Marketing attribution models: how to compare credit

Attribution models split conversion credit across touchpoints by rule or algorithm. Compare them on definitions, not on who claims accuracy.

By Bellaso · Published · Updated · 4 min read

A model is a rule, not a verdict

Every attribution model allocates credit. First-touch gives all credit to the first recorded touchpoint, last-touch to the final one, linear splits credit evenly, time-decay weights recent touches more heavily, position-based reserves fixed shares for first and last, and data-driven models learn weights from historical conversion paths. The last of these is an algorithm with data requirements, not a guarantee of accuracy.

Because a model allocates rather than measures, its output describes credit under that rule, not causal impact. A channel that looks strong under last-touch may look weak under first-touch, and neither output says whether the sale needed that touchpoint at all.

One journey, four rules

Consider one path to a $300 recognized sale: Paid social, then Organic search, then Email. Each rule allocates the same $300 differently, and the table below shows the arithmetic exactly; no estimation is involved for the rule-based models.

Time-decay requires timestamps and a stated decay rule. Data-driven credit depends on a fitted model and its inputs. This example supplies neither, so it shows no numeric split for those methods. A vendor should explain its method rather than imply one set of weights applies to every business.

Credit for one $300 recognized sale under four historical rules
ModelPaid socialOrganic searchEmail
First-touch$300$0$0
Last-touch$0$0$300
Linear$100$100$100
Position-based (40/20/40)$120$60$120

Comparing models fairly

Two attribution outputs are comparable only when the inputs match: the same conversion identity, the same lookback window, the same timezone, and the same policy on whether direct or internal traffic is excluded. A difference in any of these, not the model, may be what actually moved the numbers.

And because each model re-allocates the same conversions rather than finding new ones, allocated revenues must never be added across models. First-touch revenue plus last-touch revenue double-counts every conversion it credits.

Handle the uncredited residue explicitly too. Direct visits, missing identifiers, and consent-limited sessions land in whatever bucket the tool assigns, such as 'unattributed' or 'direct', and two models that agree on everything else can still disagree on how large that bucket is. State the exclusion policy, and check the residue's size, before reading anything into a channel's share.

What GA4 offers today, and model scope

As of its current documentation, Google Analytics 4 lists three attribution models: data-driven, paid and organic channels last click, and Google paid channels last click. First click, linear, time-decay, and position-based were removed from GA4 in November 2023, so the rule examples above are historical references, not a claim that GA4 or Bellaso supports them.

For Bellaso product scope, see Signal on the features page and the documentation rather than assuming a named model is available. For B2B reporting, deduplicate CRM opportunities or customers before allocating credit. Consent and identity gaps can leave real touchpoints unobserved.

The practical takeaway for any report you consume: every credited number sits on top of identity, consent, and window choices made upstream. A report depends on recorded or modeled inputs; identity and consent gaps still limit what you can verify. Budget reviews are safer when the credit table travels with those definitions attached.

Frequently asked questions

Which attribution model is most accurate?

No model is universally most accurate. Evaluate each on its data coverage, its validation, and the decision it informs. Data-driven attribution estimates credit from model assumptions and available data; its output is not, by itself, the result of a randomized lift experiment.

Does GA4 still offer position-based or linear models?

No. Google's current documentation lists data-driven, paid and organic last click, and Google paid channels last click; first click, linear, time-decay, and position-based were removed in November 2023.

Can I add revenue from two models to get total revenue?

Never. Models re-allocate the same conversions, so summing their outputs double-counts the business.

Does an attribution model prove a channel caused revenue?

No. A model assigns credit under a rule or fitted weights. Proving causal contribution requires methods like incrementality testing, which no credit-allocation output substitutes for.

Sources

  1. Google Analytics documentation on attribution models and the November 2023 removals

Next steps

Start a free account