# Attribution vs marketing mix modeling vs incrementality: which does a small team need?

Canonical URL: https://www.martechdevs.com/blog/attribution-vs-mmm-vs-incrementality

By [Jafrin Akter Alo](https://www.martechdevs.com/authors/jafrin-akter-alo)

Data Engineer and Co-founder, martechdevs

Published: 2025-07-14

Last updated: 2026-09-18

Use attribution to understand observed paths and operate campaigns, incrementality experiments to test a defined marketing intervention, and marketing mix modeling to estimate broader effects from aggregate historical data under explicit assumptions. A small company does not need all three immediately. Choose the method that can answer the next material decision with the data, outcome volume and operating capacity you actually have.

## Start with the decision and the alternative

“Which channel works?” is too vague to design useful measurement. Write a decision that can change an action: “Should we add $10,000 to prospecting next quarter?” or “What happens to new paid accounts if we reduce this retargeting programme?”

Specify the alternative. Additional spend versus current spend is different from advertising versus no advertising. A test that removes a channel estimates the effect of that removal under the test conditions. It does not directly tell you the return on the next dollar above today's budget.

Define the business outcome, population, period and minimum effect that matters. For SaaS, decide whether the outcome is a qualified opportunity, first paid account or retained revenue. For ecommerce, distinguish all orders from new customers and account for relevant returns. Keep the definition stable across methods.

Use the [marketing attribution models guide](https://www.martechdevs.com/blog/marketing-attribution-models) for credit-allocation rules. Here, the question is which kind of evidence the decision requires.

## What attribution can answer

Attribution assigns credit across eligible observed interactions. It can show which campaigns precede measured outcomes, where a tracked journey loses people and how a platform's bidding goals are performing.

It needs reliable event collection, identifiers, campaign information and outcome definitions. Its coverage is bounded by the interactions and identities available to the implementation. A first observed click is not necessarily the first influence on a buyer.

This makes attribution useful for operational diagnosis. If a campaign suddenly produces visits but no verified registrations, investigate its audience, landing experience and tracking. You do not need a multi-year MMM to find a broken signup flow.

The limitation is causal interpretation. A retargeting audience already contains people with demonstrated interest. Strong attributed conversion performance can reflect both advertising effects and that selection. Changing the allocation rule alone does not tell you what would have happened without the ad.

Google's [GA4 attribution documentation](https://support.google.com/analytics/answer/10596866?hl=en) describes a data-driven methodology that uses converting and non-converting paths and incorporates experimental evidence for some estimates. Avoid claiming every attribution method is merely a fixed rule. Equally, do not treat a platform's estimated credit as proof of the effect of your proposed business intervention.

## What an incrementality experiment can answer

An incrementality experiment compares outcomes under a defined treatment and a credible counterfactual. In a randomised design, assignment to treatment or control helps separate the intervention from pre-existing differences between people or markets.

The treatment might be eligibility for a campaign, a spend increase, a channel reduction or a lifecycle message. Analyse people or accounts according to the experiment's assignment and planned method. Comparing people who happened to see an ad with people who did not can reintroduce selection bias, because exposure itself is not random in many advertising systems.

For a platform-managed study, understand which intervention, population and outcome the product supports. Google's [Conversion Lift documentation](https://support.google.com/google-ads/answer/12003020?hl=en) distinguishes user-based and geography-based studies and says availability is not universal across accounts. Verify access and study requirements before writing a launch plan that depends on the feature.

Experiments have a narrower scope than the phrase “the channel's true ROI” implies. They estimate an effect for a population, intervention, time and operating environment, with uncertainty. A holiday promotion test does not automatically establish next spring's prospecting economics.

### An A/B test can answer a different question

A creative A/B test can estimate the difference between creative A and creative B. If both groups receive advertising, it does not measure the effect of the whole advertising programme against no advertising.

A landing-page experiment can estimate which experience converts assigned visitors more effectively. It cannot, by itself, establish how many of those visitors the media generated incrementally. Name the treatment and control in the report so the result is not stretched beyond its design.

## Work through lift, uncertainty and economics

Consider an invented, simplified randomised test with 10,000 eligible people assigned to each group. The outcome is whether each person makes at least one qualifying purchase during the fixed measurement window. Assume complete, comparable outcome measurement and independent assignment without cross-group spillover.

The treatment group has 600 purchasers, or 6%. The control group has 500, or 5%. The absolute difference is one percentage point. Relative lift is 20%, calculated as `(6% - 5%) / 5%`. Applied to the treatment group's size, the estimated additional purchasers are 100.

Under a simple normal approximation, the 95% confidence interval for the difference is about 0.37 to 1.63 percentage points. The estimate is not exactly 100 extra purchasers with certainty. The interval corresponds to roughly 37 to 163 for this treatment-group size under the example's assumptions.

Now shrink the example to 1,000 people per group, with 60 versus 50 purchasers. The observed rates and relative lift are unchanged, but the approximate interval widens to about negative 1.00 to positive 3.00 percentage points. A persuasive-looking 20% lift can be too uncertain to justify the proposed change.

This calculation is not a template for every advertising experiment. Cluster assignment, repeated outcomes, uneven groups, heavy-tailed revenue and platform estimation methods require the analysis appropriate to that design. Plan sample size and the minimum detectable effect before launch rather than deciding after seeing a favourable point estimate.

Suppose the larger illustrative test costs $8,000 in media attributable to the tested intervention and each additional purchaser generates $120 in measured revenue. The point estimate implies $12,000 of additional revenue, an incremental ROAS of 1.5 and an incremental acquisition cost of $80.

At a 60% contribution margin before media, that revenue contributes $7,200. Subtract $8,000 of media and $1,000 of other incremental execution costs: the point estimate is a $1,800 shortfall. Positive purchase lift is not automatically a profitable intervention. The [marketing ROI guide](https://www.martechdevs.com/blog/how-to-measure-marketing-roi) shows how to carry that contribution, the full investment and any forecast future value into the financial decision.

## When a geo experiment is feasible

A geo experiment changes marketing across geographic units and measures aggregate outcomes. Google's [original geo-experiment research](https://research.google/pubs/measuring-ad-effectiveness-using-geo-experiments/) describes assigning non-overlapping regions to treatment and control and implementing the conditions through geographic targeting.

It can be useful when the business outcome is measurable by region but individual advertising paths are incomplete. It is not automatically the easiest option for a small business. You still need enough comparable regions and outcome volume, control over the intervention and a way to limit contamination across groups.

Ask these operational questions before paying for a study:

- Can the campaign actually maintain different treatment conditions in the chosen regions?
- Can you assign business outcomes to regions consistently, including online and offline activity where relevant?
- Are important promotions, distribution changes or sales-team actions balanced or explicitly handled?
- Could customers cross regions or advertising spill across boundaries enough to weaken the comparison?
- Can the business tolerate the planned reduction, holdout or additional spend for the full period?

Google's current [Meridian GeoX design documentation](https://developers.google.com/meridian/geox/design-methodology?hl=en) includes design evaluation through simulations, minimum detectable effects and pre-test comparability. Use the actual historical outcomes to evaluate whether a commercially meaningful effect could be detected. A map with two coloured regions is not a power analysis.

The tool's [pre-test data guide](https://developers.google.com/meridian/geox/prepare-your-pretest-data) uses daily outcome and spend data by region, with one record per date and location. Do not confuse this input contract with the weekly history commonly used for an MMM.

If the feasible design can only detect a 40% improvement and the decision turns on a 5% improvement, that test is poorly matched to the decision. Reduce the scope, reconsider the design or accept that the current evidence will remain limited. Running an underpowered study does not make the uncertainty disappear.

## What MMM adds, and what it demands

Marketing mix modeling uses aggregated outcomes, marketing activity and relevant business conditions over time, often across regions. It can estimate channel response, carryover and saturation, then support budget scenarios. It does not require reconstructing every individual's path.

Its causal usefulness depends on the model and the data, not the absence of cookies. If spending rises exactly when demand rises for another reason, a naive model can assign that demand to advertising.

Google's [Meridian data-collection guidance](https://developers.google.com/meridian/docs/pre-modeling/collect-data) gives a rule of thumb of at least two years of weekly data for geographic models and three years for national models. It also recommends at least three years when only monthly data is available. These are planning guidelines, not a guarantee of a useful model or a universal rule for every MMM implementation.

Meridian's [discussion of data quantity](https://developers.google.com/meridian/docs/pre-modeling/amount-data-needed?hl=en) explains why the number of channels, controls and parameters matters. A long spreadsheet with little independent variation may tell the model less than its row count suggests. Adding old history can also make the model less relevant when the business has changed substantially.

Before commissioning an MMM, assemble:

- A consistent business outcome with a clear geographic and time definition.
- Channel spend and delivery measures at the required granularity.
- Relevant price, promotion, availability and distribution changes.
- Records of tracking, product, market and accounting-definition changes.
- A history of why budgets changed, not just the resulting spend.
- Prior experiments with their populations, interventions, periods and uncertainty.

That last budget-history question matters. A marketing manager may increase spend because a product launch or demand forecast suggests sales will rise. The analyst needs to understand that decision process to avoid assigning every coincident increase to the channel.

### Do not optimise model fit at the expense of the causal question

Meridian's [control-variable guidance](https://developers.google.com/meridian/docs/advanced-modeling/control-variables?hl=en) distinguishes confounders from variables that merely predict the outcome. It warns against selecting inputs solely to improve prediction and against controlling for mediators on the effect path.

A practical example: if the question is the total effect of advertising on purchases, website visits may be part of how advertising produces those purchases. Automatically controlling for visits can remove part of the effect you wanted to estimate. Whether a variable belongs depends on the causal question and timing, not how strongly it correlates with sales.

Ask an analyst to explain the assumed relationships and the sensitivity to plausible alternative specifications. A high fit score is not a certificate that the channel effects are correct.

### Calibrate with experiments without pretending the questions are identical

Meridian supports [calibration using experimental evidence](https://developers.google.com/meridian/docs/advanced-modeling/roi-priors-and-calibration). Its documentation explicitly notes that experiment and MMM estimates can target different effects and conditions.

A test of a 20% spend increase is not the same comparison as removing all spend from a channel. A campaign-level experiment in one market may not transfer directly to a whole channel across the country. Preserve those differences and uncertainty when using experiments to inform model priors.

## A planning interface does not remove the modelling work

Meridian's [Scenario Planner documentation](https://developers.google.com/meridian/docs/scenario-planning/meridian-scenario-planner) describes an open-beta interface for analysis reports and interactive budget optimisation. Its report-generation workflow starts with a trained Meridian model.

That is useful for letting a marketing team explore scenarios. It does not mean a founder can upload an arbitrary sales sheet and obtain validated causal answers. Someone still needs to prepare the data, choose and diagnose the model, check assumptions and explain where the scenarios exceed the evidence.

When comparing tools or proposals, separate modelling work from presentation. Ask who owns the data pipeline, how the model is refreshed, which checks can fail, and what would make the analyst refuse to recommend a reallocation.

## Choose a sequence that fits a small team

There is no defensible universal switch at $100,000 of monthly media spend. A concentrated business with many transactions and a controllable intervention may support a useful experiment below that budget. A larger business with few outcomes, unstable definitions or inseparable channel changes may still struggle.

Use this sequence as a decision process:

1. **Repair the outcome ledger and collection.** Make orders, accounts or opportunities reconcilable before treating a dashboard as evidence.
2. **Establish operational attribution.** Preserve campaign identifiers, useful journey stages and a documented business reporting view.
3. **Identify the highest-value uncertainty.** Choose the decision whose wrong answer would cost enough to justify better measurement.
4. **Design a feasible intervention.** Estimate detectable effects, operational constraints, measurement lag and cost before launch.
5. **Consider MMM when repeated allocation decisions and adequate history justify it.** Use experiments and business knowledge to challenge and inform the model.

![Measurement method selection starts with a business decision, uses attribution for diagnosis, assesses experiments and adds MMM when data and recurring decisions justify it](https://www.martechdevs.com/assets/blog/attribution-vs-mmm-vs-incrementality.svg)

You can collect qualitative evidence along the way. Customer interviews and self-reported discovery can reveal influences missing from tracked paths. They help generate hypotheses; recall and response bias mean they do not quantify incremental sales with experimental precision.

## Write the decision rule before launch

Create a one-page measurement brief naming the treatment, alternative, assignment unit, outcome, measurement period, minimum useful effect, planned analysis and decision owner. Record contamination risks, concurrent changes and the conditions that would invalidate the result.

Specify what happens if the result is positive but unprofitable, uncertain, or inconsistent across important customer groups. Commit to the analysis window and avoid repeatedly ending the test at whichever moment produces the preferred answer unless the statistical design explicitly supports that stopping rule.

For an MMM proposal, replace the experiment's assignment details with the model's causal assumptions, input history, validation plan, calibration evidence and permissible scenario range. Require uncertainty and sensitivity analysis alongside the recommendation.

Use these checks to decide whether the work is ready:

## Checklist

- [ ] The outcome measures business value and has a reliable source.
- [ ] The alternative being compared is explicit.
- [ ] The method can answer that comparison at the required scope.
- [ ] Expected uncertainty is small enough to inform the decision.
- [ ] Platform eligibility or operational targeting has been verified.
- [ ] Costs include measurement and execution, not only media.
- [ ] The analysis accounts for the actual assignment unit and design.
- [ ] Model assumptions and unsupported extrapolations are visible.
- [ ] Someone owns the decision after the result arrives.

If the first two checks fail, start with an [analytics and tracking review](https://www.martechdevs.com/analytics-tracking). If the missing capability is consistent historical business and channel data, scope the [warehouse work](https://www.martechdevs.com/data-warehouse-cdp) around those inputs. A new measurement tool earns its cost when it changes a decision with evidence the team can defend.

## Frequently asked questions

### Is multi-touch attribution dead?

No. Observed-path attribution still supports useful operational questions, provided its coverage and definitions are clear. It should not be asked to recover unobserved journeys or prove causal effects that its evidence does not establish.

### Should a small company start with geo holdouts?

Only if the design is feasible. Check regional outcome volume, controllable targeting, contamination, business tolerance and the detectable effect. A user-level or account-level intervention may fit some questions better; in other cases no affordable experiment will resolve the desired effect yet.

### Can MMM work with less than two years of data?

Feasibility depends on the model, variation, number of effects and relevant prior evidence. Meridian's published history guidance is a starting point, not a universal cutoff. A provider should explain why the available data can identify the proposed effects and how uncertain the answers remain.

### Does Meridian Scenario Planner build a reliable model for me?

Its documented workflow starts from a trained model and supports reporting and scenario exploration. Data preparation, model selection, diagnostics and causal review still need an owner.

### What if attribution and the experiment disagree?

First align the outcome, population, period and comparison. Attribution allocates observed credit; an experiment estimates a treatment effect under its design. Investigate measurement failures and uncertainty, then use the evidence appropriate to the decision rather than forcing the two numbers to match.

## Sources

1. [Google: GA4 attribution methodology](https://support.google.com/analytics/answer/10596866?hl=en)

2. [Google Ads: Conversion Lift](https://support.google.com/google-ads/answer/12003020?hl=en)

3. [Google Research: Measuring ad effectiveness using geo experiments](https://research.google/pubs/measuring-ad-effectiveness-using-geo-experiments/)

4. [Meridian GeoX: Design methodology](https://developers.google.com/meridian/geox/design-methodology?hl=en)

5. [Meridian GeoX: Prepare pre-test data](https://developers.google.com/meridian/geox/prepare-your-pretest-data)

6. [Meridian: Collect and organise data](https://developers.google.com/meridian/docs/pre-modeling/collect-data)

7. [Meridian: Amount of data needed](https://developers.google.com/meridian/docs/pre-modeling/amount-data-needed?hl=en)

8. [Meridian: Control variables](https://developers.google.com/meridian/docs/advanced-modeling/control-variables?hl=en)

9. [Meridian: Calibrate treatment priors](https://developers.google.com/meridian/docs/advanced-modeling/roi-priors-and-calibration)

10. [Meridian: Scenario Planner](https://developers.google.com/meridian/docs/scenario-planning/meridian-scenario-planner)