Media Mix Modeling for Campaigns Spanning Search, Social, and AI Channels

Media mix modeling is having a moment, and the reason has nothing to do with modeling technique and everything to do with the collapse of the tracking layer that made attribution feel solved for the last fifteen years. Third-party cookies are going away, iOS tracking limits have gutted mobile measurement, and now AI assistants are inserting themselves between the ad and the purchase in ways no pixel was built to see. A majority of brands, 71%, have already cut their reliance on user-level data. This is an industry-wide response, not a fringe response from a few privacy-conscious teams. It's the industry adjusting to a fact that was always true and just got harder to ignore: last-click was a shortcut, not a measurement. As the CPO at Kleene.ai put it, last click "places all the value on the final interaction before a conversion, which is clearly not correct."

What MMM needs to work

Media mix modeling runs on three kinds of input: historical marketing spend across every channel a brand uses (paid search, paid social, TV, radio, out-of-home, email, affiliates), sales data over that same period, and external demand variables like seasonality, pricing changes, or macroeconomic shifts that move sales independent of any ad. Kleene.ai's own guidance puts the floor at 24 months of data, and modelers generally want more than that if a business has it.

This is an econometric technique, usually Bayesian, not a dashboard pulling numbers from ad platforms and not a spreadsheet crediting whichever channel touched a customer last. The model's job is to isolate what each channel actually contributed to sales, separate from everything else moving at the same time. That includes carryover effects, since a dollar spent on TV this week might not show up in sales until next month, and it includes drawing a line between the baseline demand that would have existed with zero media spend and the incremental lift media created on top of it. The output is a channel-by-channel attribution of sales, a set of budget allocation recommendations, and scenario simulations that let a planner ask what happens to revenue if search spend drops by a meaningful amount and social rises to cover it.

How AI has changed MMM's speed, cost, and accessibility

For most of the last decade, MMM belonged to companies that could afford Nielsen or Kantar's enterprise engagements: expensive, slow, and dependent on statisticians who understood the math well enough to build and defend a model. That barrier has come down, and AI is the reason.

Bayesian machine learning fits models faster and more accurately with less data than the older regression-based approaches required. Markov Chain Monte Carlo sampling, the computational method behind modern Bayesian inference, lets modelers train models with tens of thousands of parameters where the old approach maxed out in the hundreds. Getting two years of clean spend and sales data into a single usable table used to eat most of a project's timeline; modern platforms have shrunk that to a smaller piece of the work. Scenario simulation has changed shape too: instead of three point estimates dressed up to look precise, AI-native models run large numbers of scenarios and hand back a range of plausible outcomes with the uncertainty attached, which is a more honest way to present a forecast. Some AI-powered platforms now build, test, and validate several competing model structures at once and pick the one that fits best, rather than asking an analyst to hand-tune a single specification.

Open-source tools have done as much to widen access as any commercial platform. Google's Meridian and Meta's Robyn are both free, and combined with self-service SaaS products, they've put MMM within reach of mid-sized companies that have clean historical data but no in-house econometrics team. Google has said it plans to roll Meridian out globally starting in early 2025, with access to Search and YouTube data baked in, which is a meaningfully different posture from the walled-garden approach that's defined how platforms treat their own performance data for years.

The budget landscape MMM now has to cover, including where AI channels sit today

Worldwide media ad spending is projected to reach roughly $1.17 trillion in 2026. That's the total pool MMM has to account for, and the mix inside it keeps shifting.

Paid search still leads digital budgets at around 40% of the total, with global PPC spend projected at $306 billion in 2026 and growing at 11% a year. Social is near 32%, with one major market's spend alone expected to top $121 billion in 2026. spend alone expected to top $121 billion in 2026. Display holds about 18%, and programmatic buying now handles 91.5% of all digital display, meaning almost none of it is bought the old-fashioned way anymore. Video and CTV account for roughly 10%, with a leading national market's CTV spend at $33.35 billion in 2025 and projected to grow about 14% to somewhere near $38 billion in 2026. CTV spend at $33.35 billion in 2025 and projected to grow about 14% to somewhere near $38 billion in 2026. Retail media, a $58.79 billion channel in 2025, is big enough now that it's starting to demand its own dedicated MMM treatment rather than getting folded into display.

AI chatbot ad spend, by contrast, is still small in absolute terms. It's growing, but off a small base. The bigger near-term story is AI search-adjacent advertising, which is expanding rapidly in 2026. That's the category actually disrupting how search MMM inputs get built today, because it sits inside the same auction infrastructure as traditional search while behaving nothing like it downstream, and its scale is growing rapidly even if precise figures remain unsettled.

Why conversational AI channels produce no usable inputs in a standard MMM data pipeline

A standard MMM pipeline needs three things at minimum: spend data, impression data, and sales outcomes that line up in time with the media activity that (allegedly) drove them. Search and social supply all three without much friction. AI conversation channels supply none of them in a form the pipeline recognizes.

No cookie fires inside a chat session. In most current formats, there's no click-through to a landing page. There's no platform-reported ROAS comparable to what a marketer pulls from a Search or social console, and there's no page-level pixel to record an impression, because there's often no page. One source close to this problem said: "When an assistant can research, compare, and recommend without a click, the standard last-click model breaks. A meaningful share of the influence an ad has on a purchase now happens inside the conversation, invisible to a pixel that only fires on a landing page."

That's the structural gap. Influence builds up inside the session itself, but the conversion, if it happens, happens somewhere else, later, disconnected from the moment the assistant actually swayed the decision.

Conversational intent signals versus keywords and impressions

Search has always worked off keyword signals, which are a proxy for intent that the advertiser infers and matches probabilistically. A user typing "best running shoes for flat feet" is signaling something, but the advertiser is guessing at what, matching against a keyword list, and hoping the guess is close enough. Social works off impression signals, which are even thinner: a user saw an ad while scrolling. No revealed preference, no declared need, just exposure.

A conversational intent signal is a different kind of signal. The user has stated, in their own words, what they need, often with specifics about budget, timeline, or what they're comparing it against. That's a live, first-person declaration of purchase intent, not an inferred proxy. It's a live, first-person declaration of purchase intent, and the assistant handling the conversation has access to everything the user said before and after that moment, in a way cookie-based targeting never came close to. The intent is explicit rather than guessed at, the moment of influence is identifiable within the session rather than smeared across a multi-day path, and the surrounding context is available to whatever's doing the ad matching. That's a new category of signal.

Constructing an MMM data input layer that accommodates AI conversation channels

For search and social, the inputs a modeler needs are settled: spend by day, impressions, clicks, platform-reported conversions. The actual job of the modeler is to distrust the platform's own ROAS number and use MMM to build an independent read on what those channels are really worth.

AI channels break that pattern, and the available inputs shift depending on the surface. Google's AI Overviews serve ads pulled from existing Search, Shopping, and Performance Max campaigns, so the spend data flows through the same Google Ads pipeline marketers already use, the real challenge is pulling AI Overview placements out of that pipeline and treating them as their own line item rather than letting them blend into standard search numbers. ChatGPT, through OpenAI's Ads Manager, offers impression and click data on cost-per-click formats, and OpenAI's partnership with LiveRamp gives buyers a third-party measurement relationship to lean on for this category. ChatGPT's sponsored cards show up below the model's response, labeled "Sponsored," with a headline, description, and image, while product-feed formats surface specific products directly inside the conversation. Each format produces a different kind of signal, and they shouldn't be lumped together in a data model just because they both originate from a chat interface.

When platform data is thin, or missing outright, modelers fall back on experiment design instead of platform reporting. Geo-based experiments, running AI channel spend in some markets while holding it flat in others and reading the sales differential, are a workable substitute. Time-based holdouts, pausing AI spend for a stretch and watching how baseline sales move against other active channels, work the same way. Where there isn't enough data to model AI spend as a continuous variable, modelers can fall back on treating it as a binary, on or off, and build incrementality tests around that instead of forcing a spend curve the data can't support.

Bayesian MMM has a built-in answer for channels with a short spend history: priors. A modeler can anchor the model using assumptions drawn from an analogous channel or general industry knowledge, rather than waiting for years of AI channel data to accumulate. That only works if those priors get written down and disclosed, not buried as an invisible assumption inside the model's math.

Model design choices when AI channels sit alongside search and social in the same MMM

The first real design decision is how AI search-adjacent inventory, things like Google's AI Overviews or Microsoft's Copilot ad placements, gets modeled: as a variant of paid search or as its own distinct variable. Grouping it with search runs through the same auction mechanics, sits next to the same keywords, and often comes out of the same campaign infrastructure a brand already has running. Splitting it out is arguably the stronger case: the creative format is different, the state of mind of the user encountering it is different, and the measurement signal underneath it is different. Folding it into search hides exactly the distinction a modeler needs to see. A reasonable middle path is to start grouped, simply because early volume is too small to isolate a separate coefficient, and split the two apart once spend variation is sufficient to estimate AI's contribution on its own.

Carryover deserves separate treatment too. A user who gets a product recommendation inside a conversation may not act on it for days, longer than the decay window that fits a standard paid search click. Modelers should test longer adstock decay rates for AI channels rather than assuming the same curve that fits search.

Saturation curves are likely to differ as well. AI channel inventory is still constrained relative to search or social, so the diminishing-returns shape that governs how those channels behave at scale probably doesn't transfer cleanly, and modelers shouldn't assume it does without testing it against the AI channel's own data.

There's also the question of cross-channel interaction: does AI channel activity pull volume away from paid search, or add to it? A user who gets a full answer inside a conversation may never run the search query that would have triggered a paid search ad in the old customer journey. The two channels could therefore be substitutes rather than additive, a relationship no MMM built for the search-and-social era was designed to catch, and one every modeler working across these three channels now has to test for directly rather than assume away.