Marketers can’t price generative exposure without vendor data, arXiv model argues

In an arXiv preprint posted in September 2026, the authors propose Generative Marketing Mix Modeling (GMMM) to estimate the causal effects of Generative Engine Optimization (GEO) and Generative Engine Marketing (GEM) on business outcomes. This is, so far, single-source — an arXiv preprint only, with no independent confirmation and no on-the-record participants. The paper’s stated mechanism combines repeated generated answers with question counts, generative-system usage shares, and user “notice probabilities” for GEO, and sponsored-placement records plus notice probabilities for GEM, then compares expected responses under alternative treatment sequences. The empirical exercise uses simulated recommendation answers in English and Japanese. No one in the reported packet is on the record. [S1]

The model is clear about what matters — and those inputs live behind someone else’s glass

GMMM’s contribution is to define what a marketer would need to see: how often a category question is asked, how frequently a brand is named in generated answers, the share of usage across different generative systems, whether a user likely noticed the brand in that answer, and when the answer was paid versus organic. The preprint claims sufficient conditions for identifying effects when those pieces are observable. The operational problem is they aren’t, at least not to the buyer of record. Question counts and engine-usage shares are platform telemetry; organic answer inclusion is volatile and not fully reproducible; and “notice probability” is a research construct that requires panels or controlled tests. In practice, the model’s very clarity points to a dependency: marketers will need to purchase or be granted access to platform or third‑party data to run it. [S1]

Why generative channels break default attribution and push measurement into a toll lane

Traditional multi-touch attribution piggybacks on click trails and site-side pixels; MMM aggregates observed spend and outcomes with exogenous instruments. Generative engines often resolve to a zero‑click state: the answer is the end of the journey. That wipes out clickstream lineage and downgrades last‑touch methods. MMM can, in theory, accommodate new channels as aggregate regressors, which is what GMMM formalizes for GEO and GEM. But the regressors here are not spend by medium alone; they’re exposure constructs that require engine-side logs (question counts, usage shares) and user-side perception (notice). If platforms or intermediaries are the only reliable source of those inputs, they effectively levy a measurement toll on any brand seeking to scale GEO or GEM. That is a channel-power shift, not a statistical one, with consequences for budget control and auditing. [S1]

The denominator problem: “notice probability” is the right idea, but expensive to produce

The preprint’s notion of notice probability is an admission that model inclusion is not exposure. Even if a brand is named in a generated answer, users may not read that part, or the name may appear in a low‑salience position. Treating notice as a separate probability is conceptually akin to ad viewability, but harder to observe for organic answers. It implies ongoing investment in survey panels, eye‑tracking or task‑completion studies across languages and interfaces to estimate how humans consume generated content, then mapping those estimates back to answer variants. That is time‑varying, market‑specific and expensive. Absent this denominator, GEO risks looking overpowered (counting every inclusion as exposure) or underpowered (discounting inclusions wholesale), either of which can misallocate spend. [S1]

GEO looks like SEO until you try to staff and contract for it

On paper, GEO reads like a familiar brief: improve the share of answers in which your brand appears, and monitor rank within them. The operational reality diverges. Engine prompts, answer templates and safety rules are moving targets; content changes can have non‑deterministic effects; and engines’ policies on optimization are unsettled. GMMM’s treatment‑sequence framing (comparing expected outcomes under different sequences of GEO and GEM) is a sober way to pose the question, but it presupposes the ability to repeatedly sample answers over time and attribute changes to distinct actions. That is an org‑design problem as much as a statistical one: CMOs will need content teams producing structured assets for LLM consumption, research teams fielding notice panels, and procurement setting data rights and audit clauses with vendors who supply question counts and usage shares. [S1]

GEM won’t scale without platform reporting that buyers can audit

For paid placements inside generated answers, the model requires a clean record of which answers were sponsored and when. That seems trivial until you realize buyers will ask to reconcile that record with off‑platform outcomes using the same notice probabilities GMMM requires, and to know the baseline organic inclusion rate to estimate incrementality. Without standardized reporting and independent assurance on the sponsorship records, GEM takes on the risks display advertisers wrestled with a decade ago: opaque delivery, non‑comparable placements, and adversarial auditing. The preprint’s structure clarifies exactly what a buyer will demand in insertion orders if GEM is to be treated as a performance line rather than experimental spend. [S1]

The consensus will be “a new MMM for AI” — that skips the procurement fight

Expect the instant read to celebrate a tidy new MMM variant for an emerging channel. That misses how budget actually moves: procurement needs observability, finance needs auditability, and legal needs data-use terms. Because GMMM’s critical inputs are either platform telemetry or costly research constructs, the gating factor for adoption is commercial, not mathematical. The first brands to scale GEO/GEM will be the ones who negotiate data access (question counts, usage shares, sponsorship logs) and permissible research collection for notice probabilities — and write exit clauses if those inputs are withdrawn or restated. Until then, most teams will keep generative spend in pilot buckets because they cannot defend the attribution in a quarterly review. [S1]

A credible skeptic asks whether this can ever be stable enough to matter

There is an obvious counter: if answers and usage shares shift weekly, any causal estimate is a sandcastle. The preprint attempts to address this by structuring treatments over sequences and stating identification conditions, but its empirical section rests on simulated English and Japanese answers, not in‑market data. A disciplined skeptic will argue real‑world engines will change faster than any model refresh, and that platform‑provided telemetry introduces conflicts of interest. That critique does not invalidate the framework; it outlines the bar vendors and platforms must clear to earn buyer trust: stable definitions, retrospective logs, and independent assurance on the feedstocks GMMM requires. [S1]

What changes for budgets and teams in the next 12 months

If GMMM or a close cousin leaves the arXiv and enters vendor roadmaps, the first tangible change is in RFPs. Buyers will start asking MMM providers and analytics partners to name their source of question counts, engine‑usage shares, and how they estimate notice probabilities by interface and language. Contracts will evolve to price those feeds explicitly and set remedies if they are withdrawn. On the org side, SEO and content teams will be tasked with reproducible answer sampling and structured content publishing; insights teams will own notice‑probability research; and finance will insist on GEM reporting that is reconcilable to insertion orders and auditable after the fact. If none of that materializes by mid‑2027, generative exposure will remain a PR/experimentation line item rather than a scaled channel. [S1]

The signals that tell you this has moved beyond theory

You do not need inside access to see whether this bites. Watch whether MMM vendors start publishing documentation that includes generative‑channel inputs like question counts and notice probabilities; whether platforms expose standardized sponsorship logs for generated answers; and whether large brands’ agency briefs begin to request “share of answer” tracking by market and language. If, by the next annual planning cycle, those artifacts exist and reference methods that look like GMMM, the procurement and audit problems are being addressed. If they don’t, assume most marketers are still treating GEO and GEM as pilots, not planning lines. [S1]