Omnichannel attribution: frameworks & methodology

Your analytics lead comes back with the only question that matters before anyone writes code: Shapley values or a Markov removal-effect model, run on which table, and how will we know the output is right? Picking a credit-assignment algorithm is a constrained choice, not a preference. It is constrained by how many touches you actually record per converting path, by how much of your spend never lands in that table at all, and by whether you can run an experiment to check the answer. Get the order of those constraints wrong and you will commission a model that produces confident numbers nobody can defend in a budget meeting.

Match the algorithm to the touch data you have

Run four counts on your own conversion data before you shortlist a model.

• Median touches per converting path, and the share of paths with exactly one recorded touch.
• Share of total media spend on channels that never emit a user-level touch: linear TV, out-of-home, print, sponsorship, most in-store retail media.
• Conversions per month, per channel, inside the attributionattributionA framework for assigning credit to the touchpoints that contributed to a conversion, so you can measure which channels and interactions actually drive results.View full definition → window the foundations lesson sets up.
• Share of paths that break identity mid-way, so the pre-break touches sit in a different profile.

Those four numbers decide the model for you. If most of your converting paths carry a single recorded touch, MTAMTAA method that distributes conversion credit across all marketing touchpoints in the customer journey, rather than crediting only the first or last interaction.View full definition → has nothing to divide and any multi-touch model will just re-label last-click with extra decimal places. If a channel produces a few dozen conversions a month, a machine-learned model will overfit its noise; Google Ads for years required something in the order of tens of thousands of clicks and hundreds of conversions in a 30-day window before it would fit a data-driven model at all, and only dropped those minimums around 2021. The thresholds were a reasonable statistical instinct even if the gate is gone.

And if more than roughly a quarter of your spend sits in channels that emit no touch, no touch-level model can be your planning instrument. It will hand that quarter's credit to whatever trackable channel happened to be nearby.

The four model families and what each can carry

Rule-based logic

Fast, cheap, and built on assumptions you chose rather than anything your customers did. Fine as a directional gut check, not as the basis for allocating above about $5 million. One operational note if your team lives in Google Analytics: GA4 retired first-click, linear, time-decay and position-based in 2023, and dropped last-click for cross-channel reporting too, so data-drivendata-drivenAn approach where decisions are systematically informed by data analysis rather than intuition alone.View full definition → is the only cross-channel model it offers. Any rule-based comparison your team quotes is coming from somewhere else, and you should know where.

Data-driven attribution

Two algorithms dominate. Shapley value treats each channel as a player and averages its marginal contribution across every ordering of the path. Markov chains model the path as states and compute a removal effect: delete the channel, see how much conversion probability drains out of the graph. Google's DDA is counterfactual in the same spirit, comparing paths with and without a touchpoint. Adobe sells this too: Attribution IQ in Analysis Workspace runs an algorithmic model alongside the rule-based ones on the same dataset, which is the cheapest way to see how much your answer depends on your algorithm choice. If Shapley and Markov disagree on channel ranking, the disagreement is your real finding, and it usually means thin path data rather than a modelling subtlety.

Media mix modelling

Regression on aggregate spend and outcomes over time, cookie-free, and able to absorb TV, weather, price and distribution. The constraint people underestimate is degrees of freedom: three years of weekly data is about 156 observations, and a model with a dozen channels plus seasonality, price and a promotion flag has already spent most of them. The second constraint is multicollinearity. If your channels always flight together, up in Q4 and down in January, the regression cannot separate them and will assign credit almost arbitrarily between two correlated lines. The fix is upstream of the model: stagger flight dates, vary spend by region, create the variation the model needs. Google's open-source Meridian (which replaced LightweightMMM) and Meta's Robyn are both free starting points your data science team can run without buying a black box.

IncrementalityIncrementalityThe share of results (sales, conversions, revenue) that only happened because of a marketing action, not what would have occurred anyway.View full definition → testing

Your measuring stick for everything above, and the only method that produces a number the model did not assume. Two design facts govern whether a test can work. First, in a geogeoThe practice of making your brand and content visible and citable inside AI-generated answers from tools like ChatGPT, Gemini and Perplexity.View full definition → test the unit of randomisation is the geo, so with 30 markets your effective sample is 30, not three million users. Second, required sample scales with the inverse square of the effect: moving your detectable lift from 20% to 10% costs roughly four times the data, and getting to 5% costs sixteen times. That arithmetic, not enthusiasm, decides which channels you can test this quarter. How the results then govern budget, and what holdout policy you write around them, is the playbook lesson's territory.

Validation: how you know the model is not lying

Backtest on held-out weeks and compare predicted conversions to actual. Then calibrate: Meridian accepts priors on channel ROIROIReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.View full definition → informed by experiment results, which turns a lift test into an input rather than a separate opinion. Check ranking stability by refitting on a different window; if a channel moves three places, the model is reading noise. The failure mode to watch is the self-confirming loop. Each refit trains on paths produced by the budget the previous model chose, so a model that under-credits upper funnelfunnelThe customer journey from awareness to purchase, typically Awareness, Interest, Consideration, Decision, Action, with prospects narrowing at each stage.View full definition → will see less upper-funnel data next quarter and under-credit it further. Only external experiments break that loop.

What the vendor stacks actually decide for you

Google

GA4 now defaults every cross-channel report to its data-driven model, so the model choice was made for most teams in 2023 without a procurement decision. Alongside it Google publishes Meridian for MMMMMMA statistical approach that estimates how each marketing channel and other factors drive sales, guiding budget allocation.View full definition → and its GeoX geo-experiment methodology, both free. Worth holding in mind that Google sells the media its models are grading, which is a conflict the playbook lesson tells you how to arbitrate.

Adobe

Adobe Analytics lets you swap models inside a single report, and Adobe has packaged MMM and touch-level attribution into one product line. Adobe sells attribution software, so treat its model defaults the way you would any vendor default. The genuine methodological gift is model-switching on one dataset: it turns "which model?" from a six-week project into an afternoon of sensitivity testing.

Salesforce

Salesforce (also a vendor here) exposes campaign influence models in the CRM object model, with even-distribution and custom-weighted options, plus a machine-learned variant. The reason this matters methodologically is the B2B edge case: when a deal takes nine months and involves six people at one account, your credit unit is the opportunity, not the order, and any window short enough for e-commerce truncates the first half of the path. Model at account level, extend the window to your actual sales cycle, and accept that MTA will rank channels by their ability to appear late in a long buying process unless you correct for it.

CMO action items

• Run the four diagnostic counts above in the next 30 days and write the answers down: median touches per path, single-touch share, untracked spend share, conversions per channel per month. That page determines your model, and it takes an analyst two days.
• Commission one incrementality test this quarter on a channel taking more than 20% of digital spend. Do the power arithmetic first: if the design cannot detect a lift smaller than 40%, redesign it or pick a different channel rather than running a test that can only return "inconclusive".
• Write a measurement architecture document that fixes which method answers which decision: DDA for weekly campaign optimisation, MMM for quarterly planning, experiments for annual channel strategy and for calibrating the other two. Name the owner of each model and the refit cadence.

Common mistakes that kill results

Choosing the algorithm before counting the paths. Teams buy Shapley-based tooling and then discover that most converting paths hold one touch, so the expensive model reproduces last-click and the vendor invoice buys a false sense of rigour.

Treating MMM output as precise rather than directional. A model might tell you TV drives 23% of revenue when the interval runs from 15% to 35%. Use MMM for direction and magnitude of change, and read the interval before you move nine figures.

Fitting MMM on synchronised spend. If every channel rises and falls together, the coefficients are unstable and will flip sign between refits. Vary the media plan deliberately, by geography or by flight timing, so the model has something to learn from.

Running experiments during abnormal periods. A geo holdout launched over a major promotion or a stock-out mixes your treatment effect with conditions the two groups did not share. Test in steady state, and log any exception so the next refit knows to distrust those weeks.

Key takeaways

• The model is chosen by your data, not your preference: single-touch share, untracked spend share and conversion volume per channel decide what you can legitimately fit.
• Shapley and Markov are the two mainstream MTA algorithms. When they disagree on channel ranking, that is a data-thinness signal worth more than either output.
• MMM lives or dies on variation and degrees of freedom. 156 weekly rows and a dozen collinear channels will not support a confident answer, whatever the vendor deck says.
• Validate with held-out weeks, refit stability and experiment calibration. Meridian takes experiment-informed priors, which is the cleanest way to make a lift test improve your MMM rather than argue with it.
• Watch the self-confirming loop: models retrained on budgets they shaped keep proving themselves right. External experiments are the only break in that circuit.