How to choose an attribution model

"Which channel gets credit for this sale?" has no single correct answer — it has several defensible answers depending on which method you ask, and each method is built to answer a slightly different question. The mistake most teams make is picking a model because it's the default in their ad platform, then treating its output as ground truth for budget decisions it was never designed to support.

The rule-based models, and why they persist

last-click attribution gives 100% of the credit to the final touchpoint before conversion; first-click attribution gives it all to the first. Both are simple, cheap to compute, and available in nearly every analytics tool by default — which is exactly why they're still everywhere, not because they're accurate. Last-click systematically overvalues bottom-funnel channels like branded search and retargeting, which tend to catch people who were already going to convert. First-click overvalues awareness channels and undervalues everything that closes the sale. Position-based and linear rules split the difference with an arbitrary weighting scheme, which is more defensible in a boardroom but no more empirically grounded.

Rule-based models are fine for a quick read on a single channel's path, and fine when you have too little data for anything more sophisticated. They are the wrong basis for a serious budget reallocation.

Data-driven and multi-touch attribution

data-driven attribution and multi-touch attribution use statistical or machine-learning models to assign credit across the touchpoints a converting customer actually passed through, based on patterns in your own conversion data rather than a fixed rule. This is a real improvement over last-click when it works, but it inherits two structural weaknesses: it can only attribute what it can observe, and iOS tracking restrictions, ad blockers and walled-garden data limits mean a growing share of the customer journey is invisible to it; and it is still correlational — a channel that's disproportionately present in converting paths might be a cause of conversion, or might just be where people who were already going to buy happen to show up.

Rockerbox and Northbeam both build multi-touch attribution from a brand's own warehouse or first-party data rather than a third-party pixel, which makes them more resilient to tracking loss than pixel-only tools, though neither escapes the correlation-versus-causation limit above. Triple Whale blends pixel-based tracking with post-purchase survey data for the same reason — self-reported "how did you hear about us" data fills gaps that tracking alone can't see, at the cost of relying on what customers remember and choose to say.

Marketing mix modeling: coarser, but privacy-resilient

marketing mix modeling (MMM) takes a completely different route: instead of tracking individual customer paths, it uses statistical regression on aggregate spend and outcome data over time to estimate each channel's contribution. It doesn't need cookies, pixels or user-level tracking at all, which makes it the most durable method as tracking keeps degrading — but it needs a long history of spend variation to work from, and it answers "how did channel-level spend relate to outcomes" rather than "which specific customer touchpoints mattered." Meridian (Google's open-source Bayesian MMM library) is free but requires a data scientist to build and maintain the model; Recast delivers the same Bayesian MMM approach as a managed service, aimed at teams that want the output without hiring for it.

Incrementality testing answers a different question entirely

None of the methods above tell you what would have happened without the marketing — they infer it. incrementality testing measures it directly, by holding out or varying spend in a controlled way (typically by geography) and comparing the actual outcome difference between exposed and unexposed groups. It is the only method on this list that produces a causal answer rather than a modeled or correlational one, and it is the natural check on the other three: a geo-experiment result that contradicts what your multi-touch attribution has been telling you is a signal the attribution model has a blind spot, not a coin flip to average away.

The cost is real: incrementality tests need enough scale and enough patience to run a valid holdout, and they answer "did this channel help, in this window" rather than giving you a live, always-on dashboard.

A shortlist by situation

• A Shopify or DTC brand needing warehouse-native, first-party multi-touch attribution. Northbeam is built specifically around e-commerce checkout flows and order data.
• Mid-market or enterprise advertisers who want multi-touch attribution reconciled against MMM from the same data. Rockerbox runs both from one warehouse-native pipeline rather than forcing a choice.
• An all-in-one profitability and attribution dashboard for a Shopify store. Triple Whale pairs blended ROAS with attribution and, on higher tiers, its own MMM module.
• Enough scale and a data scientist to build your own model, with no license cost. Meridian is free, open source, and Bayesian, but you own the maintenance.
• Want rigorous MMM without hiring a data scientist. Recast runs the model for you and pairs it with geo-experiments to validate it.
• Mobile app growth teams. Attribution here runs through a mobile measurement partner such as AppsFlyer, which aggregates Apple's attribution-window-limited SKAdNetwork postbacks alongside deterministic Android and web data — a genuinely different technical problem from web attribution, not just a smaller version of it.

Questions to ask before you commit

• What share of conversions can this method actually observe, given our tracking restrictions and audience?
• Has the model or attribution result ever been validated against a real holdout or incrementality test — and did it agree?
• What attribution window does the model use, and does it match how long our actual purchase cycle is?
• If we switch vendors, does historical attribution data migrate, or does the trend line reset?
• Who on our team can sanity-check the model's output when it disagrees with intuition?

Common mistakes

• Treating any attribution model's output as causal truth rather than an estimate with a specific blind spot.
• Comparing attributed revenue across channels that use different attribution windows without adjusting for it.
• Switching attribution models mid-year and comparing the new numbers to old benchmarks as if nothing changed.
• Ignoring incrementality entirely because it's slower and less granular than a live dashboard — it is the only method that answers "did this cause the sale."
• Picking a model that needs more history or scale than the business actually has, and trusting noisy early output.

For the platforms discussed here, see attribution and marketing mix modeling tools and mobile attribution tools. To run marketing mix modeling well once you've chosen it, see how to run marketing mix modeling, and for measurement once third-party cookies are gone, see how to measure marketing without third-party cookies.