What Is Marketing Attribution? Models, Limits & What Works Now

Marketing attribution is the practice of assigning credit for a conversion to the marketing touchpoints that preceded it — the search ad, the Instagram Reel, the podcast mention, the email — so that budget can follow revenue. Every attribution model is a rule for splitting that credit among the touches you can observe. None of them observes causation: attribution tells you which touchpoints appeared on the path, while incrementality asks the harder question of whether the sale would have happened without them.

The discipline has moved through three eras. Last-click ruled for a decade because it was easy. Multi-touch models promised fairness and mostly delivered complexity. And since iOS privacy changes and Safari's tracking prevention broke user-level journey stitching, the working answer has become triangulation: platform attribution for fast tactical calls, media mix modeling for budget allocation, and incrementality tests for ground truth.

How do attribution models assign credit?

An attribution model is a rule for dividing conversion credit among the touchpoints in a recorded journey. The classics: last-click gives everything to the final touch, first-click gives everything to the earliest, linear splits credit evenly, time-decay weights touches closer to the purchase, position-based favors the first and last touch, and data-driven lets a machine-learning model assign fractional credit based on patterns across thousands of observed journeys.

Same journey, six different stories:

Two things follow from that table. First, model choice rewrites your channel P&L while total revenue stays identical — switch from last-click to first-click and TikTok goes from worthless to heroic overnight, even though nothing in the world changed. Any strategy that flips when the model flips was never a strategy. Second, data-driven attribution (the GA4 default) sounds like the endgame, but a model can only learn from the journeys it observes, and observation is precisely the part that broke.

Why did user-level attribution break?

Multi-touch attribution assumes you can stitch one person's touches together across days, browsers, and devices. Four forces dismantled that assumption:

• Safari's Intelligent Tracking Prevention caps JavaScript-set first-party cookies at seven days (24 hours in some link-decoration cases), so a customer who researches this week and buys in three weeks registers as a brand-new visitor.
• iOS App Tracking Transparency requires opt-in consent for cross-app identifiers, and most users decline the prompt.
• Ad blockers and privacy extensions strip client-side pixels before they ever fire.
• Multi-device behavior splits a single buyer into strangers: research on the phone over lunch, purchase on the laptop that evening.

The platforms responded with modeled conversions — statistical estimates filling the observation gaps — which means the dashboard you optimize against is already part measurement, part inference. On your side of the fence, the highest-leverage fix is moving event collection off the browser: server-side tracking typically recovers 15–30% of otherwise-lost conversions (directional, practitioner consensus), and our GA4 vs server-side tracking comparison walks through when that build pays for itself. The other durable response is owning the relationship directly — our first-party data playbook covers the collection surfaces and activation patterns that no browser update can revoke.

Why do platforms claim more revenue than you collected?

Three structural reasons, and none of them is fraud:

Overlap. A buyer clicks a Google Shopping ad on Monday, a Meta retargeting ad on Wednesday, and converts Thursday. Google's window catches it. Meta's window catches it. Your bank account caught it once.

Window inflation. A 28-day click window credits ads a 7-day window ignores. Every platform picks defaults that flatter its own contribution.

View-through credit. Some platforms count conversions that follow an impression with no click. That influence is worth more than zero and less than a click, and the dashboard cannot tell you where in between.

The result: platform-attributed revenue summed across channels routinely exceeds real blended revenue. The antidote is keeping MER — total revenue divided by total ad spend — as the guardrail no single dashboard can inflate. We compiled the published research on how large these gaps run in our marketing attribution statistics roundup, and our free Attribution Doctor diagnoses where the double counting sits in your own account in about three minutes.

What replaced the single source of truth?

Nothing did, and accepting that is the beginning of good measurement. What works now is triangulation — three systems whose blind spots fail in different directions:

1. Platform attribution stays in the stack because it is fast and granular. It remains the right tool for relative calls: which creative, which audience, which campaign. Treat its absolute numbers as directional and its comparisons as useful.
2. Media mix modeling regresses outcomes against spend history from the top down. It needs no cookies and no user-level data, which makes it privacy-proof, and it answers the allocation question platform dashboards structurally cannot: what happens to total revenue if budget moves between channels.
3. Incrementality experiments — geo holdouts, audience splits, spend blackouts — are the causal ground truth. They are slow and they cost statistical power, so you run them quarterly on the biggest line items and use the results to calibrate the other two systems.

The three will disagree, and the disagreement is the information. When platform attribution says a channel is a star and a holdout test says it is mostly harvesting organic demand, you just learned something a single dashboard would have hidden forever.

What does a practical attribution setup look like for a mid-size brand?

A sequence rather than a software purchase:

1. Tagging hygiene first. Inconsistent campaign tagging poisons every downstream system, from GA4 reports to any future model. A shared taxonomy and our free UTM Builder get a whole team generating clean, consistent links from day one.
2. Server-side event collection. Recover the 15–30% of conversions the browser loses, and feed the recovered signal back to platform bidding algorithms, which improves delivery as well as reporting.
3. One blended scorecard. MER, new-customer count, and blended CAC reviewed weekly against the platform numbers. Divergence between the blended view and the dashboard view is your early-warning system.
4. A quarterly test calendar. One incrementality question per quarter, aimed at the biggest or most doubted line item. Retargeting and branded search are the classic first candidates because platform attribution flatters them most.
5. Lightweight MMM once you have roughly two years of weekly history — earlier if a Bayesian setup with sensible priors fits your data.

When teams engage our data and analytics practice, this stack is usually the first deliverable: a tagging audit, server-side collection, a blended scorecard everyone trusts, and a test calendar with teeth.

How does AI search change attribution?

A growing share of buying research now ends inside an AI assistant's answer rather than on a results page. When ChatGPT or Perplexity recommends your brand, the eventual visit arrives as direct traffic or a branded search — real influence carrying no referrer and no UTM. Classic attribution will systematically undercount it.

That has two practical consequences. Visibility work in AI channels — answer engine optimization and generative engine optimization — needs its own measurement, built on citation share across a fixed prompt set rather than click paths. And it strengthens the case for blended and top-down measurement generally, since MER and MMM capture value that arrives through untaggable doors.

Attribution is a vocabulary as much as a technique — the growth marketing glossary collects every definition in this series, including the metrics a good attribution readout feeds.