Marketing Mix Modeling Is Back: Google Meridian vs Meta Robyn for B2B Teams

Meta's Robyn documentation recommends 10 observations for every variable in a model. A B2B team modelling 8 channels and 4 controls on weekly data therefore needs about 120 weeks of clean history before marketing mix modeling says anything stable. That bar, not the choice between Google Meridian and Meta Robyn, decides whether the exercise is worth starting.

Google Meridian vs Meta Robyn: the short answer

Choose Google Meridian if you have weekly data split across many regions, Python skills and access to a GPU. Choose Meta Robyn if your data is national, your analysts work in R, and you want many candidate models to compare. If you have under two years of weekly history, choose neither yet.

That answer comes from reading both tools' documentation side by side rather than vendor summaries. Both are free, both are open source, and both are maintained by companies that sell advertising. Neither fact settles which one fits your team.

The table below compresses the documented differences. Every row is drawn from the official documentation of each project, opened in October 2026.

Why marketing mix modeling is back in marketing analytics

Marketing mix modeling is a regression of an outcome, such as sales or pipeline, against spend by channel and other drivers over time. It needs no user-level tracking. It needs aggregated weekly numbers, which is exactly why it has returned.

The method predates the web. Click-path attribution displaced it because clicks were cheap to collect and felt precise. That trade has reversed for most teams, and the reasons are documented rather than speculative.

The cookie did not die, and attribution broke anyway

The expected trigger never fired. On 22 April 2025 Google wrote that it would not roll out a new standalone prompt for third-party cookies in Chrome, keeping its existing settings instead. On 17 October 2025 it went further and announced the retirement of most Privacy Sandbox technologies, including the Attribution Reporting API, citing their low levels of adoption.

So third-party cookies survived and the privacy-preserving replacement for attribution did not. Attribution still degraded, because buyers now research in places that leave no click. That decline is covered in detail in the analysis of attribution and dark discovery, and this post will not repeat its numbers.

My position is that the cookie debate was always the wrong frame. Even with perfect cookies, click attribution only sees journeys that click. Aggregate models do not care how a buyer found you, which is their main advantage and, as later sections show, their main weakness.

The platforms built the tools themselves

Google announced Meridian on 7 March 2024 in limited availability. It opened Meridian to all marketers and data scientists on 29 January 2025, saying it had been tested with hundreds of brands and that over 20 certified measurement partners would support implementation. Meta's Robyn has been public on GitHub for longer and is now at version 3.12.0.

Google's 2024 announcement also cited Kantar research finding that 60% of US advertisers were using MMMs in 2023. Treat that as directional. It is a 300-person survey reported by a company launching an MMM product, so it tells you the category is mainstream, not how well it works.

It is worth stating the obvious conflict once. Both tools come from the two largest sellers of digital advertising. Both projects are open source, so the code is inspectable, but the defaults and the questions each tool makes easy to ask still reflect its author's business.

How Google Meridian works, from its documentation

Meridian's documentation describes a Bayesian model whose primary output is a posterior distribution of model parameters. In plain terms, every result arrives as a range with a probability attached, rather than a single best-fit number.

It samples that distribution with the No U Turn Sampler, a Markov chain Monte Carlo method the documentation itself calls compute intensive. JAX is the default backend, and the older TensorFlow backend is marked as deprecated.

Partial pooling is the feature that matters

Meridian's documentation says it uses partial pooling, which is hierarchical modelling across regions. Each geo gets its own media coefficients, but those coefficients borrow strength from each other. Trend, seasonality, adstock and diminishing returns are assumed common across geos.

The payoff is data volume. Google's example of 105 geos and three years of weekly data yields 16,380 data points, against 156 for the same three years at national level. For US data the guidance is to model roughly the top 50 to 100 DMAs by population, dropping the smallest geos first.

The documentation is candid that this benefit is reduced when there are relatively few geos. That sentence is the most important one in the Meridian docs for a B2B reader, and the section on B2B returns to it.

Priors, experiments and compute

Calibration in Meridian means turning experiment results into channel-specific ROI priors. The documentation calls incrementality experiments perhaps the strongest basis for those priors, and ships a builder that translates experiment results into calibrated priors. It also admits there is no single formula for that translation.

Google's related geo-experiment tool, Meridian GeoX, reached general availability in September 2026, according to trade coverage of Google's announcement. The same coverage reports a minimum of 10 geographic units for a single-cell design, with 50 or more recommended. That is secondary reporting, so verify it against Google's own pages before planning around it.

Meridian requires Python 3.11 or 3.12 and recommends at least 1 GPU. The FAQ is explicit on limits. It does not measure synergies between channels, it is designed for causal inference rather than prediction, and Google generally does not recommend running it at campaign level.

How Meta Robyn works, from its documentation

Robyn's documentation summarises its stack in one line. It uses an evolutionary algorithm for hyperparameter search, time-series decomposition for trend and season, ridge regression for fitting and a gradient-based optimiser for budget allocation.

Ridge regression shrinks coefficients toward zero without forcing any to exactly zero. Meta's stated reason for choosing it over lasso is that it is unrealistic for many media variables to have zero effect. The documentation also notes that ridge is equivalent to a Bayesian regression with a normal prior, which narrows the philosophical gap between the two tools.

Thousands of models, and the analyst picks one

The search is where Robyn differs. Nevergrad's TwoPointsDE algorithm explores adstock and saturation settings while minimising up to three errors: prediction error, a business error comparing share of spend with share of effect, and calibration error against lift tests. The documentation recommends at least 2,000 iterations and 5 trials.

The output is not one answer. Robyn returns a set of Pareto-optimal candidates, and the analyst's guide says Meta deliberately lets users choose the final model rather than picking automatically. Its stated reason is that a handful of statistical parameters cannot accurately select a model.

I think that is honest and also the tool's biggest governance risk. A model chosen by a person from a shortlist is a model that person must defend in a budget meeting. If your CFO asks why candidate 3 beat candidate 7, you need a documented answer that is not "it looked right".

Lag handling, calibration and maintenance

Robyn offers geometric and Weibull adstock, the transformations that spread a channel's effect over later weeks. The documentation states that Weibull PDF is strongly recommended for products with longer conversion windows, which describes nearly every B2B purchase.

Calibration accepts point estimates only, and the analyst's guide warns against using incrementality results whose confidence is not high enough. The documentation cites Analytic Edge, an implementation partner, finding that uncalibrated models show a 25% average difference to ground truth. That figure comes from a partner with no published method on the page, so treat it as directional.

Robyn's core is R. Its README describes the Python version as an LLM-translated beta that might encounter bugs, rewritten from R version 3.11.1. The latest tagged release on GitHub is v3.12.0, with feature documentation dated December 2024, and the model refresh feature is flagged as not yet tested thoroughly.

Marketing mix modeling data requirements, side by side

Both tools publish data guidance, and both hedge it. Meridian calls its guidance rough and directional because the true answer depends on the data. Robyn states a ratio, then softens it to a range. Read together, they agree more than the marketing around them suggests.

Meridian's most useful contribution is a worked example. A national model with 12 media channels, 6 controls and 8 trend knots has 26 parameters. Two years of weekly data gives 104 points, roughly 4 per parameter, which Google calls too low to estimate the model reliably.

Cut that to 10 parameters on three years of data and you reach about 15 points per parameter. Google's wording is that you "might be able to glean some directional information". That is the national-model ceiling in Google's own words, and it is far below what most dashboards imply.

Rows, columns and granularity

Robyn's analyst's guide sets the ratio at 1 independent variable to 10 observations, with rows about 7 to 10 times columns. It asks for at least two years of weekly data, and four to five years if only monthly data exists. For a model covering only six recent months, it recommends daily data.

Meridian asks for weekly data as best practice and at least three years if only monthly data is available, warning that monthly data can cause non-convergence. It also requires a complete dataset without missing values, and a KPI that is summable across geography and time. Media data and media spend must share the same dimensions.

The common ground is the useful part. Weekly granularity, two years minimum, few enough variables, and every week filled. If your data cannot meet those four conditions, the tool comparison is premature.

Marketing mix modeling for B2B: where both tools strain

Both tools were built by consumer advertising businesses and documented with consumer examples. Nothing in either documentation set is written for a business closing tens of deals a quarter. That is not a criticism. It is a scope boundary worth naming before you spend a quarter on implementation.

Run the arithmetic on a typical B2B setup. Take 8 paid and owned channels plus 4 controls, such as pricing changes, sales headcount, seasonality flags and a product launch. That is 12 variables, which at Robyn's 1:10 ratio needs 120 weekly observations, or about 2.3 years.

The same setup in Meridian's national counting adds trend knots. With 8 knots you have 20 parameters. Three years of weekly data gives 156 points, about 7.8 per parameter, which sits between Google's "too low" example and its "directional" one.

Your KPI is probably the wrong shape

Meridian requires a KPI summable across geography and time. Closed-won revenue is technically summable, but a single enterprise deal can be worth more than a quarter of smaller ones. One lucky week becomes the strongest signal in your dataset, and the model will try to explain it with whatever spend happened nearby.

The practical fix is to model a higher-volume, earlier outcome. Qualified pipeline created, sales-accepted opportunities or demo requests usually have enough weekly counts to vary smoothly. You trade proximity to revenue for statistical stability, and you should say so in every report the model feeds.

The lag problem compounds this. Robyn recommends Weibull PDF adstock for long conversion windows, which helps. No adstock curve, however, recovers a relationship that a nine-month sales cycle has scattered across three fiscal quarters. That is why I would never point an MMM at closed revenue for a business with cycles that long.

Geography rarely maps to B2B pipeline

Meridian's statistical edge is pooling across regions. B2B pipeline is often booked to the account's headquarters, not to where the buying committee saw your ads. A London-based champion at a company headquartered in New York becomes a New York opportunity in your CRM.

If your geo assignment is that noisy, the regional variation Meridian depends on is partly fiction. You are left with a national model, where Google itself recommends three years of data and where Meridian's distinctive advantage largely disappears.

That shifts the decision. For most B2B teams the honest comparison is a national Meridian model against a national Robyn model, and in that configuration the choice comes down to language, compute and whether your team prefers a posterior distribution or a Pareto shortlist. The commercial context for why so much of the journey happens before any tracked touch is set out in the piece on how AI assistants build B2B shortlists.

The minimum data checklist for B2B marketing mix modeling

Run this before installing either tool. Each row is drawn from one or both projects' documentation, translated into a B2B setting. A failed row does not mean never. It means fix the data first, because no modelling choice compensates for it.

Two rows deserve emphasis. Spend variation is the one most teams fail silently. Robyn's guide notes that if activity in a channel stayed constant for the whole period, the model can have difficulty, because a regression cannot attribute change to something that never changed.

Ground truth is the row that separates a model from a guess. A team with no holdout tests has nothing to calibrate against, and every channel coefficient becomes an untested opinion with a confidence interval. Designing those tests with a pre-agreed failure condition follows the same logic as setting kill criteria before a project starts.

Count your passes honestly. Six or more and a first model is worth building. Fewer than four and the better investment is a quarter of deliberate spend variation and two clean holdout tests, which will improve any future model whichever tool you choose.

Where this comparison is weakest

This is a documentation comparison. I have not run both tools on the same B2B dataset and reported the difference, which is the comparison that would actually settle the question. Nobody I could find has published that test with a stated sample, so any claim that one tool is more accurate for B2B is currently unsupported.

Documentation also lags code. Meridian's pages were updated as recently as September 2026 and are changing quickly, with agentic model-building and brand-signal features announced alongside GeoX. Some limits described here may be lifted in releases newer than this post.

The case for skipping this whole debate is strong for small teams. If your annual paid budget is concentrated in two or three channels, a disciplined holdout programme answers the budget question more directly than any MMM, at lower cost. Robyn's own analyst's guide states that the tool does not mean the model can run successfully without intervention, which is a fair warning for teams without a statistician.

The competing view I take most seriously is that B2B teams should run MMM anyway, on imperfect data, because a directional model beats a confident last-touch report. That is defensible when the output is labelled directional and used to choose experiments rather than to set budgets. It stops being defensible when a wide interval is presented as a point estimate, and that is how most MMM outputs I have seen reach a leadership deck. The general problem of outcome claims outrunning their evidence is examined in the piece on verifying case-study outcomes.

Frequently asked questions

How to start this month

Open a spreadsheet and build the weekly table before choosing any tool. One row per week for the last two years, one column per channel's spend, one for your chosen KPI and one per control. Count the empty cells and the columns. That count tells you whether marketing mix modeling is ready for you.

If the table passes, run Robyn first on a national model, because it is the faster route to a first result on data most B2B teams actually have. Move to Meridian when you can assign pipeline to regions honestly. If the table fails, spend the next quarter varying spend deliberately in one channel and running one holdout test, then build the table again. The wider context on what search and AI answers have done to measurable demand is in the piece on metrics for zero-click search.

Related reading