For about fifteen years, marketing measurement had a comfortable answer: put a pixel on the site, follow the user, read the report. That system is gone. It did not degrade gracefully — it was dismantled by regulation, browser policy and platform changes, mostly between 2020 and now.
What replaced it is not a worse version of the same thing. It is a different arrangement: several partial methods, each answering a different question, none authoritative on its own. The teams doing this well are not the ones who found a workaround. They are the ones who changed what they ask of measurement.
Why did tracking break?
Four separate forces, which is why no single fix addresses it.
Regulation. GDPR, and the wave of legislation that followed it, made the lawful basis for processing personal data a precondition rather than an afterthought. Where consent is required and declined, the data is simply not available.
Browser policy. Third-party cookies have been restricted or blocked by default in most browsers, and storage lifetimes for first-party scripts have been shortened. A returning visitor after a few weeks frequently looks like a new one.
Platform policy. Operating-system-level permission prompts sharply reduced the share of app users who can be tracked across apps, which cut the identifier supply that much of digital advertising measurement depended on.
Fragmentation. People research on a phone, buy on a laptop, and sometimes convert in a call centre. Cross-device identity was always partly guesswork, and the signals that made the guess plausible are the same ones that disappeared.
What replaced it?
Three methods, used together, each answering a question the others cannot.
| Method | Question it answers | Confidence | Speed |
|---|---|---|---|
| Platform reporting | Which ad, audience or keyword should the algorithm favour? | Low for causation | Immediate |
| Experiments | Did this activity cause incremental outcomes? | High | Weeks |
| Aggregate modelling | How should budget be split across channels? | Medium | Months of history |
The error that causes the most damage is using one method for another method's question — specifically, using platform reporting to decide budget allocation. It is the fastest number available and the least suited to the job.
Platform reporting
Still essential, and its job has narrowed. Inside a single platform, comparing two campaigns over the same period, attributed conversions are a reasonable optimisation signal. The algorithms need conversion feedback to work, and withholding it produces worse performance for no measurement gain.
What it cannot do is settle cross-channel credit. Every platform measures with its own window and its own definition of influence, so overlapping campaigns double-count by construction.
Experiments
The only method that measures causation directly. Hold back advertising from a randomly selected or matched group, run for long enough to accumulate signal, and compare. The difference is incremental effect.
Experiments cost real revenue while they run, which is the honest objection to them. The counter-argument is that the cost is bounded and known, while the cost of misallocating budget for a year on inferred credit is neither.
Aggregate modelling
Marketing mix modelling estimates each channel's contribution from aggregate historical data, using no user-level identifiers at all. That property is why a technique from the pre-digital era came back into favour.
It needs a long history and genuine variation in spend to say anything reliable. A channel whose budget has been flat for two years is nearly invisible to a model, because there is no variation for the regression to learn from.
What is modelled, and does it matter?
Yes, and it is under-discussed.
When consent is declined or an identifier is unavailable, advertising platforms increasingly estimate the conversions they could not observe, based on the behaviour of comparable measurable users. These estimates appear alongside observed conversions, usually without visible distinction.
Modelled conversions are a reasonable response to a real problem. They are also estimates with confidence intervals nobody shows you, and they are produced by a party with an interest in the number being large. Both things are true at once.
The practical response is not to reject them but to know the proportion. A view where most conversions are observed deserves more confidence than one where most are modelled, and the difference should change how hard you lean on it.
What should you actually build?
In order, because each step makes the next one worth doing.
1. A source of truth that is not a platform
Your own record of orders, revenue and new customers, by day, with as much acquisition context as you can legitimately capture. Everything else compares against this. Without it, disagreements between platforms are unresolvable because there is no independent arbiter.
2. Consent handled properly, not creatively
A consent mechanism that actually gates what it claims to, with tags respecting the choice. This is both a legal requirement and a measurement one: a consent setup that is quietly broken produces data you cannot use and liability you cannot see.
Server-side collection belongs here too, and it is worth being precise about what it does. It improves data completeness and control over what is shared. It does not create a lawful basis where none exists, and treating it as a consent workaround is a serious mistake rather than a clever one.
3. Offline outcomes fed back
If the revenue event happens in a call centre, a clinic, or a sales pipeline, the ad platforms cannot see it, and left alone they will optimise toward the last thing they can see — usually a form fill. Returning real outcomes as the optimisation event is frequently the single largest improvement available, and it is integration work rather than marketing work.
4. A standing experiment habit
One meaningful holdout per quarter on the largest channel, run long enough to detect an effect worth acting on. Not a project — a habit. The value compounds, because each result calibrates how much to trust the attributed numbers for that channel.
5. Aggregate modelling, if the history supports it
Worth doing once there is enough history and enough variation, and not before. A model fitted to insufficient data produces confident, wrong answers, which is worse than no model.
What does honest reporting look like?
Three habits, none of them technical.
Label modelled numbers as modelled. Not in a footnote. In the view, where the decision gets made.
Say when a result is directional. Not every finding is proven. Saying so is what makes the proven ones credible, and a measurement function that never admits uncertainty gets believed less, not more.
State the comparison. Almost any number can be made to look like progress with the right baseline. Naming the period and the comparison up front removes a whole class of argument.
The uncomfortable conclusion
You will not get back to a single number that explains everything. That era is over, and the businesses adapting fastest are the ones that stopped mourning it.
What is available instead is arguably better: a blended efficiency figure that cannot be gamed by attribution, causal evidence from experiments where it matters most, and modelling for the allocation questions experiments are too slow to answer. It is less tidy. It is considerably harder to fool.
Continue with the modern attribution stack, how to run incrementality tests on a small budget, or building a mix model without a data science team.