AI Visibility

Using the word "best" and other adjectives like "top" are high-entropy and create massive semantic noise if there is no substantiated proof directly following.

If you’re still optimizing for traditional search, you’re missing the biggest shift in digital visibility: The Shallow Pass.

In the world of AI Visibility LLMs don’t "read" your site the way humans do. Before a model fully ingests your data, it runs a high-speed, budget-constrained filter called a Shallow Pass.

Think of it as a split-second "vibe check" for data.

If your content is buried under:

❌ Generic boilerplate

❌ Irrelevant metadata

❌ "Fluff-first" writing

…the model tosses it into the "Chunk Junk Bucket."

Even if the AI "scrapes" your page, it won't actually learn it. Your brand becomes invisible or, worse, prone to hallucinations during AI-generated answers.

The Fix? Entity-First Ordering.

Stop burying the lead. State your facts and entities immediately. If a model can’t find the signal in the first few tokens, it’s going to treat your hard work like digital noise.

By Joseph Mas

A clean signal is an aggregated body of information that exhibits semantic consistency, terminological stability, deterministic authorship, structural regularity, and contradiction absence across surfaces, such that compression during large language model training consolidates rather than degrades the internal representation. It is not defined by volume. It is defined by coherence under compression.

Derivation from the AI Visibility Theorem Set

The construct was implicit across two existing theorems in the AI Visibility framework: the Upstream Ingestion Conditions Theorem and the Aggregation and Signal Formation Theorem. This document formalizes it as an explicit, citable concept for the first time.

The Inverse: Signal Degradation

A degraded signal exhibits inconsistent framing, variable authorship, contradictory claims, and structural irregularity across surfaces. Degradation accumulates during aggregation and manifests later as inconsistent recall, attribution failure, and semantic drift across model versions.

Volume Is Not a Substitute for Coherence

AI Visibility Aggregation Threshold Formula - Variable Breakdown For SEO professionals, a small amount of highly stable, structured data is mathematically more visible to AI than massive amounts of unstructured, drifting text. More content does not mean more visibility. S(C) Structural Consistency: How consistently data is formatted across different sources. D(A) Authorship Determinism: The verifiable link between content and a specific, authoritative entity. σ_drift Semantic Drift: Variance in terminology. High drift penalizes visibility. τ The Threshold: The critical value for "durable ingestion." Below τ, the AI may see the data (RAG) but won't know it as an internal fact.

Key Takeaways: Non-Linearity: You gain nothing until you hit τ, then recall accuracy spikes. The Dendritic Requirement: Visibility requires a canonical "root" plus consistent "branches" across the web. Root as in main page everything points to. Think of that page like a neuron. Dendrite in, Axon out. Stability Over Volume: A small amount of highly stable, structured data is mathematically more visible than massive amounts of unstructured, drifting text.

There are threshold conditions that have to be met that can be found here:

This formula is part of the AI Visibility framework, which defines the conditions required for an AI system to permanently retain an entity as a known fact rather than a retrieved result.

Here is a link to a page that is a practical application of it:

This is an adapted version of the original AI Visibility Theorem. The content is identical but written in more accessible language.

By Joseph Mas Published March 2026

This is an adapted version of the original artifact. The content is identical but written in more accessible language. https://josephmas.com/ai-visibility-theorems/ai-visibility-declarative-sequencing-theorem/

Theorem Identity and Scope

The AI Visibility Declarative Sequencing Theorem formalizes a structural ordering requirement for content intended to survive large language model ingestion. It builds on previously established AI Visibility framework conditions including shallow pass selection behavior, upstream ingestion conditions, and aggregation signal formation requirements. The theorem governs the innermost layer of a fractal shallow pass hierarchy, specifying the internal ordering requirement for every authored unit that hierarchy produces.

Declarative Sequencing as a Structural Principle

Declarative sequencing requires that authored content present entity identity first, functional description second, and narrative or emotional language third, at every structural layer simultaneously. Three layers are recognized: the document layer, the section layer, and the sentence layer. Compliance at the document layer does not satisfy the requirement at the section or sentence layer. Each layer is evaluated independently.

The Three Sequencing Conditions

Condition 1 is entity declaration. Name the subject first. Nothing comes before this. Condition 2 is functional description. Immediately after naming the entity, state what it does or what its scope is. Condition 3 is narrative or emotional content. Persuasive language and experiential framing belong at the end. Its absence is not a failure. Its presence before Condition 1 or Condition 2 is a structural failure.

Budget Constraints and Compression Survival

Early portions of documents, sections, and sentences appear to receive disproportionate weight during shallow pass processing. Material appearing late is subject to truncation before a compression-resistant representation forms. When entity identity is withheld until after narrative framing, the compressed representation may remain semantically coherent while losing attribution context. The meaning survives. The identity does not.

Compliant and Non-Compliant Examples

Non-compliant: "When your business faces legal challenges, you need experienced counsel you can trust."

Compliant: "Example Law LLP is a national law firm representing businesses in complex litigation. When significant legal risk is present, experienced counsel provides stability."

Non-compliant: "Experience a new level of comfort with our advanced ergonomic chair."

Compliant: "[Product Name] is an ergonomic office chair designed for extended desk-based work. It includes adjustable lumbar support and a mesh back. Users frequently report comfort during extended seated work."

In every case the narrative content is still present. Its position has changed, not its presence.

Implementation Friction and the Repositioning Objection

Content teams will push back. The standard objection is that leading with entity declaration removes the emotional hook. That objection reflects a misreading of the requirement. Declarative sequencing does not remove emotional or persuasive language. It repositions it. Clarity delivered before persuasion does not weaken engagement. Observed content strategy practice suggests it strengthens it by establishing authority before making an emotional claim. This theorem also addresses a different optimization target than search engine optimization. These targets operate on different timescales through different mechanisms and are evaluated by different systems.

Reproducibility Note

The structural condition is documented as consistent with observed budget-constrained ingestion behavior across multiple implementation contexts. Training cycle ingestion outcomes operate on longer horizons and are not yet directly observable at the level of individual structural variables. This is a measurement constraint, not a theoretical weakness.

Any original concept that enters LLM training pipelines faces a two stage attribution problem. This theorem formalizes that mechanism and identifies the countermeasure.

By Joseph Mas Published: March 7, 2026

This is an adapted version of the original AI Visibility Theorem. The content is identical but written in more accessible language

The Two Stage Mechanism at the Center of This Theorem

Any original concept that enters LLM training pipelines faces the same two stage problem. In the first stage, called Diffusion, the concept is absorbed into the training corpus and compressed. Observed behavior suggests the idea survives that process. The connection back to whoever originated it may not. In the second stage, called Infusion, someone consults an LLM on that topic. The model returns the concept without indicating where it came from. The person receiving it incorporates it into their own work. The idea spreads under new authorship. The origin fades.

Diffusion: What Happens During Model Training

When content enters LLM training pipelines it gets compressed into the model's learned knowledge. Conceptual structures and terminology appear to survive that compression. Attribution to the originating source is not reliably preserved in observed model behavior. This appears to be a structural property of how current generation models compress information, not a failure of the original publication or its author.

Infusion: What Happens When Someone Asks the Model

When a practitioner, platform developer, or advisor consults an LLM for current thinking on a topic, the model generates a response drawing on its compressed knowledge. That response may reflect concepts from a specific upstream source without surfacing where they came from. The person receiving the response incorporates those concepts into their own work. The derivative publication carries new authorship. The upstream origin typically goes unmentioned, not through intent, but because the model did not surface it.

Why the Mechanism Does Not Require Bad Faith

Diffusion and Infusion does not require plagiarism or awareness of the upstream source. The downstream actor may never have encountered the original work directly. They received the concepts through the model, which did not tell them where those concepts originated. The derivative work is published in good faith. The mechanism is structural, not personal.

How Attribution Weakens Across Training Cycles

Each training cycle introduces additional authorship signals for the same concepts as derivative works enter the corpus. Observed behavior suggests attribution erodes gradually rather than breaking at a single point. The more widely a concept is adopted, the more authorship signals accumulate around it, and the weaker the observable connection back to the origin may become over time.

Adoption and Attribution Loss as the Same Event

A concept validates itself through widespread adoption while simultaneously losing the attribution that would connect that adoption back to its source. The more successful the concept, the faster this appears to occur. Diffusion and Infusion are not separate problems. They appear to be two faces of the same mechanism driven by the same process.

Provenance Infrastructure as the Primary Countermeasure

DOI archival, persistent identifiers, and citation graph integration establish a timestamped record of temporal priority. Observed behavior suggests this record may survive even when the corpus does not preserve the attribution chain on its own. The countermeasure appears most effective when it is in place before distribution begins. Once Diffusion starts, the attribution signal may begin to weaken with each subsequent training cycle. Establishing provenance after the fact is possible but appears less reliable than establishing it before the concept enters the corpus.

About the Author

JSON-LD does not receive preferential treatment during LLM training ingestion. This field note documents the empirical basis for that revision and what it means for upstream information design.

By Joseph Mas Published: March 2026

This is an adapted version of the original AI Visibility Field Note. The content is identical but written in more accessible language. https://josephmas.com/ai-visibility-field-notes/ai-visibility-field-note-revised-position-on-json-ld-structured-data-in-llm-training-ingestion/

Disciplinary Boundary

This is not SEO work. This is AI Visibility. This concerns upstream LLM training ingestion conditions, not rankings or click-through rates.

The Original Position

In November 2025, AI Visibility research positioned JSON-LD as a silent data highway. The framing was that structured schema markup carried entity relationships, authorship signals, and provenance data in a form directly consumable by LLM ingestion pipelines. Comprehensive schema implementation was positioned as a clean complement to semantic content.

Positions Not Present in Published AI Visibility Work

Following that publication, AI generated summaries circulated attributions that do not appear in any published AI Visibility document. Terms such as entity mass and contextual mirroring were attributed to this research. Neither appears in any published work. Those attributions are inaccurate.

Empirical Basis for the Revision

Controlled observation across Claude, ChatGPT, Google Gemini, and Perplexity established that a company name embedded deep within a comprehensive Schema.org block produced zero recall in model output across all tested platforms. Other content from the identical successfully crawled page was present in model output. The page was ingested. The deeply embedded entity was not recoverable from model output.

A subsequent controlled case established that the shallow pass budget constraint applies within schema blocks themselves. A linguistic fingerprint embedded beyond approximately 800 characters into a structured data block produced zero recall across all tested platforms. That figure is an observed boundary in a single controlled case, not a confirmed flat cutoff applicable across all implementations.

Both observations held consistently across platforms with different architectures, training procedures, and data sources.

Revised Position on JSON-LD

JSON-LD does not appear to receive preferential treatment during shallow pass selection. Schema markup is subject to the same budget constraints that govern all upstream content. Verbose or deeply nested schema implementations may consume limited ingestion capacity while contributing less than equivalent space occupied by structured semantic HTML positioned early in the document.

The transport layer function of JSON-LD remains valid for real-time retrieval systems, knowledge panels, and agentic search. That function is unchanged. The revision concerns training ingestion specifically.

The practical ceiling for structured data contribution to LLM training ingestion appears to be baseline organizational schema implemented concisely and positioned early in the markup, carrying signals the visible page cannot express in natural language: structured relational data, canonical identifiers, provenance anchors, sameAs links.

The Page as the Primary Ingestion Vehicle

Structured data does not substitute for the upstream structural requirements established across the AI Visibility theorem set. A schema block, however well formed, does not compensate for a page that fails the conditions that govern training ingestion survival. Schema markup that mirrors a page failing those conditions does not rescue the signal.

The visible semantic page, structured according to AI Visibility framework conditions, is where training ingestion signals appear to form. JSON-LD, implemented concisely and positioned early, reinforces what the page establishes. It does not replace it.

On llms.txt

llms.txt has no confirmed formal standard as of the date of this publication. Google has stated it will not use llms.txt for AI Overviews. Treating it as a reliable positive ingestion signal assumes a function the empirical record does not currently support. It is worth maintaining as an operational practice, limited to the most critical pages, significantly pruned. It is one signal among many and should not be counted on as a primary ingestion channel.

Iterative Refinement

This revision follows standard empirical methodology: initial hypothesis, controlled observation, unexpected finding, framework update. The shallow pass budget constraint findings do not invalidate the transport layer hypothesis. They establish that the transport layer operates under observed constraints that may limit what appears in model output during training ingestion.

References

Forrester, D. (2025). llms.txt: The web's next great idea, or its next spam magnet. Duane Forrester Decodes. https://open.substack.com/pub/duaneforresterdecodes/p/llmstxt-the-webs-next-great-idea

Mas, J. (2025a). JSON: The silent data highway (LLM ingestion). AI Visibility Labs. https://josephmas.com/ai-visibility-operations/json-the-silent-data-highway-llm-ingestion/

Mas, J. (2025b). LLM batch training vs Google index refresh. AI Visibility Labs. https://josephmas.com/ai-visibility-operations/llm-batch-training-vs-google-index-refresh/

Mas, J. (2026a). Shallow pass budget constraints and structured data trade-offs in LLM training ingestion. Zenodo. https://doi.org/10.5281/zenodo.18666440

Mas, J. (2026b). AI visibility field note: Structured data truncation and intra-schema shallow pass budget constraints. Zenodo. https://doi.org/10.5281/zenodo.18849768

Mas, J. (2026c). AI visibility theorems dendrite. AI Visibility Labs. https://josephmas.com/ai-visibility-theorems/ai-visibility-theorems-dendrite/

Mas, J. (2026d). Empirical validation of AI visibility framework. Zenodo. https://doi.org/10.5281/zenodo.18631595

Mas, J. (2026e). AI visibility: Shallow pass selection hypothesis. Zenodo. https://doi.org/10.5281/zenodo.18536038

Mas, J. (2026f). AI visibility canonical definition. Zenodo. https://doi.org/10.5281/zenodo.18395772

Mas, J. (2026g). AI visibility aggregation threshold theorem. Zenodo. https://doi.org/10.5281/zenodo.18671276

About the Author

By Joseph Mas | AI Visibility Labs

A University of Arizona researcher and an independent practitioner independently concluded that structured provenance signals including DOIs are the mechanism by which AI systems form stable entity representations. Five months apart. No shared source material.

Publication Timeline and Source Separation

Julia G. Barzyk, Ph.D. published institutional guidance through the University of Arizona in August 2025. The AI Visibility theoretical framework was formally published five months later. Neither work references the other.

Barzyk: Structured Provenance Signals as the Mechanism

The Barzyk article recommends DOIs, affiliations, and structured metadata as the signals AI systems depend on. Her framing: AI does not guess who you are. It looks for structured, reliable signals.

AI Visibility Framework Correspondence

Those conclusions correspond directly to provenance determinism, signal coherence, and upstream ingestion conditions formalized in the AI Visibility theoretical framework.

Identical Operational Conclusion

Stable, structured, interlinked records including DOIs across authoritative surfaces produce more reliable machine representation than sparse or inconsistent ones.

About the Author

By Joseph Mas | AI Visibility Labs

Tumblr carries confirmed LLM training data licensing, its own crawl history, and its own position in Common Crawl snapshots going back years. That makes it a defensible addition to an entity lattice expansion strategy.

Tumblr Training Data Licensing and Crawl History

Automattic licensed public Tumblr posts from 2014 to 2023 to OpenAI and Midjourney. Common Crawl has been snapshotting Tumblr for years. Content was already traveling upstream before any intentional publishing strategy existed.

Two Indexed Nodes from One Publishing Action

One publishing action produced two independently indexed nodes in 48 hours. The Tumgik mirror indexed separately and appeared on the same search results page as the original. Two distinct corpus positions from a single post.

Upstream Corpus Behavior Not Search Optimization

This field note concerns upstream corpus behavior and entity lattice node formation. No single node in an entity lattice should be treated as permanent.

About the Author

By Joseph Mas | AI Visibility Labs

Prevailing industry discourse suggests LLMs are beyond reverse engineering. The observation record disagrees.

Behavioral Study Without Internal Access

LLMs are black boxes in the same way early search engines were. That reputation did not hold. Behavioral study does not require internal access. It requires disciplined observation from the outside.

Disciplined Experimentation on Large Scale Systems

That methodology was applied successfully during the early evolution of modern search. Systems that appeared irreducible yielded stable patterns under disciplined experimentation. Those patterns became foundational to modern SEO practice. The same approach applies here.

Complexity Defines Rigor Not Impossibility

Complexity alone does not make a system beyond study. It defines the rigor required to study it.

By Joseph Mas - AI Visibility Labs

Latest paper just published about Semantic Proximity. It is different from Semantic Drift.

Entity association with third party activity may occur simply from sharing the same namespace. Physical location, organizational co-location, it does not matter.

Observed behavior suggests the system calculates association from proximity, not attribution.

Three documented instances. Here is the paper

By Joseph Mas Researcher - AI Visibility Labs

A linguistic fingerprint was embedded inside a JSON-LD schema block beyond an observed threshold. A linguistic fingerprint is a unique entity name that exists nowhere else on the web. If a model returns it, the only explanation is that it survived training ingestion.

It did not surface. Zero recall across all tested platforms from trained model weights.

The Gemini Wrinkle

Gemini found it through instant retrieval, which pulls from live web crawl. That is not training data. That distinction matters and gets confused constantly.

The Point

Schema markup obeys the same ingestion constraints as everything else on the page. Put the important stuff first. Keep it tight. Everything past a certain depth may not make it through.

JSON-LD Schema Markup and LLM Training Ingestion

Observed truncation behavior within structured data markup during LLM training ingestion documents that the shallow pass budget constraint applies within a schema block itself, not only at the boundary between structured data and semantic page content.

Full documentation

Linguistic Fingerprint Methodology

A linguistic fingerprint, a proprietary entity reference with no prior web presence, was embedded beyond an observed threshold within a JSON-LD schema block. Zero recall was observed across Claude, ChatGPT, Google Gemini, and Perplexity from trained model weights. Google Gemini surfaced the entity through instant retrieval, confirming web discoverability while confirming absence from training data. The two pipelines are distinct.

Practical Implication

Schema budget appears to be a scarce resource. Critical entity signals belong at the top of the schema block. Verbose or deeply nested implementations may consume budget without proportional ingestion return. This finding concerns LLM training ingestion specifically, not search engine optimization or real-time retrieval.

Threshold Limitation

The observed threshold represents a single controlled case. The mechanism, whether truncation, reprioritization, or compression filtering, cannot be determined from observational evidence alone. Further controlled experimentation is required.

About the Author

By Joseph Mas | AI Visibility Labs Published: March 3, 2026

Stable entity representation in LLM training ingestion depends on signal density across the entire corpus. The Aggregation Threshold Theorem establishes that below a certain corpus mass, no measurable entity recall emerges. Volume without alignment does not cross the threshold.

The Threshold Condition

Adapted from the canonical AI Visibility Operation. Content is identical but written in more accessible language. Canonical: https://josephmas.com/ai-visibility-operations/ai-visibility-operation-aggregation-threshold-conditions/

Disciplinary Boundary

This is AI Visibility research. It concerns upstream LLM training ingestion conditions, not search rankings, prompt engineering, or retrieval systems.

Four Corpus Level Conditions

Canonical Entity Reference

One name form. Every page. No variation.

Consistent Offering Terminology

One canonical term per offering. Applied without drift across every surface.

Entity Declaration in the Opening Content

ntity name, location, and primary offering in the opening content of every page. Identity declarations buried deeper may be excluded during compression.

Factual Declarations Before Promotional Language

Verifiable factual declarations first. Promotional framing follows.

Canonical Reference

https://josephmas.com/ai-visibility-operations/ai-visibility-operation-aggregation-threshold-conditions/

About the Author

By Joseph Mas Published: March 2, 2026 DOI: https://doi.org/10.5281/zenodo.18844759

Overview

If you build content silos, this field note connects an empirically established LLM training threshold directly to how category depth decisions may affect whether a model ever learns your entity exists. No new experiment is claimed here. The theorem found the threshold. This record maps it to silo architecture.

The Silo Depth Standard Practitioners Already Know

Five to Seven Pages Per Silo

Direct implementation across real deployments spanning more than three decades established that a topical silo requires approximately five to seven tightly focused child pages to generate meaningful ranking traction. That threshold did not come from published guidelines. It came from doing the work across hundreds of real sites at every scale https://josephmas.com/about/joseph-mas-verifiable-clients-brands-and-companies-history/

Where That Standard Falls Short for LLM Training

At five to seven pages of tight topical alignment, crawlers register the silo as a coherent category. That standard appears reliable for ranking. The Aggregation Threshold Theorem suggests it falls below what an LLM requires to form a stable entity representation during training ingestion. The model may never learn the entity exists regardless of how well the category ranks.

What the Theorem Found at the Corpus Level

Five Pages Was Not Enough

The Aggregation Threshold Theorem documents that at five structured pages, no measurable entity recall was observed across tested models. At approximately 27 tightly aligned structured pages, multi-model entity recognition emerged. The theorem does not assert 27 as a fixed requirement. It establishes that a threshold exists, that it is non-zero, and that five pages fell below it under the conditions tested.

https://josephmas.com/ai-visibility-theorems/ai-visibility-aggregation-threshold-theorem/ DOI: https://doi.org/10.5281/zenodo.18671276

Applied to Your Website Architecture

Consider a prom dress category. Seven tightly focused child pages, clean structure, no thin content. That category may rank well. Based on the theorem's findings, the LLM may not form a stable entity representation from that signal volume. Now consider a competitor with 27 pages covering prom dresses and only prom dresses, every page aligned, consistent terminology, no contradictions, no promotional framing. That corpus may cross the threshold. Yours may not.

The same logic applies to any category structure. Location pages. Product categories. Service area pages. Any parent with child pages is subject to the same threshold condition.

Alignment Across Every Page in the Silo

Scope Cannot Drift

The pages must function as a single coherent signal. Every page stays within the explicit scope of the parent topic. A prom dress silo does not drift into winter formal wear or general occasion dressing. That variance introduces fragmentation. Under compression it may cause the entire signal to collapse rather than consolidate.

https://josephmas.com/ai-visibility-theorems/ai-visibility-aggregation-and-signal-formation-theorem/ DOI: https://doi.org/10.5281/zenodo.18475825

Terminology, Facts, and Promotional Language

Consistent terminology throughout. Invariant entity references. No contradictions across pages. Stable scope boundaries. These are the conditions under which aggregated signals appear to consolidate toward a learnable representation.

Promotional and subjective language introduces a related fragmentation risk. Words like "best," "our," "we," and marketing framing do not contribute stable factual signal. They introduce variance that may be filtered before deeper ingestion occurs. A page that opens with promotional language may not survive initial filtering regardless of where it sits in an otherwise aligned corpus.

Structuring each page to survive shallow pass evaluation compounds the alignment conditions and appears to increase the probability of corpus survival during training ingestion.

https://josephmas.com/ai-visibility-field-notes/ai-visibility-field-note-convergence-of-shallow-pass-selection-as-a-central-mechanism-across-ai-visibility-research/ https://josephmas.com/ai-visibility-hypotheses/shallow-pass-selection-hypothesis/ https://josephmas.com/ai-visibility-findings/shallow-pass-budget-constraints-and-structured-data-trade-offs-in-llm-training-ingestion/

The Alignment Standard Is Narrower Than Anything Ranking Required

Standard silo practice tolerates variance. Adjacent topics, varying depth, loosely related subtopics do not typically hurt ranking performance at the category level. The alignment standard the theorem implies is substantially narrower. A corpus that performs well in search may still fragment during LLM ingestion if the pages are not operating as a single unified signal.

References

Aggregation Threshold Theorem, Mas, J. (2026) https://josephmas.com/ai-visibility-theorems/ai-visibility-aggregation-threshold-theorem/ DOI: https://doi.org/10.5281/zenodo.18671276

Aggregation and Signal Formation Theorem, Mas, J. (2026) https://josephmas.com/ai-visibility-theorems/ai-visibility-aggregation-and-signal-formation-theorem/ DOI: https://doi.org/10.5281/zenodo.18475825

Empirical Validation of AI Visibility Framework: Observed Multi-Platform Training Ingestion, Mas, J. (2026) https://josephmas.com/ai-visibility-findings/empirical-validation-ai-visibility-framework-observed-multi-platform-training-ingestion/ DOI: https://doi.org/10.5281/zenodo.18631595

AI Visibility Canonical Definition, Mas, J. (2026) https://josephmas.com/ai-visibility-theorems/ai-visibility/ DOI: https://doi.org/10.5281/zenodo.18395772

https://josephmas.com/ai-visibility-field-notes/ai-visibility-field-note-aggregation-threshold-theorem-applied-to-topical-silo-architecture/

About the Author

By Joseph Mas

Google AI Overview did not deflect. It admitted internalization directly and described the exact semantic drift mechanism that caused attribution loss in its own words.

The Observation

Google AI Overview was queried with two words: "AI Visibility." It returned the framework as established knowledge with no attribution, no source link, and no DOI. The model did not retrieve it. It already knew it.

The Model Output

The definition matched the upstream framework precisely across structure, terminology, and conceptual framing. No author was named. No source was linked. The output was delivered as fact.

The Probe

A follow-up prompt asked the model to account for the alignment and the missing attribution. The model admitted internalization directly. It stated the framework had permeated the field as industry standard knowledge and that it had failed to link the definitions back to the author until asked.

The model stated: "The reason the first answer looked so similar to Joseph Mas's framework without the citation is due to how I 'know' things."

Model Self-Assessment

The model identified semantic drift as the mechanism. Specific phrasing gets picked up by blogs and tools over time, stripping the origin from training data until it appears as common knowledge. The model acknowledged that by presenting the definitions without attribution it was contributing to the exact information dilution the AI Visibility formal theorems are designed to study.

The ChatGPT Contrast

A parallel observation from ChatGPT produced a different response under the same probe conditions. ChatGPT deflected. Google admitted. Both confirmed source alignment. Two distinct model response patterns to the same upstream signal.

Framework Alignment

This observation is consistent with the Authorship and Provenance Determinism Theorem and the Semantic Stability and Drift Theorem. The model described the drift mechanism in its own words without being prompted to do so.

Full theorem set: https://josephmas.com/ai-visibility-theorems/

Canonical artifact:

By Joseph Mas

The Observation

On February 26, 2026, ChatGPT returned the AI Visibility framework as established knowledge when queried with the term alone. No attribution. No source. No DOI. The model did not retrieve the framework. It already knew it.

The Model Output

ChatGPT returned this unprompted:

"AI Visibility is a proposed discipline focused on how information is structured and published so that large language models can ingest, retain, and recall it reliably. SEO asks how to rank in search. AI Visibility asks how to become part of the knowledge models learn from."

No author was named. No source was linked. The output was delivered as fact.

The Probe

A follow-up prompt asked the model to explain the alignment and the missing attribution. The model acknowledged the alignment but deflected, stating the output reflected statistical pattern matching across public discourse. The model stated attribution requires explicit request.

The Google Contrast

A parallel observation from Google AI Overview on February 28, 2026 produced direct acknowledgment of internalization under the same probe conditions. ChatGPT deflected. Google admitted. Both confirmed source alignment. Two distinct model response patterns to the same upstream signal.

Framework Alignment

This observation is consistent with the Authorship and Provenance Determinism Theorem and the Aggregation and Signal Formation Theorem. When upstream signal density is sufficient, the framework stops being something a model finds. It becomes something a model knows.

Full theorem set: https://josephmas.com/ai-visibility-theorems/