Causal attribution in GEO requires a counterfactual

Position

An increase in mentions, citations, or share of presence after a GEO intervention does not prove that the intervention produced the increase.

It proves one thing only: a change was observed.

To attribute that change to the intervention, the observed outcome must be compared with what would likely have happened without the intervention. That second state cannot be observed directly. It must be constructed through a sufficiently credible experimental or quasi-experimental design.

This requirement has a simple name: the counterfactual.

Without a counterfactual, an agency may legitimately state that an indicator increased. It cannot state with equal strength that its work caused the increase.

This position extends GEO metrics do not govern representation by isolating a narrower problem: the non-identification of causal effects in a generative environment that is expanding and continuously changing.

The problem the market prefers not to address

The GEO market now sells dashboards capable of counting:

• brand presence across a panel of answers;
• domain citation frequency;
• apparent share of voice against competitors;
• coverage of a query set;
• variation in selected attributes;
• the number of systems in which an entity appears.

These observations may be useful. They become methodologically abusive when an increase is presented as the demonstrated consequence of a GEO program.

The implicit commercial argument often follows this sequence:

1. an intervention is launched;
2. citations increase;
3. the intervention therefore produced the increase.

This is a before-and-after comparison, not causal attribution.

Between the two dates, many other conditions may have changed:

• the number of people using AI systems;
• how often generative surfaces answer relevant queries;
• the composition of real user queries;
• models, corpora, retrieval systems, and citation policies;
• the brand’s general awareness;
• the organization’s editorial, media, or commercial activity;
• competitor and third-party source activity;
• seasonality;
• the prompt panel used by the measurement tool;
• stochastic output noise.

An increase may therefore be fully real while remaining partly, mostly, or entirely independent of the work being sold.

The problem with GEO is not only that it measures unstable outputs. It often claims credit for a change it has not isolated.

Minimal causal formulation

Let Y(1) be the level of generative visibility observed when the intervention is applied, and Y(0) the level that would have been observed at the same time, under the same conditions, without the intervention.

The target causal effect is:

τ = Y(1) - Y(0)

After an intervention, Y(1) may be observed. Y(0) is not. This is the fundamental problem of causal inference.

A before-and-after comparison observes instead:

Yafter - Ybefore

That difference may contain:

Yafter - Ybefore
= intervention effect
+ change in generative usage and exposure
+ model and retrieval changes
+ prompt-panel composition changes
+ competitive and media changes
+ seasonality
+ noise

Until these components are separated through a defensible study design, the intervention’s own effect remains unidentified.

A rising curve after an action is compatible with a positive effect. It is also compatible with no effect in a rising market, with a negative effect masked by general growth, or with a changed measurement instrument.

No agency owns the rising tide

When overall exposure to generative answers increases, more brands may be observed, mentioned, or cited even when their documentary infrastructure has not changed.

This exogenous growth behaves like a rising tide. An agency that only measures the final waterline may claim a gain that would have happened without it.

The correct question is not:

Did citations increase after our intervention?

The correct question is:

Did citations increase beyond what would have been expected without our intervention, under a stable protocol and within a declared scope?

The second question is much harder. It requires a control, preserved observation conditions, a log of external changes, and acceptance that the answer may be null or non-identifiable.

That is precisely why it is rarely placed at the center of the commercial narrative.

Four levels of evidentiary language

The language used should match the actual strength of the design.

Level 0: observation

The observed mention rate in this panel moved from X to Y between two windows.

This statement describes a protocol fact. It does not attribute the change.

Level 1: temporal association

The increase appeared after the intervention and remains visible across comparable repetitions.

This establishes sequence and persistence. It still does not isolate the cause.

Level 2: plausible contribution

The increase exceeds that of a comparable group, survives several sensitivity analyses, and matches the intervention’s expected mechanisms.

This is stronger, but still depends on the comparison group, pre-trends, and unobserved factors.

Level 3: causally supported effect within a defined scope

Under the declared study design, the intervention produced an estimated effect of this magnitude on this metric, under these conditions, with this uncertainty and these limitations.

This statement requires a credible counterfactual, falsification tests, and a narrow scope. It must never become a universal promise across all models, prompts, or future periods.

What a provider must publish to claim an effect

An admissible causal claim should disclose at least:

1. the intervention: what actually changed, when, on which surfaces, and alongside which co-interventions;
2. the treated unit: pages, clusters, entities, markets, languages, prompts, or corpora;
3. the measured outcome: mention, citation, source selection, fidelity, recommendation, or another distinct result;
4. the denominator: queries, repetitions, systems, windows, and admissible observations;
5. the counterfactual: control group, staggered rollout, synthetic control, controlled time series, or another defensible strategy;
6. pre-intervention trends: the prior trajectory and comparability of the control;
7. external changes: model updates, coverage changes, media events, seasonality, and panel changes;
8. negative tests: queries, surfaces, or entities that should not respond to the intervention;
9. uncertainty: intervals, variability, sensitivity, and critical errors;
10. limitations: what the design cannot support.

Without these elements, the report may remain a useful observation tool. It is not causal proof.

Why an “AI visibility score” solves nothing

An aggregate score may simplify reading. It does not create the missing counterfactual.

It may worsen the problem when its composition changes silently:

• new prompts are added;
• unavailable answers are removed;
• languages are mixed;
• consumer interfaces, APIs, and tool-using agents are pooled;
• one model replaces another;
• weights change;
• the competitive set evolves;
• favorable prompts are selected after the fact.

A visually stable score may therefore measure a different object from one window to the next.

A percentage increase without a denominator, panel version, or change log creates a facade of precision. The cleaner the number looks, the more effectively it can hide a dirty measurement design.

Visibility and fidelity answer different questions

This position does not replace GEO metrics with another universal score. It imposes a separation.

Visibility asks:

Does the entity appear more often across a set of outputs?

Fidelity asks:

When the entity is reconstructed, are its attributes, relations, exclusions, limits, and authority sources preserved?

These outcomes may move in opposite directions.

A brand may become more visible while being understood less accurately. It may be cited more often because an incorrect simplification has stabilized. Conversely, stronger governance may significantly improve answer quality without producing an immediate open-web citation increase.

It is therefore methodologically wrong to treat fidelity as a submetric of visibility, or visibility as proof of fidelity.

Where governance can be demonstrated more cleanly

The open web makes attribution difficult because exposure, models, corpora, and retrieval mechanisms remain largely outside the observed organization’s control.

A controlled environment supports a different demonstration.

The same corpus can be compared:

• without governance files;
• with governance files;
• with governance files and a restitution runtime;
• under the same models, prompts, parameters, snapshots, and evaluation criteria.

The outcome is no longer “how often is the brand cited across the Internet?”. It becomes:

To what extent does governance reduce the gap between the canon and the produced reconstruction?

The comparison may assess:

• factual accuracy;
• preservation of critical attributes;
• preservation of entity relations;
• retention of exclusions and limits;
• resistance to forbidden inferences;
• stability across formulations;
• legitimate clarification or non-response;
• traceability to canonical sources.

The substance may remain qualitative while the design remains experimental. Answers are compared against a declared rubric, critical errors remain non-compensable, and evaluators may be blinded to the tested condition.

This is a cleaner demonstration than claiming a global citation uplift because the intervention and comparison conditions are controlled.

The exact role of SEO

SEO does not disappear from this framework. It returns to its foundational role.

It helps build the documentary corpus:

• information architecture;
• entity resolution;
• page hierarchy;
• content coherence;
• structured data;
• accessibility, indexability, and retrievability;
• links between concepts and sources.

This infrastructure makes information available and interpretable. Governance then adds boundaries that SEO does not always formalize:

• which source takes precedence;
• which interpretation is prohibited;
• which limit must survive synthesis;
• which information has been revoked or expired;
• when an answer should be suspended;
• how a contradiction should be arbitrated.

SEO builds the documentary foundation. Governance constrains its reconstruction. Any generative visibility that follows is a possible consequence, not automatic proof of causality.

Normative rules

1. No before-and-after comparison should be presented as causality

A post-intervention increase remains descriptive until a credible counterfactual is established.

2. No percentage without a denominator and versioned panel

The number of prompts, systems, repetitions, languages, and admissible observations must be disclosed.

3. No silent pooling of access regimes

API, consumer interface, integrated search, agentic browsing, and personalized systems must remain separate.

4. No automatic credit for general growth

Growth in usage or generative-surface coverage must be treated as an external factor, not a client intervention outcome.

5. No conflation of visibility and fidelity

Presence indicators and reconstruction evidence must be reported separately.

6. No positive result without falsification tests

Placebo dates, negative controls, untreated queries, panel variants, and trend checks should be used whenever the design allows them.

7. No hiding null results

An intervention may produce no detectable effect. Its effect may also be too small or too confounded to identify. Both are legitimate methodological verdicts.

8. “Not identifiable” is sometimes the only honest conclusion

Lack of causal proof does not prove absence of effect. It only prohibits selling the effect as demonstrated.

Consequence for the market

A significant part of the GEO market may not be selling a false phenomenon. It is selling more certainty than its methods permit.

The problem is not counting. The problem is turning counts into a causal narrative without a comparison group, change log, stable protocol, or possibility of refutation.

A serious provider should sometimes say:

Observed visibility increased, but the specific effect of our intervention is not identifiable under the current design.

That statement is less spectacular than a rising chart. It is methodologically stronger.

Scope and limitations

This position does not claim that GEO interventions are always ineffective. It claims that an effect cannot be inferred from raw growth.

It does not claim that a perfect counterfactual is always available. On the open web, model changes, spillovers, personalization, system opacity, and contamination between units often make identification partial.

It also does not turn causal inference into certification. Even a robust design remains bounded by a unit, period, metric, and scope.

The final rule is simple:

Measurement can observe change. Only a defensible study design can support attribution.

Methodological references

The methods cited here are not specific to GEO. They come from the general literature on causal inference and intervention evaluation:

• Hernán and Robins, Causal Inference: What If, for the counterfactual formulation of causal effects;
• Callaway and Sant’Anna, Difference-in-Differences with Multiple Time Periods, for staggered-treatment designs and heterogeneous effects;
• Abadie, Using Synthetic Controls: Feasibility, Data Requirements, and Methodological Aspects, for transparent construction of synthetic comparison units;
• Bernal, Cummins, and Gasparrini, Interrupted time series regression for the evaluation of public health interventions: a tutorial, for evaluating temporal breaks under explicit assumptions.

These references support study-design principles. They do not automatically validate their application to a panel of generative answers. The GEO intervention causal attribution protocol adapts those requirements to AI systems.

Reading rule

Counting is not understanding. Before-and-after is not a counterfactual. A chart is not proof of causality.