Attribution & ROI
Can marketing mix models measure AI search? What the exposure series needs first
Finance teams fund TV and podcasts on the strength of a marketing mix model's estimate. AI answers also move buyers who never click, so the same model is the natural place to look for AI search.
Updated
Questions this guide answers
- What is a marketing mix model?
- Can marketing mix models measure AI search?
- What data does an MMM need to measure AI search?
- How can I test whether my AEO changes worked?
Direct answer
Yes, in time. A marketing mix model (MMM) estimates how much each channel contributed to sales from a few years of weekly history, without tracking clicks, which suits AI answers that influence buyers who often never click. Google's Meridian and Meta's Robyn, 2 widely used open-source MMM tools, both have a place for unpaid activity, which is where AI visibility fits.
The catch is the input: a weekly exposure series, meaning how often AI answers show your brand. It needs a method that stays the same over time, enough answers each week that random noise doesn't swamp it, separation from the demand that also drives sales, and history. Robyn's guide asks for "a minimum of two years of historical weekly data," and every AI exposure report from Google, Microsoft, and OpenAI started in 2026.
So start the series now, from a fixed set of unbranded buyer questions asked daily on each engine, and save the platform reports alongside it. While it grows, use page-level or prompt-level experiments to test whether your changes move AI visibility; they need weeks, not years.
What a marketing mix model is
An MMM is a statistical model fitted to your own history. Week by week, you give it how much you spent on each channel or how much exposure each one had, outside factors (controls) such as season, price, and overall category demand, and the sales or leads you got. It estimates how much of the outcome each channel explains, and what more or less spend would do.
Two assumptions shape the estimate. Effects carry over, so a week of exposure keeps working for a while afterward, and returns diminish, so each extra unit adds less than the one before. Meridian calls these Adstock and Hill effects. Because the model works from weekly totals, it needs no user tracking and no clicks.
The price is data. The model can only tell channels apart if their levels moved differently over time, and it needs a lot of history: Google researchers note that "three years of national weekly data is only 156 data points" for a model that often has 20 or more channels.
Why click reports miss AI search
When an AI answer names you and the buyer searches your name later, analytics records a branded search or a direct visit, not AI, so click-based reports see only part of AI's effect, as the web traffic attribution guide explains.
An MMM sits above the attribution stack rather than replacing it: referrals, surveys, and buyer interviews show that AI influence exists, and the model estimates its size.
How AI search goes into the model
In both tools, AI visibility fits the slot for unpaid channels. Meridian calls it organic media, "media activities that have no direct cost," and Robyn calls it organic variables, listing "SEO improvements" among its examples.
With no spend attached, the model reports a contribution to sales, not a return; add your content and staff costs to get one for your ROI business case. The harder part is the exposure series, which needs 4 things.
| What the series needs | What goes wrong without it | How to meet it |
|---|---|---|
| A frozen method | Changing the prompts, engines, or counting rules moves the series on that date, and the model reads the step as a change in exposure | Keep prompts and engines fixed and report each engine separately; when the set changes, run old and new side by side or mark the break |
| Enough answers per week | Random noise in an input shrinks its estimated effect toward zero, so AI search looks weaker than it is | Keep enough answers per prompt and engine each week, and track the margin of error (how many runs) |
| Demand kept apart | When more people ask, platform counts rise, and so do sales, so the model credits AI for demand | Add category search volume as a control, or ask the same unbranded questions daily, weighted by demand you estimate separately |
| History | Too few weeks for a stable estimate | Start now, and save the platform reports on a schedule (Search Console, Bing, and Business Hub) |
Platform counts or a fixed question panel
Two kinds of exposure series exist. Platform counts come from the platforms: Search Console counts impressions of your links in Google's AI features, Bing counts AI citations, and OpenAI's Business Hub counts answers that reference an online business's website. They measure real users, but each covers 1 platform, counts links rather than brand names, and rises and falls with how many people ask.
That last point is a trap Google's MMM researchers describe for paid search: ads follow searches, and searches follow demand that also drives sales (Chan and Perry), so a naive model "leads to over-estimates of ROAS," the return on ad spend (Chen et al.). Meridian's remedy is to add category search volume as a control. Platform counts need the same.
A fixed question panel asks the same questions on each engine every day, so it doesn't move with how many people ask, and it records brands named in the text, not only links. It measures how often you appear per question, a rate rather than a count of the people who saw you. Weight each question by its share of demand and keep the weights fixed; Meridian rescales each series by its median, so yours doesn't need absolute volumes. What it can't show is a change in how many people ask, which is why the platform counts belong alongside it.
The panel also needs many answers, because answers change across days: in a 2026 study of Google AI Overviews on trending queries, when a query showed an AI Overview on 2 different days, the 2 answers cited exactly the same set of URLs only 0.24% of the time.
The harder problem: AI visibility follows demand too
A fixed panel removes one bias but not the deeper one. Engines mention brands that have more coverage, more reviews, and more people talking about them. A brand that grows for other reasons, such as a launch or a TV campaign, can also gain AI mentions, and a model that doesn't separate the two credits AI search with demand it reflected rather than created.
Meridian's documentation describes the same pattern for search volume: TV and other channels drive branded searches, so branded search volume can sit between them and sales, and Meridian's full-funnel design models it as a result of those channels before using it as a cause; AI visibility can be modeled the same way. Also build the series from unbranded questions, because questions that already name you mostly show demand you had, as branded search does.
Experiments that work while the series grows
Experiments, also called incrementality tests, answer a narrower question than an MMM: did a specific change move AI visibility, or the branded search and leads that follow it. Regional holdout tests rarely fit, because a change to your site reaches every market at once.
- Page-level split tests. Randomly split comparable pages, change only one group, and compare citations and AI impressions by page. Pinterest published this design for SEO: it splits pages, not users.
- Prompt-level rollouts. Randomly choose which buyer questions get new content first, and compare their mention rates with the rest over many runs.
- Untouched controls. Track questions you didn't work on, and 2 or 3 competitors on the same questions. If everything moves together, the engine moved, as performance-based contracts also check.
- Time-based tests. When a change is global, compare branded search or leads after it with a forecast of what would have happened without it, built from series the change didn't touch. That is the approach behind Google's CausalImpact.
How SolCrys fits
SolCrys measures the answer side. It reruns a frozen set of buyer prompts on ChatGPT, Gemini, Google AI Overviews, Perplexity, and Claude, depending on plan, daily by default on paid plans. For each answer it records whether you were named, how the answer positioned you, and which URLs it cited, and it exports those dated records as CSV or serves them to your own agents over MCP. Each prompt carries a demand band you can use as a weight. SolCrys doesn't connect to GA4, Search Console, or your CRM and doesn't fit the model; you join the series to outcomes in your own analysis.
Sources
- Google Meridian: Organic media and non-media treatment variables
- Google Meridian: Control variables
- Google Meridian: Input data
- Meta Robyn: Analyst's guide to MMM
- Meta Robyn: Features
- Chan and Perry, Google: Challenges and Opportunities in Media Mix Modeling (2017)
- Chen et al., Google: Bias Correction for Paid Search in Media Mix Modeling (2018)
- Sørensen, University of Houston: Bias of the OLS estimator when the regressor is measured with error (2025)
- Xu, Iqbal, and Montgomery, Washington University in St. Louis: Measuring Google AI Overviews (2026)
- Pinterest Engineering: Demystifying SEO with experiments (2015)
- Brodersen et al., Google: Inferring causal impact using Bayesian structural time-series models (2015)
- Google Search Console Help: Generative AI performance report
FAQ
Is AI share of voice a good input for an MMM?
Not on its own. Share of voice divides your mentions by everyone's, so it rises when a competitor drops out even if buyers saw you no more often. Build the exposure series from your own mention rate on each question, and keep share of voice as a competitive read.
Free ChatGPT visibility check
See where AI answers skip your brand — then fix it, free
Start a free workspace with your domain: 10 buyer-intent prompts through ChatGPT show where you are mentioned, cited, or skipped, and who gets recommended instead. A free content audit in the same workspace hands you the first fix to ship.
Free · No credit card · About 5 minutes
Related guides
Attribution & ROI
AEO Actions to Revenue
Visibility metrics are not revenue. This guide is the operational playbook for tracking AEO actions through to conversion events - GA4 limits, survey-based attribution, prompt-set recovery estimates, cohort comparison, and how to combine them.
Attribution & ROI
What AEO Does to Your Web Traffic
AI shapes far more buying than it sends clicks, so analytics undercount AEO. What GA4, Search Console, and Bing show, what they miss, and how to attribute it.
Measurement
How Many Runs Until Your AI-Visibility Number Is Trustworthy?
A single AI check is one sample, not a measurement. How many repeated runs and prompts you need before an AI-visibility number is trustworthy, with the published evidence on sample size and confidence intervals.
Measurement
Your AI Visibility Score Looks Fine. Here's How It Hides the Questions You're Losing.
A blended AI visibility score averages prompts and engines, so branded prompts, changes to the tracked prompts, and engine averaging can hide the buyer questions you lose. How to read it per prompt.
Buyer Guides
Performance-Based AEO Contracts: 5 Questions to Ask Before You Sign
Pay-for-performance AEO ties an agency's fee to an AI visibility threshold or a revenue share. Five questions decide whether that threshold means anything.
Attribution & ROI
Prove AI Influence with Closed-Won Interviews
Your dashboard says branded search. Your buyers say ChatGPT. The fastest defensible way to prove AI influence on pipeline is a monthly closed-won interview — the script, the sample size, the CRM field, and the 1-in-X readout that gets AEO funded.
Attribution & ROI
The Free AI Reports: What Search Console, Bing, GA4, and Merchant Center Show, and What None of Them Do
Search Console, Bing Webmaster Tools, GA4, and Merchant Center now report AI search for free. What each one shows, what it misses, and how to cover the gaps.
Attribution & ROI
AEO ROI Business Case
A practitioner framework for estimating and reviewing AEO ROI with finance. Includes the AEO Revenue Model formula, three attribution methods, a 5-slide deck structure, and a 12-month measurement template.