Your Base Demand Is Hiding ChatGPT. Here Is How MMM Finds It.

A marketing science lead at a $200M DTC brand told us last quarter: "Our base demand has grown by 14 points over the past two years. Brand health tracking has not moved. None of our existing channels explain it. I think we are losing the plot."

They were not losing the plot. They were losing ChatGPT.

Their marketing mix model was doing exactly what MMMs do when a channel is missing from the spec: silently absorbing its contribution into the unexplainable intercept. Every dollar of revenue that ChatGPT, Perplexity, Claude, and Gemini were driving had to go somewhere in the decomposition. The model put it in the base. Base demand grew. Brand equity surveys could not explain it. The team's instinct that something was off was correct. The instrument was not equipped to find it.

This is the most common AI search measurement failure in 2026. It is not a measurement absence. It is a measurement misattribution. The model is reporting numbers that look fine until you ask the model to explain itself, and then the explanation falls apart.

MMM Was Built for This Problem

Marketing mix modeling was invented in the 1960s and 1970s by consumer-packaged-goods companies for exactly the situation marketers face today. TV, radio, and print could not be measured at the user level. You could not put a cookie on a TV viewer. You could not attach a UTM to a magazine ad. The CPG industry built MMM because aggregate time-series measurement was the only mathematically honest option.

Then digital came along. User-level tracking made attribution feel solved. MMM was relegated to the back office of CPG and pharma while DTC brands ran on platform-attributed ROAS and multi-touch attribution. For about a decade, this worked. User-level signal was abundant. Cookies were everywhere. Mobile IDs were stable. The dark funnel was small enough to ignore.

That decade is over. Apple's ATT broke mobile attribution. Third-party cookies broke web attribution. And now AI assistants intercept the entire research phase of the funnel before any tracked touchpoint fires. The conditions that made user-level attribution viable have collapsed simultaneously. The methodology that was built for exactly these conditions is the one that is coming back.

Why Multi-Touch Attribution Cannot Fix This

Vendors will tell you that data-driven attribution, Shapley value algorithms, and Markov chain models are the modern answer. They are not. Or rather, they are the modern answer to a different question.

Shapley attribution is mathematically rigorous within the touchpoints in the journey data. The keyword is within. AI assistant interactions produce no journey touchpoint. They do not fire your pixel. They do not pass a referrer. They do not appear in your CRM. The Shapley algorithm cannot assign credit to a touchpoint it cannot see, regardless of how sophisticated the algorithm is.

The result is that every AI-influenced conversion gets credited to whatever closing channel happened to be active. A buyer asks ChatGPT for vendor recommendations, makes a mental note, types your brand into Google two weeks later, clicks your branded paid search ad, and converts. Your data-driven attribution gives 100% credit to branded paid search. ChatGPT did most of the work and gets none of the credit.

This is not a configuration error. It is the architectural ceiling of user-level attribution. No amount of algorithm tuning can recover credit for touchpoints that were never recorded.

What MMM Does Instead

MMM does not need touchpoints. It needs a weekly time series of channel exposure and a weekly time series of business outcome. It regresses outcome on exposure with appropriate transforms (adstock to capture carryover, saturation to capture diminishing returns) and decomposes the outcome into channel contributions.

For AI search, the exposure proxy is share of voice in AI assistant responses. How often is your brand mentioned in the AI responses to a defined set of category prompts. The series is measurable. We measure it weekly. Feed it into the MMM the same way TV GRPs or paid search impressions go in, with adstock and saturation transforms tuned for the channel's actual carryover behavior.

What happens next is the part that surprises people. Base demand drops. Often by 8 to 15 percentage points. The contribution does not disappear. It moves out of the residual and into a discrete AI search channel with its own coefficient, its own response curve, and its own budget allocation line. The mystery of the growing base resolves into a measurable channel that has been doing work the whole time.

The Spec Details That Matter

We have walked enough teams through this transition to know which choices break the model and which choices make it work. Three matter most.

Adstock half-life. AI search carryover is longer than paid search and shorter than TV. Two to four weeks geometric half-life is the right starting prior. The single most common failure mode is borrowing the paid search prior (essentially zero carryover), which guarantees an undervalued AI coefficient.

Saturation curve. AI recommendation saturates relatively quickly once a brand is consistently top-three in category prompts. Hill function with the half-saturation point in the middle of your observed AI share of voice range produces realistic response curves. Linear response specifications send the optimization toward infinite AI spend recommendations, which is a tell that the spec is wrong.

Multicollinearity with PR and branded search. Real, but tractable. PR drives AI visibility, AI visibility drives branded search, branded search closes the conversion. All three need to be in the spec with informative priors on their adstock structures. Omitting any one causes the others to absorb its credit and overstate. The Bayesian framework handles this cleanly when the priors are set thoughtfully.

What the Decomposition Usually Reveals

The pattern we see across brands adding AI search to MMM for the first time is consistent enough to predict. Base demand contribution drops 8 to 15 percentage points. The new AI search variable picks up most of that drop, typically 7 to 14 percent of decomposed revenue. Branded search contribution drops slightly because some of what looked like branded search effectiveness was the AI-driven lift in branded search query volume. Paid social and other lower-funnel channels stay roughly flat.

The response curve on the new AI variable usually shows the brand operating well below the inflection point. Steep marginal returns. Significant additional value available per dollar of additional AI visibility investment. This is the response-curve evidence that says the channel is under-invested.

The optimization then recommends reallocation. The standard pattern: 8 to 15 percent of saturated branded search and retargeting moves into AI visibility inputs (PR, content production, Wikipedia-class authoritative source work, MCP integration, structured product data). The reallocation is uncomfortable because it cuts channels that look healthy on platform-attributed ROAS. The MMM is saying that what looks healthy is over-attributed, and that the next marginal dollar produces more in AI than in branded search.

The Causal Calibration Step

MMM coefficients are correlational on their own. The validation step is to run a geographic lift test on AI visibility inputs every two to three quarters. Pause PR and content syndication in matched DMAs for eight to twelve weeks. Measure the lift in branded search, direct traffic, AI share of voice, and converted revenue in test versus control regions.

The result is causal. If the MMM coefficient and the lift test agree within confidence interval, the model is calibrated. If they disagree, the lift test is ground truth and the MMM spec needs to be revisited. Brands that present "MMM says AI contributes X" without ever running a lift test calibration are presenting correlation as if it were causation, which finance will catch the first time it scrutinizes the methodology.

For the full operational playbook on adding AI visibility to MMM, including data formats, prior recommendations, and validation workflows, see our practitioner guide: Marketing Mix Modeling for the AI Search Era.

Why This Is Happening Now

Three forces are pushing brands back toward MMM at the same time. User-level signal continues to deteriorate (privacy regulation, browser changes, AI assistant adoption). AI search is rising fast enough to be material rather than rounding-error in most categories (typically 8 to 18 percent of converted revenue in well-instrumented brands as of 2026). And the open-source MMM tooling, Robyn from Meta, LightweightMMM from Google, PyMC-Marketing, has matured to the point that brands without dedicated econometrics staff can run production-quality models.

The combination is the comeback story. MMM did not get cleverer. The conditions just shifted back to where MMM was the right tool. Brands that figure this out before competitors do are buying themselves a measurement advantage that compounds quarter over quarter as AI search continues to grow.

The Window

Only 23% of brands running MMM include AI search as a discrete channel as of Q1 2026. The other 77% have a growing base-demand mystery and a measurement framework that cannot solve it. The brands that resolve the mystery, reallocate budget toward the channel the model now sees, and build calibration discipline around the AI variable are establishing a multi-quarter lead that the laggards cannot quickly close.

The lead is not in the tooling. Robyn, LightweightMMM, and PyMC-Marketing are open source. The lead is in the methodology discipline: locking the prompt set, governance the visibility data, running the calibration tests, and translating MMM output into operational reallocation decisions. None of this is mysterious. All of it requires committing to do the work.

The brands that will look smart in 2027 are the brands sitting down with their marketing science teams in May 2026 and asking the simple question: why is our base demand growing, and what is it hiding.