Marketing mix modeling explained: incrementality and measurement in the AI era

Attribution is no longer enough because it can only share out credit among the touchpoints it manages to see, and it sees fewer every year. Three shifts have weakened it: consent, buying journeys that run through AI answers, and platforms that each claim the same sale for themselves.

Consent. In Spain, analytics and advertising cookies need consent, and the cookie guide from the AEPD, the Spanish data protection authority, requires rejecting to be as easy as accepting. Anyone who declines drops out of user-level measurement. Chrome is not the cause, since it keeps third-party cookies: signal is lost through consent and through Safari and Firefox restrictions.

AI-shaped journeys. AI Overviews have been live in Spain since March 2025, and AI Mode in Spanish since October 2025: part of the comparison before a purchase now happens inside an answer, with no click to record. According to Pew Research Center (July 2025, a panel of 900 US adults), users clicked a traditional result in 8% of visits with an AI summary, against 15% without one; Google disputes the methodology.

Double counting. Every platform counts conversions with its own windows and models, and sees only its own part of the journey. If someone watches a video on a social network and later clicks a search ad, both can book the whole sale.

Marketing mix modeling (MMM), spelt marketing mix modelling in British English, is a statistical analysis that relates your sales to everything that could have moved them (spend by channel, prices, promotions, seasonality) to estimate how much each factor contributed. It works on aggregate series, usually weekly, and follows no one.

The model separates the baseline, meaning what you would sell anyway through brand, seasonality or market demand, from what each marketing lever adds. To do so it builds in two ideas: advertising keeps working after the day it runs (carry-over, or adstock), and each extra euro returns less than the one before (saturation).

A marketing mix modeling example: a skincare brand with an online shop invests in paid search, paid social, creators and email. In November, its social platform dashboard reports record sales. The MMM notes that November sells more every year, that a promotion was running and that branded search rose after the creator campaign, and it shares the credit among those factors.

Because it uses no identifiers or cookies, it does not depend on consent or on a click, and it can include channels attribution cannot see, such as television, outdoor or sponsorship. The trade-off is speed: it is refreshed monthly or quarterly, and it needs history.

The three techniques do not compete; they answer different questions. Attribution tells you which touchpoints preceded a purchase; MMM, how much each channel weighs in the total; and incrementality, whether a specific activity caused sales that would not otherwise exist.

What is incrementality in marketing? It is the causal effect of a marketing activity: the difference between what happened with the campaign and what would have happened without it. Only an experiment with a control group measures it directly. Side by side, the three techniques look like this:

• Attribution. Question: which ads did the buyer touch? Data: user-level tracking, with consent. Cadence: almost daily. Limit: it cannot see what cannot be tracked, nor separate out sales that would have happened anyway.
• Marketing mix modeling. Question: how much does each channel add, and where does the next euro work hardest? Data: aggregate series of spend, sales and external factors. Cadence: monthly or quarterly. Limit: it needs history and variation, and its estimates carry a margin of error.
• Incrementality test. Question: does this campaign cause new sales? Data: a control group or region with no exposure. Cadence: weeks per test. Limit: each experiment answers a single question.

MMM is no longer the preserve of big brands, because its main tools are now free and open source. Meridian, Google's MMM, has been available to every advertiser since 29 January 2025, according to the company's announcement. Robyn is Meta's open-source marketing mix model.

Free software does not make the model free. The work lies in gathering and cleaning the data, deciding which variables go in, setting reasonable assumptions and checking that the result holds up. Between Meridian and Robyn, the choice matters less than the team that will maintain it and the quality of the history.

It is worth remembering who publishes these tools: Google and Meta sell advertising. The advantage of open source is that anyone can inspect what the model does and on which assumptions. Keep code and data in your own systems, and check conclusions against experiments.

An incrementality test compares what happens with advertising against what happens without it, in comparable groups over the same period. It is the most direct way to learn whether a campaign generates new sales or merely claims ones that were coming anyway.

Two designs are common. A geo test pauses or runs the campaign in some regions and leaves comparable ones unchanged; in Spain, provinces or autonomous communities can serve as the units. A holdout sets aside a group of people who are deliberately shown no ads.

Incrementality testing vs A/B testing: an A/B test compares two versions of an ad or a page among people who do see them, and tells you which performs better; it is the working tool of conversion rate optimisation. An incrementality test compares exposure with no exposure, and tells you whether advertising there achieves anything at all.

The platforms have lowered the entry bar. For incrementality testing in Google Ads, Google said in November 2025 that Bayesian methods allow tests on budgets from about $5,000; for incrementality testing on Meta, the platform offers an incremental attribution setting. These are tools from the companies selling the space: useful in Google Ads and paid social, provided you check them against your own sales and confirm in your account which minimums apply in your market.

Triangulation means using attribution, MMM and experiments together, so that each covers the others' blind spots. That is what the IAB's best-practice guide Modernizing MMM, published in December 2025, recommends.

Each method has a job. Experiments calibrate the model with a causal reading; the model gives the overall picture, including channels with no click; and attribution guides daily adjustments. When they disagree, the disagreement is the most valuable part: if a platform reports a high return and the geo test detects barely any effect, the campaign is probably harvesting demand that already existed.

No method is infallible. In the IAB's State of Data 2026 report (more than 400 US decision-makers; reported by PPC Land in early 2026), up to 75% said attribution, incrementality tests and MMMs fall short on rigour, timeliness, trust or efficiency, and none believed all their paid channels were well represented in their MMM. That is why our marketing measurement service treats the model as a hypothesis to be tested, not a verdict.

An MMM is only as good as the data it is fed: spend by channel, real sales or leads, and external factors. The same IAB report puts it this way: measurement success is architectural, and AI adds no value to MMM until data quality is fixed.

MMM itself needs no cookies, but attribution, experiments and platform bidding do rely on sound data collection. That is the work of analytics, first-party data and server-side tracking. In 2026, the baseline in Spain includes:

• Consent Mode v2: it passes four consent signals to Google. Since March 2024, European Economic Area traffic without those signals has not fed audiences or remarketing.
• Enhanced conversions: web and leads are now managed as a single setting in Google Ads.
• Google tag gateway: it serves Google tags from your own domain, and has been generally available since May 2025.
• Meta Conversions API: it sends server-side events alongside the pixel, with deduplication.
• Reconciled sales: orders, bookings or leads taken from your CRM, alongside a calendar of prices, promotions and seasons.

Your brand's presence in AI answers can now be measured with official tools, but only as visibility, not as sales. It is a brand indicator, alongside branded search, not a channel to credit with conversions.

Search Console provides generative AI performance reports for AI Overviews and AI Mode, available to every site since 31 August 2026: they show impressions, but no clicks or CTR. Bing Webmaster Tools launched its AI Performance report in public preview in February 2026, showing citations of your pages in Copilot.

Advertisers are already moving: in the IAB's 2026 Outlook Study (28 January 2026, more than 200 US brands and agencies), 86% said they are changing, or expect to change, their measurement because of AI, and 48% already measure brand visibility and citations inside AI platforms.

The paid side is a separate matter. As of October 2026, ChatGPT has shown ads in Spain since 24 August 2026, on the Free and Go plans only and on a contextual, non-personalised basis; ads inside AI Overviews are not available in Spain, and ads in AI Mode are being tested only in the United States. Our ChatGPT Ads guide covers the detail; organic visibility is the job of GEO.

An MMM needs a continuous history of spend and sales, usually weekly, long enough for the model to see every season more than once, and with real variation in spend. If you always invest the same in the same channels, the model has nothing to learn from.

Seasonality makes it both more necessary and more demanding. In hospitality, travel and tourism, sales rise in season with or without advertising, and only a model that knows the pattern can isolate what the campaign adds; the same holds for consumer brands and ecommerce with sales periods and promotions.

It is not the first step for everyone. Start with properly configured attribution and a couple of simple tests if:

• Your spend history is short or has gaps.
• One channel takes almost the whole budget.
• Spend is so small that its effect is lost in the normal ups and downs of sales.
• Nobody is going to use the result to decide the budget.

The useful output of an MMM is not a report but a better-informed decision. The model estimates a response curve for each channel, showing where the next euro still pays and where saturation begins.

Those curves feed scenarios: what might happen if you move part of the budget from one channel to another. An honest scenario is expressed as a range: it is an estimate, not a guarantee. Before a major budget line moves, the change is tried at small scale with a test, and the result goes back into the model to calibrate it.

For this to work, it has to speak the language of whoever decides: real sales, contribution by channel and uncertainty on a single page, as part of the digital strategy. It is one part of the shift we describe in the future of digital marketing. If you want to know whether your data can support a model, tell us about your case.