Causality Engine vs Meta Robyn: Who Has the Data Team

The short answer: two or three people in marketing, and nobody on the payroll who writes R. That store should pick Causality Engine. The first finding is free, and the full read of every channel in your GA4 export is €99 once, refundable within 30 days (pricing, checked on 27 October 2026). Meta Robyn is free, open source and capable. It suits a company with in-house analysts, at least two years of weekly data and a habit of running lift tests to calibrate against. What decides it is not which model is cleverer. It is who runs it.

Disclosure: I founded Causality Engine and own it, and this page recommends it. Every claim about it links to Causality Engine's own pages.

The staffing test:

1. Price the first study in weeks. Robyn's analyst's guide to MMM gives typical timings for a first study. Scope takes 1 to 2 weeks, data collection 4 to 6, data review 1 to 2, modelling 4 to 8, analysis and recommendations 2 to 4. Add them up, for example: 1 + 4 + 1 + 4 + 2 = 12 weeks at the quick end. At the slow end, 2 + 6 + 2 + 8 + 4 = 22. Only then does the first budget move.
2. Check the rows you hold. The same guide asks for at least two years of weekly data, and one independent variable per ten observations. So 104 weekly rows support about 104 ÷ 10 = 10 variables, for example. Those ten have to cover every paid channel, organic activity, promotions and seasonality. Ten fills up fast.
3. Name the person who will choose. Robyn returns a set of candidate models on a Pareto front and asks the analyst to select one. Someone owns that call, defends it to finance and refreshes it. Write that name down. If the line stays blank, the read fits better: Causality Engine returns one figure per channel with a data-health score, per its how-it-works page.

Robyn, in Meta's description

Robyn is an experimental, open-source marketing mix modelling package from Meta Marketing Science, per its welcome page. It fits with ridge regression, tunes with Meta's Nevergrad library, splits out trend and seasonality with Prophet, and ends with a budget allocator. Meta built it for granular datasets with many variables, especially from digital and direct-response advertisers. It needs no personal data, cookies or pixel. The R package installs from CRAN under the MIT licence, with a one-time Python install of Nevergrad, per the installation page and GitHub. The beta Python API still needs R underneath.

Where the two diverge

One read, or a front of candidate models

Robyn's guide recommends a high iteration count, for example 2,000 iterations across five trials for geometric adstock. It shows Nevergrad discarding most of the solutions it explores and keeping the few on the first Pareto fronts. That is the strength: a wide search that reduces the analyst's bias. It is also a job description. Compare candidates, check ROAS convergence, pick one. The read has no menu. Each channel gets one estimate and a grade, and when the grade drops under 90 the how-it-works page says to move budget in steps rather than all at once.

When the data is too thin

Robyn's ten-rows-per-variable rule is a self-check you can run before you start. Causality Engine applies its own threshold and says no out loud. Its home page says that when spend is too small for the model to find a signal, it declines. The case-studies page shows one. At OFFFTRACK, the channel gap sat inside the confidence interval at the brand's spend, and the advice was to skip the continuous plan and come back when paid spend grows. A thin dataset is not a mistake the store made. Small budgets give any model little to see. For a small store, being told to wait is a useful answer.

Whose experiment pins it down

Robyn's home page says it calibrates against ground-truth methods such as geo tests, Facebook lift and MTA. The guide strongly recommends doing so, and names Meta Conversion Lift and Meta GeoLift. Causality Engine's incrementality page grants that a Meta lift study is a real randomised experiment. It is also Meta's experiment, on Meta's inventory, graded by Meta. It says nothing about the channels around it.

Cost: licence, hours, euros

Checked on 27 October 2026. Causality Engine's pricing page: €99 per read, one-time. Pro, at €299 a month billed monthly, adds continuous reads with automated GA4 ingestion and unlimited seats. Enterprise is a custom 90-day pilot. All prices exclude VAT. Robyn is free under the MIT licence, and its analyst's guide states the in-house cost plainly: no ongoing fee apart from headcount and time. It requires in-house analytical or data-science capability.

Robyn's guide names two methods as the ground truth to calibrate against: Meta Conversion Lift, which Meta runs on its own inventory, and Meta's GeoLift. By default, then, the method that pins down your Meta numbers was designed by the company that sold you the impressions.

That is no reason to skip calibration. It is a reason to add one test the platform did not design, such as a regional holdout you plan and read yourself.

Doing it free

Three free steps carry most of the decision. First, run Robyn's demo script on the simulated data in the package, per its GitHub README. If setup defeats the team, the staffing question has its answer. Second, count your weekly rows and variables against the one-in-ten rule. Third, put your GA4 export through the free first finding on the start page, which the home page says runs in your browser without the file leaving it. Pay the €99 when the free finding shows enough hiding under Direct to change a budget. Take Pro when you run eight or more channels and budgets move weekly. For the wider route without a modeller, see marketing mix answers without a data team.

Choose Causality Engine if

• Your job is this quarter's channel split, and the team has no analyst to run and defend a model.
• You have under two years of clean weekly data, or too many channels for the rows you hold.
• You want the tool itself to say when the signal is too weak to act on.
• You want Direct's credit returned to the channels behind it.

Choose Meta Robyn if

• Your job is an in-house mix model, and you employ analysts who work in R.
• You hold two years or more of weekly data with real variation in spend, offline media included.
• You run lift studies or geo tests regularly and want them built into the model.
• You want full control of the code and no licence fee.

Causality Engine's MMM page makes the same split in its own words: Robyn and Google's Meridian are free if someone on the team can build, feed and maintain the model.

The read, if the team is you

Causality Engine, which I founded, is built for the case where nobody will maintain a model: take the free finding on the start page, and if it earns it, the €99 read, described step by step on the how-it-works page. The glossary keeps a short entry on Causality Engine.

FAQ

Is Meta Robyn free to use?

Yes: MIT-licensed and on CRAN, per its GitHub page. The cost is analyst time, which its guide counts in weeks for a first study.

Does Robyn need Meta ad data?

No. It models whatever media you supply, with no pixel or personal data, per its home page. For calibration, its guide names Meta Conversion Lift and GeoLift.

Can Causality Engine and Robyn run together?

Yes. The read takes only the GA4 export, so it runs beside any model. Causality Engine's home page calls it a layer, not a migration, and a Robyn team can use it as an outside check on digital channels.