Last updated: May 2026

The attribution problem no one wants to admit

Every marketer has lived this moment. Facebook reports a 4x ROAS. Google Analytics shows conversions climbing. Your dashboard is green. Success feels certain.

Then the uncomfortable realization creeps in. Many of those customers were already planning to buy. They bookmarked the site last week. They searched your brand name yesterday. The ad captured credit, but it did not create the sale.

This is the structural flaw in attribution. It measures correlation, not causation. It is good at counting what happened and poor at estimating what would have happened anyway.

The data backs this up. The median ecommerce ROAS in 2024 was 2.04x, meaning half of all brands were operating below a 2:1 ratio on a reported basis. On a true incremental basis, those numbers tend to be much lower. Stella's 2025 DTC benchmarks, drawn from 225 incrementality tests, show a median iROAS of 2.31x across paid channels, with branded Google Search at just 0.70x. The implication is that a meaningful share of platform-reported revenue is not actually caused by the ads taking credit for it.

The antidote is incrementality testing. It is not a better way to attribute. It is a different question entirely. It asks what your marketing truly caused.

Key idea to hold on to: attribution tells you what happened. Incrementality tells you what you caused to happen.

‍

What is an incrementality testing platform?

An incrementality testing platform is software that answers the only question that matters for profitable growth:

"Would this sale have happened without the ad?"

Rather than crediting clicks or views, the platform runs controlled experiments. It compares markets that receive exposure to markets that do not, then calculates the additional revenue and orders caused by the campaign. This is the incremental lift. It is the cleanest way to separate signal from noise.

Attribution vs. incrementality

[Insert original comparison image]

The goal is not to replace attribution. The goal is to ground decisions in causal truth when real money is on the line.

‍

Why incrementality alone is not enough in 2026

This is the most important shift in the measurement space over the last 18 months, and most platforms have not caught up to it.

Incrementality testing answers the channel-level question well. It tells you what Meta, Google, or CTV truly drove during the test window. But a single test is a snapshot. It does not tell you what to do next month, or how to allocate across all channels at once, or how the curve shifts as you scale.

That is what Marketing Mix Modeling is for. MMM provides the always-on view, accounts for cross-channel interactions, and produces the response curves that drive allocation decisions.

The two methodologies are not competing approaches. They are complementary, and the industry consensus has converged on this. The accepted best practice is to run an MMM continuously and use periodic incrementality tests to calibrate it. A 2025 Sellforte study illustrates why: an MMM-estimated ROI of 3.91 was validated by an incrementality test that returned 4.00, with a 90% confidence interval of 2.91 to 5.09. The model and the experiment agreed because the model had been calibrated against real causal evidence.

Without that calibration, MMMs drift. They are mathematically sophisticated guesses based on historical correlations. With incrementality calibration, they become evidence-grounded decision systems.

This is why a serious measurement program needs three things working together:

‍

1. Periodic incrementality tests to establish causal ground truth on specific channels and tactics
2. A continuously running MMM that translates that truth into allocation decisions across the full mix
3. Always-on incrementality monitoring to catch when channel performance shifts between formal tests

‍

If your current vendor only does one of the three, you are paying for a partial answer.

‍

How incrementality testing platforms work

The concept is simple. The execution requires rigor. Here is the flow used by serious teams.

Step 1: Location selection

Most tests are won or lost here. If your test and control regions do not naturally behave the same way before the test, no model will rescue the results afterward.

What strong location selection looks like:

• 90 to 120 days of pre-period revenue data analyzed per region
• Correlation analysis run across candidate pairs or clusters
• A/A tests that simulate a treatment with no actual changes
• Validation against objective fit metrics:
  
  • R² above 0.7 is a healthy target for many brands
  • MAPE below 20 percent indicates reasonable predictive accuracy
  • Correlation coefficient above 0.8 signals strong co-movement

A practical example:

• Poor match: Seattle vs. Miami. Different climate, seasonality, and demand pulses produce weak fit.
• Strong match: Portland vs. Austin. Similar patterns yield high R² and low MAPE.

Stella's own benchmark research found that test fit quality is the single strongest predictor of statistical significance. Tests with MAPE below 0.15 and R² between 0.85 and 0.94 reached 100 percent significance. Tests with looser fit failed at meaningful rates, regardless of duration or budget.

If markets do not move together in the past, differences during the test are not credible evidence of lift. The integrity of your test begins with correlation, not with modeling.

Step 2: Experimental design

Pick the design that answers your question with the least noise.

Geo holdouts

• Turn ads on in test markets and hold them out in controls
• Useful for new channels or tactics that were not previously live
• Typical duration is four to eight weeks where purchase cycles are longer

Inverse holdouts

• Pause ads in test markets and keep them live in controls
• The most common design because most teams start by measuring their largest active channels
• Shorter duration is common since the business impact is immediate and visible

Scale tests

• Three-cell design to estimate marginal returns at higher spend
• Cell A: baseline
• Cell B: scaled up, for example 150 to 250 percent of baseline
• Cell C: scaled up more aggressively, for example 200 to 500 percent
• Produces a response curve that shows where diminishing returns begin

‍

Start with an inverse holdout on your biggest channel. You will learn more in two to four weeks than months of dashboard debate.

Step 3: Implementation approaches

This is where platforms diverge, and the trade-offs are not trivial.

Automated implementation

• Connect directly to ad platforms and auto-apply regional exclusions
• Convenience is high
• Cost is high due to integration maintenance and account liability
• Risk of contamination if a new campaign launches without exclusions

Manual implementation

• Platform identifies optimal test and control regions
• Your media team applies targeting and exclusions directly in the ad platforms
• Keeps cost low and control high
• Reduces the risk of silent errors from automation

The most expensive part of incrementality is a bad test, not the software. Human control with a clear checklist often yields cleaner experiments.

Step 4: Data upload and modeling

Once the test and any cooldown window complete, the platform needs clean, daily, region-level data.

Core inputs:

‍

• Date and region
• Revenue by region
• Orders by region
• Tested channel spend by region
• Optional confounders such as other channel spend, promotions, or stock-outs

How most platforms handle this

Most enterprise platforms connect directly to ad accounts and automate data pulls. This reduces manual work but is a major driver of price. Maintaining deep integrations and taking on account-side responsibility pushes the monthly cost into the $4,000 to $12,000 range, with Measured starting at roughly $50,000 per year for incrementality alone according to public industry sources. It also creates a failure mode. A junior buyer launches a campaign without exclusions and the test is compromised before the data even enters the system.

How Stella handles this

Stella uses a clean Google Sheet or CSV upload. Data can come from Shopify, your analytics platform, or ad exports. This keeps the product affordable, preserves flexibility across any channel, and keeps operational control with your team. For most brands running studies regularly, the return on heavy integrations is limited. For all brands, the integrity of the test is paramount.

What happens after upload in Stella:

• Multiple models evaluate your data rather than a single method
• Weighted Synthetic Control, BSTS, and GeoLift all run in parallel
• The platform automatically surfaces the approach with:
  
  • The highest R²
  • The lowest MAPE
  • The tightest iROAS interval
  • Strong statistical significance
  • Stable cumulative impact behavior

Different models can produce different answers on the same dataset. Hiding that variance with a one-model approach creates false certainty. Showing multiple candidates and selecting based on fit creates trustworthy outcomes.

Step 5: Reporting and post-treatment

A credible report is both statistical and practical.

Expect to see:

• Incremental revenue and incremental orders
• iROAS, calculated as incremental revenue divided by tested spend
• Confidence intervals showing the plausible range of iROAS
• Cumulative impact rising during treatment and flattening when lift stops

Many teams add a short post-treatment window. For a geo holdout, you turn ads back off. For an inverse holdout, you turn ads back on. For a scale test, you return spend to baseline. This is a useful validity check. If results normalize as expected, your causal story is stronger.

‍

Use cases that deliver real value

Incrementality testing is not just a method. It is a decision engine across common marketing questions.

Channel-level measurement. Quantify the true lift from Meta, Google, TikTok, or CTV when attribution is inflated. Stella's 2025 benchmarks show wide variance: Tatari CTV at 3.30x median iROAS, Google Performance Max at 2.98x, Meta at 2.92x, and Pinterest at 2.96x. Branded Google Search came in at 0.70x, a result that surprises most marketers and routinely shifts six and seven figures of annual budget.

Campaign optimization. Separate branded from non-branded search. Compare bidding strategies head to head.

Creative testing. UGC versus polished production across matched markets, measured at the business outcome rather than the platform metric.

Upper funnel validation. YouTube prospecting, linear TV, and podcasts rarely look good in click-based analytics. Geo holdouts give them a fair read.

Budget scaling. Map response curves and move dollars away from saturation, toward channels with room to grow.

Finance alignment. Translate iROAS into EBITDA so marketing and finance agree on what is profitable.

Each use case helps a team shift from vanity metrics to business outcomes.

‍

Build vs. buy: should you DIY in R or Python?

Teams with technical talent often consider building their own workflow. There are open-source options, including GeoLift in R, CausalPy in Python, and Google's recently released Meridian and GeoX libraries. These are valuable tools. They are also easy to misuse.

Where DIY usually struggles:

• Market selection is harder than it looks and matters more than the model
• Model selection requires judgment that comes from running many tests
• The operational stakes are high when reallocating hundreds of thousands of dollars
• The real time cost lands between 40 and 60 hours per test when done carefully
• Open-source MMM tools like Meridian require a Bayesian statistics background to interpret correctly

If your team runs one complex study per year and has the in-house expertise, DIY can be a learning exercise. If your team needs repeatable decisions and guardrails, a dedicated platform is usually safer and cheaper in total cost.

‍

The Stella difference: incrementality, MMM, and always-on, in one platform

The market has bifurcated into two camps. On one side, enterprise platforms like Measured and Haus deliver rigorous incrementality testing at enterprise prices, often $50,000 to $150,000 per year, with MMM as a separate or add-on service. On the other side, low-cost MTA tools like Northbeam and Triple Whale offer attribution-style measurement that does not actually answer the causal question.

Stella sits in a different position. The full platform includes:

• Incrementality testing. Geo holdouts, inverse holdouts, and scale tests with multi-model validation across Weighted Synthetic Control, BSTS, and GeoLift.
• Bayesian MMM. Always-on, automatically calibrated by your incrementality tests, with response curves and budget optimization across the full media mix.
• Always-on incrementality. Continuous monitoring that flags when channel performance shifts between formal tests.

All three are included in Stella Professional at $3,000 per month, flat. No spend tiers. No per-channel fees. No separate MMM contract. Compare that to a single incrementality module from a legacy vendor at $4,000 to $12,000 per month, often with the MMM sold separately at a similar price.

The reason the bundle matters is straightforward. Incrementality alone tells you what one channel did during one window. MMM alone tells you what the historical correlations imply, with no causal grounding. Always-on monitoring alone catches changes but cannot explain them. The three together form a complete decision system. Tests calibrate the model. The model translates calibrated truth into allocation. Always-on monitoring catches the drift between tests and triggers the next experiment.

This is the system that the largest brands have spent years building internally. Stella packages it for mid-market ecommerce brands at a tenth the cost.

‍

Best practices that protect validity

• Validate at least 90 days of clean pre-period data
• Confirm you can easily exclude regions in each ad platform
• Plan for at least three weeks of test time, longer for long purchase cycles
• Monitor spend delivery daily and document anomalies such as promos or outages
• Focus on the confidence interval, not just the point estimate
• Add a short post-treatment window to double-check the causal story
• Calibrate your MMM with every completed test, not just the first one

Your first test may show lower iROAS than attribution suggests. That is not a failure. That is informative truth.

‍

Conclusion: from measurement to growth

Attribution is useful for operations and daily reporting. It is not designed to answer the causal question that drives profitable growth. Incrementality testing does exactly that. It does not make your marketing look better. It makes your decisions better.

The 2026 measurement landscape has matured past the question of whether to do incrementality testing. The new question is how to operationalize it as a continuous program rather than an annual project. That requires three capabilities working together: incrementality tests for ground truth, MMM for allocation, and always-on monitoring to keep the system honest between tests.

Most platforms force you to choose one. Stella delivers all three.

If you want to stop arguing with dashboards and start reallocating with confidence, begin with one clean inverse holdout on your largest channel. The clarity you gain will change how you plan budgets for the rest of the year.

Run your first study free today. No credit card required.

‍