MMM vs Attribution vs Incrementality: When to Use Which

Use attribution to make day-to-day campaign decisions, marketing mix modelling (MMM) to split budget across channels over months, and incrementality testing to prove a channel is actually causing sales rather than just taking credit for them. Attribution describes the path a customer took, MMM estimates what statistically moves your sales line, and incrementality answers the only question that really matters for spend: what would have happened if you had not run the ad. Most growing brands need all three — attribution weekly, incrementality quarterly, MMM once or twice a year.

The three methods, defined without jargon

Attribution — usually multi-touch attribution, or MTA — assigns credit for a conversion to the ads and clicks that came before it. Your Meta Ads Manager reports, Google Ads dashboard and Google Analytics 4 (GA4) reports are all attribution. It works by tying one person’s touchpoints to one person’s purchase.

Marketing mix modelling (MMM, sometimes written as media mix modelling) is a statistical model built on aggregate history — typically two to three years of weekly spend, sales, pricing, discounts, distribution and seasonality. It never sees an individual person. It asks a simpler question: when we spent more on this channel, did sales rise, after accounting for everything else that was happening?

Incrementality testing is a controlled experiment. You deliberately withhold advertising from one group — a set of cities, a randomised user holdout, or a public service announcement (PSA) placebo ad — and compare their sales against a matched group that did see the ads. The gap is the lift you genuinely caused.

Comparison table: MMM vs attribution vs incrementality

	Attribution (MTA)	MMM	Incrementality test
Question it answers	Which ads did buyers touch?	How should I split budget across channels?	Did this spend cause extra sales?
Data it uses	User-level clicks, views, conversions	Aggregate weekly spend and sales history	Test vs control group outcomes
Decision speed	Daily to weekly	Quarterly to annual	2–8 weeks per test
Granularity	Campaign, ad set, keyword, creative	Channel and sometimes sub-channel	One channel or tactic at a time
Covers offline and brand media?	No	Yes (TV, print, sponsorships, OOH)	Yes, if you can vary it geographically
Proves causation?	No	Correlation, with causal assumptions	Yes — this is its whole point
Rough cost in India	Free in platform tools; ₹1–5 lakh a year for a paid tool	₹15–60 lakh a year via an agency; far less with open-source tools plus an analyst	Mostly the revenue you give up during the holdout, plus analyst time
Biggest weakness	Double-counts and over-credits the last touch	Needs long history and spend variation; slow to react	Answers one narrow question, then goes stale

Cost figures above are rough market estimates for Indian brands, not quoted rates — treat them as orders of magnitude.

Attribution: fast, granular, and reliably over-optimistic

Attribution’s strength is speed and detail. Nothing else tells you at 9am that a particular creative is burning money. If your decision is “which ad do I pause today,” attribution is the right tool and MMM is useless to you.

Its weakness shows up the moment you add the numbers. Take a D2C skincare brand spending ₹12 lakh a month: ₹6 lakh on Meta, ₹4 lakh on Google Search, ₹2 lakh on influencers. Meta claims 900 orders, Google claims 700, the influencer tracking links claim 200. That is 1,800 claimed orders — against 1,400 actual orders in Shopify.

Nobody is lying. Each platform counted the same buyer, and each is using a different window and a different definition of a “view.” Attribution also cannot see the buyer who was already going to purchase and simply clicked the branded search ad on the way — the single most over-credited line item in most Indian D2C accounts.

Signal loss has made this worse. Safari and Firefox block third-party cookies by default, and Apple’s App Tracking Transparency prompt cut mobile app tracking sharply. India is somewhat insulated because it is overwhelmingly an Android market — Android’s share of Indian mobile usage sits around 95% — but India’s Digital Personal Data Protection (DPDP) Act, passed in 2023 and now moving into enforcement, pushes in the same direction: less consented user-level data over time.

MMM: the budget-setting tool

MMM works at the altitude attribution cannot reach. Because it uses aggregate data, it can price channels attribution never sees: television, IPL sponsorships, out-of-home in Bengaluru, print in a regional daily, even a distributor push. It also handles the messy stuff — a Diwali sales spike, a Big Billion Days price war, a competitor launch.

The output that matters is marginal return: what the next rupee earns, not the average rupee. A model might tell an apparel brand that Google Search returns ₹4.20 per rupee on average but only ₹1.30 on the next rupee, because search demand is nearly saturated, while connected TV returns ₹1.90 on average and ₹1.85 on the next rupee. That is a clear signal to move spend, and attribution would never surface it.

The catch is data. A credible MMM wants roughly two to three years of weekly data and genuine variation in spend. If you have spent an identical ₹8 lakh a month on Meta for two years, no model can tell you what ₹12 lakh would do. As a rough rule of thumb, MMM starts earning its keep somewhere above ₹1–2 crore of annual media spend across three or more channels. Open-source options such as Meta’s Robyn and Google’s Meridian have lowered the entry price considerably.

Incrementality: the truth serum

An incrementality test is the only method that produces a number you can defend to a CFO, because it is a real experiment rather than a model.

The most practical version for Indian brands is a geo test. Suppose the same skincare brand picks 20 cities, matches them into 10 similar pairs on historic sales, then turns Meta off in one city from each pair for four weeks. Test cities deliver 13,500 orders; control cities deliver 12,000. The 1,500-order gap is the incremental result of ₹15 lakh in spend — a true cost per acquisition of ₹1,000, against the ₹450 the platform reported.

That is not a disaster; it is information. It means roughly a third of platform-reported orders were genuinely caused by the ads, and every future Meta forecast should be discounted accordingly. Meta’s open-source GeoLift package and Google’s geo-experiment tooling both automate the matching and the maths.

The trade-off is scope and shelf life. A test answers one question about one channel at one spend level, and the answer decays as creative, audiences and competition change. Budget for two to four tests a year on your largest channels rather than one heroic study.

A decision framework you can actually use

1. “Which creative, keyword or audience do I change this week?” → Attribution. Speed beats precision here, and the errors are roughly consistent across options you are comparing.
2. “How do I split next quarter’s or next year’s budget?” → MMM, if you have the spend and history. If you do not, use a series of incrementality tests to build the same picture channel by channel.
3. “Is this channel worth anything at all? Should I cut it?” → Incrementality. Never kill a channel on attribution data alone, and never defend one on it either.
4. “Should we spend on brand, TV or sponsorships?” → MMM, supported by geo tests. Attribution structurally cannot measure media without a click.
5. “Our reported ROAS looks great but the P&L doesn’t.” → Incrementality, immediately. That gap is the classic symptom of over-credited retargeting and branded search.

Use them together, don’t pick a winner

Mature measurement teams run all three and let them correct each other — often called triangulation. The pattern is simple: run an incrementality test, then use the result to calibrate the other two.

If a lift test shows Meta’s true contribution is 35% of what Ads Manager reports, apply that 0.35 factor to your daily attribution reporting so the team optimises against something closer to reality. Feed the same experiment result into your MMM as a prior or a validation check — if the model says Meta’s return is 2.8 and your test says 1.1, the model is wrong and needs rebuilding.

Common mistakes

• Adding platform-reported conversions together. The total will always exceed your actual orders. Always reconcile against one source of truth — your order system.
• Treating MMM output as causal proof. It is a correlation model with causal assumptions baked in. Validate it with an experiment.
• Running a geo test without matched controls. Comparing Mumbai to Indore is not an experiment; it is a coincidence with a spreadsheet.
• Ending tests too early. Short tests miss delayed purchases and produce noise. Two weeks is a bare minimum; four or more is safer for considered purchases.
• Building MMM at the wrong scale. Below roughly ₹1 crore of annual media, you will spend more on the model than you could ever save with it.
• Changing the attribution window mid-quarter. Your results will “improve” for reasons that have nothing to do with marketing.
• Ignoring the holdout cost. Withholding ads costs real revenue. Price that in, and treat it as research spend.

What this means for you

• Pick your tool by decision speed: daily decisions use attribution, quarterly budgets use MMM, existential “is this working” questions use incrementality.
• Reconcile platform-reported conversions against your order system every month and track the gap as a standing metric. A widening gap is an early warning.
• Book one incrementality test on your largest channel this quarter. A matched-city geo test on four weeks of spend is achievable for most brands over ₹50 lakh in annual media.
• Turn the test result into a discount factor and apply it to your attribution dashboard, so the team optimises against calibrated numbers rather than raw platform claims.
• Hold off on MMM until you have two years of weekly data, meaningful spend variation and three-plus channels. Before that, sequential lift tests give you more truth per rupee.
• Whatever you use, write down the assumption you are making. “We assume branded search is 20% incremental” is a testable statement; “ROAS is 4.2” is not.

Frequently asked questions

What is the difference between MMM and MTA?

Multi-touch attribution (MTA) uses individual user-level data to credit specific ads for specific conversions, so it is granular and fast but blind to offline media and prone to double-counting. Marketing mix modelling (MMM) uses aggregate weekly spend and sales history to estimate each channel’s contribution, so it covers every channel including TV and print, but it is slower and less granular. MTA guides campaign management; MMM guides budget allocation.

Is incrementality testing worth it for a small budget?

Yes, but scale the method to the budget. A brand spending ₹3–5 lakh a month cannot run a 20-city geo test, but it can run a two-week full-channel pause and compare order volume against the preceding and following periods, or use a platform’s built-in conversion lift study. The result is rougher than a formal experiment, but it is still closer to the truth than a platform dashboard.

How often should you rebuild a marketing mix model?

Refresh the data quarterly and rebuild the model annually, or sooner after a structural change such as entering a new category, a major price shift or a new sales channel. Between rebuilds, validate the model against any incrementality tests you run — a large disagreement means the model needs work before you trust its budget recommendations.

Which measurement method should a D2C brand start with?

Start with clean attribution plus a monthly reconciliation against your actual orders, then add incrementality testing on your biggest channel once monthly spend passes roughly ₹5 lakh. Add MMM last, when you have two to three years of weekly history and spend across several channels. Building MMM first is the most common expensive mistake in marketing measurement.