MMM for the Automotive Sector: TV GRP, Dealer Network & Geo Hierarchical Modelling

Introduction: The Automotive Attribution Puzzle

If you lead marketing at an automotive distributor, you have probably felt that you are fighting alone in the attribution world. Other sectors can say "the customer clicked, bought, ROAS is 4"; in our world it does not work that way. A person's decision to buy a new vehicle takes 3 to 6 months. During that period, dozens of touchpoints chain together — a TV ad, a programmatic display, a YouTube test-drive video, a Google search, an OOH billboard, an Instagram showroom story, the dealer's SMS reminder, the showroom visit, the test drive, and the salesperson's follow-up call. Then the invoice is issued; the last-click system gives all the credit to the digital channel that received the final click.

The MMMytics team carries more than 20 years of marketing experience in automotive and FMCG; a significant share of that time was spent at major media agencies managing the investments of Turkish brands. With our distributor customers, we have lived the same scene many times: the panel says "Google sold 1,200 cars this quarter"; in reality, half of those 1,200 people would never have walked into a showroom without the upper-funnel TV burst. Last-click does not show this. This post explains the automotive interpretation of Marketing Mix Modelling (MMM) — why you need a geo hierarchical model, how we feed TV GRP data into the model, and the actual incremental impact a budget shift produces.

Automotive Sector Specifics

Five core behaviours separate automotive from other sectors; each one changes the attribution model.

TV GRP is still dominant. In new model launches, TV continues to take 35–45% of the budget. The Turkish consumer's perception of a new model is still built around the evening news, prime-time drama breaks, and weekend sports broadcasts. Programmatic TV and OTT are taking share — but classical linear TV has not given up the volume lead. The average mix observed with distributor customers: 40% TV + 30% digital + 20% OOH + 10% dealer local media. You cannot measure this mix with last-click; TV's effect on showroom traffic does not leave a click.

Geographic heterogeneity. Istanbul, Ankara, and Izmir are one market; the other Anatolian cities are a different market. Dealer density, digital consumption, and financing penetration in Istanbul cannot deliver the same channel performance in Konya, Trabzon, or Gaziantep. Running automotive MMM as a single national model — pooling all cities into one bucket — distributes the contribution of large cities to the rural areas, and vice versa. The wrong answer comes out on both sides.

Dealer performance variance. You cannot expect two dealers of the same brand to deliver the same result with the same investment. Showroom location, sales-team experience, local CRM discipline, and even whether the storefront is clearly visible from the street, all create differences. The model must see this variance, so that "your dealer X is performing badly" is not the conclusion before "the national campaign reached low in region X" is checked.

OOH is local. Billboards above the highways touch showroom traffic directly — but only in their physical neighbourhood. The billboard on Eskişehir Yolu in Ankara has no effect on the dealer in Adana. In automotive, OOH's geographically short-range effect must be represented in the model as a geo variable; otherwise the wrong conclusion "OOH does not work nationally" is drawn.

Q4 FX-driven purchasing. The Turkish automotive market has its own season: weekly Central Bank FX moves trigger "let's buy before the rate climbs further" panic purchases. In Q4, this tendency compounds; 30–35% of annual sales can be compressed into the last quarter. This sales explosion is the effect of the macro variable, not the media investment. If you do not enter FX into the model as a control variable, you will attribute the Q4 sales explosion to a channel; you will also shift the budget to the wrong place.

Geo Hierarchical Bayesian Model

A single national model does not work in automotive. The solution is the geo hierarchical Bayesian model — a multi-level Bayesian structure where city-level parameters are connected to a national-level structure.

The intuitive version: we partition Turkey into 12–15 large markets (3 metropolises + the larger Anatolian cities + the East/Southeast cluster). Each market has its own channel-performance parameters. At the same time, all markets are linked to a national "brand average," because a brand's TV ad essentially runs nationally. The Bayesian structure resolves these two levels as follows: each city learns its own posterior distribution but starts from the national-level posterior. So if the national-level prior says "TV adstock around 4 weeks," Konya inherits that prior — then, as Konya's own data arrives, it updates its posterior toward its own reality. Cities with thin data borrow strength from the national structure; data-rich cities learn their own model.

The practical benefits are twofold:

1. National vs dealer-specific traffic separation becomes clear. The Mecidiyeköy dealer in Istanbul receives September showroom traffic from two components: Istanbul's share of the national TV burst + the Mecidiyeköy dealer's local SMS campaign + local OOH. The geo hierarchical model separates these two components. The dealer manager gets the answer "how much impact did I generate in my own territory"; the distributor's marketing leader gets the answer "in which city did my national campaign work harder."

2. The model keeps working for thin-data dealers. A dealer selling 200 cars per year produces a noisy weekly sales series; if you fit a model to this dealer alone, posteriors come back with wide credible intervals, and you cannot make useful decisions. In a hierarchical structure, this dealer borrows strength from the national level; its estimates remain reasonable.

Because the Google Meridian framework supports geo-specific priors, we build this structure naturally. In the MMMytics Engine, "geo hierarchical mode" is enabled by default for the automotive sector template; your analysis starts seeing 12–15 markets instead of a single national view. You invoke this structure with one click in Model Studio; the result file reports each market's own posterior — in a format your dealer managers can read as "this is my market."

Another practical benefit: when a campaign is unevenly distributed across cities, the geo hierarchical model sees it. When the distributor plans the national campaign, large agencies typically allocate the budget by population share — but sales share is not always equal to population share. You may have a model with high sales share in Anatolia; the national TV plan there is already saturated, and underinvested in the metropolises. The geo hierarchical posterior surfaces this misalignment numerically — making it possible to read a city-by-city ROI table.

How TV GRP Enters the Model

The crucial part of the automotive model is the TV GRP data. GRP (Gross Rating Point) is the total reach-weighted score a TV campaign generates against a target audience. In Turkey, Kantar Media (a TIAK member) provides weekly panel data — how many GRPs you collected by channel, day, daypart, and target audience.

We apply three transformations on the way into the model:

1. Reach-frequency curve. The same GRP could come from two different reach profiles: 5 million people once each, or 1 million people 5 times each. The effect is different. We feed not "raw GRP" but GRP decomposed into reach and frequency components. In a new-model launch, reach matters; in reminder campaigns, frequency matters. The model sees the distinction.

2. Adstock parameter. The effect of a TV ad does not end in the broadcast week; it carries into subsequent weeks. Adstock models how a campaign's effect decays week over week. In Turkish automotive, the observed half-life averages 3–4 weeks — slightly shorter than the European market, because the 4-week burst tradition keeps stacking the "new" message on top of itself. The default Meridian adstock prior is calibrated for the Turkish market; we do not search parameters from scratch.

3. Saturation curve (diminishing returns). As GRPs increase, the sales contribution does not grow linearly; beyond a point, every additional GRP brings less incremental sales. This saturation point is critical. It is common to observe that when a distributor's 3,500 GRP monthly plan is reduced to 3,000 GRP, sales remain the same — the curve was already saturated, the spend was excessive. In the MMMytics Channel Intelligence dashboard, each channel's saturation curve is drawn live; "from this point, diminishing returns begin" can be read numerically.

Anonymous Case Study: "Brand A"

A distributor in the top 5 of the Turkish automotive market — call it "Brand A" in this post, anonymous. In Q3 2025, it was running a typical automotive mix: 40% TV + 30% digital + 20% OOH + 10% dealer local media. Around 40,000 annual vehicle registrations, mid-budget segment with a sedan + SUV portfolio. The distributor's CMO saw, in the last-click panel, that digital was dominant; nobody had a clear answer to the question of TV's actual contribution.

We set up a geo hierarchical Bayesian MMM. Into the model went 104 weeks of sales (2 years), Kantar GRP data, digital spend (Google + Meta + DV360), OOH spend (Posterscope inventory report), dealer-specific media, Central Bank FX, CPI, and seasonality variables. MCMC with 4 chains, R-hat 1.05 (clean convergence), result in 22 minutes.

The findings were clear:

• TV overspend by 15%. The frequency side was overloaded; the incremental contribution of the 4th and 5th repetition was near zero. The reach saturation point was around 65 GRP/week; weekly 80+ GRP demolition budgets did not generate extra returns.
• Digital underinvested. Performance digital was 25% away from its saturation point, still on the increasing-returns curve. Each marginal 1 million TL of digital produced 18% measured incremental contribution.
• OOH strong in metropolises, weak in rural areas. The geo hierarchical model surfaced this clearly; 30% of the national OOH budget was going to non-metropolitan segments where the incremental effect was not statistically distinguishable from zero.
• Dealer local media high lift. Low absolute budget but high return; we recommended scaling it up.

Our recommendation: a 1 million TL shift in the next quarter — from TV frequency to digital + dealer local media. We did not cut the TV budget; we just lowered the frequency GRPs that exceeded the saturation point and redirected those savings to channels with higher marginal returns. Four weeks after the shift, incremental sales rose 23% — the kind of progress sitting around the sector average that we knew could be realised. The CMO could now tell the CFO "15% of our 100 million TV budget was idle; shifting it produced a 23% sales lift." This number was cross-validated with the showroom traffic figures: weekly visitors across the dealer network were up 18%, test-drive bookings up 21% — model output and showroom reality spoke in the same direction.

Note: this case represents a pattern we have observed in the Turkish automotive sector; the specific customer name has not been disclosed. The narrative uses a scenario close to sector averages.

CTA: Let's Build Your Sector's Model Live

Building MMM for an automotive distributor is a process that starts from a ready template and is then calibrated with your brand's actual data. At MMMytics, the automotive sector template — geo hierarchical mode on, TV GRP integration ready, FX control variable flowing automatically from the Central Bank, seasonality priors built around the Turkish calendar — comes by default. When you reach out via Request a Demo, we set up a live model in a 30–60 minute session with your data; we read your channel-by-channel incremental ROI together in the Channel Intelligence dashboard; we evaluate next quarter's budget scenarios together in the Budget Optimiser.

To learn about our methodology and team, see the About page; to see plans, the Pricing page. The time to measure real incremental ROI in automotive is now.