Why MMM Needs a Rebuild for AI Search
The standard MMM built between 2020 and 2024 has a structural blind spot: it does not include AI search as a channel. The result is that ChatGPT, Claude, Perplexity, and Gemini referrals are absorbed into the base demand intercept along with brand equity and word of mouth. The model says base demand has been rising; in fact, AI search has been doing more of the work, and the brand is systematically underinvesting in it.
The fix is mechanical, not exotic. Add a weekly AI visibility variable to the model, give it the same adstock and saturation treatment as any other media input, and refit. The harder work is the data plumbing and the calibration discipline that follows.
Step 1: Establish the AI Visibility Series
The model needs a weekly numeric series that proxies AI exposure. The standard choice is LLM share of voice across a fixed prompt set covering your category, use cases, and competitive comparisons. The series should be available at national level for top-line MMM and at DMA or region level for any geographic lift testing that calibrates the model.
Lock the prompt set before measurement starts. Changing the prompts mid-stream contaminates the time series and makes the AI variable look like it moved when in fact only the measurement instrument changed. Presenc AI exports weekly SOV with a locked prompt set as the default workflow precisely to prevent this failure mode.
Step 2: Decide Where the AI Variable Sits
The AI visibility series is not "spend." It is "exposure," more analogous to organic search impressions than to paid spend. Modelers have two clean choices: enter it as a media-equivalent variable with its own coefficient, adstock, and saturation curve, or enter it as a covariate in the base-demand equation. The first is preferable when the goal is to value AI search as a discrete channel; the second is acceptable when the data history is too short to support a full media-style treatment.
Either way, do not enter it as a control variable demoted to "other." That hides the channel's effect in a category bucket and makes budget recommendations impossible.
Step 3: Pick Adstock and Saturation Parameters
The carryover from AI exposure to outcome is longer than for paid search and shorter than for TV. As a starting point, use a geometric adstock with half-life of two to four weeks, and a Hill or S-curve saturation with prior on the half-saturation point in the middle of the observed range. Bayesian MMM frameworks like Robyn and LightweightMMM will refine these from data, but the priors matter when the AI visibility series is short.
The carryover prior reflects the consideration cycle in the category. Short-cycle consumer purchases need shorter half-life; long-cycle B2B needs longer. The single biggest mistake is borrowing the paid search prior, which understates AI carryover and produces undervalued AI coefficients.
Step 4: Refit and Validate the Decomposition
Refit the model with the AI variable added. Two checks before trusting the output: first, the contribution of the AI variable should be plausible given the visibility movement in the period (a 30 percent SOV jump should not produce a 0.5 percent contribution change). Second, the model fit (MAPE on holdout periods) should improve. If it does not, the AI variable is not informative and the spec needs revisiting.
Pay particular attention to whether the base demand contribution dropped after adding the AI variable. If it did, the new variable is doing what it is supposed to do: pulling AI-driven revenue out of the unattributable bucket and into a discrete channel.
Step 5: Calibrate Against Incrementality Tests
MMM coefficients are correlational on their own. The validation step is a periodic geographic lift test on the AI visibility input, typically by pausing PR and content syndication in matched regions for eight to twelve weeks. The lift in branded search, direct traffic, and AI-attributed referrals in test regions, relative to controls, is the causal anchor.
Compare the lift test result to what the MMM coefficient implies for the same intervention. If the two agree within confidence interval, the model is calibrated. If they disagree, the test is the ground truth and the model spec needs to be updated. A mature measurement program runs this cycle on every material channel on a rolling basis.
Step 6: Refit Cadence and Governance
Refit the MMM on at least a quarterly cadence, with Bayesian updating between refits if the framework supports it. Track coefficient stability quarter over quarter; sharp jumps without a corresponding business reason are usually a data hygiene issue (a new variable, a change in tagging, an outlier event not flagged).
Document the AI visibility methodology in the same governance pack as the rest of the MMM. Finance and the board need to understand that the AI variable comes from a specific prompt set, a specific platform list, and a specific measurement cadence. Any change to those inputs is a model change and should be documented.
Step 7: Use the Model to Reallocate Budget
With the AI variable in the model and the response curve estimated, the budget allocation step is straightforward. Most brands find that the model recommends moving budget from saturated lower-funnel channels (especially branded search and retargeting) into the inputs that drive AI visibility (PR, content, structured data work, Wikipedia, MCP integration). The recommendations are uncomfortable for organizations whose comp plans reward lower-funnel ROAS, which is part of why governance discipline matters.
How Presenc AI Helps
Presenc AI provides the weekly AI visibility series that MMM needs, at national and regional granularity, with prompt-set governance built in. Exports are formatted to drop directly into Robyn, LightweightMMM, PyMC-Marketing, and commercial MMM platforms. For marketing science teams running MMM with the AI variable for the first time, Presenc also provides a baseline historical series of at least 52 weeks so the variable is informative from day one.