Attribution & Measurement

MTA vs MMM vs Incrementality: Which to Use When

MTA shows journeys, MMM sets channel budgets and incrementality proves cause. Here is what each needs and how to choose.

Simran Rawat14 Oct 202617 min read
MTA vs MMM vs Incrementality: Which to Use When

Use multi-touch attribution (MTA) to see which tracked touchpoints appear in buyer journeys, media mix modelling (MMM) to estimate each channel's contribution to total sales, and incrementality testing to prove whether a campaign caused extra sales. Fix tracking first, run holdout tests next, and add MMM once you have at least two years of weekly data.

When comparing MTA vs MMM vs incrementality, the short answer is that each method answers a different question. Use multi-touch attribution (MTA) to see which tracked touchpoints sit in buyer journeys, media mix modelling (MMM) to estimate what each channel adds to total sales, and incrementality testing to prove whether a campaign caused extra sales. Start with clean tracking and holdout tests, then add MMM when you have two or more years of weekly data.

Most brands do not fail at measurement because they picked the wrong method. They fail because they asked a method a question it cannot answer, such as asking an attribution report to prove cause or asking an MMM to steer this week's creative. This guide sets out what each method needs, how to match a method to a decision, and how to tell whether your data can support it, using the documentation Google and Meta's open-source teams publish.

This guide covers the report step of Nurdd’s influencer marketing measurement hub: forecast, vet, attribute and report, with a free tool for each.

Key takeaways

  • Match the method to the decision. Creative and campaign steering needs MTA, budget splits across channels need MMM, and "is this channel worth the money" needs an incrementality test.
  • MTA describes, it does not prove. It shares credit among touchpoints it can see, so it cannot say what would have happened without a channel.
  • MMM is limited by data points, not by budget. Robyn asks for a minimum of two years of weekly data, and Meridian shows that two years can still be too little when a model has many parameters.
  • GA4 no longer has linear, time decay, first click or position-based models. Google's help page says they were removed as of November 2023.
  • Small brands should test before they model. Reliable tracking, a weekly reconciled report and simple holdouts come before any MMM project.
  • Count parameters before you commission an MMM. The worked example below shows the arithmetic in about two minutes.

What is the difference between MTA and MMM?

Multi-touch attribution works at the level of the individual journey. It follows the ads, emails and visits a converting user interacted with and divides the credit among them, so its output is granular (down to the ad) and updates as conversions arrive.

Media mix modelling works at the level of the whole business. It is a statistical model, usually fitted to weekly data, that relates total sales to spend by channel while controlling for factors such as seasonality, price and promotions. Its output is coarse, but it can include channels that user-level tracking cannot see, such as TV, podcasts and much of creator marketing.

In one line each: MTA answers "which touchpoints showed up before people bought?" and MMM answers "if I move budget between channels, what happens to total sales?"

Modern MTA is more careful than the old fixed rules. According to Google Analytics Help, November 2026, the data-driven model compares the key-event probability of users who were exposed to an ad with that of similar users in a holdback group. It is still built only from the touchpoints Google can observe, which is why it describes journeys rather than proving cause.

For a deeper look at rules-based credit, see our guide to last-touch versus multi-touch attribution.

What does incrementality testing measure?

Incrementality testing measures the sales a campaign caused by comparing a group that could see the ads with a group that could not. The difference between the two groups is the lift. It is the only one of the three methods built around a control group, which is why it is the closest thing to proof a marketer can get.

There are three common designs, and they differ in how much trust they earn and how much volume they need.

DesignHow it worksStrengthWeakness
Platform lift studyThe ad platform randomly splits people into a test group and a control group that is held out from seeing the adsRandomised at person level, so the comparison is cleanMeasures one platform at a time and needs enough conversions to detect a difference
Geo testSpend is switched off or changed in some regions and compared with similar regions; the open-source GeoLift project from Meta Open Source supports geo-level analysisWorks across channels and does not depend on user-level trackingNeeds regions that behave alike and a period long enough to see an effect
Time-based holdoutA channel is paused for a period and sales are compared with a baselinePossible with almost no toolingThe weakest design, because seasonality or promotions can mimic an effect

Platforms publish their own volume requirements for lift studies and these change, so read the current requirements in your ad account before you plan a test rather than relying on a figure quoted in an article. Meta's learning phase also matters here, because large edits during a test can restart learning and blur the result.

How do MTA, MMM and incrementality compare?

FactorMulti-touch attribution (MTA)Media mix modelling (MMM)Incrementality testing
Question answeredWhich tracked touchpoints appear before a conversion, and how should credit be shared?What does each channel contribute to total sales, and how should the budget be split?Did this campaign or channel cause sales that would not have happened otherwise?
Data neededUser-level click and visit data from analytics and ad platforms, with consistent UTMsWeekly sales, spend by channel and control variables; Robyn's guide says a minimum of two years of weekly dataA test group and a control group, and enough conversions in both to detect a difference
Time to first resultAs soon as tracking is live and conversions are recordedAfter the history exists and a model has been fitted and checkedThe length of the test plus time to read it
GranularityCampaign, ad set, adChannel or channel groupThe one campaign, channel or region tested
Sees offline and untracked channels?NoYes, if spend and timing are in the dataYes, if the test design reaches them (for example a geo test)
Proves cause?NoEstimates contribution; strength depends on data and checksYes, when well powered and well designed
Main limitsMisses views without clicks, dark social and offline channelsCoarse, slow to react, unreliable with too few data points per parameterAnswers one question at a time; underpowered tests are noisy

Google and Meta both publish open-source MMM projects: Google's Meridian and Meta's Robyn. Both are documented for analysts, which means they are free to download but not free to run, because someone has to prepare the data and check the model.

Which method answers which marketing question?

The fastest way to choose is to start from the decision, not the method. This table maps ten decisions that marketers face to the method that fits them best. It is our own framework, built from the way each method is defined in the documentation above, not a published standard.

Decision you faceBest-fit methodWhy
Which ad or creative to scale this weekMTA and platform reportsNeeds ad-level, fast feedback
Whether to raise spend on one channel by a large amountIncrementality test on that channelYou need proof that the current spend causes sales
How to split next year's budget across five channelsMMMCompares channels on one scale across total sales
Whether TV, podcast or creator spend is workingGeo test or MMMUser-level tracking cannot see these channels
Whether platform ROAS matches your store's revenueIncrementality testA controlled experiment is the referee
Whether retargeting is capturing sales that would happen anywayHoldout testRetargeting is a classic over-credited channel
Why last week's results changedMTA plus a diagnosis of account changesShort-term question, short-term data
Whether your MMM is rightIncrementality tests used to calibrate itRobyn's guide recommends calibrating with experiments
Which touchpoints assist the final clickMTAThis is what path-based credit is for
How a launch affected brand searchTime-based holdout or geo testCompares a baseline with the exposed period or region

If a decision in your calendar is not in this table, ask three questions: is it about a single campaign or the whole mix, does it need an answer this week or this quarter, and does it need proof of cause or only a ranking. Single, fast and ranking-only points to MTA; whole-mix and slow points to MMM; proof of cause points to a test.

How much data does media mix modelling need?

Neither major open-source MMM project gives a spend threshold. Both talk about data points.

  • Robyn (Meta's open-source MMM): the Analyst's Guide to MMM says a model will need a minimum of two years of historical weekly data, and strongly recommends more than two years if only monthly data is available.
  • Meridian (Google's open-source MMM): the amount of data needed page works through how many parameters a model has compared with how many data points you can feed it.

A worked example from Google's documentation

Meridian's documentation describes a national model with 12 media channels, six controls and eight knots, which makes 26 parameters. Two years of weekly data gives 104 data points, or four data points per parameter, and the page states that this sample size is too low to estimate the model reliably.

The lesson is that two years of weekly data is not a pass mark on its own. What matters is how many things the model has to estimate compared with how many weeks you have.

Run the parameter count on your own brand

You can repeat the same arithmetic in a few minutes. Meridian counts channels, controls and knots as parameters (ignoring adstock and Hill parameters for simplicity), then divides your weekly data points by that total. The table below applies the same counting convention to three illustrative setups. The figures are our own arithmetic for illustration, not thresholds published by Google or Meta.

SetupChannels + controls + knotsParametersData points per parameter at 104 weeksAt 156 weeks
Meridian's example12 + 6 + 8264.06.0
Grouped into five channels5 + 4 + 4138.012.0
Three channel groups3 + 5 + 21010.415.6

Reading the table: the same two years of history can look hopeless or workable depending on how many channels you model separately. Grouping channels, using fewer knots and trimming controls all raise the ratio, at the cost of less detail per channel. Meridian's own page also discusses adding geo-level data as an alternative, which multiplies the data points without waiting for more calendar time.

The Meridian page says the best check is to run the model and look at how wide the credible intervals are. That is a better test than any rule of thumb, but it only works once you have paid for the data preparation, so use the parameter count as a cheap pre-check.

What does GA4 data-driven attribution need?

Google's current attribution help page states that the first click, linear, time decay and position-based models are no longer available as of November 2023. If an older guide tells you to fall back to linear attribution in GA4, that option is gone. The data-driven model remains, alongside last-click-based options.

The page also says conversions can be reattributed for up to 7 days after the conversion, so the most recent numbers in a report can still move. When you compare weeks, leave the latest few days out or mark them as provisional.

Data-driven attribution only sees what your tagging records. Inconsistent UTM values split one campaign into several rows and hide the pattern the model is trying to learn. Our UTM structure guide for influencer campaigns gives a naming convention that keeps reports clean.

When should you use incrementality testing?

Use a test whenever a decision depends on whether a channel causes sales, not just appears near them. Four situations come up most often.

  1. Before a large budget increase. Test whether current spend is incremental first, because scaling a channel that mostly captures existing demand wastes money faster.
  2. When platform reporting and your store disagree. A controlled test is the referee. Our post on why Meta ROAS can look great while the bank account disagrees explains the usual causes.
  3. For channels with weak tracking. Creator posts, podcasts and retargeting are classic cases where attribution either misses sales or over-credits them.
  4. To calibrate an MMM. Robyn's guide recommends running incrementality studies on an ongoing and regular basis so that their results can calibrate the model.

Checklist: before you start a holdout test

  1. Write the decision the test will inform, for example "raise retargeting spend or not".
  2. Choose the primary metric (orders, revenue, new customers) and write it down before launch.
  3. Check your conversion volume against the platform's current lift-study requirements, or choose a geo or time-based design if you fall short.
  4. Fix the budget and avoid large edits to campaigns during the test window.
  5. Check there is no big promotion, price change or stock-out planned that would hit the test group and the control group unevenly.
  6. Decide in advance what lift would change your decision, and what result would mean "retest".
  7. Run the full duration even if early numbers look exciting.
  8. Record the result and the design so a later MMM can use it for calibration.

Step seven is the one most teams skip. Stopping early because the number looks good is the same as choosing the answer first.

Which method suits a small brand?

A small brand should use the simplest setup that gives one trustworthy weekly number, then add causal tests as spend grows. The choice depends less on spend bands than on three things you can check today: how many conversions you record, how many weeks of clean history you have, and how many channels you run.

Your situationStart withAdd when ready
Tracking is patchy and UTMs are inconsistentFix the pixel, the Conversions API and UTMs; reconcile platform reports with store orders weeklyGA4 data-driven attribution once key events are recorded reliably
Tracking is clean, with one or two paid channelsGA4 data-driven attribution alongside platform reportsTime-based or geo holdouts on your largest channel
Your conversion volume meets the platform's lift-test requirementsA platform lift test before each major budget changeMulti-cell tests that compare strategies
Two or more years of weekly data, several channels including offline or creator spend, and a parameter count that gives a workable ratioAn MMM, with channels grouped to keep parameters lowRegular lift tests to calibrate the model

The first row matters most. A model built on broken tracking produces a confident wrong answer, and no later method repairs it. If you want a look at what "clean" means for a Meta account, our 30-minute Meta ad account audit is a practical starting point.

How do you set a target so you know what working means?

Before any method, set a break-even target. A channel with a platform ROAS of 3 can still lose money if margins are thin, and a channel at 1.5 can be profitable if margins are high. Our Google + Meta ROAS calculator works out per-channel and blended return against break-even, and the ecommerce KPIs that predict profit better than ROAS explains what to track beside it.

Then compare your numbers with published category medians. The marketing benchmarks hub lists each headline figure with its source and period, and the campaign benchmarker lets you check your own campaign against them. Benchmarks do not prove cause, but they tell you whether a result is unusual enough to be worth a test.

How do MTA, MMM and incrementality work together?

Each method covers a weakness of the others, so a mature setup runs them on different clocks. The cadence below is our suggested operating rhythm, not a standard from any platform.

CadenceMethodOutputDecision it feeds
WeeklyMTA and platform reports, reconciled with store ordersOne trusted number per channel and the top and bottom adsCreative and campaign steering
QuarterlyHoldout or lift test on the channel with the most money or the most doubtLift result with a stated confidence levelScale, hold or cut that channel
Once or twice a yearMMM, calibrated with the test resultsChannel contribution and a budget splitAnnual and half-year budgets

Our marketing measurement framework shows how influencer, paid and AI search reporting fit into a single system, and how Nurdd works describes how the four capabilities connect.

A 90-day plan for a brand starting from scratch

  1. Weeks 1 to 2: audit tracking. Check the pixel, Conversions API, Google Ads conversions and UTMs. Result: platform numbers you can compare with store orders.
  2. Weeks 3 to 4: set break-even targets. Use margins to set a ROAS floor for each channel. Result: a definition of working.
  3. Weeks 5 to 6: build the weekly reconciliation. Put platform revenue, GA4 revenue and store revenue side by side. Result: a gap you can explain.
  4. Weeks 7 to 12: run one holdout. Pick the channel where platform reporting and store revenue disagree most. Result: one causal answer and a recorded design.
  5. Throughout: log weekly data. Keep a clean weekly sheet of sales, spend by channel, promotions and price. Result: the history an MMM will eventually need.

How do you measure creator marketing with these methods?

Creator marketing is where the methods diverge most. Click-based attribution undercounts creator posts, because many viewers buy later by searching for the brand (see dark social and word-of-mouth attribution), and promo codes can leak to coupon sites. MMM only sees creator spend if it is large and steady in weekly data.

That makes incrementality the most useful check: compare sales in the period or region where a creator wave ran with a baseline, as set out in our influencer incrementality guide. Track codes, links and post dates for each creator, as described in influencer performance attribution, so the test has clean inputs. The influencer ROI calculator projects revenue and ROAS from a creator budget before you spend.

What mistakes make attribution and testing unreliable?

  • Treating attributed conversions as caused conversions. Attribution assigns credit; it does not say the sale would not have happened anyway.
  • Building an MMM on too little history. Meridian's own example shows four data points per parameter is too low.
  • Changing campaigns mid-test. Budget swaps and new creative can restart learning and blur the result.
  • Choosing the success threshold after seeing the number. Decide in advance what lift would change your plan.
  • Following outdated GA4 advice. Linear, time decay, first click and position-based models were removed in November 2023.
  • Using one method for every question. A weekly steering number and an annual budget split need different tools.
  • Ignoring tracking quality. Every method inherits the errors in its inputs.

How can Pulse help with attribution and measurement?

Pulse by Nurdd is a web dashboard that connects to Meta Ads and Google Ads through read-only OAuth, explains campaign performance in plain language, diagnoses what changed and recommends specific actions without requiring you to build reports. It gives a brand the weekly steering layer of the stack above, so the numbers used to decide when to run a holdout test or commission an MMM are read consistently across both platforms.

The Pulse page on nurdd.club, describing Meta and Google Ads intelligence. Source: Nurdd.
The Pulse page on nurdd.club, describing Meta and Google Ads intelligence. Source: Nurdd.

Pulse is paid, has no free trial and is in private access. Book a demo and you will see Pulse read your own Meta and Google Ads performance, with a plain-language explanation of what changed and what to do next, and we will talk through which of the three methods your current data can support.

For brands that want a team to run the work, Nurdd's paid social advertising service audits tracking, sets break-even ROAS targets and reports monthly against published benchmarks. For creator campaigns, Nia by Nurdd attributes product-level sales to each creator, which gives an incrementality test cleaner inputs. You can also see the Pulse overview and the marketing budget planner before you book.

How we researched this

We reviewed Google's GA4 attribution documentation, the Robyn and Meridian open-source documentation, the GeoLift project site, and three comparison articles that currently rank for this topic (from ClickZ, Caliber Mind and Qry) to see what they cover and leave out. Documentation was read in November 2026 and checked on the publication date. We did not run models, lift tests or hands-on trials for this article, and the decision table, parameter-count table and cadence are our own frameworks built from those sources. Meta's help pages on Conversion Lift could not be re-opened during this update, so we removed the spend and conversion thresholds that earlier versions of this post attributed to them.

Frequently asked questions

Is MMM better than multi-touch attribution?

Neither is better in general, because they answer different questions. MMM estimates each channel's contribution to total sales and can cover offline and untracked channels, but it is coarse and needs long history. MTA is fast and granular but only sees tracked touchpoints and cannot prove cause.

How much data do I need for marketing mix modelling?

Meta's Robyn guide says a model needs a minimum of two years of weekly data. Google's Meridian documentation shows two years can still be too little for a model with many channels, so group channels or add geo-level data.

Can a small brand run incrementality tests?

Yes, but the design depends on volume. A platform lift study needs enough conversions in the test and control groups, and the platform publishes the current requirements. Below that, a geo or time-based holdout on your largest channel still gives a directional answer if other variables stay steady.

Does GA4 still offer linear attribution?

No. Google's help page says the first click, linear, time decay and position-based models have not been available since November 2023. GA4 now offers data-driven attribution and last-click-based options.

What is the difference between attribution and incrementality?

Attribution assigns credit for conversions to the touchpoints that preceded them. Incrementality measures how many conversions happened because of the ads, by comparing a group that could see them with a group that could not. Attributed conversions can include sales that would have happened anyway.

Which should I do first: MTA, MMM or incrementality?

Fix tracking and set up reliable attribution first, because every other method depends on clean inputs. Add an incrementality test on your biggest or most doubtful channel next. Commission an MMM when you have the weekly history and a workable parameter count.

Can incrementality tests calibrate an MMM?

Yes. Robyn's guide recommends running incrementality studies on an ongoing basis so that their results can calibrate the model. Record each test's design and result so they are usable later.

Why do my platform numbers and store revenue disagree?

Platforms credit conversions to their own ads using their own windows, and several platforms can claim the same sale. Reattribution can also move recent numbers, as GA4's 7-day note shows. A weekly reconciliation against store orders, followed by a holdout test, is the practical fix.

Want a measurement plan that fits the data you actually have? Book a demo with Nurdd and see Pulse read your own Meta and Google Ads account.

See it with your own goal

Leave your work email and pick a time.