Meridian’s GeoX update makes the calibration assumption worth reading

GeoX integration makes experiments easier to carry into a model while keeping counterfactual and positive-support assumptions open to review.

A marketing model can absorb an experiment result without preserving the question that experiment answered. Meridian’s recent GeoX and calibration work makes that handoff easier to automate, but the documentation also exposes conditions that deserve human attention. For an agency evaluating AI-assisted measurement, the practical task is to preserve the experiment’s counterfactual and uncertainty as the result becomes an input to a wider model.

This source-based analysis was checked on 16 October 2026. The Netherlands bicycle-cooperative exercise below is fictional and unrun. No dataset, notebook, model or campaign was executed, and no budget recommendation or measured advertising return is offered.

A release chronology, rather than a single launch headline

An official maintainer [announced Meridian 2.0 and GeoX general availability on 4 September 2026](https://github.com/google/meridian/discussions/1819). Google’s [measurement-suite article on 10 September](https://blog.google/products/ads-commerce/data-strength-updates/) subsequently described global general availability and the use of experiment results with marketing mix modelling. The later blog is therefore not evidence that GeoX first became available on that date. The earlier maintainer announcement belongs in the chronology.

The [GeoX overview](https://developers.google.com/meridian/geox) describes an open-source, publisher-agnostic geographic experimentation solution that can operate independently or help calibrate marketing mix models. That is a description of intended scope, not proof that a proposed experiment is feasible for a particular advertiser. A team still needs an appropriate outcome, comparison design and usable data before interpreting a result as evidence for its own decision.

The current [calibration guide](https://developers.google.com/meridian/docs/advanced-modeling/set-custom-priors-past-experiments) documents automated translation of experiment results into model priors. It warns about a mismatch between experiments against a non-zero-spend counterfactual and the model’s zero-spend counterfactual. It also notes that the default positive-support baseline discards probability below zero, which can affect uncertain or negative estimates. Those are documented modelling boundaries, not a software failure reproduced by this article.

There is a longer history behind the update. Google’s [29 January 2025 announcement](https://blog.google/products/ads-commerce/meridian-marketing-mix-model-open-to-everyone/) described Meridian’s open availability, aggregated-data approach and use of experiment-informed priors. The new integration should not be described as the invention of experimentation or marketing mix modelling. Its practical interest is the increasingly explicit connection between tools and the assumptions carried across that connection.

Preserve the question attached to an experiment

Before importing a result, write down the intervention and its alternative. Was the experiment comparing advertising with no advertising, or more advertising with less? Which locations, products and outcome period were included? A numerical estimate without those details can look portable while answering a different question from the one the receiving model is designed to address. The proposed review should expose that mismatch before model fitting.

Keep the estimate and its uncertainty together. A point estimate alone is an incomplete handoff for a decision involving uncertain effects. The analyst should retain the experiment report, outcome definition, comparison and limitations rather than transcribe only the most favourable number into a presentation. This is a proposed evidence discipline, not a claim that a particular advertiser has omitted those records.

The calibration documentation’s warning is especially useful when results are uncomfortable. A model configuration can constrain which effects its distributions can represent. A reviewer should therefore inspect the chosen support and resulting distribution rather than assume a positive-looking summary establishes a positive campaign effect. That observation is an analytical implication of the documented boundary, not a computed result or advice to select a different prior solely to obtain a preferred answer.

Ask a qualified analyst to justify the configuration using the measurement question and evidence. Preserve alternative explanations, data limitations and sensitivity checks, including cases where the result remains inconclusive. Do not use an AI assistant to fill a missing assumption with plausible prose. If a reviewer cannot explain how a key result entered the model, the workflow should remain a draft analysis rather than become a confident allocation recommendation.

Fictional exercise: bicycle servicing across regions

Imagine a fictional Netherlands bicycle cooperative evaluating promotion of its servicing appointments. This exercise contains no actual sales, geographic units or experiment output. The proposed business question is whether a specific advertising intervention creates additional completed appointments. The analyst would need to establish suitable comparisons and data permissions before any real evaluation; the mere existence of GeoX does not establish that this cooperative can run a valid study.

The exercise’s draft records distinguish booked appointments from completed paid work and retain cancellations separately. They also identify proposed changes in workshop capacity and local promotions that could complicate interpretation. Those conditions are invented for learning. They show the kind of context a model handoff needs, not an empirical explanation for an observed change in bicycle demand or an actual cooperative’s operations.

The team then considers a hypothetical report comparing a higher advertising level with a lower one. It does not assign a numerical lift. Before importing anything, the analyst flags the counterfactual for specialist review because the calibration documentation describes a different baseline for its intended experiment inputs. The safe exercise outcome is a held handoff, not an invented compatible estimate or an automatic conversion of marginal effect into total effect.

A separate proposed review asks how an uncertain or negative result would be represented under the chosen model configuration. The analyst would examine the configured support, original estimate and resulting distribution together. This article does not run that calculation. Its exercise deliberately leaves space for a specialist to determine whether the configuration faithfully represents the evidence and whether any alternative is defensible.

The final hypothetical packet contains the experiment question, outcome definition, report version, proposed receiving model and unresolved assumptions. No budget is moved. The agency can explain what is ready for analysis and what needs review without pretending that a successful file import is a successful causal evaluation. That distinction also preserves useful work when the correct immediate decision is to stop.

Keep automation useful and accountable

A sensible workflow can automate clerical transfer while reserving interpretive decisions for a named person. The person should be able to trace the model input back to the original experiment and explain any adjustment. Record the software version used for a future authorised run, because documentation observed today does not prove how an older installed version behaves. This article has inspected sources rather than a local installation.

For cross-platform agency work, maintain comparable business definitions before comparing channel estimates. A sale, a visit and an appointment are not interchangeable simply because each appears in a conversion column. Preserve the boundaries of each study and avoid using a general product description as evidence that all channels or markets have equivalent measurement quality. Worldwide tool availability is not worldwide validity of every proposed experiment.

Separate the output’s technical completion from its decision status. A model may finish while an assumption remains unresolved. A useful human-facing report says which evidence supports the conclusion, which limitations remain and what decision is being held. It should not expose raw tool failures to a client or disguise a missing review as a successful recommendation. The underlying diagnostic record can remain restricted and appropriately redacted.

Meridian’s update is newsworthy because its own documentation makes the experiment-to-model boundary inspectable. The strongest agency deliverable is therefore more than an automated model run: it is an understandable account of what was measured, how uncertainty survived the handoff and why the result supports, or does not yet support, the next decision. That remains useful even when the honest conclusion is inconclusive.