MMM 101 in Turkey: Bayesian Marketing Mix Modelling for Turkish Brands

Introduction: Intuition or Evidence?

In 2026, marketing budgets at Turkish brands sit in the middle of a paradox. On one side are channel intuitions built up over years: "TV always works for us," "we keep growing performance digital month over month, so it must be working," "OOH has to run during the holidays." On the other side is a measurement world that is starting to crack: third-party cookies are deprecated, iOS 14.5 ATT (App Tracking Transparency) broke the attribution chain, and GA Consent Mode V2 silenced ad cookies in consent-denied mode. Last-click attribution no longer tells you where the budget actually went; the "last click wins" logic measures click order, not real contribution.

The conclusion is straightforward: in 2026, before a CMO can tell a CFO "we spent 50 million on TV, and here is the return," there must be a model that genuinely makes the case provable. The name of that proof is Marketing Mix Modelling (MMM) — the bridge between intuition and financial discipline. This post explains what MMM is, why the Bayesian approach matters, and why the specific dynamics of the Turkish market require a different model — at a clarity that CMO, CFO, and Analytics Lead can read together.

What Is MMM?

Marketing Mix Modelling (MMM) is a statistical method that decomposes the contribution of marketing spend to sales. In its classical form, it is a regression model: the dependent variable is sales (or some other KPI — applications, sign-ups, visits), and the independent variables are weekly spend per channel, price, distribution, macro variables, and seasonal factors.

MMM is not new. Its origins trace back to the 1960s, when Procter & Gamble first tried to measure the effects of television advertising. It grew up as the marketing version of econometrics; in the FMCG (fast-moving consumer goods) golden era, it was nourished by panel data. In 2017, Google published "Bayesian Methods for Media Mix Modeling with Carryover and Shape Effects" (Jin et al. 2017), and the field jumped a generation: the model was no longer a single point estimate but a tool that also reported its uncertainty.

The point where MMM diverges from last-click is fundamental: it does not look at user-level data. It needs neither cookies nor user identifiers. It works with monthly or weekly aggregate data — total channel spend, total impressions, total GRPs. This structure aligns naturally with data protection regimes like KVKK (Turkey's data protection law, similar to GDPR); no PII (personally identifiable information) is required. In a cookieless world, last-click is dead — MMM keeps working.

The model output is clean: each channel's incremental contribution (how much sales would drop if that channel were absent), the channel's saturation point (if you double the budget, do sales double, or are marginal returns diminishing?), and adstock (how the effect of a TV burst decays gracefully across subsequent weeks).

The Bayesian MMM Difference

Classical regression gives you a single number: "TV contribution is 22.4%." Is this number correct? What is the margin of error? The classical model cannot say so elegantly; at best it produces a confidence interval, and the interpretation of that interval is counter-intuitive for most marketing leaders.

Bayesian MMM instead reports a posterior distribution — the probability distribution of a parameter after the data has been observed. Not a single number, but a distribution. It says "TV contribution is most likely 22.4%, with a 95% credible interval between 18% and 27%." The strength of this language is that you can tell the CFO "of the 100 million TL we put into TV, there is a 95% probability that the real contribution is between 18 and 27 million." Intuitive, directly tied to a decision.

Bayesian's second superpower is the prior concept. The model does not have to learn from scratch; sector knowledge, prior campaigns, or market dynamics enter the model as priors. In a market like Turkey with a tradition of 4-week TV bursts, you give a sensible prior for how TV adstock should behave; the model updates this prior with the data and arrives at the posterior. That is "make existing knowledge the starting point" instead of "learn from zero" — which makes the model far more robust when data is thin (short time series, low channel variance).

The runtime method for Bayesian models is MCMC (Markov Chain Monte Carlo — the model runs four parallel scenario chains and produces a probability distribution for each parameter). At MMMytics, we run 4 chains, 1,000 burn-in steps + 1,000 sampling steps; we verify convergence with the R-hat ≤ 1.1 rule. If R-hat is above 1.1, the model is not trustworthy — we move to a different model or recalibrate the priors.

Turkey-Specific Market Dynamics

Applying a global MMM library to a Turkish brand as-is — that is, running it with default settings — will almost always produce the wrong answer. The reason is that the Turkish market diverges from the global average on several fundamental behaviours:

Religious and national calendar. Ramadan is the most aggressive sales period of the year for FMCG and retail; the 17:00–19:00 broadcast slots before iftar are a primed market. The Eid al-Fitr and Eid al-Adha holidays fragment weekly data — one week with 4 working days, one with 6. Black Friday landed in Turkey between 2018 and 2020; today it is Q4's leading peak for e-commerce. The "Christmas peak" assumption baked into a standard MMM library does not hold in Turkey; seasonality priors must be built around the local calendar.

FX and macro variables. Central Bank (TCMB) FX rates and CPI clearly influence the sales behaviour of a Turkish brand. In automotive, you see Q4 FX-driven panic buying; in durable goods, sales explosions after high monthly inflation; in retail, end-of-month deferral waves. If these variables do not enter the model as control variables, the sales lift caused by FX movement is wrongly attributed to a channel. At MMMytics, weekly FX + CPI is automatically pulled from the Central Bank, and the model isolates these as macro controls.

The TV burst tradition. Turkish advertisers prefer 4-week heavy TV bursts: 4 weeks on-air, off-air, then back on-air. Not continuous low frequency, but waving high frequency. This puts pressure on the classical adstock parameter; carryover must be sensitive to a 4-week periodicity. The default Meridian adstock prior is calibrated for the US/EU markets — calibrating the TV adstock prior for Turkey is a custom job.

Post-cookie independence. With GA4 Consent Mode V2 in consent-denied mode, the third-party ad cookie is dropped; the Meta Pixel attribution chain breaks. Because MMM works on aggregate data, it is unaffected by this rupture — quite the opposite, the independent measurement power of MMM grows in this new world. When the cookie dies, MMM lives.

KVKK and auditability. For Turkish brands, data cannot leave the EU region. Global MMM SaaS products running in US regions conflict with KVKK Article 9 (cross-border data transfer). MMMytics keeps all Turkish customer data in the EU region; every data processing step is logged as a KVKK Article 12 record in the audit log.

How MMMytics Solves It

MMMytics is built on the open-source Google Meridian framework. The reason Meridian is preferred is that it is not a black box. The academic paper (Jin et al. 2017) is published; the mathematical foundation is auditable; community contributions keep evolving it; it is the reference implementation of the Bayesian MMM literature. Unlike closed systems that arrive with "patented proprietary algorithm" claims, you can audit the decision-making process of a model built on Meridian.

On top of Meridian, MMMytics adds four layers:

1. Turkey seasonality priors out of the box — Ramadan, Eid al-Adha, Black Friday, school start, Q4 FX wave come configured as the model's starting parameters. The customer does not calibrate from scratch.
2. Central Bank FX + CPI feed — weekly macro data flows automatically as a control variable in the model. The customer does not upload CSV files manually.
3. Sector templates — for Automotive, FMCG, and Retail, a default channel list + adstock prior + saturation curve come as starting parameters. For sectors not in the list, a Custom Sector option starts the calibration.
4. Command Board → CEO Brief PDF — the model output is converted into a one-page Turkish-language summary (CEO Brief) for non-analysts; the live Command Board dashboard compares budget scenarios like a scenario sandbox; the Budget Optimiser derives 10%/20%/30% budget scenarios live from the same model's posterior.

The typical flow: weekly sales + channel spend uploaded to the Data Hub, sector template chosen in Model Studio, the MMMytics Engine runs the model with MCMC (10–30 minutes), Channel Intelligence reads channel-by-channel incremental ROI, the Budget Optimiser optimises the next quarter's budget, and the Report Generator produces the CEO Brief PDF.

CTA

If you want to see Turkey's true marketing incremental ROI, Request a Demo — we will set up your sector template live. For detailed documentation and methodology, visit the About page. For pricing plans, see the Pricing page.