01 Definition
What the AI Visibility Evidence Model is.
The AI Visibility Evidence Model is a reference model that orders the factors behind AI visibility by strength of evidence. It defines five factors, Topical Relevance, Machine Access, Entity Consistency, Extractability and Independent Corroboration, and assigns each a documented evidence grade based on peer-reviewed research, official platform documentation and transparent industry data.
The model exists because the field lacks exactly this: a single, verifiable reference that separates what research demonstrates from what marketing claims. Every factor in the model carries its grade and its sources, so every statement on this page can be checked against the primary literature listed in the source register below.
02 Purpose and Boundary
A map of the evidence, not a methodology.
The AI Visibility Evidence Model maps what the evidence shows works. The AEO Mastery Framework describes how richresults.ai implements it. The model is descriptive: it reports the state of the research. The framework is prescriptive: it defines a working method. Neither replaces the other.
The model is not a ranking system, not a score and not a promise of results. AI visibility is a distribution across repeated, non-deterministic answers, and no factor in this model guarantees a citation. What the model provides is priority: it tells you which work is supported by evidence, which work is hygiene, and which work the evidence contradicts.
03 The Evidence Scale
Four grades, defined before use.
Grade A: peer-reviewed and controlled, or independently replicated. Grade B: controlled with limited transferability to open production systems, or an official platform statement. Grade C: correlational or triangulated across independent datasets, without causal proof. Grade D: unsupported or contradicted by evidence.
The scale is deliberately conservative. A factor rated C is not unimportant; it means no one has isolated its causal effect in production, and anyone claiming otherwise is claiming more than the data supports. Grades move as evidence accumulates, and this page carries a visible update date for that reason.
04 The Five Factors
Factor 1: Topical Relevance
Content that directly addresses the actual question is the strongest documented driver of citation. In the largest controlled citation study to date, 252,000 trials across six language models, topic match to the query and position in the model context were the dominant factors, and off-topic content was practically never cited first [1]. No entity work, no markup and no authority signal compensates for content that does not answer the question being asked.
Evidence grade: A. Peer-reviewed, controlled, convergent across models [1, 2, 9].
Factor 2: Machine Access
A source that crawlers cannot reach cannot be retrieved, and a source that cannot be retrieved cannot be cited. Machine Access covers crawl permissions for the relevant bots, index presence, server-side rendering of the main content and firewall configurations that do not silently block AI crawlers. Platform documentation is explicit on both sides of this gate: OpenAI requires OAI-SearchBot access for search visibility [12], and Google states that its AI features run on the same index and ranking systems as classical search [11].
Evidence grade: A, as a gate. Access is a binary precondition. It creates no advantage; its absence creates total invisibility [11, 12, 13].
Factor 3: Entity Consistency
Consistent entity signals decide whether relevant content is attributed to the right source. Structured data, stable identifiers, canonical name strings and connected external profiles reduce ambiguity in how systems resolve who is speaking. The documented effect is error reduction and correct attribution, not a visibility boost: Google states that no special markup is required for its AI features [11], and controlled work shows knowledge-graph grounding reduces entity disambiguation errors in the lab. The transfer of that effect to production citation rates is inferred, not demonstrated.
Evidence grade: B to C. Official platform statements and lab evidence for disambiguation; no causal proof as a citation driver [11].
Factor 4: Extractability
Passages that begin with the statement and carry dense, verifiable evidence are used more strongly once a document is retrieved. The position of information in the model context changes outcomes causally [3, 4], and pages with concrete numbers, definitions, comparisons and procedures show substantially higher influence on generated answers than pages without them [7]. The boundary is equally documented: these effects apply after retrieval, question-and-answer formatting alone does not help [7], and content rewriting tricks show no reliable effect and are frequently harmful under competition [2].
Evidence grade: B. Causal for position effects [3, 4], controlled for evidence density in fixed contexts [6], correlational in production [7], bounded by [2] and [9].
Factor 5: Independent Corroboration
Mentions by independent third parties are the strongest correlated predictor of AI visibility and the only factor with a causal foundation for what models know without searching. Controlled research shows that the frequency and spread of an entity in training data causally determines what a model knows about it [5]. Comparative analyses report a systematic bias of AI search toward earned media over brand-owned content [8], and vendor datasets with disclosed methodology, including Ahrefs and Muck Rack, triangulate the same direction. The boundary matters: only genuine, independent mentions carry this weight. Self-published corroboration networks and purchased mentions replicate one voice in many costumes and add nothing a model treats as confirmation.
Evidence grade: C, with a causal foundation for the parametric layer [5]. Correlationally strong and triangulated [8]; causally isolated in production: not yet, by anyone.
05 What the Evidence Contradicts
Graded D, with sources.
llms.txt as a visibility lever. Google states it does not use llms.txt for search or AI features, and no major platform has confirmed it as a production signal [11].
Content-level rewriting tricks. The broadest controlled benchmark found most conversational optimization methods ineffective and frequently harmful to citation ranking, while classical retrieval position dominated [2].
Schema as a citation switch. Structured data clarifies and disambiguates. As a direct citation driver for AI features it is officially not required [11], and no causal evidence exists for language-model citations outside search pipelines.
AI ranking positions as a metric. Answers vary heavily between identical runs. Position within a generated answer is noise; visibility is a share across repeated measurements with confidence intervals, not a rank.
Manufactured mentions. Corroboration works because it is independent. Producing it yourself removes the property that makes it work [5].
06 Method Note
How this model was compiled.
Every claim in this model was verified against its primary source: the papers, official documentation and datasets listed in the register below. Vendor-origin data is used only where the methodology is disclosed, and it is labeled as such. The synthesis was additionally stress-tested through blind reconstruction: independent AI systems were asked the same evidence question without access to this model, and their convergence on the same core sources served as a corroboration check.
Three limits apply to everything on this page. Production systems are non-deterministic, so no single observation proves an effect. Models and retrieval methods change without notice, so findings carry dates. And no one, including the best published research, can causally isolate a single intervention inside a live answer engine. Where this page says the evidence ends, it ends.
07 Source Register
Primary sources, numbered as cited above.
- Vishwakarma, Kumar, Jamidar (2026). What Gets Cited: Competitive GEO in AI Answer Engines. SIGIR 2026. arXiv:2605.25517. Peer-reviewed; authors affiliated with Sprinklr.
- Puerto et al. (2025). C-SEO Bench: Does Conversational SEO Work? NeurIPS 2025 Datasets and Benchmarks. arXiv:2506.11097.
- Liu et al. (2024). Lost in the Middle: How Language Models Use Long Contexts. TACL 2024. arXiv:2307.03172.
- Hsieh et al. (2024). Found in the Middle: Calibrating Positional Attention Bias. ACL 2024 Findings. arXiv:2406.16008.
- Kandpal et al. (2023). Large Language Models Struggle to Learn Long-Tail Knowledge. ICML 2023. arXiv:2211.08411.
- Aggarwal et al. (2024). GEO: Generative Engine Optimization. KDD 2024. arXiv:2311.09735. Effects conditional on fixed-context settings; see [2] and [9].
- Zhang, He, Yao (2026). From Citation Selection to Citation Absorption. Preprint. arXiv:2604.25707.
- Chen, Wang, Chen, Koudas (2025). Generative Engine Optimization: How to Dominate AI Search. Preprint, University of Toronto. arXiv:2509.08919. Acknowledges support by ktau.ai.
- Martinez (2026). Optimizing Visibility in Generative Engines: A Critical Survey. Preprint. arXiv:2607.14035.
- Jaźwińska, Chandrasekar (2025). AI Search Has a Citation Problem. Tow Center, Columbia Journalism Review. cjr.org.
- Google Search Central: AI features and your website. developers.google.com.
- OpenAI: OAI-SearchBot documentation. openai.com/searchbot.
- Google AI for Developers: Grounding with Google Search. ai.google.dev.