Adapted from Duane Forrester's essay on Duane Forrester Decodes. Aurora extends his framing for marketing teams running AI-search visibility programs.
When a brand disappears from ChatGPT or loses share in Perplexity, the default marketing response is to produce more content. That is a retrieval-layer fix applied to what is increasingly a different kind of problem entirely. The result is wasted budget, missed quarters, and a creeping sense that the work is not connecting to the outcomes anymore.
Forrester's contribution is a clean three-layer model of where AI visibility actually breaks. Each layer fails differently. Each layer has a different owner. Each layer needs a different fix.
Layer 1: Retrieval
This is where AI systems pull external content into an answer. Retrieval-augmented generation, the layer most teams already work on. Success here looks like classical search optimization: crawlable, parseable, chunk-friendly content. Structured headings, self-contained answers, schema markup, clean technical implementation. Most marketing teams have versions of this work underway.
The layer's structural limit is what Microsoft Research called the connect-the-dots failure: plain RAG retrieves chunks but cannot reason about relationships between them. For questions requiring synthesis across multiple sources, retrieval alone breaks down, and the model fills the gap with confident-sounding hallucination.
The question this layer answers: can the model retrieve our content at all, and is it the right content for the query?
Layer 2: Knowledge graph (the relationship layer)
The knowledge graph layer decides how your brand is represented as an entity: what category you occupy, what other entities you connect to, whether you are treated as a recognized member of your category or as one fuzzy candidate string among many. Google's Knowledge Graph, Microsoft's Satori, and the open graph built on Wikidata and schema.org collectively define this.
Clean, well-defined entities get cited consistently. Undifferentiated tokens scattered across the web get pattern-matched against many candidates and lose. Producing more content does almost nothing if entity definition stays fuzzy. The fix is structural: schema markup on owned properties, consistent naming and identifiers across the open web, presence on high-trust nodes (Wikidata, review platforms), and the slow accumulation of brand mentions in contexts the graph treats as authoritative.
The question this layer answers: are we a clean, defensible entity in our category, or are we pattern-matched against twelve competitor candidates?
Layer 3: Context graph (governed enterprise retrieval)
This is the layer most marketers have not yet named. A context graph has the same structural shape as a knowledge graph — entities, relationships, typed connections — but it is grounded differently. A knowledge graph models the world. A context graph models one specific organization's data, decisions, policies, and operational reality. The cleanest framing: a knowledge graph is the library; a context graph is the operating manual written by the people who run the place.
What makes the context graph distinct is that governance lives inside the graph. Policies, permissions, validity windows, and authorization rules are nodes the graph itself queries, not external documentation applied at the edges. The result is governed retrieval: queries return information already filtered through current authorization and applicability.
This went from invisible infrastructure to enterprise procurement vocabulary when Google introduced the Knowledge Catalog at Google Cloud Next '26 — a unified, dynamic context graph of a customer's business that grounds enterprise agents in that customer's own data. Gartner projects that 40% of enterprise applications will feature task-specific AI agents by the end of 2026, up from less than 5% in 2025. Those agents — handling procurement, competitive intelligence, vendor evaluation — will not reason about your brand from the open web. They will reason about you from inside their company's context graph.
The question this layer answers: when an agent inside a customer company reasons about our brand, what does it find, and is the version it finds the version we would want it to act on?
Where Aurora plays
The three layers each need different evidence, and Aurora is structured to provide each in turn.
Retrieval layer. Aurora runs scheduled prompts against ChatGPT, Perplexity, Gemini, Google AI Mode, Google AI Overviews, Copilot, and Claude, captures the raw responses, and extracts citations and brand mentions. That tells you whether the retrieval layer is finding you — per prompt, per engine, per competitor.
Knowledge-graph layer. Aurora's source intelligence rolls up which domains the engines treat as authoritative in your category, exposing the third-party surfaces that feed entity definitions. Brand-visibility scoring catches the paraphrased mentions and aliases that pure citation tracking misses — the surface area where entity reinforcement is happening.
Context-graph layer. This layer is upstream and harder to measure directly because it lives inside customer companies. The leverage is to arrive at the context graph in a clean state: consistent category positioning, reliable structured data, robust third-party signals. Aurora's competitor and source-intelligence views surface where your brand is fragmented across owned and earned media, which is the upstream cause of an ambiguous representation downstream.
Stop solving the wrong layer
Most teams concentrate on retrieval-layer optimization while losing ground on knowledge-graph work and remaining absent from context-graph conversations. The teams that win in 2026 are the ones that figure out how to operate across all three responsibility zones rather than perfecting their work on just one.
The diagnosis comes first. The fix follows the diagnosis. More content is sometimes the right answer. It is rarely the answer when the underlying issue is fuzzy entity definition or governed-retrieval invisibility.
