Attribution Vendors That Explain Their Methodology Openly

An attribution vendor explains its methodology openly when it states, in writing you can read before you buy, what the model assumes, whether each number is experimental or observational, what the confidence interval is, which of your channels it cannot measure at your spend, and what it would report if your advertising had no effect at all. A black box is any vendor that answers those with an accuracy percentage instead. The Price of Being Found audited the public materials of 31 commercial measurement vendors between 14 August and 2 September 2026 and found none with a published validation against randomised experiments, which is the reason the questions matter more than the marketing.

What open actually means

Open is not open source. No commercial vendor publishes its estimator, and none needs to. Open means the four properties of a defensible number are visible on the vendor's own pages and in its reports:

1. The source of the number, and whether that source sells you media.
2. The coverage: how much of your real orders the data could see.
3. The design: experimental, quasi-experimental, or observational, and if observational, labelled as a description rather than a causal claim.
4. The interval, and the smallest effect the design could have detected.

A vendor that shows all four on a sample report is explaining its methodology. A vendor that shows a single confident number per channel is not, whatever the method behind it.

The thirteen questions

The book's list is addressed to any vendor, including this site's publisher. The ones that separate open from opaque fastest:

• What is my minimum detectable effect under your design, with my data, before we start?
• Have you validated this method against randomised experiments on real data, not simulated, not backtested? Show the study, the sample, and the discrepancies.
• What is my coverage, the share of my true conversions your system observes, and how did you compute it?
• When your number disagrees with my commerce platform's orders, which is wrong, and how would we find out?
• Does your estimate come from a model fitted on observational data, on experimental data, or both?
• Is there an LLM anywhere in the causal estimate, as opposed to the presentation of it?
• What result would this system produce if my advertising had no effect at all? Have you tested that?
• What is the confidence interval, not the point estimate?
• Which of my channels are not measurable at my current spend, on your method?
• What would falsify your model?

The book's note on the measurability question: a vendor who says all of your channels are measurable has just failed it. Ask your attribution vendor for a placebo test covers the placebo question in depth, and how to vet an attribution vendor before you sign turns the list into a checklist.

Why "AI-powered" is not an answer to any of them

Two of the largest vendors of measurement tooling are the platforms themselves, and their own documentation is instructive. The book quotes Google's Meridian page on assessing a fitted model: directly validating the quality of causal inference is difficult and requires well-designed experiments, and since you are using an MMM, experiments are likely not practical. A method that cannot be validated by its own documentation is not made more open by a machine-learning label. The same applies to any vendor whose estimate comes from a model fitted on observed journeys: the better it fits, the more faithfully it may be reproducing the platform's targeting rather than the advertising's effect, which is the selection trap the Facebook experiments the book quotes demonstrated.

An LLM in the presentation layer is fine and useful. An LLM in the estimate is a question mark, and the honest vendor says which it is.

What this site publishes, for comparison

Causality Engine's method is proprietary causal inference, counterfactual estimation on aggregated first-party data, and it is not machine-learning attribution and not an LLM. The how it works page states the assumptions: complete and accurate spend and sales data, enough natural variation in the window to identify each channel's effect, and no unobserved external factor moving sales in lockstep with one channel. Every estimate carries an interval that widens where an assumption is uncertain, the read states which channels fall below its fit floor rather than inventing a number, and a methodology document is available on request. It is bound by the same audit finding as everyone else, and the placebo question applies to it too.

What to do this week

• If you have to defend the number: send the ten questions above to your current vendor in writing. Answers that exist in September are a specification; answers that arrive with the December report are a rationalisation.
• If you own the budget: ask only the measurability question. The list you get back is the list of channels the vendor is currently describing with false precision.

The interactive demo shows what a report with intervals, coverage and a stated design looks like on a sample store, with no signup.

As of 9 September 2026. The 31-vendor audit, the vendor questions and the Meridian quotation are from The Price of Being Found (Edition 2.10), Chapter 17, with the book's stated limitations: the audit covered public materials only.