Attribution & Measurement: What Actually Tells You What's Working

Hero photo by Johannes Plenio via Pexels.

Here is the single most useful fact in marketing measurement, and almost no dashboard will tell it to you: across 2,226 real Meta advertising experiments, standard 7-day last-click attribution explained just 19% of the variance in what those experiments actually measured (Gordon, Moakler & Zettelmeyer, arXiv:2304.06828). The number your attribution report shows you and the number a randomised experiment would show you are, most of the time, not the same number.

That is not an argument for giving up on measurement. It's an argument for understanding what each layer of it can and can't tell you. Attribution and measurement is a stack: an experiment tells you the truth but is expensive; an attribution model tells you a story cheaply; conversion tracking is the raw input feeding both; and first-party data is the fuel that keeps all of it working as third-party signal disappears. This is the map of that stack — what each layer is for, where it lies to you, and which of our deeper guides to read for each.

Key Takeaways

• Attribution is not truth. The best-calibrated method in the largest study to date still disagreed with a randomised experiment's verdict in 8–12% of campaigns; last-click disagreed in 12–20%.
• Incrementality — did this spend cause conversions that wouldn't have happened anyway — is the question every model is trying to approximate. Experiments answer it directly; models only estimate it.
• Your measurement is only as good as its input: iOS opt-in tracking gaps run from 14% to 50% by app category, so conversion tracking has to be built for the gap, not in denial of it.
• The whole stack now runs on first-party data, and the rules governing it differ by market — Consent Mode v2 doesn't even apply to Singapore, India, or Indonesia.

Start With the Question Every Model Is Really Asking

Strip away the jargon and attribution has exactly one job: to answer "did this marketing cause a sale that wouldn't have happened otherwise?" That property has a name — incrementality — and it is the only thing worth paying for. A conversion that would have happened anyway, credited to an ad, is a rebate you paid yourself.

The problem is that incrementality is invisible. You never see the version of the customer who didn't see your ad. So the entire discipline is a series of increasingly good attempts to estimate an unobservable counterfactual. A randomised experiment estimates it by force — hold out a group, compare. An attribution model estimates it by assumption — apply a rule for splitting credit and hope the rule tracks reality. Everything else in this guide follows from that one distinction.

Attribution Models Tell a Story, Not the Truth

An attribution model is a rule for assigning credit across the touchpoints on the way to a conversion. Last-click gives everything to the final touch. Data-driven models distribute it using observed patterns. They are cheap, always-on, and useful for pacing day to day — and they are not measurement of cause.

The largest test of this gap comes from the PIE study, which compared attribution against the ground truth of 2,226 randomised Meta experiments. Judged on how well it tracked the incremental conversions per dollar those experiments actually found, industry-standard 7-day last-click scored an R² of 0.19. An experiment-calibrated method (PIE) scored 0.88.

An R² of 0.19 means last-click captures less than a fifth of the real variation in what works. It isn't random — it's directionally useful — but treating its numbers as fact is how budgets get misallocated with total confidence.

The consequence isn't abstract. When you turn attribution into a decision — scale this, cut that — how often does it point the wrong way? In the same study, last-click disagreed with the experiment's own verdict in 12–20% of campaigns. The calibrated method narrowed that to 8–12%, but never to zero.

Up to one campaign in five getting the opposite answer from last-click is the real cost of mistaking a model for a measurement. None of this means attribution is useless — it means you should hold its outputs as hypotheses to be tested, not verdicts. Which model to actually pick (and why GA4 quietly has three, not the seven most guides still teach) is its own decision.

→ Deep dive: What Attribution Model Should You Use? — what GA4 actually offers in 2026, and the Shapley-value myth that won't die.

Experiments Are the Gold Standard — Use Them on the Big Decisions

If models only estimate incrementality and experiments measure it, why not experiment on everything? Because randomised tests cost real money and traffic, and they answer one question at a time. The discipline isn't "experiment always" — it's experiment on the decisions big enough to be worth the cost, and let calibrated models fill in between.

Practically, that means reserving geo-holdout tests, conversion-lift studies, and platform incrementality experiments for the questions where being wrong is expensive: is this whole channel pulling its weight? Is a brand campaign doing anything? Would we lose sales if we cut this line? Those are worth a holdout. The daily "which ad set is up 8%" question is not — that's what your always-on model is for. The measurement-maturity gap is stark: only 52% of senior marketing leaders can prove marketing's value and get credit for it, rising to 62% among those who meet regularly with their analytics team (Gartner, 2024). Proof comes from experiments; stories come from dashboards.

None of It Works Without Clean Conversion Tracking

Every model and every experiment is downstream of one thing: whether the conversion event actually reached your measurement system in the first place. This is the least glamorous layer and the one that quietly breaks everything above it.

The gap is structural, not incidental. iOS App Tracking Transparency opt-in rates range from 14% for education apps to 50% for sports apps, against a ~35% average (Adjust, 2025). That means a large, category-specific share of conversions never gets deterministically observed — and no attribution model can credit an event it never saw. Server-side tracking (Conversions API and equivalents) is now table stakes for narrowing that gap, though it supplements client-side tracking rather than replacing it. Get this layer wrong and every number above it is confidently computed on top of a hole.

→ Deep dive: How to Set Up Conversion Tracking That Actually Works — the iOS gap, server-side setup, and what Meta's Event Match Quality score really measures.

The Whole Stack Now Runs on First-Party Data

The final layer is the fuel. As third-party cookies and cross-app identifiers degrade, the durable input to all of the above is data your customers give you directly — and the rules governing it are not global boilerplate. A common mistake is to apply European consent tooling everywhere: Google Consent Mode v2 is scoped to the EEA, UK, and Switzerland, and is not required for Singapore, India, or Indonesia traffic. Each of our core markets is on its own clock — India's DPDP Rules were notified in November 2025 on a phased runway; Indonesia's PDP Law is in force but its implementing regulation is still unsigned. Building your measurement on a compliant, well-structured first-party foundation is what keeps the rest of the stack working next year.

→ Deep dive: First-Party Data Rules in Southeast Asia & India — what actually governs your data in each market, minus the vendor mythology.

How the Layers Fit Together

Layer	Question it answers	What it costs	How much to trust it
Experiment (incrementality)	Did this spend cause conversions?	High — holdouts, time, traffic	Highest — it's the ground truth
Attribution model	How should credit be split, roughly?	Low — always on	Directional; a hypothesis, not a verdict
Conversion tracking	Did the event reach my system at all?	Setup + maintenance	Foundational — everything above depends on it
First-party data	Will any of this still work next year?	Ongoing governance	Strategic — the durable input

The mistake that wastes the most money is trusting a lower row as if it were the top row — reading a last-click dashboard as if it were an experiment. Read down the stack for daily pacing; measure up the stack when the decision is big enough to matter.

Frequently Asked Questions

What is the difference between attribution and incrementality?

Attribution assigns credit for a conversion across the touchpoints that preceded it, using a rule (last-click, data-driven, etc.). Incrementality asks a stricter question: would that conversion have happened anyway without the ad? Attribution is a cheap, always-on estimate; incrementality — measured by a randomised experiment — is the ground truth attribution is trying to approximate. In the largest study to date, standard last-click attribution explained only 19% of the variance in measured incrementality.

Is last-click attribution accurate?

Not as a measure of cause. Across 2,226 Meta experiments, 7-day last-click attribution achieved an R² of just 0.19 against actual measured incrementality, and disagreed with the experiment's decision in 12–20% of campaigns (Gordon, Moakler & Zettelmeyer, arXiv:2304.06828). It's useful for day-to-day pacing and directionally informative, but treating its numbers as truth is how budgets get misallocated. Use it as a hypothesis, and test the big decisions with experiments.

What is incrementality testing?

Incrementality testing measures the causal effect of marketing by comparing a group exposed to ads against a randomly held-out group that wasn't. The difference in conversions is the incremental lift — the sales the marketing actually caused. Because it randomises exposure, it estimates the counterfactual directly rather than assuming it, which is why it's considered the gold standard. It's costly, so it's best reserved for high-stakes decisions like whether an entire channel is pulling its weight.

Why don't my attribution numbers match my actual sales?

Because attribution models estimate an unobservable counterfactual and conversion tracking has real gaps. Two things are usually happening: the model is crediting conversions that would have occurred anyway (inflating apparent performance), and tracking is missing events it never observed — for example, iOS opt-in rates run as low as 14% in some app categories. The fix isn't a better model alone; it's cleaner conversion tracking underneath and incrementality experiments on top.

Do I need Google Consent Mode v2 in Singapore, India, or Indonesia?

No. Consent Mode v2 is scoped to the EEA, UK, and Switzerland, verified against Google's own published policy — it is not required for Singapore, India, or Indonesia traffic. Those markets have their own frameworks on their own timelines: India's DPDP Rules were notified in November 2025 on a phased runway, and Indonesia's PDP Law is in force though its implementing regulation is still unsigned. Applying European tooling everywhere is a common and unnecessary mistake.

What should I actually measure to know marketing is working?

Measure incrementality on the decisions big enough to justify an experiment, and use a well-tracked attribution model for everything in between. Practically: keep conversion tracking clean (including server-side), run holdout or lift experiments on whole channels and major campaigns, and treat dashboard attribution as a hypothesis. Only 52% of senior marketers can prove marketing's value — the ones who can are disproportionately the ones running experiments rather than reading dashboards.

Conclusion

Measurement isn't one number from one dashboard — it's a stack, and each layer has a different relationship to the truth. Attribution tells a cheap, always-on story that captures less than a fifth of what experiments find. Incrementality tells the truth but charges for it. Conversion tracking decides whether any of the data is real, and first-party data decides whether it'll still work next year. The teams that measure well aren't the ones with the fanciest model — they're the ones who know which layer they're looking at, and never mistake a story for a fact.

How this post was compiled. The central attribution-versus-incrementality figures (R²=0.19 vs 0.88; 12–20% vs 8–12% decision disagreement across 2,226 experiments) are taken directly from the peer-reviewed PIE study (Gordon, Moakler & Zettelmeyer, arXiv:2304.06828), verified against its published abstract. The iOS opt-in range (14–50%) is from Adjust's 2025 data; the "52%/62% can prove value" figures are from Gartner's 2024 survey as reported by Business Wire; the Consent Mode v2 scope and regional privacy timelines are verified against Google's policy and each market's own regulations. Where widely-cited figures (such as a "2.9x first-party data uplift") had no openable methodology, we left them out. Coact is a performance marketing agency working with ecommerce and app businesses across Singapore, India, and Indonesia.

Continue Learning

• What Attribution Model Should You Use? — GA4's real model set and the Shapley myth
• How to Set Up Conversion Tracking That Actually Works — the input layer, done properly
• First-Party Data Rules in Southeast Asia & India — the compliant foundation
• Why Is My CAC Increasing? — what broken measurement hides
• Talk to our team about building measurement you can actually trust.