The Meridian Mirage: Why Open-Source MMM Is Quietly Wrecking UK Media Budgets

Over the past eighteen months, UK marketing directors have staged a quiet rebellion against multi-touch attribution. Google Analytics 4's data-driven model lost the room, server-side CAPI setups proved fiddly, and third-party cookies finally completed their drawn-out funeral. In their place stepped open-source Media Mix Modelling (MMM)—spearheaded by Meta’s Robyn and, more recently, Google’s lightweight Bayesian library, Meridian.

Suddenly, econometrics is no longer an eye-watering £60,000 project outsourced to a boutique consultancy in Mayfair. It is a set of Python notebooks run by an in-house performance lead on a Tuesday afternoon. On paper, it sounds like absolute enlightenment: privacy-safe, macro-level media evaluation that measures the holistic sales impact of every pound spent.

There is just one glaring problem. Without rigorous calibration, open-source MMM is not uncovering marketing truth. It is simply building mathematically complex alibis for your most bloated, non-incremental campaigns.

The Baseline Hallucination: Why Branded Spend Looks Untouchable

Every MMM relies on a foundational calculation: separating your organic baseline sales (what your brand would sell if you turned off every paid channel tomorrow) from the marginal lift generated by paid media. Get the baseline wrong, and the entire house of cards collapses.

When in-house teams feed two years of weekly spend and revenue data into Meridian without strict variable controls, the model commits what econometrists call endogenous bias. In plain English: it mistakes coincidence for causation. When demand naturally spikes during Black Friday or payday weekends, brands increase their branded search, retargeting, and Performance Max budgets. The algorithm sees ad spend rise at the exact same moment transactions surge, and concludes that the ads caused the volume.

We recently audited an omnichannel UK retail account whose internal Meridian model claimed their Google Brand Search was driving a staggering 11.4x ROAS, recommending they double the budget. When we forced an isolated holdout test, the genuine incremental ROAS was 1.15x. The model was giving £25,000-a-month credit to paid clicks that customers were executing merely as a navigational shortcut.

The 'House Edge' Baked into Default Priors

Bayesian modelling works by combining prior beliefs (priors) with observed data to reach a posterior probability. It is an exceptional mathematical framework, but it comes with a major catch: whoever sets the priors controls the output.

Google built Meridian to be accessible, which means it comes pre-packaged with default assumptions about saturation curves, adstock decay rates, and channel effectiveness. Unsurprisingly, those defaults assume high carry-over value for persistent, algorithmic reach products—namely YouTube, Demand Gen, and broad Performance Max setups.

If your internal team runs these open-source models out of the box without heavily penalising paid capture channels, you are letting the ad network grade its own homework. It is the econometrics equivalent of asking a casino pit boss to inspect the roulette wheel.

The Counter-Weight: Grounding Python Scripts in Geo-Lift Reality

Does this mean MMM is dead on arrival? Absolutely not. It means that uncalibrated MMM is dangerous fiction. The only thing that turns an open-source regression model into an actionable budgeting tool is continuous, real-world incrementality experiments.

• Matched-market geo testing: Darken your upper-funnel Meta or PMax spend across the Midlands or Yorkshire for three weeks while keeping spend flat in the North West and South East, then measure the actual revenue divergence.
• Calibrated Bayesian priors: Take the empirical lift percentage from that real-world dark test and feed it directly into Meridian as a hardened prior constraint.
• Strict unbundling: Never lump branded search, dynamic product retargeting, and prospecting into single channel buckets in your data prep. Treat high-intent capture as a separate operational cost.

When you anchor your statistical curves to empirical lift tests, the model ceases to be an academic guessing game. It becomes a calibrated financial map.

A Sensible Operating Rhythm for UK Growth Teams

If you are using or evaluating Meridian, Robyn, or lightweight MMM packages in 2026, stop treating the outputs as gospel truth handed down from Mount Sinai. Shift your internal team to this practical workflow:

1. Run the model quarterly, not weekly: MMM is a strategic capital allocation tool, not a daily bidding dashboard. Evaluating weekly adstock swings on small data sets introduces noise that ruins creative decision-making.
2. Budget 5% to 10% of total media for holdout testing: You cannot measure incrementality for free. Allocate a portion of your quarterly ad spend specifically to matched-market holdouts and regional blackout tests.
3. Triangulate with customer self-attribution: Pair your macro MMM outputs with simple, post-purchase surveys ('How did you first hear about us?'). When the econometric model aligns with what living, breathing buyers report, you can scale budgets with genuine conviction.
4. Challenge the baseline relentlessly: Calculate what your business would generate organically if your entire paid estate was turned off for 14 days. If your MMM calculates a baseline that feels suspiciously low, your agency or team is over-crediting paid channels.

Open-source data science tools have brought sophisticated econometrics to brands spending £50k a month, not just £5m. That is an enormous competitive advantage. But mathematics without real-world friction is just expensive theatre. Calibrate your model with cold, hard incrementality tests—before it convinces your board to spend thousands chasing conversions you already owned.