Attribution software: seven criteria for comparing options
Put validation first: ask whether a vendor's numbers have been checked against randomised experiments. Then compare input data, method, price model, privacy, exports and time to first answer, and send the gaps to the vendor as written questions.
€99 once, excl. VAT. Full refund within 30 days, no questions asked. You keep the read.
By Joris van Huët, Founder & CEOPublished 6 min read
Start with validation, then compare six other things. In 663 Facebook experiments (Gordon, Moakler and Zettelmeyer, Marketing Science, 2023), a machine-learning method with over 5,000 user-level features estimated a median upper-funnel lift of 83% where the randomised experiments measured 29%. Ask every vendor whether its numbers have been checked against a randomised test, then compare input data, method, price model, privacy, exports and time to first answer.
What are the seven criteria, and what does a pass look like?
| Criterion | Ask the vendor | A pass looks like |
|---|---|---|
| Validation | Has any number been compared with a randomised test, and can I see it? | A named experiment or calibration with its result. Accuracy claimed with no method fails. |
| Input data | What must be installed or exported before the first number? | A short list you can verify, such as a file export or a script you can read. |
| Method | Does it credit touchpoints, estimate contribution over time, or read a test? | A plain answer naming the question the number answers. |
| Price model | What do I pay at my order volume, billed how, for how long? | A price you can read without a call, and terms you can leave. |
| Privacy | Which personal data reaches you, where is it processed, and in what role? | A data processing agreement and a named processing location. |
| Exports | Can I take inputs and outputs out as files and recompute the totals? | Yes, in a format that opens without the vendor's software. |
| Time to first answer | How long from sign-up to a number I can check against my orders? | A stated time, and a first number that reconciles to your store. |
Why does validation come first?
An estimate that has never met an experiment has no known error. In the same paper the experiments put median lifts at 29%, 18% and 5% for upper, middle and lower funnel outcomes, while double/debiased machine learning gave 83%, 58% and 24%, and the authors conclude that they are unable to reliably estimate an ad campaign's causal effect. The data covers Facebook advertising and is richer than most advertisers can access, so it shows what can go wrong, not how any named vendor performs.
An experiment is validated by design but answers a narrower question. Google Research describes geo experiments as randomly assigning non-overlapping geographic regions to a control or treatment condition, and Haus says it uses random stratified sampling to assign test and control groups. The result covers the channel and window tested. Calibration joins the two ideas: Google's Meridian documentation describes calibrating a marketing mix model with experiments, and Meta's Robyn says it calibrates models against ground-truth methods such as geo-based tests and lift studies.
How do you score two vendors in an afternoon?
Send both vendors the same seven questions in writing and score each answer:
- Pass: the answer points to a document, a file or a test you can check yourself.
- Partial: the answer is a claim with nothing to check.
- None: no answer after a reminder.
No answer on validation counts as a fail, and a vendor's statement about its own accuracy counts only with the comparison behind it. Then run the check no vendor can do for you: compare the orders it credits for the last full month with your Shopify orders for the same dates. A gap you can explain (refunds, cancellations, time zones, buyers who declined tracking) passes. An unexplained gap fails.
What can you pre-fill from vendors' public pages?
These entries come from each vendor's own pages as read on the dates shown; a missing figure means the page lists none.
- Northbeam: multi-touch attribution on a tracking script, with managed onboarding. Starter is from $1,500 and Professional from $3,500 a month, priced on pageviews, with Growth and Enterprise quoted; Starter is month-to-month, while Professional and Enterprise are annual terms (northbeam.io/pricing, as read on 2026-09-08).
- Wicked Reports: click-based multi-touch attribution on a pixel, with cart and CRM integrations. Plans are tiered by annual revenue, from $499 to $999 a month, with Enterprise from $4,999 (wickedreports.com/pricing, as read on 2026-09-08).
- Polar Analytics: a BI platform with multi-touch attribution on its own first-party pixel, which does not work retroactively and needs about two weeks of active tracking, per its help center. The Shopify App Store lists Core from $750 a month, priced on online GMV, with discounts for annual terms (apps.shopify.com/polar-analytics, as read on 2026-09-26).
- Rockerbox: multi-touch attribution, marketing mix modelling and control-group incrementality testing, with no public price and integration-based onboarding.
- Haus: geo experiments, with Causal Attribution and Causal MMM calibrated on them. It lists no price, and each of its four plans starts with a demo (haus.io/pricing, as read on 2026-09-26). By its own guidance a test runs 2 or 3 weeks for demand capture such as search and 6-8 weeks for Meta reach campaigns.
- Recast and Measured: Recast does Bayesian marketing mix modelling and Measured does geo-based incrementality testing; both are quoted through sales, with no public pricing page.
Pricing pages are a thin source for validation, privacy and exports, so send those three as written questions; the EDPB's guidelines on controller and processor frame the privacy answer.
Causality Engine reads one file, the Attribution paths export from GA4 as a CSV, and shows what each channel caused next to what last-click gave it, with a data-health score from 0 to 100 and a next step for every channel. A read takes 1 to 2 minutes after upload and costs €99 once, excluding VAT, refundable within 30 days, no questions asked; Pro is €299 a month. The read is observational, so it runs no experiment, it works on GA4's channel groups rather than campaigns, and direct connections to ad platforms are on the roadmap, not live.
Sources, 30 September 2026: Close Enough? (Gordon, Moakler and Zettelmeyer, Marketing Science, 2023, peer-reviewed); Measuring Ad Effectiveness Using Geo Experiments (Vaver and Koehler, Google Research, 2011); Haus incrementality experiments; Meridian (Google); Robyn (Meta Marketing Science); Polar help center and Haus test-length guidance (vendor-published, not independently audited); the vendor pricing pages linked above; and the last entry's own pricing, comparison and integrations pages (causalityengine.ai).
Related answers
Frequently asked questions
What should I ask an attribution vendor first?
Ask whether any of its numbers have been compared with a randomised experiment, and ask to see the comparison. A peer-reviewed study in Marketing Science reported significant relative measurement errors for two observational methods, even with rich user-level data, against Facebook experiments.Can I compare attribution software by price alone?
No. Price models differ: some vendors publish monthly tiers, some price on revenue or data volume, and some quote only through sales. Compare what it costs to reach a first number, including setup time and contract term, and check that number against your Shopify orders.Which attribution vendors publish their prices?
It varies. On the dates read, Northbeam, Wicked Reports and Polar Analytics publish figures, while Haus, Rockerbox, Recast and Measured list no price. Check the vendor's own page, because prices change without notice.
Go deeper: Incrementality testing, explained.
Sixty-second versions of these ideas: Causality Engine on YouTube Shorts.
Keep reading
Terms in this article
- Attribution PlatformAttribution Platform is a software tool that connects marketing activities to customer actions. It tracks touchpoints across channels to measure campaign impact.
- Causal AttributionCausal Attribution uses causal inference to determine which marketing touchpoints genuinely cause conversions, not just correlate with them.
- Control GroupControl Group is a segment of an audience intentionally not exposed to a marketing campaign, used to measure the campaign's true causal impact.
- IncrementalityIncrementality measures the true causal impact of a marketing campaign. It quantifies the additional conversions or revenue directly from that activity.
- Incrementality TestingIncrementality Testing measures the additional impact of a marketing campaign. It compares exposed and control groups to determine causal effect.
- Machine LearningMachine Learning involves computer algorithms that improve automatically through experience and data. It applies to tasks like customer segmentation and churn prediction.
- Marketing MixThe marketing mix is the set of actions a company uses to promote its brand or product. It traditionally includes product, price, place, and promotion.
- Multi-Touch AttributionMulti-Touch Attribution assigns credit to multiple marketing touchpoints across the customer journey. It provides a comprehensive view of channel impact on conversions.