Should each country have its own MMM?

For brands with very different country dynamics, yes; for brands with similar dynamics across operating countries, hierarchical pooled is better. The decision depends on the variability of effect sizes across countries: high variability favors separate models, low variability favors pooled.

How do we handle countries with short data history?

Through the hierarchical structure. Newer markets get priors derived from established markets, with the data updating the priors as it accumulates. This produces usable inference for new markets immediately rather than requiring 52 weeks of dedicated history per market.

Which AI platforms matter in international markets?

ChatGPT and Perplexity in most Western markets; Claude for technical audiences; Gemini globally with strong presence in Google-strong markets; Baidu Ernie and Qwen in China; Naver HyperCLOVA in Korea; Sarvam and other India-specific models in India. The weighting depends on the brand's audience in each country.

How often should multi-country MMM refit?

Country-level refits monthly with full hierarchical refit quarterly. The country-level refits update country posteriors against new data; the hierarchical refit allows the cross-country structure to evolve. This is more frequent than typical single-country MMM but is justified by the operational complexity of managing multiple country budgets.

How to Run Multi-Country MMM

The Multi-Country MMM Problem

Global brands operate in many countries with different market dynamics, different channel availability, and different AI platform adoption. A separate MMM per country is operationally heavy and loses the cross-country learning that should improve identification. A single global MMM aggregates across heterogeneous markets and loses country-specific actionability. The answer is hierarchical MMM with country-level coefficients that partially pool toward global priors.

Step 1: Build the Country List

Inventory the countries where the brand operates with sufficient data history (at least 52 weeks of weekly data per country). Group countries by similarity for hierarchical specification: developed markets with similar channel mix versus emerging markets with different channel mix. The grouping matters because the partial pooling will shrink country effects toward group-level rather than fully global means.

Step 2: Choose the Hierarchical Structure

Three patterns. Country-specific: separate MMM per country, no pooling. Global pooled: one MMM with country dummies, full pooling. Hierarchical: separate country MMMs that share priors via a hyperprior structure. Hierarchical is the modern recommendation; PyMC-Marketing and custom Bayesian implementations support it natively, while Robyn and LightweightMMM require workarounds.

Step 3: Handle Channel Availability Differences

Some channels are not available in some countries (TikTok not in India until recently, retail media networks specific to certain markets). The MMM spec needs to handle missing channels gracefully: either set the spend to zero in countries without the channel, or use the hierarchical structure to only model available channels per country. Both work; the choice depends on framework support.

Step 4: Weight AI Visibility by Country

AI platform adoption varies by country. ChatGPT is dominant in most Western markets; Claude has higher penetration in technical audiences; Perplexity has uneven global adoption; Baidu and Qwen are relevant in China; Naver and HyperCLOVA in Korea. The AI visibility variable should be weighted by country-specific platform usage rather than using a global weighting.

Step 5: Run the Refit With Country-Level Holdouts

Validate the hierarchical model with country-level holdout: hold out the most recent eight weeks from each country and check MAPE on the holdout. Country-level holdout MAPE should be comparable across countries within the group; sharp differences indicate either spec issues or genuinely different market dynamics that the hierarchical structure should accommodate.

Step 6: Translate to Country-Level Budget Recommendations

The optimization step needs to handle country-level budget constraints (regional budget caps, country team allocations) alongside the global constraint. Most teams run a two-stage optimization: global allocation across countries, then within-country allocation across channels. Some run a joint optimization, which is more rigorous but operationally complex.

How Presenc AI Helps

Presenc AI provides AI visibility data segmented by country with country-specific platform weighting. The data supports hierarchical MMM specifications and accommodates country-by-country differences in AI platform adoption. For brands operating in many markets, Presenc closes the AI visibility data layer for the multi-country MMM stack.