Marketing attribution practices compared: evidence over guesswork

What to take away

Attribution practices side by side: last-touch to MMM

Marketing attribution practices are crediting rules, not measurements of cause. Most practitioner work runs through the models below. They differ in what they assume, what they cost and what they can prove.

Google Analytics 4, Adobe Analytics and Triple Whale ship last-touch and rule-based multi-touch views. Those views are convenient and reproducible. They are also silent about what would have happened without the spend.

Data-driven models sit in between. GA4 assigns credit from your own conversion paths once the property clears a documented conversion threshold; smaller properties fall back to last-click. Adobe Analytics sells a comparable model inside an enterprise contract, quoted per account rather than from a list price.

Northbeam and Rockerbox price ecommerce multi-touch attribution on custom quotes. Small Shopify stores typically pay hundreds of dollars a month; larger brands pay thousands.

Meta's Robyn and Google's Meridian are open-source marketing mix modeling packages, so the software is free and the analyst time is not. Nielsen and Kantar sell MMM as a managed service, with engagements typically in the six figures a year.

Where each practice breaks

Last-touch overstates branded search and retargeting, because both sit near the purchase. First-touch overstates whichever channel opened the journey. Linear, time-decay and position-based splits move the problem rather than solving it: change the rule and the ranking changes, with no new evidence.

Data-driven models inherit whatever tracking missed. Consent-gated, cross-device and offline conversions enter as absences, not as flagged errors. On iOS, AppsFlyer, Adjust and Branch lean on Apple's AdAttributionKit and modeled conversions after App Tracking Transparency removed deterministic identifiers.

Incrementality tests answer a narrower question well. Hold back a channel in a matched set of geographies, or split audiences at random, then compare conversion rates. Meta Conversion Lift, Meta's open-source GeoLift package and Google Ads conversion lift run this kind of test.

The result holds for that test, that period and that population. Extending it to next quarter is an assumption, not a finding.

MMM covers channels that no pixel sees, including television, out-of-home and sponsorship. It needs enough historical variation in spend to separate one channel from another. When budgets move together every quarter, the model cannot tell them apart, and the intervals widen to say so.

Most marketing attribution mistakes start here: a team treats a descriptive view as causal proof, then defends a budget cut it cannot support.

Reconcile before you compare

Pick one governed outcome total: orders, qualified leads, closed revenue. Count it once, with stated rules for reversals and maturity. Every allocation view must tie back to that total, or the comparison is between two different numbers.

Then test the model against an adverse case. Remove a channel identifier, delay an outcome by a week, shorten the lookback, reverse a sale. Confirm totals still reconcile and quality warnings fire. Run the same checks each cycle with a marketing attribution checklist.

NIST's Privacy Framework starting guide describes a voluntary process for identifying and managing privacy risk. Use it to assign owners and record data processing. It does not establish legal compliance or approve a marketing use.

CISA's business-system logging guidance explains why organizations retain and review event records. Attribution needs the same trail for access changes, source defects and corrections. Whether that trail exists depends on the marketing attribution tools in place. Logging supports investigation; it does not validate a metric.

Worked example: choosing under a fixed budget

A team has one quarter and a modest testing budget. Write the arithmetic in your own figures. For scale, a vendor-run geo holdout typically costs tens of thousands of dollars per test, and a consultancy MMM build typically starts in the six figures.

Let C be the cost of one geo holdout test, M the cost of one MMM refresh, and B the budget available. If C is small enough that B / C covers several channels, test the two channels you are most likely to cut.

If M fits and C does not, run MMM and treat its output as a hypothesis to test later.

Record the rejected option and the constraint that rejected it. Constraints change, and the note is what lets a successor reopen the question.

Common questions

Should one model be used companywide?

A shared model improves consistency, but different decisions may need separate, clearly named views. Name the owner and the decision each view serves.

How should algorithmic attribution be reviewed?

Check the eligible data, the method, drift, version history and coverage. Then compare the output with an incrementality test before anyone acts on it.

When should a model be retired?

Retire it when its decision, data, ownership or product support ends. Preserve the records, and note what replaced it and why.