GEO needs exposure data before it can claim revenue

TL;DR: GEO only becomes measurable when teams model AI-answer exposure, user notice, and business response together, not when they count brand mentions in screenshots.

What does GMMM measure that SEO dashboards miss?

The primary source here is “Generative Marketing Mix Modeling: A Causal Inference Framework Linking GEO and GEM to Business Impact”, listed on arXiv in cs.AI and cs.LG. The useful move in the paper is simple: it treats generative AI answers as a marketing channel with missing instrumentation.

That sounds obvious until you try to measure it.

Classic marketing mix modeling has inputs like spend, impressions, clicks, conversions, seasonality, pricing, and competitor activity. AI answers break that pattern. A user may ask ChatGPT, Gemini, Claude, Perplexity, or another system for “best project management software for a 20-person agency.” The answer may mention your company, bury it, omit it, or compare it unfavorably. The platform may not send a click. The user may still remember the name and convert later.

Standard analytics mostly see the bottom of that funnel. GMMM tries to estimate the missing upper part.

For GEO, the framework combines repeated generated answers, question counts, shares of use across generative systems, and notice probabilities. For GEM, it combines records of sponsored placements with notice probabilities. That last phrase matters. A mention is not the same as an impression, and an impression is not the same as being noticed.

Can AI-answer visibility be tied to business impact?

The paper’s answer is yes, under stated causal conditions. Not “count mentions and declare victory.” GMMM compares expected business responses under alternative treatment sequences, then defines conditions under which those effects can be identified.

That is the grown-up version of the GEO conversation.

A lot of AI search advice right now is still stuck at the visibility layer: get cited, structure pages better, appear in answer engines, monitor share of voice. Useful, but incomplete. If revenue goes up after a brand appears more often in generated answers, was that because of GEO work, paid campaigns, pricing, distribution, seasonality, news coverage, or product changes?

The paper does not magically remove that mess. It gives teams a way to state the mess clearly. You need repeated answer sampling. You need estimates of how many people ask relevant questions. You need platform usage shares. You need some model of whether people actually notice the brand inside an answer. For sponsored generative placements, you need placement records plus the same attention problem.

The empirical section uses simulated answers for product recommendations in English and Japanese. That is a good testing ground, but it also tells you not to overread the claim. Product recommendation is only one class of query. Simulated answer environments are not the same as messy live user behavior. Multilingual testing is valuable, but English and Japanese do not cover the full spread of language, culture, ranking behavior, and platform market share.

So I read GMMM less as “the answer to GEO attribution” and more as a measurement spec the market needed.

What should marketing teams instrument first?

The first practical step is building an answer exposure panel. Pick the queries that matter commercially. Run them repeatedly across the generative systems your buyers actually use. Store the full answers, positions, surrounding context, competitors mentioned, and whether the brand appears as a recommendation, neutral option, warning, or citation.

Then connect that panel to demand data at the right grain. Weekly may be enough for some categories. Daily may matter during campaigns or launches. The trap is pretending query screenshots are analytics. They are samples. Useful samples, but still samples.

Notice probability is the hardest part. A brand named first in a short answer is different from a brand buried in a ten-item list. A brand in a direct recommendation is different from a brand in a citation footnote. Teams will need human ratings, eye-tracking studies, user surveys, or behavioral proxies. None are perfect. Better an explicit imperfect estimate than an invisible assumption.

Practitioner’s Take: If I were building this inside a marketing org, I would start with one product line, 50 to 100 high-intent questions, three to five AI systems, and a weekly answer collection job. I would score visibility and likely notice before trying to claim revenue lift. The catch most teams will miss: GEO measurement is not an SEO rank tracker with a new coat of paint. It is attribution under uncertainty, and the uncertainty needs to be modeled, not ignored.

Related: Ashe runs the SEO practice at Lucky Domains, which builds search visibility the durable way: foundations first, then pages worth ranking.