Top Platforms for Measuring Incremental Revenue by Channel

A practitioner's guide to choosing the right incrementality tool for your DTC brand, from geo-tests to MMM to fast historical causal reads.

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

The fastest way to lose an argument in a budget meeting is to bring platform-reported ROAS. Meta says it drove the sale. Google says it drove the sale. Your last-click GA4 report says email drove the sale. Add them up and you have attributed 180% of your revenue to paid media, which is a physical impossibility.

That 180% has a name. It is your claim ratio: every platform's claimed conversions added up, divided by the orders you actually shipped. Anything above 1 is the amount by which your suppliers collectively believe they did more work than exists. It costs nothing to compute, and it is the single most effective defence against a scoreboard you do not own, because the party that sells you the advertising is also the party measuring whether it worked, and nobody audits the result.

The question that matters is different from any of those dashboards. It is: what would have happened if we had turned this channel off? That gap between "revenue with the channel" and "revenue without it" is incremental revenue, and almost none of the numbers you get by default measure it.

This is a guide to the platforms that actually try to measure that counterfactual, sorted by the job they do. There is no single winner across all situations, because a Meta lift study, an MMM, and a historical causal read are not interchangeable kinds of evidence. Picking the wrong one for your stage wastes weeks.

What "incremental" actually means (and why attribution isn't it)

Attribution distributes credit among the touchpoints it can see. Incrementality compares what happened against an estimate of what would have happened without the channel. Those are different math.

Incremental revenue = Revenue with the channel − Revenue under the counterfactual
Incremental ROAS   = Incremental revenue ÷ Incremental media spend

The hard part is the counterfactual, not the credit rule. Rockerbox itself draws this line: attribution assigns credit among observed touchpoints, while incrementality tests compare against a baseline that would have happened anyway. Meta is even blunter about it, warning that Conversion Lift results should not be compared directly with Ads Manager numbers because they measure different things.

If you take one idea from this article: brand search and retargeting usually show gorgeous attributed ROAS precisely because they reach people who were already going to buy. That is exactly why they are the first channels you should be suspicious of, not the last.

The measurement jobs, and which platforms fit each

There are five distinct jobs. Most brands need two or three, sequenced over time.

Controlled experiments give the strongest causal evidence. You randomize users or geographies, suppress or increase spend, and measure the difference. This is the gold standard when you have the scale to run it cleanly.

Media mix modeling (MMM) looks at historical spend and revenue across all channels and estimates each channel's contribution, plus diminishing returns. Good for ongoing budget allocation, weaker on isolated causality.

Historical causal modeling applies causal inference to data you already have, without turning anything off. Fast, low-disruption, model-based. Good for a baseline and for deciding what to test next.

Native publisher lift studies validate one ad platform using its own randomized holdouts. Meta and Google both offer these.

Build-your-own frameworks like Google Meridian and Meta Robyn are open-source MMM if you have a data team.

Here is how the main candidates map to those jobs.

Platform	Primary method	Best-fit role	Key caveat
Measured	Geo-matched and first-party split tests	Enterprise brands needing managed, repeatable cross-channel experiments	Needs geographic or audience scale and a testable channel
LiftLab	Geo experiments feeding an Agile MMM	Finance-sensitive brands building a recurring test-to-budget loop	Better for brands with real media scale and analytical ownership
Northbeam	MTA + MMM + automated incrementality testing	Teams wanting one operating layer for daily optimization plus lift tests	Verify channel eligibility and test availability during evaluation
Rockerbox	MTA, MMM, and managed incrementality tests	Brands wanting expert-supported tests rather than fully self-run ones	Test results don't replace all attribution reporting
Fospha	Daily MMM across DTC and marketplaces	Omnichannel brands split between owned store and Amazon/TikTok Shop	Outputs stay model-based; depend on data quality and validation
Google Conversion Lift / GeoX	Native user and geo experiments for Google Ads	Validating Google's incremental contribution	Answers the Google question only
Meta Conversion Lift	Randomized test and holdout for Meta	Publisher-specific validation of Meta campaigns	Not comparable with Ads Manager attributed conversions
Triple Whale Incrementality	Self-serve GeoLift and Meta lift	Shopify teams wanting an accessible test interface	A dashboard doesn't remove power and design constraints
Causality Engine	Historical causal inference on a GA4 export	Fast per-channel baseline before investing in controlled tests	Model-based estimate; validate high-stakes calls with experiments
Google Meridian	Open-source Bayesian MMM	Brands with data-science resources	A framework, not a turnkey product
Meta Robyn	Open-source semi-automated MMM	Technical teams building internal MMM	Requires real modeling and engineering effort

Where Causality Engine fits, and where it doesn't

Most brands reading this are not going to spin up a managed geo-testing program next Tuesday. You need a defensible number before your next budget meeting, and you cannot go dark on Meta for three weeks to get it.

That is the gap Causality Engine is built for. You upload a GA4 export, and it runs causal inference across your history to estimate per-channel incremental contribution, with a confidence interval on every number. Setup takes a couple of minutes with no pixel, no code, and no developer. The one-time causal read costs €99 and is refunded if it does not move a single budget decision, which lowers the stakes of finding out whether your paid social is pulling its weight.

I want to be precise about the evidence class here, because it is easy to oversell. A historical causal model is not a randomized holdout. It makes stronger assumptions than a clean geo-test, and for very high-stakes scaling decisions you should still validate with an experiment where feasible. What it does well is give you a fast, low-disruption baseline that tells you which channels deserve a formal test and which ones are obviously fine or obviously wasteful.

The results tend to be blunt in a useful way. In the published causal read for Me Gorgeous, a Dutch DTC brand, about €2,000 a month of Meta spend showed no incremental contribution. The Two Sisters ran a causal read that kept Pinterest live after last-click told them to kill it. Neither of those conclusions comes out of a standard attribution dashboard, because a dashboard cannot see the counterfactual.

For a smaller or mid-sized DTC brand on Shopify and GA4, this is the most sensible first move. It is cheaper than a managed experiment, faster than standing up MMM, and it produces a number you can actually defend rather than a platform-inflated one. When you outgrow it, the results feed naturally into a testing program.

One more thing, and it applies to us as much as to every vendor in the table. 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. The read states its counterfactual and its interval, and it is a model on observational data, not an experiment. Ask us, and every other vendor, which of your channels are not measurable at your current spend. The honest answer is a list.

Before you pick anything: define the decision

The most common failure I see is buying a tool before writing down the decision it needs to inform. Do that first.

Write the question in operational terms:

• Should paid social go up, down, or stay capped?
• Is brand search creating revenue or capturing demand you already had?
• Does retargeting produce new orders or reclaim buyers who'd have converted anyway?

Then lock a primary KPI (incremental revenue, contribution margin, or new-customer revenue), a decision threshold (for example, scale only if the lower confidence bound of iROAS clears your margin-adjusted hurdle), and an analysis unit (user, DMA, city, week). A test optimized for clicks cannot answer a revenue question, and dressing it up afterward doesn't fix that.

Getting your revenue baseline right

Whatever platform you choose, garbage joins produce confident-looking nonsense. Use your commerce system as the financial source of truth. It is the one number in this whole exercise that the seller of your media does not produce. Shopify supports exporting filtered orders and transaction history directly, and you want net revenue reconciled by date and geography: gross sales, discounts, refunds, returns, new versus returning customer.

GA4 is fine for behavioral and campaign context, but treat it carefully as a revenue source. Google's own documentation notes GA4 can include modeled key events when direct observation isn't possible due to privacy or cross-device limits. Before you trust any channel-level incrementality number, run a simple reconciliation:

Shopify net revenue
− GA4 purchase revenue
− payment/refund timing differences
− timezone differences
− excluded channels or stores
= explainable residual

If the residual isn't understood, stop. A model will happily draw tight confidence intervals around a broken join.

Choosing the evidence design

Match the design to what you can actually do.

Randomized user holdout is strongest when the platform can assign people to test and control and observe outcomes at the user level. Watch for audience overlap and identity loss.

Geo experiments work when a channel can be targeted by DMA or region and revenue aggregated at the same level. Google names three primary designs: go-dark (suppress existing media), holdback (withhold a new tactic from control geos), and heavy-up (increase spend in treatment geos). The risk is spillover, national campaigns, and too few independent markets. Our geo-testing guide walks through when this is worth the operational cost.

Two arithmetic checks belong before any geo test, and no vendor can waive them. First, the floor: multiply the 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 your design can detect. For a typical DTC brand with six months of history and an eight-week test, that is about 8%. If the expected effect is smaller, the test cannot answer the question and a null result will be read as an answer. Second, count your units. One treated region against N regions in total means the smallest p-value a placebo-based design can physically return is 1 in N. Twelve Dutch provinces give 0.083, which cannot clear the 0.05 most people report against; the forty COROP regions give 0.025. Work that out before you design the test, not after.

MMM suits always-on cross-channel budgeting when you have enough historical variation and can include promotions, seasonality, and pricing. Meridian's docs list media, spend, KPI, and control variables as core inputs. Its weakness is that it's observational; historical spend variation is often correlated with demand, so MMM is not automatically causal.

Historical causal modeling fits when you can't safely suppress a channel and need a directional baseline fast. Treat the output as provisional and strong enough to prioritize, not to bet the quarter on without follow-up.

Combining evidence without fooling yourself

The worst thing you can do is average platform ROAS, MTA ROAS, MMM ROAS, and lift-test iROAS as if they were four measurements of the same quantity. They use different populations, time windows, outcomes, and counterfactuals. Averaging them produces a number with no meaning.

The defensible pattern is to use each method for its job. Experiments set causal anchors. MMM interpolates across the full mix. Attribution diagnoses journeys and creative. Historical causal reads give you a fast baseline and tell you what to test. LiftLab explicitly runs this loop, feeding each completed experiment back into its MMM to sharpen future budget calls. Meridian and Robyn both support using experiments to calibrate model priors.

Whatever you run between anchors is a model, and confidence decays from the date of the last anchor. Put that date on the dashboard next to the numbers, so everyone reading them knows how long it has been since anything in the system last touched the world.

What I'd do first

If you're a Shopify DTC brand under real time pressure, sequence it like this:

1. Run a historical causal read on your GA4 history to get a defensible per-channel baseline this week. This is where Causality Engine earns its keep.
2. Take the two most suspicious channels from that read, usually brand search and retargeting, qualify them against the floor above, and design a geo or publisher lift test to confirm.
3. Only build or buy MMM once you have a recurring budget-allocation decision and enough clean history to feed it.

Most brands skip straight to step three, spend two months on data engineering, and still can't answer whether Meta is incremental. Start with the fast baseline, then earn your way into the heavier machinery.

FAQ

Is Causality Engine a replacement for geo-testing?

No, and it doesn't claim to be. It's a historical causal model, which makes stronger assumptions than a randomized holdout. Use it to get a fast, defensible baseline and to decide which channels are worth the operational cost of a formal experiment. For high-stakes scaling decisions, validate with a geo or publisher lift test where you can.

Why can't I just trust the ROAS in Meta and Google Ads Manager?

Because those numbers count conversions the platform can claim credit for, not conversions the platform caused. They overlap, so they sum to more than 100% of your revenue, and they can't see what would have happened without the ad. Meta says as much itself: its Conversion Lift metric uses a scaled holdout and isn't the same as Ads Manager attribution.

How much scale do I need to run a proper incrementality test?

More than most calculators suggest. Meta publishes minimum conversion and holdout requirements for a well-powered lift test; check the current thresholds in Ads Manager before you plan one, and expect them to need meaningful spend and time. For geo-tests, the number of markets matters less than having enough independent, balanced markets with low spillover, and the placebo floor of one in N applies regardless. If you're small, a historical causal read is often the more realistic starting point, and some channels will not be measurable at your scale by any method. Manage those on judgement, openly.

What data do I actually need to get started?

For a historical causal read, a GA4 export is enough, and setup takes a couple of minutes with no pixel or code. For experiments and MMM you'll also need media spend and exposure data by campaign and geography, plus a warehouse to hold one clean fact table. Reconcile Shopify net revenue against GA4 before trusting any channel-level output.

Sources and further reading

• The Price of Being Found (Causality Engine, Edition 2.10, September 2026): the claim ratio (Chapter 9), the measurability floor and the placebo floor (Chapters 15 and 16), the vendor audit (Chapter 17), the experimental anchor (Chapter 18)
• Incremental ROAS explained
• When geo-testing is worth it
• MMM explained for DTC brands
• 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.