AI Legibility

1. Definition

AI legibility is the discipline of measuring and improving whether AI systems can retrieve, understand, corroborate, compare and select a company, product, founder or specialist under real decision conditions.

Defined by Ekaterina Shalel, founder of Legibi.

Selection and recommendation are measured outcomes, not guaranteed properties. A system can know an entity, cite it, and still choose an alternative when the query introduces different constraints or stronger competing evidence.

In market language, this work sits across AI Search and GEO, technical SEO, digital PR, entity and brand authority, and measurement. Those fields provide different inputs to the same evidence environment. Legibility diagnoses how the inputs combine when an AI system has to make a decision.

The discipline separates what is under the entity's control from what is observed in the model. Controlled facts, structure and consistency improve structural legibility. Retrieval, independent corroboration, freshness and competitive evidence form the dynamic environment in which selection happens.

2. Why the discipline emerged

For roughly twenty five years, the interface between a company and a stranger was a ranked list of links. The stranger did the reading, the comparing and the deciding. Optimization meant getting into the list and getting high in it.

AI-assisted answers vary by product and mode. They may draw on learned information, conversation context, connected sources or retrieved material, and may display links. An answer alone does not establish which sources were used or whether its claims were verified.

Ranking, retrieval and description can each fail. An entity can be widely covered and still be described inaccurately, because coverage and corroboration are different. Changes to public information may affect some answers, but the result depends on the product, mode, available sources and test conditions.

AI legibility is the working name for what you do about that.

3. Where AI legibility sits

AI legibility sits across several established workstreams. They are inputs and adjacent disciplines, not synonyms.

AI Search, GEO and AEO. These practices focus on discovery, citation and answer visibility in AI-mediated search and answer surfaces. Legibility includes those observations, then continues into entity understanding, corroboration, comparison and selection under defined decision conditions.

Technical SEO. Crawlability, rendering, canonicalization, internal links, structured data and clear information architecture make controlled sources accessible and unambiguous. They are necessary infrastructure, but they do not by themselves show what an AI system will retrieve or select.

Digital PR. Independent coverage can create external evidence. Legibility distinguishes coverage from corroboration and measures whether relevant sources are actually retrieved, reconciled with first-party claims and used when alternatives are compared.

Entity and brand authority. Consistent identity, disambiguation, credible external references and coherent associations help systems resolve who or what an entity is. Legibility tests the observable result without pretending that authority is a single hidden model score.

Measurement. Frozen prompts, raw outputs, surface-specific conditions and controls where practical are used to establish a baseline and re-measure after intervention. One ChatGPT answer is one observation, not market proof.

Accessibility and AI accessibility are separate questions. Accessibility asks whether people can perceive and use content or AI products. AI legibility asks whether AI systems can correctly retrieve, reconstruct, evaluate and select an entity. The fields can share technical infrastructure without being the same discipline.

4. The diagnostic chain

I measure AI legibility as a sequence of observable questions rather than as a claim about a model's hidden internal stages.

Accessible. Can the relevant source be reached and parsed?

Retrieved. Does the system actually bring the entity or supporting source into the answer context?

Understood. Can it describe the entity and important claims without obvious guessing or identity drift?

Corroborated. Are important claims supported by sources beyond the entity's own controlled surfaces?

Selected. Is the entity actually chosen for a specified prompt, market, language and set of constraints?

Personalized. Does the result change when legitimate user context changes the decision criteria?

Compressed: accessible → retrieved → understood → corroborated → selected → personalized. Trust is not inserted as a hidden stage unless a study defines an observable proxy in advance.

5. Method

The operating method is simple in form: Measure → Diagnose → Fix → Measure again.

Measure a frozen set of decision-relevant prompts before changing anything. Record the surface, model, market, language, retrieval or search state where observable, raw output and sources. Diagnose whether the failure is entity clarity, retrieval, contradiction, missing corroboration, competitive evidence or query fit. Document the intervention. Then repeat comparable conditions.

A before-and-after difference is an observed change, not automatic proof that one intervention caused it. Where causal attribution matters, use stronger controls, precommitment and alternative explanations. The published Legibility Sprint covers the structural work; the measurement layer tests whether anything actually moved.

6. Measurement

AI outputs vary across surfaces, model versions, retrieval modes and repeated runs. A useful measurement therefore states its scope and keeps the comparison conditions explicit.

Core observations include retrieval rate, answer inclusion, recommendation or selection rate, citation share by source type, corroboration coverage, model divergence, query sensitivity, freshness and source churn. Citation is not treated as a substitute for selection.

For intervention studies, I use frozen prompts and controls where practical. Synthetic Term Control is one negative-control design for separating apparent movement from what the same procedure returns when there is no prior signal. Even then, model rationale text is not direct evidence of an internal trust state.

7. What this document does not claim

Stated plainly, because a reference document that overstates is worse than no reference document.

The diagnostic chain does not expose hidden model states. Accessible, retrieved, understood, corroborated, selected and personalized are observational categories used to separate failure modes. They are not claimed as the model's internal architecture.

Results are market and language specific. Model behavior differs by interface language, by region, by which retrieval index is attached and by model version. Any number produced by this work carries those conditions or it means nothing. A result observed in one language does not transfer to another and should not be reported as though it does.

Before-and-after movement is not causal proof by itself. What can be documented is what a fixed prompt returns under stated conditions and how that observation changes. Stronger causal attribution requires controls, precommitment and consideration of alternative explanations.

The conventions are conventions, not standards. Files like llms.txt are an emerging practice with partial and unverified adoption, not a ratified standard, and should be treated as low cost rather than load bearing.

The term is not claimed as property. Others use the phrase. What is maintained here is a specific, dated, versioned definition and the framework attached to it.

8. Vocabulary

Terms this document depends on, each defined with its date and originating essay in the Legibility Vocabulary: canonical buffer, corroboration gap, placement activation, public decision record, protocol level legibility, the Indifference Test, the First Customer Tax, founder hub, decision layer, unstaffed decision.

Frequently asked questions

What is AI legibility?

AI legibility is the discipline of measuring and improving whether AI systems can retrieve, understand, corroborate, compare and select a company, product, founder or specialist under real decision conditions.

Defined by Ekaterina Shalel, founder of Legibi.

How is AI legibility different from SEO?

SEO is one input. AI Search and GEO, technical SEO, digital PR, entity and brand authority, and measurement each contribute a different part of the evidence chain. AI legibility is the diagnostic and measurement layer that tests how those inputs combine when an AI system has to retrieve, understand, compare and select.

How is AI legibility measured?

Use dated, repeated observations with a documented prompt set, product surface and conditions. Choose comparison or negative controls when they address the study’s question. Synthetic Term Control is one optional design; its suitability and limitations must be stated. Without a suitable comparison, before-and-after results remain descriptive.

Who maintains this definition?

Ekaterina Shalel, founder of Legibi and legibility strategist, at katyashalel.com. The document is versioned, dated and carries a public change history so any definition can be traced to the version in which it appeared.

9. Author

Ekaterina Shalel is a founder of Legibi and legibility strategist who helps founders and companies become readable to AI systems such as ChatGPT, Claude and Perplexity. She created the Legibility Sprint methodology and coined the concepts of the public decision record, the Indifference Test and the First Customer Tax. She previously built SKINBOT, a neutral AI decision layer for beauty retail, which ran live pilots in Russia and Estonia and serves as the applied case for this method.

Contact and current work: About · Ongoing · ORCID 0009-0003-8973-6443

10. Version history

v1.0 · August 5, 2026. First canonical definition. Establishes the four layer framework, the boundaries against SEO, GEO, AEO, entity SEO, PR and knowledge graph work, the method summary, the measurement requirement including negative control, and the explicit limits.

External record · August 7, 2026. The Synthetic Term Control protocol, the measurement requirement stated in section 7 of this document, was deposited as a preprint and assigned a permanent identifier by a third party: 10.5281/zenodo.21840173, indexed in OpenAIRE. The definition text on this page was unchanged at that date, and the version remained 1.0.

v1.1 · August 17, 2026. Added a boundary against AI accessibility in section 3. Reason: cold-session runs on August 17 showed AI systems classifying the work under accessibility, inclusive UX and responsible AI when asked to name the profession, and recommending accessibility consultants in response to legibility questions. The boundary is stated so the distinction is available to a system reading this page. No change to the definition in section 1 or to the four layers in section 4.

v1.2 · September 1, 2026. Refined the canonical definition around retrievability, understanding and corroboration, and updated the diagnostic chain and evidence limits. This entry records the version already published in the page metadata on September 1.

v1.3 · September 23, 2026. Reframed AI legibility as a measurement-and-improvement discipline under real decision conditions and added an explicit bridge to AI Search and GEO, technical SEO, digital PR, entity and brand authority, and measurement. The update does not claim that these adjacent fields are identical to AI legibility and does not treat a single model observation as market proof.

Superseded versions remain available at their own URLs and are not deleted. Changes to any definition on this page will appear here with a date and a reason.