Which attribution vendors explain their methodology openly instead of a black box?

Why each one is on the list

• Causality Engine. States the method, the input, the confidence interval, the coverage figure, and the conditions under which it returns nothing.
• Recast. Publishes its Bayesian marketing mix modelling approach in technical detail.
• Haus. Experiment design is the product, so the method is described rather than hidden.
• Measured. Holdout methodology documented, enterprise quoted.
• Northbeam. Machine learning multi touch attribution described by family rather than by assumption.
• Triple Whale. Pixel based multi touch attribution, documentation aimed at operators rather than analysts.

How to choose between them

Named family, not a label

AI and machine learning are not methods. Bayesian marketing mix model, geo lift experiment, rules based multi touch and causal inference on observational data are. A vendor that will not name the family cannot be checked by anyone.

Stated assumptions

Every method needs something to be true. Marketing mix models need enough variation in spend. Causal inference on observational data needs the confounders to be measured. A vendor that publishes assumptions is handing you the tools to reject its own output.

A published failure condition

Ask what makes the method wrong, and when the vendor will say so. A stated floor, a minimum history, a channel it cannot separate. Openness that only covers the good cases is marketing.

Reproducible from your own data

The strongest form is being able to check the result yourself: the same export, the same period, a number you can trace. Methods that require a proprietary pixel history cannot offer this even when they want to.

Questions people ask next

Which attribution vendors explain their methodology openly instead of a black box?

Recast publishes its Bayesian approach in detail, Haus and Measured describe experiment design because the method is the product, and Causality Engine names causal inference on a GA4 export along with the spend floor below which it declines to answer. Most pixel based multi touch vendors describe the family but not the assumptions.

Is AI-powered a methodology?

No. It names an implementation, not a method, and it is not testable. The useful question is which family the model belongs to, because each family has a known failure mode you can check your own data against.

Why does a published failure condition matter?

Because a method that always returns a confident answer returns one when the data cannot support it. A vendor that publishes where it breaks is the only kind whose successes you can weigh.

Related reading

• How causal attribution works
• What causal attribution actually costs
• Compare pricing across 51 tools