How AI Visibility Works: The 4 Layers Behind Every AI Citation

Four layers. Three mechanisms. One framework.

AI now handles 56% of the search volume that traditional search engines do worldwide. In the US, that number is 34% and growing at 300% year over year. Most of that usage is on mobile apps, invisible to standard web analytics.

Brands are being recommended, described, and cited by AI systems millions of times a day. Most have no idea what’s driving those recommendations or how to influence them.

The industry response so far has been to recycle SEO tactics with a new label. “AEO” or “GEO” gets attached to the same playbook, and the conversation stays at the surface: optimize your content, add schema, get listed on directories.

That advice isn’t wrong. It’s incomplete. It treats AI visibility as a single thing you do, when it’s actually a stack of four distinct layers, each powered by different mechanisms, each requiring different actions.

I’ve spent the last several months mapping this stack. My team and I catalogued research from over 88 sources, analyzed citation behavior across ChatGPT, Gemini, Perplexity, and Claude, ran hundreds of direct platform tests, and tracked how the engineering architecture of retrieval systems actually determines what gets cited and what doesn’t.

This article isn’t a technical breakdown of how LLMs work. It’s a map of the architecture and dynamics that determine whether AI recommends your brand or skips it. This is built for marketers and business operators who need to understand what’s actually happening and what to do about it. Enough with all the “GEO is SEO” and “URGENT: New GEO Tactic” BS…..Let’s lay out what is actually happening…

AI visibility is not one thing

When someone asks ChatGPT, “What’s the best CRM for small businesses?” or Perplexity, “Who should I hire for web design?”, the AI doesn’t just search the web and pick a result. It runs through a sequence of operations that most marketers never see.

First, it resolves who you are. Entity resolution happens before retrieval. The system needs to confirm that your business exists, that your name maps to a real entity, and that the structured data about you is consistent. This is the knowledge graph layer.

Second, it assesses how confidently it can describe you. If the model has seen your brand mentioned across press coverage, authoritative domains, and consistent sources, it can describe you with confidence. If it hasn’t, it hedges or skips you entirely. This is the training layer combined with retrieval.

Third, it decides whether to recommend you. For category queries (“best X in Y”), AI pulls from editorial listicles, review sites, and comparison content. If you’re not in those sources, you’re not in the answer. This is the retrieval layer.

Fourth, and only fourth, it decides whether to cite your content. If you’ve published content that’s structured like an answer, dense with data, and positioned in the first 30% of the page, AI retrieves it as a source. This is also the retrieval layer, but targeting your owned content rather than third-party placements.

These four operations map to four distinct layers.

The four layers of AI visibility

Layer 1: Entity establishment

AI confirms you exist as a real entity. This is the knowledge graph layer. Directories, Google Business Profile, schema markup, Wikidata. It’s entity resolution: the system confirming your name, location, category, and attributes before it even begins to retrieve content.

ChatGPT uses directory listings for 48.7% of its local sources. Entity resolution happens before retrieval. No entity, no citation path. This is the foundation everything else builds on.

Layer 2: Entity depth

AI describes you accurately and confidently. This layer is fed by press coverage, earned media, brand mentions on authoritative domains, and consistent information across sources. The more sources that say consistent things about you, the more confidently AI puts your name in an answer.

Research confirmed in February 2026 (the superficial alignment hypothesis) shows that what AI learns during pre-training is its permanent knowledge layer. Post-training doesn’t add new knowledge. It just surfaces what’s already there. What the model “knows” about your brand is determined by the content that existed at training cutoff. Getting into training data is the durable competitive advantage.

Layer 3: Category citation

AI recommends you when someone asks for options in a category. When a user asks “best X in Y,” AI pulls from editorial listicles, review sites, and comparison content. We tested 100s of category queries across 10 verticals on ChatGPT, Gemini, and Perplexity. Zero press releases were cited. All 300 platform-query combinations cited editorial content instead. Sites like G2 for SaaS, NerdWallet for finance, Wirecutter for consumer products, Yelp for local, and Healthgrades for healthcare. The specific sites change by vertical. The pattern doesn’t.

This is where commercial value lives. If your brand isn’t in the third-party editorial content that AI retrieves for category queries, you’re not in the recommendation.

Layer 4: Informational citation

AI cites your content as a trusted source when answering topic questions. This requires content structured for how AI retrieval works: answer-shaped, data-dense, front-loaded. Research across 1.2 million ChatGPT responses found that 44.2% of citations come from the first 30% of text. An 800-word page gets 50%+ grounding coverage from AI, while a 4,000-word page gets just 13%.

Recent research from McGill NLP shows that embedding models are shifting from encoding queries to encoding answers. Content that already reads like the model’s response gets a structural retrieval advantage at the architecture level. This isn’t a behavioral correlation. It’s an engineering shift.

How the layers connect

Each layer strengthens the one below it. Weak L1 doesn’t block L3, but it limits how confidently AI recommends you. The stronger the foundation, the more each layer above it compounds. Not every business needs all four layers. A local restaurant lives in L1 and L3. A research institution might skip L3 entirely and focus on L4. The framework is a map, not a checklist. The value is knowing which layers matter for your business and why.

Three mechanisms power the four layers

Behind the four layers are three distinct mechanisms through which AI learns about and surfaces your brand. Understanding which mechanism powers which layer is the difference between random tactics and a coordinated visibility strategy.

Knowledge graph (K): Entity resolution

This is structured entity data across directories, Google Business Profile, schema markup, Wikidata, and industry platforms like G2 or Crunchbase. AI uses this to resolve “what entity is this?” before it retrieves anything. This mechanism powers Layer 1 primarily. Without it, AI can’t confirm you exist, and the rest of the stack has nothing to build on.

Training (T): Permanent memory

This is what AI learned during pre-training from the entire web, frozen at a cutoff date. When AI answers without searching, it’s pulling from this layer. Your brand is either in there or it isn’t. Press coverage, earned media mentions, and brand references on authoritative domains feed this mechanism because that content gets absorbed into the model’s training data.

Retrieval (R): Real-time search

This is AI searching the web in real time and pulling passages from pages to build answers. This is the mechanism most people think of when they think of AI visibility. It powers Layers 3 and 4 directly, and contributes to Layer 2. If your content isn’t retrieved into the candidate set, it cannot be cited.

A single action often triggers multiple mechanisms. A press release on Yahoo Finance’s /news/ path hits all three: the /news/ URL gives it retrieval trust (R), the brand mention feeds training data (T), and the entity consistency across the placement reinforces knowledge graph resolution (K). One action, three mechanisms, three layers affected.

How different actions flow through the stack

The framework becomes practical when you trace how specific actions move through the layers. Not every action hits every layer. Understanding which layer(s) a given tactic affects, and which it doesn’t, is how you build a complete AI visibility strategy instead of hoping one tactic does everything or just thinking “it’s all just SEO, bro”.

A Clutch or G2 profile with consistent entity data and client reviews touches L1 (entity resolution), L2 (third-party confirmation), and L3 (category listicle placement). A well-structured pricing page touches L2 (ChatGPT cites first-party pricing pages to verify training data) and L4 (retrieved as a source for product comparison queries).

An answer-shaped blog post with front-loaded findings touches L4 only. Pure retrieval. But it works best when L1 through L3 are already strong. A blog post from a brand that AI doesn’t recognize gets deprioritized in retrieval.

NAP consistency across 15 directories touches L1 only. Foundation layer. But it’s the layer everything else compounds on. Entity resolution happens before retrieval. No entity, no citation path.

Reddit is unique: it’s the only platform that feeds all three mechanisms simultaneously. A brand mention on Reddit feeds retrieval (threads get pulled in real time), training (Reddit corpus is in every major LLM’s training data), and knowledge graph (community consensus shapes entity associations). One platform, three mechanisms.

No single action covers all four layers. AI visibility is a stack, not a tactic. Press builds depth. Content earns citations. Directories establish existence. Reddit feeds everything. A complete strategy coordinates across all four.

Why this matters now

AI search is not a future concern; it’s now a required part of your consideration if you’re marketing online. Worldwide AI sessions are 56% of the size of traditional search, and 83% of that usage is happening on mobile apps. These are apps that are largely invisible to standard web analytics. In the US, AI usage grew 300% year over year in 2025. ChatGPT alone commands 20% of search-related traffic worldwide.

The data also shows that traditional search isn’t dying or being eaten by AI search at all. Total search volume (engines plus AI) increased 26% worldwide and 16% in the US. The pie is getting bigger, and AI visibility is adding to it, not replacing it.

This doesn’t mean, though, that the brands being recommended in AI answers are necessarily the ones ranking on Google. Top-10 Google ranking gave a 76% chance of appearing in AI Overviews in July 2025. By March 2026, that dropped to 38%. Two out of three AI citations now come from pages users would never see on page one of Google.

The gap is widening. The brands that understand or are present in AI visibility will have an intentional strategy or simply be strong in The Stack (four layers, three mechanisms, coordinated actions across all of them). These will be (and are) the brands that AI recommends…… May The Stack Be With Them. The brands that treat it as “SEO with a new name” will be the ones asking why they’re invisible.

What comes next

This framework is how I’ve organized what we see. It’s not mine, it’s how the data shows up when you scrutinize the mechanisms. We’re mapping as we go, and the territory is new, and the data keeps moving. By no means am I claiming this is the final word. What I am claiming is that the data supports this structure, and that nobody else has mapped it this way, to my knowledge.

Over the coming days and weeks, I’m publishing the full breakdown of each layer and mechanism. I’ll show how entity establishment actually works at the architecture level, what training data means for your brand’s permanent AI footprint, and how retrieval determines what gets cited and what doesn’t.

After this, we’ll make it more practical for businesses and marketers and look at how the framework applies across verticals: local businesses need different things than B2B SaaS companies, and both differ from ecommerce and professional services.

Quite honestly, I’m excited. There are true distillable insights and tactics coming out of data if you can apply critical review, data analysis, and insight extraction.

Every claim traces back to evidence. 108 research entries across 88 sources and counting. We call it The Data Sieve.

AI visibility is a new discipline. It shares DNA with SEO but operates on different architecture, rewards different content structures, and requires a different strategic framework. This is my attempt to define what that framework actually is.

I’ll be posting each piece of the stack here.

This article was originally published on X by Aaron Haynes. Aaron is the CEO of Loganix, a visibility + SEO platform for brands and agencies.