as information rather than errors.

Method one: platform reporting, read with suspicion

Every advertising platform reports the conversions it can claim, using its own attribution rules and its own view window. This is genuinely useful for in-platform decisions, such as which creative to retire or which audience is saturating, and genuinely misleading when totalled across platforms.

Two disciplines make it usable:

• Send conversions server to server rather than relying on the browser alone. Server-side event forwarding recovers events that would otherwise be lost to blocking and storage limits.

• Deduplicate rigorously. The same conversion sent by both browser and server, counted twice, produces confidence in a number that is simply wrong.

  
  

Method two: modelling, for the whole picture

Marketing mix modelling went out of fashion when click tracking was reliable and has returned for exactly the reason you would expect. It works on aggregate spend and outcome data, needs no user-level tracking at all, and is therefore immune to almost everything in the list above.

What changed recently is access. Open-source implementations are now published and maintained by the major platforms themselves, notably Meridian from Google and Robyn from Meta, which puts a technique that used to require a specialist consultancy within reach of an in-house analytics team.

Two cautions are worth stating plainly. A model built on your own spend and your own outcomes is a genuinely useful instrument. A model published by a company that sells advertising will always deserve a second look at its assumptions, particularly around how it treats the channels its publisher happens to sell. Run more than one where you can, and compare.

Method three: experiments, for causality

Modelling tells you what correlates. Experiments tell you what caused what, and they are the only method that does.

The most practical form for most companies is a geographic holdout. Split comparable regions, turn a channel off in some and leave it running in others, and measure the difference in outcomes. It is inconvenient, it requires holding your nerve while a region goes dark, and it produces the only number in your reporting that nobody can argue with.

Run one per quarter on the channel you are least certain about. Over a year that is four questions permanently answered.

How the three fit together

Different jobs, different time horizons

• Platform data answers this week's question. Which ad, which audience, which placement.

• Modelling answers the quarterly question. How should the budget be split across channels.

• Experiments answer the annual question. Is this channel actually causing anything.

  
  

Problems start when one is asked to do another's job, most often when platform data is used to set the budget split.

Start with the boring part

Almost every measurement project that fails does so before any of the above. It fails on definitions.

If marketing counts a form submission as a lead and sales counts a qualified opportunity as a lead, no amount of modelling will reconcile the two. If three teams have three definitions of a conversion, the data warehouse will faithfully store all three and the dashboard will average them into something meaningless.

Before touching a tool, get the following written down and signed by both marketing and sales:

1. What counts as an enquiry, and what does not.

2. What makes an enquiry qualified, in criteria a person can apply consistently.

3. Which system holds the record of truth for each stage.

4. How long after a first touch a conversion may still be attributed to it.

5. Who reviews the numbers, and how often.

   
   

This takes a fortnight and saves a year. It is also the step most often skipped, because it produces no visible output and cannot be bought.

A sequence that works

For a team starting from a broken measurement position, the order matters more than the ambition.

• Weeks 1 to 2. Agree definitions. Audit which events fire, from where, and whether any are duplicated.

• Weeks 3 to 6. Implement server-side conversion forwarding on your main channels. Verify deduplication with a manual test rather than trusting the setup screen.

• Weeks 6 to 10. Join platform data to CRM outcomes so cost per qualified enquiry exists as a real number rather than an estimate.

• Quarter 2. Run the first geographic holdout on your least certain channel.

• Quarter 3 onwards. Begin modelling once you have enough clean history, which in practice means at least a year of weekly spend and outcome data.

  
  

Note what is absent from that list. There is no new platform purchase in the first three months. The most common measurement mistake is buying a tool to solve a definition problem.

What good looks like

You will know the stack is working when three things become true. The finance team stops asking why the platform numbers do not match the revenue. The marketing team can say which channel it would cut first and why. And a disagreement between two data sources produces a question rather than an argument.

Perfect attribution is gone and is not coming back. Confident spend decisions are entirely achievable without it, which is a considerably more useful goal.

Frequently asked questions