You Can’t Measure What AI Says About You — Until Now, Maybe

Every dollar you’re spending on AI visibility is flying blind. No impressions, no click-through, no attribution — just vibes and hope. A new causal inference framework claims to fix that. Before you rebuild your measurement stack, read the fine print.

What happened

A team of three researchers (Kato, Honma, and Kato) identified a gap that every growth marketer bumping into GEO (Generative Engine Optimization) already feels in their gut: standard marketing data pipelines have no field for “how often did a generative engine mention your brand, and did the user actually notice?” Their proposed solution — Generative Marketing Mix Modeling (GMMM) — is a causal inference framework that attempts to plug that hole for both organic AI placements (GEO) and paid AI placements (GEM). For the GEO side, GMMM ingests repeated generated answers, question volumes, share of voice across different AI systems, and what the authors call “notice probabilities” — an estimate of whether a user actually registered your brand in the output. For GEM, it folds in sponsored placement records alongside those same notice probabilities. The framework then compares expected business responses across alternative “treatment sequences” — i.e., what would have happened if your brand appeared more or less often — and derives sufficient conditions for causal identification. Empirical validation was run on simulated product recommendation queries in both English and Japanese, not on live business data.

Cold read

The word “simulated” in the last sentence of the abstract is doing enormous heavy lifting here, and founders should not miss it. There are no real business outcomes validated — no revenue lift, no conversion delta, no brand search uplift from an actual campaign. The core variable, “notice probability,” is not directly observed anywhere in practice; it’s a model input, which means the causal chain is only as good as your guess at how often users actually register a brand entity in LLMs output. The framework also requires “repeated generated answers” as an input — meaning you need to be running systematic query sampling at scale before this is useful, which is non-trivial and not free. Most critically, the paper establishes sufficient conditions for identification, a theoretical result — it does not demonstrate that real marketing data in the wild will satisfy those conditions. The dual-language simulation (English and Japanese) is a thoughtful touch but tells you nothing about robustness to different product categories, AI system behavior changes, or the chronic problem that LLM outputs shift with model updates.

What it means for you

• Signal maturity: 2/5 — theoretically grounded but entirely pre-empirical on real business data
• Who gets hurt: Mid-market CMOs who buy a GEO SaaS tool that ships “GMMM-powered attribution” before anyone has validated it on live revenue data
• What breaks if this is true: The entire cottage industry of AI visibility metrics built on proxy signals (mention counts, citation rate vs. mention rate) loses its excuse for not connecting to business outcomes — and vendors will be forced to show causal lift, not just share of voice
• Why it might not land: Notice probability is unobservable in any standard data pipeline; without a credible way to measure it, the model is a theoretical skeleton with no empirical muscle
• Watch for: A major marketing analytics platform (think Meridian, Robyn, or a Nielsen successor) adopting this framework with a real-world validation study — that would be the signal this moves from paper to practice

Forecast as of 2026-09-11

By Q3 2027, at least one marketing mix modeling vendor will ship a product explicitly referencing GMMM-style GEO attribution — but fewer than three will have published a peer-reviewed or audited validation against actual revenue outcomes, leaving the “notice probability” assumption still unresolved in practice.

Source: Generative Marketing Mix Modeling: A Causal Inference Framework Linking GEO and GEM to Business Impact — Masahiro Kato, Daiki Honma, Taka Kato. https://arxiv.org/abs/2609.11915v1