An anti-guide to MMM vs MTA on a small budget: how to pick the wrong model, and what to run instead

    On a small budget, a cheap experiment often beats a model you cannot trust

    Published
    11 min read
    By Bram Valdenmere
    Updated

    What most small brands get wrong about MMM vs MTA is the framing. They treat it as a choice between two measurement products and assume one of them must fit. In practice, at €10,000 to €50,000 of monthly ad spend, there is a good chance neither works as advertised.

    Multi-touch attribution needs tracking coverage that privacy changes have eroded, and marketing mix modelling needs volume and variation that a small advertiser rarely has. Picking the model with the nicer dashboard is how teams end up confidently reallocating budget on numbers that are mostly noise.

    So this is an anti-guide: the wrong ways to choose, why they fail, and the conditions under which each approach starts to earn its cost.

    What each model actually claims to measure

    Multi-touch attribution (MTA) works at the level of individual users. It collects the touchpoints a converting customer had, such as a paid social click, an email open, a branded search, and splits credit for the sale among them. Rule-based versions use fixed weights (linear, time decay, position based); data-driven versions estimate weights from paths that did and did not convert. Its question is: given this sale, which tracked touches deserve credit?

    Marketing mix modelling (MMM) works on aggregates. It regresses a business outcome, usually weekly revenue or orders, on weekly spend per channel plus controls such as seasonality, price, promotions and holidays. It models carryover (adstock: an ad's effect decays over following weeks) and saturation (each extra euro buys less). Its question is: how much revenue would we lose if this channel's spend dropped, holding the rest constant?

    Those are different questions. MTA describes paths; it does not by itself say whether a touch caused the sale or merely sat on the path of someone who would have bought anyway. Branded search and retargeting are the classic examples: they appear on many converting paths because they target people already close to buying. MMM is closer to a causal estimate, but only if spend varied enough and independently enough for the model to separate channels.

    Dimension MTA MMM
    Unit of analysis User path Weekly or daily totals
    Typical minimum history A few weeks of tracked paths 2 to 3 years of weekly data
    Sees offline, TV, podcasts No Yes, if spend is recorded
    Affected by iOS and cookie limits Heavily Barely
    Answers "is it incremental?" Weakly Moderately
    Granularity Campaign, ad, keyword Channel, sometimes campaign type
    Cost to start Often included in analytics tools Free open-source code, high analyst time
    Main failure mode Credits touches that were not causal Unstable estimates from too little variation

    Wrong choices small teams make again and again

    • Buying an MTA tool after Apple's App Tracking Transparency, introduced in iOS 14.5 in April 2021, cut off much of the cross-app data it depends on, and then trusting its channel split as complete.
    • Running an MMM on 14 months of data because that is all the store has, which leaves too few weeks to separate seasonality from spend.
    • Fitting an MMM when spend per channel barely moved: if Meta was always €8,000 a month, the model cannot learn what happens at €4,000 or €12,000.
    • Using MTA to justify branded search and retargeting budgets, the two channels most likely to look good on paths without being incremental.
    • Comparing the MMM's return on ad spend with platform-reported ROAS and picking whichever is higher.
    • Treating open-source MMM code as a product: Meta's Robyn, Google's Meridian and PyMC-Marketing are good libraries, but each needs someone who understands priors, adstock and model diagnostics.
    • Changing the budget split every week based on a model that was fitted on weekly data, so the next refit sees noise caused by your own reactions.

    The pattern behind these mistakes is the same. Each picks a method for what it outputs rather than for whether the inputs it needs exist in the business.

    Why a small budget breaks both models

    Start with MMM, because the arithmetic is unforgiving. Say a store has 2 years of weekly data: 104 observations. It runs five channels. A reasonable model estimates, per channel, an effect size, an adstock decay rate and one or two saturation parameters, so roughly 3 to 4 parameters each, 15 to 20 in total, before seasonality, trend and promotions add another 8 to 12.

    That leaves around 3 or 4 observations per parameter. Bayesian MMMs cope by leaning on priors, which is legitimate, but then the output partly reflects the priors you chose. With little data, the model mostly hands your assumptions back to you with a confidence interval.

    Variation is the second problem. A model learns a channel's effect from weeks when its spend changed and others did not. Small teams tend to run steady budgets and move every channel together for Black Friday and January sales. Correlated spend makes channels statistically inseparable, and the model splits credit almost arbitrarily between them.

    MTA fails differently. It does not need much history, but it needs to see the journey. Safari's Intelligent Tracking Prevention caps the life of some first-party cookies set by scripts to 7 days, so a customer who clicks on Monday and buys ten days later on the same iPhone can look like a new visitor.

    Cross-device journeys break the chain entirely, and view-through exposure on social or video is invisible to click-based paths. A small DTC brand where a large share of traffic comes from iPhones can easily find that its attribution tool sees only part of the journeys behind its orders, and it cannot tell which part.

    There is a strong counterargument, and it deserves its best form. MTA, even partial, is fast and granular: it tells you that one ad set converts at twice the rate of another this week, which is exactly the decision a small team makes every day. MMM cannot work at that grain.

    For optimisation inside a channel, comparing creatives or keywords on similar audiences, MTA's bias is roughly the same across the options being compared, so the ranking holds even when the totals do not. Where MTA loses is the cross-channel budget question, which is the question small brands most often ask it.

    A decision tree for spend under €50,000 a month

    1. Is more than 70 percent of your spend in a single platform, such as Meta or Google? If yes, use that platform's own lift or experiment tools for the big question, and in-platform reporting for ranking ads; skip both MTA and MMM for now. If no, go to step 2.
    2. Do you have at least 2 years of weekly sales and spend data, with records of promotions and price changes? If no, go to step 5. If yes, go to step 3.
    3. Did each major channel's spend vary by at least 30 to 50 percent at different times from the other channels? If no, MMM will not separate them; go to step 5. If yes, go to step 4.
    4. Do you have, or can you buy, 3 to 5 days a month of analyst time from someone who has fitted an MMM before? If yes, an open-source MMM is reasonable, calibrated with at least one test from step 5. If no, go to step 5.
    5. Can you switch a channel off, or cut it sharply, in some regions or for some weeks? If yes, run a geo holdout or on/off test on your largest uncertain channel before buying any model. If no, use simple MTA for in-channel decisions, add a post-purchase "how did you hear about us" survey, and revisit this tree in six months.

    Most small brands end at step 5, and that is the useful answer. A test that turns off paid social in a quarter of your regions for four weeks answers "is this channel incremental?" more directly than either model can at this scale.

    A worked example: one channel, two numbers

    Take an illustrative store spending €20,000 a month: €12,000 on Meta, €5,000 on Google search, €3,000 on retargeting. Monthly revenue is €90,000. The MTA tool, data-driven, assigns €38,000 of revenue to Meta, an attributed ROAS of 3.2. Google's search reporting claims €21,000 from €5,000. Retargeting claims €14,000 from €3,000, a ROAS of 4.7, which looks like the best channel.

    The team runs a geo holdout on retargeting: it pauses retargeting in regions that normally produce 25 percent of revenue for 4 weeks. Revenue in the paused regions drops by 2 percent against the forecast built from the other regions, roughly €450 a week on a base of about €22,500 a week.

    Scaled to all regions, retargeting's incremental revenue is close to €7,800 a month, not €14,000. The incremental ROAS is about 2.6, still positive, but well below what MTA reported, and below Meta, which a later test puts at 2.9.

    The decision the test enables is modest: move €1,000 from retargeting into prospecting and keep watching. It is also the kind of decision small brands get wrong for years when they read attributed ROAS as truth. These numbers are invented, but the gap they illustrate, retargeting looking far better on paths than in a holdout, is the pattern incrementality tests commonly expose.

    Signs you have outgrown tests and need a model

    Holdout tests have their own ceiling. Each one answers a single question about a single channel for a single period, and a store running five channels cannot test all of them every quarter without leaving revenue on the table. Three signals suggest the business is ready to add an MMM to the testing routine rather than replace it.

    The first is channel count combined with spend. Once four or more channels each take at least 10 percent of a budget above roughly €50,000 a month, the interactions between them start to matter more than any single channel's lift, and tests one at a time cannot capture that. The second is offline or upper-funnel spend: podcasts, out-of-home, connected TV or retail media that no click path will ever show.

    MMM is the only one of the two approaches that sees those channels at all. The third is a history of deliberate variation. If the team has already run two or three tests and shifted budgets on purpose, the weekly data now contains the independent movement a model needs to learn from.

    Even then, the sensible order is to calibrate. Fit the model, compare its estimate for the channel you already tested with the test result, and adjust priors until the two roughly agree. Robyn supports calibration against experiment results directly, and Meridian accepts priors informed by experiments. A model that contradicts your own clean test by a factor of two is telling you about its assumptions, not about your marketing.

    What not to do at this stage is switch off MTA. It keeps doing the one job it is good at, ranking ads and keywords inside a channel, and the model and the tests handle the budget split between channels.

    Questions small teams ask about MMM and MTA

    Can I build an MMM in a spreadsheet? A linear regression of weekly sales on lagged channel spend can be built in a spreadsheet, and it is a useful first look. It will not model saturation or carryover properly, so treat the coefficients as directional and never use them to set budgets above the spend range you observed.

    Does GA4 data-driven attribution count as MTA? Yes, it is a multi-touch model that assigns fractional credit from observed paths, but only for touchpoints Google Analytics can see. Since 2023 GA4 offers only data-driven and last-click models for new reports, after it removed first-click, linear, time-decay and position-based options.

    How long should a geo holdout test run? Long enough to cover at least one full purchase cycle plus the ad carryover, which for most stores means 3 to 6 weeks. Shorter tests mostly measure noise; longer ones risk contamination from promotions and seasonality.

    Is a post-purchase survey reliable enough to use? It is biased toward memorable channels such as podcasts and influencers and against search, but it captures exposure that click tracking misses. Use it as one input next to tests, and keep the answer options identical across months so trends are comparable.

    The harder question sits underneath all of this. If the honest answer for a small brand is "test first, model later", at what spend does a model start paying for itself: when the cost of one wrong reallocation exceeds the analyst days, or only when the business can afford to vary spend on purpose just to make the model learnable?

    Related Articles