Sketchweb Microblog

Key Takeaways

• Generative Marketing Mix Modeling, or GMMM, links AI visibility to business outcomes such as sales, leads, and conversions.
• The framework measures more than brand mentions. It also considers query demand, AI platform share, and whether users notice the brand.
• A 2026 study collected 2,240 AI answers and found large differences in brand visibility by model and language.
• GMMM accounts for delayed impact through carryover and diminishing returns through saturation.
• The research proposes a measurement framework. It does not prove that GEO automatically increases revenue.

Generative Marketing Mix Modeling, or GMMM, is a proposed framework for measuring how visibility in generative AI answers may affect business results.

Instead of asking only, “How often does ChatGPT mention our brand?”, GMMM asks a harder question:

How much meaningful AI exposure did a brand receive, and what business impact did that exposure cause?

A September 2026 research paper by Masahiro Kato, Daiki Honma, and Taka Kato introduced the framework. It connects Generative Engine Optimization (GEO), Generative Engine Marketing (GEM), user attention, AI query demand, and Marketing Mix Modeling.

The result is an important shift in GEO measurement. AI visibility becomes an intermediate metric rather than the final goal.

What Is Generative Marketing Mix Modeling?

Generative Marketing Mix Modeling is a causal measurement framework designed to estimate the business impact of organic and paid visibility inside generative AI answers.

It extends traditional Marketing Mix Modeling, often called MMM.

Traditional MMM helps companies estimate how different marketing channels contribute to outcomes such as:

• sales
• revenue
• leads
• conversions
• sign-ups
• other business KPIs

GMMM adds generative AI exposure to that measurement system.

The framework separates this exposure into two main categories:

1. GEO, or Generative Engine Optimization
2. GEM, or Generative Engine Marketing

The distinction matters because one represents changes to source content, while the other represents paid placements.

What Is GEO?

Generative Engine Optimization is the process of changing source material in ways that may affect how generative AI systems use, mention, cite, or describe that information.

The GMMM paper gives examples such as:

• adding FAQ content
• revising product documentation
• improving information available to generative systems

GEO is different from traditional SEO.

SEO mainly focuses on visibility within search engine results and related search features.

GEO focuses on visibility or representation inside generated answers.

For example, a company could improve product documentation. Later, an AI assistant may mention that company’s product more often when users ask relevant questions.

However, a higher mention rate does not automatically prove business impact.

That gap is what GMMM tries to address.

What Is GEM?

Generative Engine Marketing refers to paid placements within generated AI experiences.

In simple terms, GEO deals with organic influence on generated answers.

GEM deals with sponsored visibility.

For GEM, the paper proposes measuring factors such as:

• marketing spend
• sponsored placements shown
• expected number of placements
• probability that users notice those placements

GMMM can then place GEO and GEM beside existing marketing channels inside a broader measurement model.

Why AI Visibility Percentage Is Not Enough

Many current GEO measurement systems focus on metrics such as:

Our brand appeared in 30% of tested AI answers.

That number can be useful.

However, it does not tell you how much real exposure the brand received.

Consider two AI questions:

A simple prompt-level visibility calculation says the brand appeared in 50% of questions.

Yet almost all real demand may come from the question where the brand never appears.

Therefore, prompt occurrence alone can give a distorted picture of market exposure.

GMMM tries to solve this by adding several missing pieces.

How GMMM Measures GEO Exposure

The research describes GEO exposure as the expected number of generated answers in which a target property appears and is noticed.

In plain English, the idea can be simplified to:

Expected GEO exposure = relevant AI questions × AI system share × brand occurrence probability × notice probability

Each part answers a different question.

1. Relevant AI question volume

How many relevant questions do people ask?

A prompt with large real-world demand should carry more weight than a rarely used prompt.

2. AI system share

Which generative system receives those questions?

Users may split their activity across different AI systems. Therefore, visibility on one system does not represent the whole market.

3. Occurrence probability

How often does the chosen brand, product, source, or property appear in generated answers?

Because AI responses can vary between runs, repeated measurements are important.

4. Notice probability

Does the user actually notice the brand or property when it appears?

A brand buried near the end of a long answer may have a different effect from a prominent recommendation.

Together, these factors create a more meaningful exposure estimate.

A Simple GMMM Example

Suppose people submit 100,000 relevant AI questions during one month.

A brand appears in 30% of relevant answers.

Assume users notice the brand in half of those appearances.

A simplified estimate would be:

100,000 × 30% × 50% = 15,000 noticed occurrences

Now imagine the brand appeared only 20% of the time before a GEO change.

The earlier estimate would be:

100,000 × 20% × 50% = 10,000 noticed occurrences

The difference is:

15,000 - 10,000 = 5,000 additional noticed occurrences

This example illustrates the framework. It is not a result reported in the research paper.

The next challenge is much harder.

Did those additional exposures cause more leads, sales, or conversions?

That is where Marketing Mix Modeling and causal inference enter the picture.

How Marketing Mix Modeling Fits Into GEO

Marketing Mix Modeling uses aggregated data to estimate how marketing activities relate to business outcomes.

Modern MMM systems can include channels such as:

• paid search
• social advertising
• television
• display advertising
• video
• offline media
• other marketing activity

GMMM proposes adding generative AI as another measurable input.

The basic flow becomes:

1. Measure GEO or GEM activity.
2. Estimate meaningful AI exposure.
3. Apply carryover effects.
4. Apply saturation effects.
5. Model business outcomes.
6. Compare treatment scenarios.
7. Estimate incremental business impact.

Two parts of this process are especially important: carryover and saturation.

What Does Carryover Mean?

Carryover means that an AI exposure can affect a business result after the period in which the exposure happened.

For example, someone could see a software product recommended in an AI answer today.

They may not buy it immediately.

Instead, they might:

1. remember the name
2. search for it later
3. visit the website
4. compare alternatives
5. sign up two weeks later

The original AI exposure may still have influenced the final conversion.

Therefore, GMMM does not assume that an AI mention on Monday must produce a sale on Monday.

Traditional MMM already uses similar methods to model delayed advertising effects.

GMMM extends that logic to generative AI.

What Does Saturation Mean?

Saturation means that every additional exposure may not create the same amount of additional business impact.

Imagine a brand moving from 100 AI exposures to 10,000.

That increase could make a large difference.

But moving from 1,000,000 exposures to 1,009,900 may have a much smaller marginal effect.

In other words:

More exposure does not always create proportionally more value.

The GMMM framework uses a saturation transformation to represent this diminishing-return pattern.

This prevents the model from assuming that business results will rise forever in a perfectly straight line with AI exposure.

Why Causal Inference Matters

Correlation is not enough to prove that GEO worked.

Imagine that AI visibility rises in May.

Sales also rise in May.

It is tempting to conclude:

GEO increased sales.

However, many other factors may have changed at the same time.

For example:

• demand may have increased
• prices may have changed
• paid campaigns may have launched
• seasonal demand may have started
• traditional search traffic may have grown
• competitors may have reduced activity
• landing page conversion rates may have improved

GMMM therefore focuses on a counterfactual comparison.

It asks:

What would the business result have been with GEO compared with the expected result without that GEO treatment?

The difference between those scenarios represents the estimated treatment effect.

Why GMMM Compares Complete Treatment Sequences

Carryover creates another complication.

A GEO change introduced in January could affect:

• January
• February
• March
• later periods

Therefore, simply comparing May visibility with May sales can miss much of the story.

The paper instead defines treatment effects using complete sequences.

Conceptually, the model compares:

Business performance with the GEO treatment over time

against:

Business performance without that GEO treatment over time

The evaluation window can even continue after active treatment ends so that delayed effects remain included.

This is one of the biggest differences between simple AI visibility tracking and causal business measurement.

What the 2026 GMMM Study Tested

The researchers also collected AI responses to examine occurrence probabilities.

Their dataset contained 2,240 complete AI answers.

The experiment used:

• 2 generative AI models
• 56 product recommendation questions
• 28 English questions
• 28 Japanese questions
• 20 answers for every question and model combination

The target brand was Glasp.

Importantly, the researchers did not place Glasp inside the questions or system instructions.

They then measured how often Glasp appeared in the generated answers.

GPT-5.6 Luna vs GPT-4o: What the Study Found

Across all collected answers, Glasp appeared:

• in 33.8% of GPT-5.6 Luna answers
• in 27.8% of GPT-4o answers

Looking only at the overall number might suggest that GPT-5.6 Luna provided stronger visibility.

However, the language-level data told a more complicated story.

GPT-4o showed the higher Glasp occurrence rate for English questions.

GPT-5.6 Luna showed a much higher rate for Japanese questions.

This finding has a major implication for GEO measurement:

A single AI visibility score can hide important differences between models, languages, questions, and user contexts.

AI Visibility Can Change by Model and Language

Suppose a dashboard reports:

AI visibility: 34%

That raises several follow-up questions.

Thirty-four percent across which:

• models?
• languages?
• topics?
• query types?
• markets?
• time periods?
• user intents?

The GMMM experiment demonstrates why segmentation matters.

A brand can perform well on one model while performing poorly on another.

The result can also reverse when the language changes.

Therefore, useful GEO reporting should avoid treating AI visibility as one universal number.

Mention Frequency Is Not the Same as Prominence

The study also examined where Glasp appeared among candidate brands.

When Glasp appeared, GPT-4o placed it first more often than GPT-5.6 Luna.

The reported first-position rates were:

• GPT-4o: 60.5%
• GPT-5.6 Luna: 39.4%

This creates another important distinction.

A brand may be mentioned frequently but rarely appear prominently.

Therefore, a mature GEO measurement system may need several separate metrics:

These metrics answer different questions.

They should not be treated as interchangeable.

Why Referral Traffic Cannot Measure the Full GEO Effect

Referral traffic remains useful, but it measures only one type of behavior.

A user may read an AI answer, notice a brand, and never click the cited website.

Later, that person could:

• search for the brand directly
• visit through organic search
• type the website address
• buy through another channel
• contact a sales team

In those cases, the AI answer influenced the user without creating an immediate referral session.

The GMMM paper specifically distinguishes noticed exposure from referral traffic.

This is a key reason why GEO attribution is difficult.

AI systems can influence decisions before a measurable website visit occurs.

Plug-In, Cut, and Joint GMMM Models Explained

The researchers discuss three broad ways to handle uncertainty in estimated AI exposure.

Plug-in approach

The plug-in method uses the best estimated exposure value directly.

For example:

Estimated exposure is 15,000, so use 15,000 in the response model.

It is the simplest option.

Cut approach

The cut approach carries uncertainty from the measurement stage into the business model.

However, later business data do not update the original measurement parameters.

In simple terms, information flows forward but not backward.

Joint approach

The joint approach allows the measurement and response parts to influence each other.

It is more integrated.

However, more complexity did not automatically create better results.

The Most Complex Model Was Not Always Better

One of the useful findings from the simulations was that there was no universal winner between the modeling approaches.

The joint method sometimes improved certain coefficient estimates.

However, it did not consistently improve estimates of the GEO treatment effect.

Under one simulated condition using estimated occurrence probabilities, the joint approach performed worse than the cut approach.

Meanwhile, accurately estimating carryover and saturation explained much of the improvement over simpler approaches that fixed those values.

The practical lesson is straightforward:

A more complex model is not automatically a more accurate GEO measurement system.

Correctly modeling how exposure works over time may matter more.

What the Research Does Not Prove

The study has an important limitation.

Its 2,240 collected AI answers did not directly measure the causal effect of GEO on real business revenue.

The answers were collected under the source state that existed during the experiment.

Therefore, they help estimate questions such as:

How often does Glasp appear under these conditions?

They do not directly answer:

How much did a specific Glasp GEO change increase visibility?

Nor do they answer:

How much additional revenue did that change generate?

The paper is clear that measuring a GEO treatment effect requires stronger evidence.

In particular, analysts need variation between source states.

What a Real GEO Business Impact Test Would Need

A stronger real-world GMMM study could compare a source before and after a clearly defined GEO treatment.

For example:

Before GEO

• original documentation
• original FAQ coverage
• baseline AI occurrence probability
• baseline business outcomes

After GEO

• revised source material
• updated FAQs or product information
• new AI occurrence probability
• later business outcomes

However, a simple before-and-after comparison can still be misleading.

A stronger design may also require:

1. control markets or product groups
2. staggered treatment
3. repeated AI answer collection
4. reliable query-demand estimates
5. AI system usage shares
6. user-attention studies
7. business outcome data
8. controls for other marketing activity

Randomized or well-controlled treatment variation can provide much stronger evidence than ordinary prompt monitoring.

A Practical GEO Measurement Framework for Marketers

Most teams are not ready to build a full causal GMMM system today.

Still, the research provides a useful roadmap.

Step 1: Define the property you want to measure

Do not begin with a vague metric such as “AI presence.”

Choose something specific.

For example:

• brand occurrence
• product occurrence
• citation of your domain
• first-position recommendation
• inclusion in a comparison
• mention of a product benefit

Step 2: Build relevant query clusters

Group prompts around real customer needs.

Possible clusters include:

• product discovery
• comparisons
• alternatives
• buying questions
• troubleshooting
• brand research
• category education

This prevents one type of prompt from dominating the entire visibility score.

Step 3: Test repeatedly

Generative answers can vary between runs.

A single answer therefore provides weak evidence about occurrence probability.

The GMMM experiment used 20 answers for every question-model combination.

The right number for a commercial study will depend on the desired precision and available resources.

Step 4: Segment by model and language

Do not combine everything too early.

Measure results separately for important:

• AI systems
• languages
• markets
• topics
• intents

The Glasp experiment shows why this matters.

Step 5: Weight visibility by real demand

A high occurrence rate for an unimportant prompt may create little real exposure.

Whenever reliable data exist, weight visibility according to relevant demand.

Step 6: Consider user attention

Not every mention creates the same amount of exposure.

Position, presentation, context, and answer length may influence whether a user notices a brand.

Step 7: Connect exposure to business data

Next, compare AI exposure with relevant outcomes such as:

• leads
• sign-ups
• demos
• purchases
• revenue

Other channels and market conditions should also remain in the model.

Step 8: Test incrementality

Finally, compare business outcomes under meaningful treatment and control conditions.

This moves measurement from correlation toward causal inference.

From AI Visibility to Business Impact

The biggest idea in GMMM is not a new AI visibility score.

It is a new measurement chain.

The framework can be summarized as:

GEO activity → generated answers → brand occurrence → user attention → meaningful exposure → carryover and saturation → business outcome → estimated incremental impact

That chain changes how marketers should think about GEO.

A mention is not a conversion.

A citation is not revenue.

A visibility percentage is not ROI.

Each metric describes only one stage between content optimization and business impact.

What GMMM Means for GEO Tools

Many GEO platforms currently focus on the first few stages of measurement.

A typical dashboard might report:

Brand visibility: 42%

A GMMM-inspired system could eventually go further.

This table is an example of how the framework could be operationalized. It is not a dashboard presented in the research paper.

The important change is moving from visibility reporting toward incrementality measurement.

Did You Know?

The GMMM study found that the model with the highest overall Glasp occurrence rate was not the model that placed Glasp first most often. This shows why presence and prominence should be measured separately instead of being combined into one AI visibility score.

Conclusion

Generative Marketing Mix Modeling offers a framework for moving GEO measurement beyond prompt tracking.

Instead of stopping at “How often does AI mention us?”, GMMM connects query demand, AI system usage, occurrence probability, user attention, carryover, saturation, and business outcomes.

The 2026 research also shows why simple AI visibility scores can be misleading. Glasp’s occurrence rates changed substantially between GPT-5.6 Luna and GPT-4o, and the pattern reversed across English and Japanese questions.

However, the study does not prove that GEO automatically increases revenue. Its larger contribution is methodological.

It shows what marketers may need to measure before they can credibly move from AI visibility to causal business impact.

FAQs

What is Generative Marketing Mix Modeling?

Generative Marketing Mix Modeling, or GMMM, is a framework for estimating the business effects of organic and paid exposure inside generative AI answers. It combines AI answer measurements with query demand, platform usage, user attention, traditional marketing data, carryover, saturation, and causal comparisons.

How is GMMM different from GEO tracking?

GEO tracking usually measures metrics such as mentions, citations, rankings, or visibility percentages. GMMM goes further. It tries to estimate meaningful exposure and then connect that exposure to sales, leads, conversions, or other business outcomes while accounting for competing explanations.

Can AI referral traffic measure GEO ROI?

Not by itself. Referral traffic measures users who click from an AI experience to a website. Some people may see and remember a recommendation without clicking. They could later visit through search, direct traffic, or another channel. GMMM therefore treats noticed exposure as different from referral sessions.

What did the GMMM study find about GPT-5.6 Luna and GPT-4o?

Across 2,240 answers, Glasp appeared in 33.8% of GPT-5.6 Luna answers and 27.8% of GPT-4o answers. However, GPT-4o had the higher occurrence rate for English questions, while GPT-5.6 Luna had a much higher rate for Japanese questions.

Does the research prove that GEO increases sales?

No. The paper proposes a framework for estimating causal GEO effects, but its collected Glasp answers were mainly used to estimate occurrence probabilities under one source state. Measuring incremental business impact requires treatment variation, market data, attention estimates, business outcomes, and suitable controls.

References

• https://arxiv.org/abs/2609.11915
• https://arxiv.org/abs/2311.09735
• https://collaborate.princeton.edu/en/publications/geo-generative-engine-optimization/
• https://blog.google/products/ads-commerce/meridian-marketing-mix-model-open-to-everyone/
• https://www.pymc-marketing.io/en/stable/