How to Choose an Attribution Platform for Incremental Sales

A buyer's guide to the five questions that separate causal measurement from credit reallocation, with a decision table by ad spend and data maturity.

Updated 8 September 2026 · Joris van Huët, founder, Causality Engine

The fastest way to sort attribution vendors is to ask one thing: does the tool estimate what sales would have happened without the advertising, or does it just hand out credit to whatever showed up before a purchase? Almost everything else is downstream of that. If a platform cannot describe its counterfactual, it is reporting correlation dressed up as measurement, and you will pay for that gap the next time you move budget.

I have sat through too many budget meetings where two dashboards disagreed by a factor of three on the same channel. Meta claimed a number, GA4 claimed another, and the last-click report told a third story. Nobody in the room could say which was right, so the loudest opinion won. That is not measurement. That is folklore with a login screen. And it is not an annoyance to smooth over in a spreadsheet, either. The three numbers answer three different questions: what the platform could associate with its own inventory, what the analytics tool could resolve to a source, and what the commerce platform recorded as orders. Only the last one is a fact about money, because it is the one number the seller of the media did not produce.

Below are the five questions I actually ask vendors, the two questions they least want to hear, the differences between the main methods, and a decision table so you can match a tool to your spend and data reality instead of buying the most expensive thing with the best sales deck.

The five questions to ask any attribution vendor

1. What exactly is your counterfactual?

Make the vendor finish this sentence out loud: "For this channel, the counterfactual is the sales that would have occurred if..."

Good answers are specific. The audience had not seen the ad. A geo had not received the campaign. The channel had received less spend. A weak answer is a shrug, or worse, a confident description of how their model assigns credit, which is a different thing entirely.

This matters because the counterfactual defines the claim. Haus builds its counterfactual from treatment and holdout groups, so the comparison is a group that saw the ad against a group that did not. Causality Engine builds its counterfactual by modeling a lower-spend world from the brand's own sales history and natural variation in the data. Both are causal in intent. They are not the same kind of evidence, and a serious vendor will tell you which one you are getting.

Follow-ups worth asking: Is it randomized, quasi-experimental, or modeled from observational data? Is it calculated per channel or lumped together? Can you see the baseline sales the incremental number is measured against? What assumptions have to hold for it to be credible?

2. How do you handle confounding?

Confounding is the quiet killer. Something else moves both your spend and your sales at the same time, and the model gives credit to the wrong thing. Promotions. Seasonality. A product launch. Branded search, where high-intent shoppers were going to buy anyway and happened to click your ad on the way.

Ask which variables are in the model, how promotions and launches are represented, and what happens when two channels move together. Haus notes that traditional MMM can be vulnerable to multicollinearity and uses experiment results directly for tested channels. Measured ingests incrementality tests as causal priors. Sellforte applies calibration multipliers to pull attribution toward incremental ROAS.

For any observational tool, including Causality Engine, do not accept the idea that statistical adjustment scrubs out all confounding. It does not. The right framing is that the platform accounts for the confounders it can measure and is upfront about the assumptions and the leftover uncertainty. If a vendor claims their model removes bias entirely, that is a sales claim, not a statistical one.

3. Does it report uncertainty, not just a point estimate?

A report that says "Facebook drove 84,000 euro" with no range is selling you false precision. The number is an estimate. Estimates have error bars.

Ask to see confidence or credible intervals on incremental sales and ROAS. Ask whether an interval crossing break-even means "cut this" or "we can't tell yet." Causality Engine reports a confidence interval on every channel estimate, and an interval that crosses 1.0x incremental ROAS is a signal the channel may not be pulling its weight. Haus explains interval width plainly: narrow means precise, wide means uncertain. Recast can take a prior test's point estimate and its standard error as a model input.

Then ask for the floor. An interval without the minimum detectable effect of the design is incomplete: "12% lift, interval 4 to 20, from a design that could detect 8%" is a defensible sentence, and the last clause is the one that decides whether the first two mean anything.

Here is the real test. Ask the vendor to show you an inconclusive result, not another glossy case study. A platform you can trust in a budget meeting is one that can say "the evidence is too weak to move money right now." If every read comes back with a clean, confident recommendation, be suspicious.

4. Can the method be explained to a CFO?

Your finance lead needs six things: what was spent, what sales occurred, what would likely have happened without the spend, the estimated incremental contribution, the plausible range, and the budget action that follows.

Ask for a one-page explanation with the intervention, the counterfactual, the outcome metric, the adjustment variables, the incremental math, the uncertainty range, and the assumptions. If the vendor answers with "AI-powered" or "proprietary models," that describes their engineering, not their claim about your sales. Push back.

The cleanest way I have found to explain any of this to finance: "The tool does not ask which ad got credit. It asks how many sales changed when the marketing changed, compared to the best estimate of what would have happened otherwise." Measured's split is also useful here: incrementality for causal validation, MMM for planning, attribution for tactical execution.

5. What data does the platform need?

Ask for two datasets: the minimum to get a first read, and what a reliable production setup actually requires. These are rarely the same.

The requirements vary a lot. Causality Engine runs causal inference on a GA4 export with no pixel, no SDK, and no code project, and it accepts Shopify data when you have it. One read is 99 euro, Pro is 299 euro per month, and the company is registered in Utrecht. Recast wants daily spend by channel plus a business KPI, recommends around two and a half years of history, and treats a promotional calendar as required when promos move the business materially. Sellforte needs marketing activity and business outcomes, with GA4 or Adobe used to calibrate. Measured leans on a large integration library across media, spend, and sales. Haus is experiment-led and needs enough geo or audience volume to build clean holdouts.

If a vendor needs a warehouse, a data engineer, and six weeks of onboarding before you see a number, that is a fine tradeoff for a large advertiser and a terrible one for a brand doing 40,000 euro a month.

The two questions vendors least want

Ask these of every vendor on this page, including us.

What is my minimum detectable effect under your design, with my data, before we start? The floor is arithmetic, not a vendor guideline. Multiply a channel's share of revenue (spend divided by revenue) by the return you would honestly defend for it; that is roughly how much total revenue would move if the channel stopped. Compare it with the smallest lift the design can distinguish from noise, which for a typical DTC brand with six months of history and an eight-week test is about 8%. If the first number is smaller than the second, the test cannot answer the question, however good the tool.

Which of my channels are not measurable at my current spend, on your method? A vendor who says "all of them are measurable" has just failed the arithmetic. The honest answer for most advertisers is a list, and it is not short. Question two is the one that separates a measurement partner from a measurement salesperson.

How the methods actually differ

The word "attribution" gets stretched over five very different things. Keeping them straight is most of the battle.

Method	What it does	Counterfactual	Best buying question
Rule-based attribution (last click, linear, position)	Assigns credit using fixed rules	Usually absent; reallocates observed conversions	"What proves credited conversions wouldn't have happened anyway?"
Data-driven attribution / MTA	Algorithmic credit across touchpoints	Modeled journey, dependent on tracking coverage	"Does it estimate a treatment effect or just allocate credit?"
Traditional MMM	Relates media changes to sales over time	Modeled sales path with different media inputs	"Which experiments anchor the causal read?"
Incrementality testing	Compares treated vs untreated groups	The control outcome: what happened without the ads	"How are treatment and control built and protected from contamination?"
Causal inference on observational data	Estimates effect from historical variation, no new holdout	Predicted outcome under different spend	"Which confounders are handled, and how wide are the intervals?"
Causal / test-calibrated MMM	MMM anchored by experiment results	Model-based scenario, anchored by tests where available	"Which outputs are tested, calibrated, or still observational?"

Two things people get wrong here. First, "data-driven" does not mean causal. A machine learning model that allocates credit across observed touchpoints is still working with correlation, and under selection a model that fits the observed data better may be reproducing the platform's targeting more faithfully rather than recovering the advertising's effect. Measured says as much: attribution generally shows correlation, incrementality aims for causation. Second, plain MMM is not experimental. It becomes causal when experiments calibrate it, which is exactly why Haus and Measured draw a line between ordinary MMM and their causal versions.

The honest hierarchy of evidence, strongest to weakest for a given decision: a clean randomized holdout, then test-calibrated MMM, then observational causal inference, then MMM without tests, then data-driven attribution, then a fixed rule. Cost and effort run roughly the same direction. The trick is buying the least expensive method that is strong enough for the decision in front of you.

A decision table by spend and data maturity

No vendor I reviewed publishes a spend cutoff where one method suddenly becomes valid, so ignore anyone who tells you "under 100,000 euro use X." Use bands and match to what you can actually feed the tool.

Situation	Data you have	Start with	Vendors that fit	Main warning
Low or exploratory spend	GA4 and order history, no warehouse	Causal inference on existing export	Causality Engine	It's observational, not a holdout test. Read wide intervals conservatively.
Low to moderate spend	GA4 plus clean daily spend and sales	One-off causal read, then a targeted test if a big decision stays unclear	Causality Engine, Haus	Don't buy an always-on system before the data justifies it.
Moderate spend	A year-plus of structured data, promos documented	MMM or causal inference, calibrated by tests where possible	Recast, Causality Engine, Haus	A precise-looking model can still be confounded. Ask which channels have direct test evidence.
Moderate to high spend	Reliable channel data, ability to run geo tests	Incrementality testing plus model scaling	Haus, Measured, Sellforte	Tests need volume, duration, and clean control groups.
High spend, many channels	Warehouse, offline/retail sales, test history	Test-calibrated MMM with ongoing experiments	Measured, Sellforte, Haus, Recast	Demand a channel map: tested, calibrated, observational, or platform-reported.
High spend, granular optimization	Mature attribution plus transaction data	Causal calibration of granular attribution	Sellforte, Measured	Campaign-level precision may be inherited from a broader channel estimate. Ask how uncertainty carries through.
Large omnichannel portfolio	Sales across ecommerce, retail, marketplace, offline	Triangulated measurement	Measured, Sellforte, Haus	Make sure it measures the outcome finance trusts, not just online conversions.

Where I would start, and what I would look for

If you run a Shopify or DTC brand with GA4 history and no data science team, start with a causal read on what you already have before you commit to anything monthly. Causality Engine's 99 euro read exists for exactly this moment. You get per-channel incremental ROAS with confidence intervals off a GA4 export, no pixel, no onboarding project, and a refund if the read does not move a single budget decision. That is the cheapest way to find out whether a channel you have been funding on faith is actually moving sales. In the published causal read for Me Gorgeous, a Dutch DTC brand, roughly 2,000 euro a month of Meta spend showed no incremental contribution. Another brand kept Pinterest alive that last-click wanted to kill.

The honest caveat, which the vendor states plainly: this is causal inference on your history, not a randomized holdout. The counterfactual is modeled from your own natural variation, not created by withholding ads from a control group. For most brands under mid-market scale, that tradeoff is exactly right. You get an always-available read from data you already own, and you keep the option to run a formal test later when a single decision is worth the setup.

One more thing, and it applies to us as much as to every vendor named here. Between 14 August and 2 September 2026 we audited thirty-one commercial measurement vendors for a published validation of their method against randomised experiments, with the sample, the design and the discrepancies disclosed. We found none. Hold Causality Engine to the same question, and put the two questions above to us before you weight the read.

Move up the ladder when the money justifies it. Once you are spending enough that a wrong call costs real revenue and you can build clean holdouts, Haus and its experiment-led approach earns its keep. At full omnichannel scale with offline and retail in the mix, Measured and Sellforte are built for triangulation across everything. Recast fits when you want Bayesian planning and forecasting off a long, clean history.

The pattern I keep coming back to: buy the smallest method that answers your actual question, insist on a counterfactual, an interval and a floor, and never accept a number you cannot explain to the person controlling the budget.

FAQ

Is causal inference on a GA4 export as good as a randomized holdout test?

No, and any vendor claiming otherwise is overselling. A holdout test gives you direct evidence by comparing groups that did and did not see ads. Causal inference on observational data estimates the same effect statistically, using your sales history and natural variation. It is weaker evidence but far cheaper and always available, which makes it the right first move for most brands before they invest in formal testing.

Why do my ad platform and my attribution tool disagree so much?

Ad platforms count conversions that touched their surface, which double-counts across channels and rewards clicks that would have converted anyway. A causal tool asks a different question entirely: how many sales changed because of the spend. The gap between the two numbers is usually the over-attribution you have been paying for. Track it as a standing metric: all platform-claimed conversions divided by the orders you actually shipped.

Can I trust a confidence interval that crosses break-even?

That is the tool doing its job. An interval crossing break-even ROAS means the evidence cannot confirm the channel is incremental at current spend. The right move is not always to cut it. Sometimes it means you need more data or a proper test before deciding, and sometimes the channel is simply below the floor at your scale. Treat "we don't know yet" as a valid and useful answer.

What is the cheapest way to check whether a channel is wasting money?

A one-time causal read on your existing GA4 data. Causality Engine prices this at 99 euro, refunded if it does not move a budget decision, which is far below the cost of an enterprise measurement contract or a formal incrementality test. It will not replace ongoing measurement, but it is the fastest way to find obvious waste before you commit to anything larger.

Sources and further reading

• The Price of Being Found (Causality Engine, Edition 2.10, September 2026): the scoreboard (Chapter 9), the measurability floor (Chapter 15), the vendor audit and the twelve questions (Chapter 17), what a defensible number looks like (Chapter 19)
• Causal attribution for ecommerce brands
• Incrementality testing without geo holdouts
• When geo-testing is worth it
• Sixty-second versions of these ideas on YouTube Shorts

Vendor prices and features quoted in this article were taken from each vendor's own website on 8 September 2026 and may have changed since. Check the vendor's pricing page before relying on a figure.