---
title: "Marketing mix modelling, attribution or incrementality: which to use"
url: https://marketingsignals.com/insights/marketing-mix-modelling-attribution-incrementality
type: Article
date: 2026-11-02
author: "Simon Rattray"
---

# Marketing mix modelling, attribution or incrementality: which to use

Marketing mix modelling, attribution or incrementality testing? What each answers, the data and spend it needs, and how to combine them in one plan.

## Key takeaways

- Marketing mix modelling, attribution and incrementality testing answer different questions, so the useful decision is which to use for which budget choice, not which one wins.
- GA4 now offers only data-driven and two last click models, and none of them can credit exposure that never produced a tracked visit.
- Incrementality tests are the closest marketing gets to proof, but each needs enough conversions to read and answers one question for one period.
- Marketing mix modelling needs years of weekly data with real variation in spend; free tools such as Google Meridian and Meta Robyn cut licence costs, not the data and analyst work.
- Use the model to set the channel split, tests to calibrate it, and attribution to optimise campaigns inside each channel day to day.

Should you invest in marketing mix modelling, rely on attribution, or run incrementality tests? The question usually arrives after a budget argument. GA4 says one thing, the Meta dashboard says another, and finance has a third number based on what happened to total revenue.

Each of those numbers answers a different question, and most measurement arguments start when someone treats one of them as the whole truth. This guide sets out what each method answers, the data and spend it needs, how quickly it reports and where it misleads, so you can decide what your business needs now.

The short version: they are not rivals. A model sets the budget split, experiments check that spend causes the sales it claims, and attribution runs the day-to-day optimisation. The rest of this article explains why, and what it takes to run each one well.

## Why the choice matters to the business

The cost of the wrong measurement is misallocated budget, and it compounds. Click-based attribution favours channels that sit close to the purchase, such as brand search and retargeting, because they collect the last click. Spend higher up the funnel that created the demand looks weak and gets cut, and a year later new customer acquisition slows without an obvious cause.

The opposite mistake costs money too: funding a channel on the strength of platform-reported conversions that would have happened anyway. Our piece on [testing whether brand bidding is incremental](https://marketingsignals.com/insights/brand-bidding-incrementality-test) covers one of the most common versions of that question.

Getting it right gives the board a defensible answer to the question it actually asks: if we moved £100,000 from one channel to another, what would happen to revenue and margin?

## Attribution modelling: fast, granular and partial

[Attribution models](https://marketingsignals.com/glossary/attribution-models) share credit for a conversion between the touchpoints a tool recorded on the way to it. In GA4 the choice is now narrow. The first click, linear, time decay and position-based models were removed in November 2023, leaving [data-driven attribution and two last click models](https://support.google.com/analytics/answer/10596866): paid and organic last click, and Google paid channels last click.

Data-driven attribution is a real improvement on last click. Google says it evaluates converting and non-converting paths and uses a counterfactual approach. But it can only credit what GA4 records. Direct visits receive no credit under any model, the lookback window for most key events [defaults to 90 days](https://support.google.com/analytics/answer/10597962), and exposure that never produced a tracked visit, such as a TV slot, a podcast read or most social video views, is outside its sight. Visitors who decline analytics cookies and journeys that cross devices remove more of the path.

So attribution is the right tool for daily and weekly decisions inside a channel: which campaigns, keywords, audiences and creative to push. It is the wrong tool for deciding how much each channel should get in the first place.

## Incrementality testing: the causal answer, one question at a time

[Incrementality testing](https://marketingsignals.com/glossary/incrementality-testing) withholds a channel, campaign or audience from a control group and compares its outcomes with a group that keeps seeing it. The difference is the lift the spend caused. Because it is a controlled experiment, it is the closest marketing gets to proof.

Platforms offer this built in. Google's [Conversion Lift](https://support.google.com/google-ads/answer/12003020) splits test groups by user or by geography, though Google notes it is not available in every account. Geo experiments, where spend is switched off or increased in chosen regions, work for any channel, including offline media.

The constraints are volume and patience. A test needs enough conversions in both groups to show a difference larger than normal week-to-week noise, which rules out small channels and short windows. Holding back spend has an opportunity cost while the test runs. And each test answers one question for one period; it will not tell you the right budget across ten channels.

## What marketing mix modelling answers, and what it needs

Marketing mix modelling (MMM) is a statistical model of total sales over time, usually by week, that estimates how much each channel contributed alongside price, promotions, seasonality and other factors. It needs no user-level tracking, so it covers TV, out of home and offline sales as readily as paid search, and cookie consent does not affect it.

It is also the method that most needs history and variation. Google's Meridian documentation describes its own data guidance as rough and directional, but its worked example is telling: two years of weekly data is too small a sample to estimate a national model with 26 parameters reliably. It also warns that [insufficient variation in media spend](https://developers.google.com/meridian/docs/pre-modeling/amount-data-needed) weakens national models. A channel that has run at the same weekly budget for two years gives a model nothing to learn from.

Regional data helps. Meridian's documentation recommends [geo-level data when possible](https://developers.google.com/meridian/docs/pre-modeling/geo-selection-national-data), because each week then provides several observations rather than one. For a UK retailer that means sales and spend by region, which finance and media teams can often produce if asked early.

MMM is slow to build and to refresh, and it estimates average effects over months. It will not tell you which ad group to pause on Tuesday.

## Open-source MMM: Google Meridian and Meta Robyn

Two free, open-source tools have lowered the cost of building a model in-house. [Google Meridian](https://blog.google/products/ads-commerce/meridian-marketing-mix-model-open-to-everyone/) became available to everyone in January 2025. Its documentation says its hierarchical structure was designed for geo-level modelling; it accounts for reach and frequency rather than impressions alone, and it lets you feed incrementality experiment results in to calibrate the model. Google's geo experiment library, [Meridian GeoX](https://developers.google.com/meridian/geox), is now available alongside it for running those experiments.

[Meta's Robyn](https://github.com/facebookexperimental/Robyn) describes itself as an experimental, semi-automated MMM package from Meta Marketing Science. It is an R package with a Python API in beta, uses ridge regression with automated hyperparameter search, includes a budget allocator, and treats [calibration against experiments](https://facebookexperimental.github.io/Robyn/docs/features) as one of its optimisation targets. Meta says it is built for granular datasets with many independent variables, which suits digital and direct-response advertisers.

Free software is not a free model. Both tools need an analyst who understands the assumptions, clean weekly data going back years, and time to check that the outputs make commercial sense. The licence costs nothing; the people and the data work are where the cost sits.

## How we combine all three

We rarely recommend one method on its own for a business with several channels. This is the sequence we follow.

1. Start with the decision, not the method. List the budget decisions leadership needs to make in the next year: the annual channel split, whether to enter a new channel, whether brand search or retargeting is worth what it costs. Each decision points to a method, and some need nothing beyond what you already have.
2. Fix the attribution foundations. Check that conversions and revenue are tracked consistently across GA4 and the ad platforms, that campaigns are tagged, and that platform revenue can be reconciled with the finance figure. Every later method depends on this data.
3. Test the biggest open questions first. We prioritise by spend multiplied by doubt: a large budget whose effect nobody can prove comes before a small one. Agree the success measure and the smallest lift worth detecting before launch, not after.
4. Build the model once you have the history. In our experience a model becomes worth building with two to three years of weekly data, real variation in spend and ideally regional figures. If spend has been flat, plan deliberate variation, such as a regional pause or a burst, so the next model has something to learn from.
5. Calibrate the model with the test results. Feed experiment results into the model so its estimate for a tested channel agrees with the causal evidence. Where model and test disagree sharply, investigate before trusting either.
6. Set the split from the model and run channels on attribution. Use the model's response curves to set channel budgets for the quarter, then let GA4 and platform attribution steer campaigns inside each channel. Refresh the model quarterly or half-yearly and plan the next tests around its least certain estimates.

## Which to use at your stage

- Mostly search and shopping, few channels: attribution for optimisation, plus tests on brand search and any campaign type you doubt.
- Several paid channels including paid social and video, with a growing budget: add a regular programme of lift or geo tests on the largest lines.
- Many channels including offline media, with several years of history: a model for the annual and quarterly split, calibrated with tests, with attribution running campaigns day to day.

Our [e-commerce visibility playbook](https://marketingsignals.com/playbooks/the-e-commerce-visibility-playbook) shows where this fits in a retailer's wider search plan. The playbooks for [B2B lead generation](https://marketingsignals.com/playbooks/the-visibility-playbook-for-b2b-lead-generation) and [SaaS](https://marketingsignals.com/playbooks/the-visibility-playbook-for-saas) cover longer sales cycles, where offline conversions and pipeline value need importing before any of these methods can see revenue.

## How to measure the measurement

A measurement programme needs its own scorecard, or it turns into an expensive research project. We track four things.

- Decisions changed: how much budget moved because of a test or the model, and what happened to revenue afterwards.
- Agreement between methods: whether attribution, tests and the model point the same way for the same channel. Persistent disagreement is a finding in itself.
- Model credibility: how wide the uncertainty ranges are, and whether the model predicts a recent period it was not trained on.
- Blended efficiency: total marketing spend against total revenue or new customers, which no attribution setting can flatter, alongside platform-reported [ROAS](https://marketingsignals.com/glossary/roas).

Report upwards on one page: what we tested, what we learned, what budget moved and what the business result was. The principle behind [measuring your agency on revenue rather than rankings](https://marketingsignals.com/insights/measure-your-agency-on-revenue-not-rankings) applies to measurement work as much as to media.

## When each method misleads, and the mistakes we see

- Building a model on flat spend. If budgets barely moved, the model cannot separate channel effects from seasonality, and the output reflects its assumptions more than your data.
- Testing channels too small to read. A test that cannot detect a plausible lift wastes the holdout period, so size it before launch.
- Treating data-driven attribution as proof of incrementality. It shares credit among recorded touchpoints; it does not show that any of them caused the sale.
- Adding up platform-reported conversions. Each platform counts by its own rules and windows, so the totals can exceed your actual sales.
- Running a model once. Media costs, competition and the mix change, so a model that is not refreshed and recalibrated drifts out of date.

## Questions to ask your team or agency

- Which budget decisions in the next twelve months will this measurement inform?
- Which of our largest spend lines has never been tested for incrementality, and what would a test cost in held-back spend?
- Do we have the weekly history, regional data and spend variation a model needs, and if not, when will we?
- Where do attribution, tests and the model disagree, and what are we doing about it?

We design and run measurement plans like this as part of our [integrated search](https://marketingsignals.com/service/integrated-search) work, alongside the paid and organic programmes they measure.

## Frequently asked questions

### What is the difference between marketing mix modelling and attribution?

Attribution shares credit for each conversion between the tracked touchpoints that preceded it, such as ad clicks and visits, so it suits day-to-day campaign optimisation. Marketing mix modelling uses aggregate weekly sales and spend to estimate each channel's contribution, including channels with no click data such as TV, so it suits setting budget splits. Attribution is fast and granular; marketing mix modelling is slower but covers the whole mix.

### How much data do you need for marketing mix modelling?

There is no fixed minimum, but more history and more variation both help. Google's Meridian documentation calls its own guidance rough and directional, and gives an example where two years of weekly national data is too small to estimate a 26-parameter model reliably. It recommends geo-level data where possible and warns that spend with little variation weakens national models. Most businesses need several years of weekly data.

### What is incrementality testing in marketing?

Incrementality testing is a controlled experiment that measures how many conversions a channel, campaign or audience actually caused. One group is exposed to the activity and a control group is not, by user or by region, and the difference in outcomes is the incremental lift. It is the most reliable causal measure available, but it needs enough conversion volume and answers one question per test.

### Is Google Meridian free to use?

Yes. Google Meridian is an open-source marketing mix model that Google made available to everyone in January 2025, and its documentation says the library is free and open sourced on GitHub. The software costs nothing, but building a reliable model still needs clean weekly data, ideally by region, and an analyst who understands the method and can check the results.

### Which attribution models does GA4 offer?

GA4 offers three attribution models: data-driven, paid and organic last click, and Google paid channels last click. Google removed the first click, linear, time decay and position-based models in November 2023. Direct visits receive no credit under any model, and the default lookback window for most key events is 90 days, with 30 or 60 days available.

## Sources

- [Google Analytics Help: Get started with attribution (models available; rules-based models removed November 2023)](https://support.google.com/analytics/answer/10596866)
- [Google Analytics Help: Select attribution settings (lookback windows)](https://support.google.com/analytics/answer/10597962)
- [Google Ads Help: About Conversion Lift](https://support.google.com/google-ads/answer/12003020)
- [Google Meridian: Amount of data needed (last updated 28 October 2026)](https://developers.google.com/meridian/docs/pre-modeling/amount-data-needed)
- [Google Meridian: Geo selection and national-level data (last updated 28 October 2026)](https://developers.google.com/meridian/docs/pre-modeling/geo-selection-national-data)
- [Google Meridian: General product information (FAQs, last updated 28 October 2026)](https://developers.google.com/meridian/docs/faqs)
- [Google Meridian GeoX](https://developers.google.com/meridian/geox)
- [Google blog: Meridian is now available to everyone (29 January 2025)](https://blog.google/products/ads-commerce/meridian-marketing-mix-model-open-to-everyone/)
- [Meta Robyn on GitHub](https://github.com/facebookexperimental/Robyn)
- [Robyn documentation: Features (calibration, ridge regression, budget allocator)](https://facebookexperimental.github.io/Robyn/docs/features)

Visibility Playbooks

## Related Playbooks

More Insights

## More articles we think you’ll like