AI Recommendations Depend on Brand Depth

Getting cited in AI search results has become a popular benchmark for brand visibility. But citations alone fail to explain why certain brands dominate ChatGPT, Google AI Mode, Perplexity, and other AI-powered search engines.

Citations are merely symptoms of visibility, not the systems that generate them. AI platforms consistently favor brands with a strong semantic presence across training data, online reviews, media coverage, search systems, and interconnected web entities.

This is why Generative Engine Optimization (GEO) presents two distinct visibility challenges happening simultaneously. You must build long-term brand weight inside AI systems while also creating content that survives modern retrieval pipelines.

AI recommendations are shaped during both retrieval and synthesis. Brand depth is what increases your odds in both stages.

GEO means playing two games at once

Each layer influences visibility in its own way.

Game 1: Parametric weight

Brands function as coordinates within an LLM’s embedding space. Their position is defined by the density and consistency of signals found in training data.

This parametric weight accumulates slowly over months and years through consistent brand presence across the web. Inconsistent messaging makes the brand’s vector fuzzy, which reduces recall and confidence in the model.

A brand with little parametric weight is functional, forgettable, and interchangeable. You cannot easily alter what a model has already internalized during training, so most efforts must focus on future training cycles.

Focusing exclusively on citations for months neglects the structural foundation that eventually makes those citations unavoidable.

Game 2: Retrieval survival

When a system like Google AI Mode or ChatGPT Search activates its retrieval pipeline, does your content actually make it through?

Approximately 85% of brand mentions in AI search come from external domains, not the brand’s own website. Every major AI search system starts with retrieval, but each handles it differently.

Perplexity retrieves, ranks, and embeds citations into the context window before the LLM generates a single token. The model synthesizes answers from retrieved evidence rather than directly from training data.

Google AI Mode decomposes a single query into 8 to 12 parallel subqueries across the live web, Google’s Knowledge Graph, and specialized data sources before synthesizing a response. Google calls this query fan-out.

ChatGPT search expands a query into five or six semantic variations, retrieves 35 to 42 candidate URLs, disqualifies 83% before extraction, and synthesizes three to five citations in the final response. Retrieval is typically skipped only for nonfactual prompts like creative writing or basic math.

In fan-out systems, you compete across 8 to 12 parallel subqueries simultaneously.

Citations are receipts

Only 6% to 27% of frequently mentioned brands are also top-cited sources. Models can know a brand without citing it.

Citation frequency tracks output presence, not the retrieval and synthesis decisions that surfaced the brand in the first place. Optimizing for citations focuses on the receipt rather than the underlying driver.

Brand depth, built through density, consistency, and cross-source coverage, is what makes a brand the statistically low-risk answer before a citation is ever generated.

Brand depth: How human brains and LLMs default to the familiar

The human brain operates similarly to LLMs. We manage a massive volume of daily decisions by relying on mental frameworks and heuristics built over time.

This idea is rooted in predictive processing theory, which describes the brain as a forecasting engine that uses past information to minimize errors.

LLMs and human cognition handle ambiguity in similar ways. Both prioritize information that is most densely established within their respective systems.

| Brand element | Human brain | LLM | | — | — | — | | Memory and recall | Episodic and emotional, triggered by sensory cues | Statistical frequency and co-occurrence density in training data. High frequency increases recall | | Brand identity | Sensory and visual: logo, typography, packaging | Semantic proximity: adjectives, reviews, articles associated with the brand name. A coordinate in embedding space | | Building trust | Social proof, word-of-mouth, personal trial | Parametric authority: training data weighted toward high-authority sources | | Handling mistakes | Forgiveness through empathy; apology can repair | Data permanence: models consolidate patterns, not intent. Negative signals persist until newer data outweighs them | | The recommendation | Impulsive and bias-driven: scarcity, FOMO, halo effect | Synthesis-weighted: shaped by what’s most densely represented in parametric memory and retrieved sources simultaneously |

Getting technical about branding with brand depth

AI models and Google’s Knowledge Graph learn from many of the same trusted websites. AI models learn by identifying which words frequently appear together, while Google uses that same information to build a network of connected facts.

Google’s systems specifically evaluate entity salience, entity coherence, and inter-entity relationship density.

Entity salience

How prominent and distinct your brand is within a specific topic cluster. Entity salience influences citation probability.

Google asks: How prominent is this brand within a topic cluster?

LLMs ask a similar question at inference time: Which entities have enough statistical weight to surface when a topic is queried?

Low salience means you’re retrievable only through exact branded queries. High salience means you appear when the topic comes up, not just when your name is searched.

Google evaluates salience through systems like RepositoryWebrefLatentEntities, which maps the latent entities a brand co-occurs with, and RepositoryWebrefKGCollection.

Entity coherence

The consistency of your brand’s identity across all retrieved contexts.

Inconsistent naming, conflicting positioning, and contradictory dates signal that an entity is unreliable. LLMs trained on that same corpus learn a fragmented, low-confidence representation.

The model fills gaps created by entity incoherence, leading to brand drift. The model’s version of your brand slowly diverges from reality because the training signal was never stable enough to anchor it.

Inter-entity relationship density

The strength and number of connections between your brand and other authoritative entities, including products, concepts, and proofs.

Inter-entity relationship density influences associative retrieval paths.

In agentic systems like Deep Research, AI Mode, and Perplexity Pro, each reasoning step is a retrieval event. Relationship density determines whether your brand survives hop two and hop three.

A brand that only exists at the center of its own graph disappears the moment the query moves one step sideways. GlobalLinkInfo and LatentEntity in Google’s Content Warehouse map these inter-entity edges.

The RAG layer is where site quality becomes a gate

Mark Williams-Cook documented a site quality score in December 2024. The score uses a 0-to-1 scale, and sites scoring below roughly 0.4 are not retrieved as candidates, regardless of optimization efforts.

That matters because retrieval eligibility influences which entities and sources repeatedly enter AI systems in the first place. Brand integrity becomes an infrastructure problem. You cannot optimize your way into LLM citations if you have not first built the entity coherence and relationship density that make your brand consistently retrievable.

Why AI systems repeatedly surface Black Honey

The more co-occurrences you have, the higher your mutual information score, and the more often you appear in answers.

Clinique’s Black Honey lipstick is a good example of how this works in practice because of its strong entity depth:

• Concept: Co-occurs with “universally flattering” and “my lips but better” (MLBB) value propositionsBecause of this density, AI systems repeatedly surface Black Honey when answering questions about universally flattering lipstick.High recall: AI models are more likely to recall and mention Clinique Black Honey across a wide range of relevant queries, including “best universally flattering lipsticks,” “viral makeup trends,” and “iconic ’90s beauty.”High authority: The depth of co-occurrence, including historical evidence, cultural context, and product variants, provides AI models with sufficient information to generate detailed, authoritative, and multifaceted answers.

Building for retrieval, recall, and recommendation

Preference is what survives. Build for the layer that determines synthesis weight and for what happens inside the retrieval funnel.

When your brand is specific, consistent, and densely connected across topical clusters, it becomes easier for AI systems to retrieve, synthesize, and recommend.

Focus on what survives the retrieval funnel

Specific, data-rich, hard-to-reproduce content gets retrieved and cited. Academic literature refers to this as adaptive retrieval.

Generic, predictable content gets skipped because the model can generate it on its own.

| Low entropy gets ignored | High entropy gets cited | | — | — | | “Our coffee is smooth and delicious.” | “The Gesha variety from Hacienda La Esmeralda in Boquete, Panama. Grown at 1,700 meters. Water at 94 C. Brew ratio 1:16.” |

The second version anchors named entities, including a variety, an organization, a location, and quantitative values. These are details the model cannot plausibly generate without a source.

Actionable tip: Add high-density assets, including company history, team bios, and ISO certifications, designed to serve as grounding data for retrieval-augmented generation (RAG) systems.

Build AI navigation maps

Your website functions like a knowledge graph. AI systems use internal links to build a semantic map of your domain.

Embed links that define logical relationships between entities and create clear paths for crawlers to follow. Structure links around the user’s decision journey, which often mirrors AI retrieval paths:

• Topic → Subtopic (broad context)

Avoid orphan pages

Pages with no meaningful incoming anchors are likely demoted in processing. They do not accumulate siteAuthority or NavBoost signals.

The fix is to give these pages strategic internal links that connect them to the graph, or delete them. If a page is not worth linking to, is it worth human or bot attention?

Visibility starts before the citation

Citation frequency studies are symptom trackers, not diagnostic tools. They can tell you that certain brands appear more often. They cannot reliably explain whether that visibility comes from training data, RAG retrieval, entity salience, or category dominance.

Build the thing that causes citations, not the thing that imitates them.