Measuring GEO's revenue impact with a marketing mix model

To measure GEO's effect on revenue, do not use the share of AI answers that name your brand as it is. Multiply that share by question volume, each engine's share of use, and the probability that a user notices the name to get expected exposures per market and period, then pass it through carryover and saturation like any other medium before it enters a marketing mix model. Masahiro Kato of the University of Tokyo and Mizuho-DL Financial Technology and coauthors propose this framework, GMMM, in Generative Marketing Mix Modeling^[1], posted to arXiv in September 2026. The evidence is simulation, and the only real data collected are occurrence rates for one brand measured across 2,240 GPT answers^[1].

We read the arXiv v1 manuscript (54 pages, not yet peer reviewed) through its appendices and checked the numbers against its tables^[1]. The brand whose answers the paper collects is Glasp, and the referral analysis in the appendix uses the AEO natural experiment data^[2] published by glasp.co. The paper carries no conflict of interest statement.

We build a product that repeatedly measures brand mentions in AI answers, so we have an interest in this topic. Our previous post, Calibrating MMM with experiments, covered how to feed experiment results into an MMM. This post covers the step before that: how to build the AI answer exposure input itself.

Why AI search exposure cannot enter a marketing mix model as it is

Brand mentions inside AI answers are not recorded anywhere as exposure counts per market^[1]. A marketing mix model (MMM) relates an aggregate response such as sales to media inputs observed across markets and periods. Impression logs do not record mentions that nobody paid for, and a platform's record of sponsored placements does not show whether users noticed them.

The paper separates two treatments. Generative engine optimization (GEO) changes answers by changing source material, such as adding an FAQ section or revising product documentation^[6]. Generative engine marketing (GEM) pays for sponsored placements inside generated answers^[1]. Neither leaves an impression log the way established media do.

ChatGPT referral sessions measure a different event, too. Users who read an answer without following a link never show up as referrals^[1]. Using "what share of collected answers named us" in an MMM therefore needs a measurement step first, one that moves the share onto the market and period scale.

The GEO input: from occurrence rates to expected exposures per market

In GMMM the GEO input is the expected number of answers in market $i$ and period $t$ that contain the brand and are noticed by the user^[1]. The formula is:

Here $N_{iqt}$ is how many questions in cluster $q$ arrive in market $i$ during period $t$, ${\omega }_{ipt}$ is the share of those questions handled by generative system $p$, ${\pi }_{qpt}(z)$ is the probability that an answer contains the brand, and ${\lambda }_{qpt}$ is the probability that a user notices it when it does. $z=1$ is the state with GEO applied and $z=0$ the state without it^[1].

This input is positive even without GEO. GEO does not create the input from nothing; it changes the input by shifting the occurrence probability $\pi$^[1]. Occurrence probabilities come from a hierarchical logistic model with question-cluster effects, and notice probabilities come from a binomial model fitted to a user study that records who noticed the name.

The GEM input has the same shape. A Poisson model links spending to the number of sponsored placements actually shown, and that count is multiplied by a notice probability^[1]. Both inputs then pass through the same geometric carryover (adstock) and Hill saturation transforms as established media before entering the response model^[4].

A single occurrence rate can replace the input only in a narrow case. The Hill midpoint absorbs the scale only when the product of question volume, system share, and notice probability is the same constant multiple in every market and period^[1]. When question mix or system share differs across markets, one occurrence-rate series cannot represent each market's input.

Defining the GEO effect: compare complete treatment sequences

GMMM defines the GEO effect as the difference in expected revenue between the complete treatment sequence with GEO and the complete sequence without it^[1]. Because carryover makes period $t$ revenue depend on earlier inputs, switching GEO off changes the input sequence in every affected period.

$a_G$ and $a_P$ select the GEO and GEM sequences, and $T_E$ is the evaluation window. Besides the path through generated answers (${\beta }_G{H}^G$), the response model includes a term for the GEO state itself (${\beta }_{D}Z$)^[1]. A source change can also move revenue outside the answers, for example through conventional search traffic or landing page conversion.

The authors show that under four assumptions (consistency, sequential exchangeability, positivity, and interference only through specified aggregates) the longitudinal g-formula identifies ${\Delta }_G$^[1]. They also give a condition under which the total effect remains identified even when the GEO state and the GEO input move together so closely that ${\beta }_D$ and ${\beta }_G$ cannot be separated.

AI answer exposure differs by engine and language for the same questions

In the real collection, the ranking of models by brand occurrence flipped with the language of the question^[1]. The authors asked 56 product recommendation questions across seven use cases, including web highlighting and AI summarization (28 in English, 28 in Japanese), to GPT-5.6 Luna and GPT-4o, 20 times each, for 2,240 answers. Every call required web search, and the brand name appeared in neither the questions nor the system instruction.

Table 1. Share of 2,240 GPT answers in which Glasp appeared, collected September 3 to 4, 2026. "Named first" is the share in which Glasp was the first of 11 candidate brands mentioned. Source: Kato et al. 2026, Table 1^[1].

Overall, GPT-5.6 Luna was 5.70 points higher (95% interval 3.12–8.27 points)^[1]. On English questions GPT-4o was 6.79 points higher, and on Japanese questions GPT-5.6 Luna was 18.75 points higher. When Glasp did appear, GPT-4o named it first more often.

The lesson is about measurement design. One overall occurrence rate does not even preserve the ordering between models, and without weights that reflect real demand for each question, the number is only the average of a fixed question panel^[1]. We covered repeated measurement and question mix in how many measurements to trust AI search visibility.

Simulation: estimating carryover and saturation drove accuracy

In the simulations, what changed the accuracy of the GEO effect estimate most was estimating the carryover and saturation parameters from revenue data instead of fixing them^[1]. The setup has 5 markets, 6 product clusters, and 36 periods. Two clusters start GEO in period 19, two in period 23, and two never receive it, and each condition was replicated 120 times.

With carryover and saturation fixed at their prior means, the coefficient vector RMSE was 1.207, and estimating them from revenue data cut it to 0.945^[1]. A joint posterior that also lets revenue data update the measurement parameters gave 0.950, no different. The joint posterior showed no consistent advantage in any condition, and in the condition using real occurrence rates its relative RMSE for the total effect was 17.480%, slightly worse than 17.326% for the cut posterior that keeps the measurement distribution fixed.

The total effect was more accurate than its parts. In the simulation the path outside answers (${\beta }_{D}Z$) made up 66–75% of the total effect, yet that term and the answer exposure term had relative RMSEs of about 33% and 48%, far above the 17.5% for the total^[1]. Their estimation errors ran in opposite directions (correlation −0.63) and cancelled when added. A decomposition such as "X% of the effect came through AI answer exposure" is much less reliable than the total effect.

Adding a randomized experiment estimate to the likelihood did little. An aligned estimate from the same population moved the total effect's relative RMSE from 17.430% to 17.424%, and a deliberately shifted estimate raised bias from 1.2% to 4.2%^[1]. Another route for putting experiments into the model is Google's ROAS prior calibration^[3].

Is rising ChatGPT referral traffic a GEO effect: two traps in before-after comparisons

Rising ChatGPT referrals alone do not establish a GEO effect, and two simulations in the paper's appendix show why in numbers^[1]. The first trap is platform growth. When treated and untreated clusters share the growth of the generative platform, comparing treated units before and after treatment gave a bias of 0.081, and 95% intervals covered the truth 0% of the time.

Difference-in-differences (DiD) with untreated clusters as the comparison group cut the bias below 0.001 and the RMSE to 0.008 on the same data, with 95.4% coverage over 500 replications^[1]. We covered experiment designs with comparison groups, baselines, and measurement intervals in how to design AI search visibility experiments.

The second trap is controlling for branded search. In a simulation with a direct GEO effect of 0.70 and an effect through branded search of 0.88, a model without branded search recovered the total effect of 1.58 as 1.579^[1]. A model that controlled for branded search volume estimated only 0.700 and erased the whole path through branded search.

How estimates moved on real ChatGPT referral data

Applying segmented regression to glasp.co's public referral data, the effect estimate moved from 1.18x to 2.40x depending on which week was taken as the start of treatment^[1]. The data come from a field study of one domain that applied AEO interventions to hundreds of thousands of YouTube question-and-answer pages in January 2026, with weekly referral indices for treated and control groups from July 2025 through May 2026^[2]. GMMM's authors modeled the log ratio of treated to control referrals with a trend plus post-treatment level and slope changes and estimated the immediate ratio.

Table 2. Immediate referral ratio (treated to control) by the week taken as the start of treatment. The intervals for the first two dates include 1. Source: Kato et al. 2026, Appendix F, Table 18^[1].

A change in measurement also moved the estimate. The original study reports that bot filtering changed in mid-March, and allowing both a level and a slope change at that point lowered the immediate ratio to 1.337–1.455, while using only weeks before March 10 gave 1.337 (95% interval 0.937–1.906), which includes 1^[1][2]. A placebo test over 19 pre-treatment dates placed the estimate at rank 0.150, meaning changes of that size occurred before treatment as well. The authors write that this diagnostic does not by itself establish a treatment effect.

The authors also note that GMMM cannot be run on this data. It lacks contemporaneous question counts, generated answers, and notice observations, so the GEO input cannot be built^[1]. How much referral metrics miss of AI search's contribution is covered in AI recommended you, but GA4 can't see it.

Limitations

The GEO effect and the business response in this paper all come from simulation, and there is no validation on real company data^[1]. The limitations the authors state, or that the manuscript shows, are:

• The real collection covers occurrence rates under one source state at one time. There are no observations of a state change before and after GEO.
• The simulations treat question counts and system shares as known and omit established media channels. A real application adds uncertainty from estimating question volume and from shifting shares.
• Notice probability needs a user study. The simulations assume 50 notice observations per cell.
• The collection spans English and Japanese, two GPT models, one brand, and a little over one day.
• The causal reading rests on sequential exchangeability and positivity. An observational rollout needs pre-treatment controls and a contemporaneous comparison group.
• Saturation and effects that change over time are sometimes not distinguishable in standard MMM data, an ambiguity the authors cite^[5].
• This is a v1 manuscript that has not been peer reviewed, and it has no conflict of interest statement.

What we take from it at TRAIL

In the paper's GEO input formula, the part our product covers is the occurrence probability $\pi$. TRAIL Search asks a question set to the engines selected in a plan on a regular schedule and records mentions, rank, and cited URLs per engine, question, and check time. Share of voice is computed only from questions without brand names, measurements are repeated under matched conditions, and failed measurements are kept apart from real non-mentions.

What we do not measure is just as clear. Question volume per market $N$, engine share of use $\omega$, notice probability $\lambda$, and revenue data are outside our measurement. Our numbers alone therefore cannot state GEO's revenue effect; they are raw material for one measurement module in a framework like GMMM. We laid out a staged view of brand mentions in the brand mention stage model.

Four rules from this paper carry straight into practice:

• Weight the question set to match real demand. A fixed panel's average occurrence rate does not even preserve the ordering between models.
• Use pages or product lines without GEO as a comparison group instead of comparing before and after. This keeps platform growth out of the effect.
• Do not add branded search volume as a control. It erases the effect GEO produces through branded search.
• Read referral metrics as a lower bound. Users who read an answer and leave never appear as referrals.

If the ROAS reparameterization in the previous post meant "make the quantity you want a model parameter," this paper fills the step before it: "build AI answer exposure into a market-level input." Why a channel effect should not be read as one number is in Bayesian MMM credible intervals, and an MMM implementation that takes reach and frequency as inputs is in Google's Meridian documentation.