Marketing Measurement for DTC Brands: Multi-Touch Attribution vs Marketing Mix Modeling vs Incrementality Testing

Marketing measurement is how a brand works out which marketing caused its sales. There are three main methods: multi-touch attribution follows individual clicks to orders, marketing mix modeling estimates each channel's effect from spend and sales over time, and incrementality testing runs controlled experiments to measure what ads actually caused.

Each method answers a different question, needs different data and fails in different ways. This guide explains how multi-touch attribution (MTA), marketing mix modeling (MMM) and incrementality testing work, compares them side by side, and covers which ones a DTC brand can realistically run at each stage. It also walks through designing a lift test, turning its result into an incremental ROAS you can use every day, the data an MMM needs, and a 90-day plan to put it together. If you're new to attribution models, start with our guide to first click vs last click attribution.

What is marketing measurement?

Every measurement method is trying to answer the same question: which sales happened because of a given ad, campaign or channel, and which would have happened anyway? Meta's March 2026 measurement announcement frames it the same way and calls incrementality experiments "the gold standard in measurement".

The three methods come at that question from different directions:

---

How does multi-touch attribution work?

Multi-touch attribution tracks the marketing touchpoints each customer had before buying, usually ad clicks, email clicks and other tagged visits, and assigns credit for the order across them. Rules-based models split credit with a fixed formula (first click, last click, linear, time decay, position-based). Data-driven models learn the split from your conversion paths: Google Analytics 4's data-driven attribution, for example, compares converting and non-converting paths to estimate each touchpoint's contribution.

Our guide to multi-touch attribution for small ecommerce brands covers the models and how to choose between them.

Strengths

• It's granular. MTA can credit individual ads and campaigns, which no other method does.
• It's fast. Reports update daily, so you can act on them inside a campaign's life.
• It uses data you already collect. UTM parameters, click IDs and order data are enough to start.

Weaknesses

• It shows correlation, not causation. A customer who clicked a retargeting ad and then bought may have bought anyway. In a study of 15 advertising experiments at Facebook covering 1.6 billion ad impressions, Gordon, Zettelmeyer, Bhargava and Chapsky (Marketing Science, 2019) found that observational methods often failed to reproduce the effects the randomized experiments measured, even with extensive demographic and behavioral data.
• It only sees what it can track. Ad views, podcasts, word of mouth and TV don't leave clicks. Device switches, blocked scripts and browser cookie limits break the paths it can see: Safari, for instance, caps cookies set by scripts at seven days.
• Definitions differ between tools. Until 2026, Meta counted shares, saves and likes as clicks for attribution. Its March 2026 update narrowed click-through attribution to link clicks so Ads Manager lines up better with tools like Google Analytics. Two MTA reports can disagree without either being wrong.

How to read MTA without fooling yourself

MTA is still the only daily, ad-level view you have, so the goal is to read it with its biases in mind rather than throw it out:

• Compare first click and last click side by side. A channel that looks strong under first click and weak under last click is opening journeys that something else closes. Cutting it on last-click numbers starves the channels that close.
• Split new and returning customers. Returning customers would often have bought anyway. A campaign whose credited orders are mostly returning customers is likely to be less incremental than its ROAS suggests.
• Treat brand search and retargeting with suspicion. Both sit next to the purchase by design, so click attribution over-credits them. They are the first candidates for a lift test.
• Check the total against your store every week. If attributed revenue across channels is rising while total store revenue is flat, the attribution is reallocating credit, not finding growth.
• Apply an incrementality factor once you have one. After a lift test, discount each channel's attributed results by what the test showed (see the worked example below).

---

How does marketing mix modeling work?

Marketing mix modeling is a regression model built on aggregate data. It takes weekly sales and weekly spend by channel, adds factors that also move sales (seasonality, promotions, pricing, holidays), and estimates how much each channel contributed. Most models also account for carryover, where an ad keeps working after the week it ran, and diminishing returns as spend rises.

You don't need a vendor to try it. Google publishes Meridian and Meta publishes Robyn, both open source. Our guide to marketing mix modeling for DTC brands covers how the models work, the tools and how to check whether a model can be trusted.

Strengths

• It needs no user-level tracking, so cookie limits, ad blockers and device switches don't affect it.
• It covers channels clicks can't, such as TV, podcasts, out-of-home and organic social.
• It's built for budget questions: what happens to total sales if you move spend from one channel to another.

Weaknesses

• It needs a lot of history. Robyn's analyst guide recommends a minimum of two years of weekly data. Google's Meridian docs show why that can still be thin: two years of weekly data is 104 data points, which Google calls too few to reliably estimate a national model with 12 channels, six controls and eight time parameters.
• It needs your spend to vary. If a channel's spend barely changes week to week, the model can't separate its effect from everything else.
• It works at channel level. Google doesn't recommend running Meridian at campaign level and points to data-driven multi-touch attribution for more granular digital insight.
• It's slow. A model works on weekly data and is rebuilt periodically, so it can't tell you which ad to pause this week.

MMM is the modeled version of a top-down view many brands already track. If you watch marketing efficiency ratio (MER), total revenue divided by total ad spend, you're looking at the same aggregate data an MMM starts from.

What data a marketing mix model needs

The model is only as good as its weekly inputs. Before building one, collect this for every week of history:

Missing weeks, changed definitions and promotions nobody wrote down are the most common reasons a model attributes sales to the wrong cause.

Average ROAS vs marginal ROAS

The most useful output of an MMM is not each channel's average return but its response curve: how revenue changes as spend rises. That gives you marginal ROAS, the return on the next dollar, which is what a budget decision is actually about. Using illustrative numbers for one channel:

The average still looks healthy at $40,000, but the last $10,000 only returned $10,000. If your break-even ROAS is above 1.0, that extra spend loses money even though the channel's reported ROAS is 2.2.

Create spend variation on purpose

A model can only learn a channel's effect from weeks when that channel's spend changed. If budgets move together or stay flat, it can't separate them. Build variation into your plan: raise or lower one channel at a time for a few weeks, stagger changes across channels instead of moving them all at once, and write down every change with its dates so the model's inputs match what happened.

---

How does incrementality testing work?

Incrementality testing measures lift with an experiment. You split an audience or a set of regions into two groups, show ads to one, withhold them from the other (the holdout or control), and compare sales. The difference is the incremental effect: sales the ads caused that wouldn't have happened otherwise.

The common designs:

• Conversion lift tests. The ad platform randomly splits people into test and control groups. Meta's Conversion Lift works this way.
• Geo experiments. Regions instead of people are split, which works for channels that can't target individuals. Meta's GeoLift, named in Robyn's guide, and Google's Meridian GeoX support this design.
• On/off tests. Pause a channel for a period and watch total sales. They're cheap but easy to confuse with seasonality.

Lift tests are different from Meta's incremental attribution setting. That is an attribution model in Ads Manager that uses machine learning to predict whether a conversion was caused by an ad, and optimizes delivery toward those conversions. It's an estimate built from experiments, not an experiment you run.

Strengths

• It measures cause, not correlation. With random assignment, the only systematic difference between the groups is the ads.
• It's the reference the other two methods are checked against. See the calibration section below.

Weaknesses

• One question per test. Each test measures one campaign, channel or region over one period, and results age as creative, audiences and seasons change.
• The holdout costs sales. People or regions in the control group don't see your ads while the test runs.
• It needs volume. A small brand may not get enough conversions in each group to detect a real but modest lift.
• It only counts what the platform receives. A platform lift test compares conversions it can see in each group, so missing purchase events make a real effect harder to detect.

---

How to design an incrementality test

A lift test is only useful if it can answer a question you will act on. Our incrementality testing guide has the lift and iROAS formulas, test types, tools and which channels to test first. Seven steps cover the design:

1. Start from one decision. "Should we keep spending on retargeting?" is testable. "Is Meta working?" is not. Write down what you will do if the test shows a large lift, a small lift or none.
2. Choose the design that fits the question. A platform conversion lift study fits when you are testing that platform's ads and it supports one; Meta's lift studies, for example, hold out a percentage of people who won't see the ads. A geo holdout fits when you want to measure store orders directly or test a channel that can't target individuals. An on/off test is the last resort.
3. Measure the outcome you care about. Platform lift studies count the conversions that platform receives. A geo test can measure your store's own orders by shipping region, which also captures sales the platform never saw.
4. Size the test before you start. Whether a test can detect a real effect depends on four levers: how many conversions each group gets, how long the test runs, how much of the audience is held out, and how big the effect is. More of the first three makes a smaller effect detectable. Run a power analysis first; Meta's open-source GeoLift includes one for choosing markets and duration.
5. Run whole weeks and cover a purchase cycle. Partial weeks mix weekday and weekend behavior, and a test shorter than the time customers usually take to buy will miss part of the effect.
6. Hold everything else still. Avoid big promotions, price changes and budget shifts in other channels while the test runs. If one is unavoidable, apply it to both groups.
7. Read the result as a range. A test reports an estimate with uncertainty around it. "Between 450 and 750 incremental orders" is the honest result; plan for the low end.

Worked example: from lift result to incremental ROAS

Here is how one test turns into numbers you can use every day. The figures are illustrative.

The factor is what makes the test useful after it ends. Next month, if attribution shows this campaign at a ROAS of 2.0, its likely incremental ROAS is about 2.0 × 0.6 = 1.2. Compare that with your break-even ROAS, not the attributed figure, before deciding whether to scale. Re-test when creative, audiences or seasons change, because the factor drifts.

Running a geo holdout on a DTC budget

You don't need a measurement vendor to run a basic geo test:

1. Pick the regions. States or metro areas work. Group them into test and control sets whose sales moved together in the months before the test; GeoLift's market selection does this for you.
2. Use your own orders as the outcome. Shopify records a shipping address on every order, so orders and revenue by region are data you already have.
3. Change the channel in the test regions only. Most ad platforms let you exclude locations. Turn the channel off (or on, or up) in the test regions and leave the control regions alone.
4. Compare against what would have happened. Use the control regions and the pre-test period to estimate what the test regions would have done without the change; the gap is the lift.
5. Watch for spillover. People travel, ship gifts to other states and see national campaigns, so some effect leaks across regions. Larger regions and channels that target by location reduce it.

MTA vs MMM vs incrementality: how they fit together

The three methods aren't competitors. They work at different speeds and levels of detail, and each covers another's blind spot:

The way they connect is calibration. Robyn's guide strongly recommends using experiment results such as Meta Conversion Lift or GeoLift to calibrate an MMM, and Google's Meridian docs offer GeoX geo experiments for the same purpose. You can calibrate MTA the same way, informally: if a lift test shows retargeting drives half the sales last click gives it, discount retargeting in your attribution reports by a similar amount.

A practical loop looks like this:

1. Run click attribution daily to manage campaigns and creative.
2. Run a lift test on your largest channel, or the one you trust least.
3. Compare the test's incremental result with what attribution credited, and note the gap.
4. Once you have enough history and channels, build an MMM and calibrate it with the test results.
5. Re-test as spend, creative or seasons change.

---

Which measurement does a DTC brand need?

Most brands add methods in order, as their data and budget allow:

A 90-day measurement plan

For a brand with several paid channels that has relied on platform reporting so far, a first quarter could look like this:

Start collecting the weekly MMM inputs from week one. By the time you have two years of clean history, building a model is a much smaller project.

Two mistakes are common at every stage. The first is treating platform-reported results as the answer. Each platform counts the orders it can link to its own ads, so their totals can add up to more than you sold; our post on why GA4 and ad platform numbers don't match explains the gaps. The second is buying a more advanced method before the basic data is clean.

---

Every method depends on the same data

None of the three methods can fix missing or wrong inputs:

• MTA needs every click captured and tied to the right customer across devices and browsers. Lost click IDs and blocked scripts turn real journeys into "direct" orders.
• MMM needs accurate weekly revenue and spend. Platform-claimed revenue overstates sales, so the model should start from your store's actual orders.
• Lift tests can only count the conversions a platform receives. Purchases that never reach Meta or Google shrink both groups' numbers.

That's where server-side tracking comes in, and it's the part of the stack Upstack works on. To be clear about scope: Upstack doesn't build marketing mix models or run incrementality tests. It provides the click-level layer and the data the other methods depend on:

• The Attribution board credits Shopify orders under First click, Last click or Any click, with windows from 1 day to Lifetime, from visits recorded on your own site, as described in how Upstack attributes orders. Each order is credited once under First or Last click, so channel totals add up to attributed revenue.
• Clicks are matched across every browser and device that identity resolution has linked to a customer, so a phone click can earn credit for a laptop order.
• The Attribution board exports orders, revenue, spend, CPA and ROAS by source as CSV, which you can use as an input to your own models.
• Purchases are sent server-side to Meta, Google and TikTok, so the platforms running your lift tests aren't relying on the browser pixel alone to see your orders.

---

Frequently asked questions

What does incrementality mean in marketing?

Incrementality is the share of results that happened because of your marketing and wouldn't have happened without it. If a campaign is credited with 100 orders but a holdout test shows 60 of those buyers would have bought anyway, the campaign's incremental result is 40 orders.

Is marketing mix modeling better than multi-touch attribution?

Neither is better overall. MMM is better for splitting budget across channels, including ones that leave no clicks. MTA is better for daily decisions about individual ads and campaigns. Brands with enough data use both and check each against lift tests.

What's the difference between incrementality and attribution?

Attribution assigns credit for orders to the touchpoints that came before them, whether or not those touchpoints changed anything. Incrementality measures how many orders the marketing caused, usually by comparing people or regions that saw ads with ones that didn't.

Can a small brand run marketing mix modeling?

It can try, but the results are often too uncertain to act on. Robyn's guide recommends at least two years of weekly data, and Google's Meridian docs show that two years can be too little when you model many channels. Small brands usually get more from click attribution, MER and a few well-designed lift tests.

How long should an incrementality test run?

Long enough to cover at least one typical purchase cycle and to collect enough conversions in each group to detect the effect you care about, in whole weeks. There is no universal length: a power analysis based on your conversion volume and holdout size gives the answer for your account.

How big should a holdout group be?

It's a trade-off. A larger holdout makes a smaller effect detectable sooner, but withholds ads from more people while the test runs. Smaller holdouts need longer tests. Use a power analysis to find the smallest holdout that can detect the lift you need to make your decision.

What is incremental ROAS?

Incremental ROAS (iROAS) is the revenue your ads caused divided by what you spent, measured with a lift test rather than credited by attribution. It is usually lower than attributed ROAS because attribution also credits sales that would have happened anyway.

What is marginal ROAS?

Marginal ROAS is the return on the next dollar of spend, as opposed to average ROAS across all spend. Because returns diminish as spend rises, a channel can have a healthy average ROAS while its marginal ROAS is below break-even. Marketing mix models estimate it from each channel's response curve.

What is unified marketing measurement?

Unified marketing measurement means using attribution, marketing mix modeling and incrementality testing together, with experiments calibrating the models, rather than relying on one method. It's the approach described in the "how they fit together" section above.

---

Key takeaways

• Multi-touch attribution is fast and granular but shows correlation and misses what it can't track.
• Marketing mix modeling covers every channel and suits budget planning, but needs two or more years of weekly data with varied spend.
• Incrementality tests measure cause and are the benchmark for calibrating the other two, one question at a time.
• An incrementality factor from a lift test turns daily attribution into an estimate of incremental ROAS; budget decisions should rest on marginal, not average, ROAS.
• Most DTC brands should start with click attribution against store revenue, add lift tests, and add MMM once they have the history and channel mix.
• All three methods depend on complete, accurate conversion data.

Request a demo and we'll show you how complete your conversion data is before you invest in modeling.