What Is RAG?

Andrew McPherson

Author Introduction

I’ve spent my career as a CIO and CTO, so I’ve seen first hand where “traditional” AI projects succeed, stall, or quietly fail in the real world. In this article, I unpack Retrieval-Augmented Generation (RAG) in plain language, so you can separate hype from value and design AI that actually serves your customers.

Outline

What RAG is and why it matters for B2B
How AI platforms retrieve and cite your content
The three stages of RAG processing explained
Why uncited brands lose pipeline to competitors
Structuring content as RAG-ready retrievable chunks
Entity disambiguation through structured data markup
Freshness signals and direct answer optimisation
Common RAG optimisation pitfalls to avoid

Key Takeaways

RAG determines which brands AI platforms cite
Uncited companies lose influence in AI-mediated discovery
Content must work as standalone retrievable chunks
Entity disambiguation increases retrieval confidence significantly
Freshness timestamps directly affect retrieval ranking
FAQ schema produces high-quality directly answerable chunks
Contextual prefixes can reduce retrieval failures by 67%
Share of Model is the new competitive visibility metric

RAG (Retrieval-Augmented Generation) is the technical architecture that allows large language models (LLMs) to ground their responses in external sources rather than relying solely on training data. When you ask ChatGPT, Perplexity, Google AI Overviews, Claude, or Gemini a question about your industry or your company, RAG is the system that determines whether your content gets retrieved, cited, and displayed in the response.

At its core, RAG separates the knowledge retrieval process from the text generation process. Instead of generating answers purely from what the model learnt during training – which becomes outdated quickly and leads to hallucinations – RAG systems search external knowledge bases, retrieve relevant passages, inject those passages into the generation context, and cite the sources. This architecture explains why some B2B companies achieve high Citation Authority (their content appears frequently in AI responses with attribution) while others remain invisible despite publishing substantial content.

The foundational research behind RAG was published by Lewis et al. at Meta AI Research in 2020, demonstrating that retrieval-augmented models produced more factual and specific language than purely generative models across knowledge-intensive tasks. For B2B companies competing for AI visibility, understanding RAG is not optional. It is the mechanism that determines whether your brand becomes a cited authority or an unattributed mention.

Why RAG Matters for B2B Companies

Traditional search engine optimisation (SEO) focused on keywords, backlinks, and crawlability. AI visibility optimisation requires understanding how RAG systems retrieve, rank, and cite sources. The distinction matters because RAG introduces new ranking factors: semantic relevance, chunk quality, entity disambiguation, and source verification.

If your content is not optimised for RAG retrieval, AI systems will answer questions about your industry, your capabilities, and your competitors without ever citing you. Your potential customers will receive AI-generated recommendations that exclude your brand entirely – not because you lack expertise or market presence, but because your content structure does not match RAG retrieval patterns.

This affects every B2B category. When a manufacturing VP asks an AI assistant “What industrial automation providers offer predictive maintenance for CNC machines?”, RAG determines which companies get cited. When a CFO queries “Which ERP systems integrate with NetSuite and support multi-currency consolidation?”, RAG selects the sources. When a procurement director searches “What logistics companies provide cold chain distribution?”, RAG decides whose content is authoritative enough to cite.

The Three RAG Consequences for Brand Visibility

Cited sources gain trust and click-through. Users perceive cited brands as authoritative and click through to verify claims or explore further. Citation converts passive mentions into active engagement.

Uncited sources become invisible. Even when AI systems use your content to formulate responses, without attribution you receive no traffic, no brand recognition, and no competitive advantage. Your insights become commodified without credit.

Misattributed sources suffer reputation damage. When RAG systems retrieve your content but attribute it to competitors or generic industry descriptions, you lose both visibility and authority. This happens when entity disambiguation fails – the AI cannot definitively link content to your brand entity.

Share of Model Is the New Competitive Metric

Share of Model (SoM) – the percentage of relevant AI responses that mention or cite your brand – has become a measurable competitive metric. Companies with high SoM dominate AI-mediated discovery. Companies with low SoM become invisible in the channels where B2B buyers increasingly start their research: conversational AI interfaces, AI-powered search engines, and autonomous AI agents evaluating vendors. CiteCompass measures Share of Model across the full buyer journey to diagnose where visibility breaks down.

RAG is not a future consideration. Google AI Overviews, ChatGPT search, Perplexity, Claude, and Microsoft Copilot all use RAG architectures today. If your content strategy does not account for RAG retrieval mechanics, you are optimising for a previous era’s discovery mechanisms.

How RAG Works: The Three-Stage Technical Process

RAG operates in three distinct stages. Understanding each stage reveals specific optimisation opportunities for B2B companies seeking to improve their AI visibility.

Stage 1: Query Understanding and Expansion

When a user submits a query, the RAG system does not search for literal keyword matches. Instead, it transforms the query into a semantic representation (typically a vector embedding) that captures intent, context, and related concepts.

For example, when a buyer asks “What CRM systems integrate with Salesforce and support custom objects?”, the RAG system interprets the semantic vector encompassing: CRM software category, Salesforce integration requirement, custom object support, platform compatibility, and API extensibility. The system may also expand the query with synonyms and related entities. “CRM” might expand to include “customer relationship management”, “sales automation”, and “contact management”.

This expansion explains why exact keyword targeting is less critical in RAG optimisation than semantic clarity and entity disambiguation. Your content needs to signal what concepts and entities it addresses, not just what keywords it contains.

Stage 2: Semantic Retrieval and Ranking

The RAG system searches its indexed knowledge base – which includes your website, documentation, competitor content, third-party reviews, industry publications, and other authoritative sources – for passages with vector embeddings similar to the query embedding. Retrieval uses multiple signals to determine which passages surface.

Semantic similarity measures how closely a passage’s meaning aligns with the query intent, calculated by measuring the cosine similarity between the query vector and passage vectors. Higher similarity scores increase retrieval likelihood.

Chunk quality determines whether a passage is self-contained and coherent. RAG systems chunk content into retrievable units, typically 200-500 words or logical sections. Well-structured content with clear headings, topic sentences, and logical flow produces higher-quality chunks. Dense paragraphs without structure produce poor chunks that lack retrieval context.

Source authority evaluates whether the source demonstrates trustworthiness. RAG systems weight passages from sources with strong entity recognition, verified authorship, third-party citations, and consistent factual accuracy. Research from Google’s REALM project (Guu et al., 2020) demonstrated that retrieval-augmented pre-training with entity linking significantly outperformed prior methods on open-domain question answering benchmarks by 4-16% absolute accuracy.

Freshness determines whether information is current. Passages with recent dateModified timestamps rank higher for queries requiring up-to-date information. A pricing page updated last week will outrank a pricing page unchanged for two years, even if content is similar.

Entity linking assesses whether the passage clearly identifies entities (companies, products, people, concepts). Content that uses structured data (JSON-LD schema) to explicitly define entities ranks higher because RAG systems can verify entity relationships with higher confidence.

The system retrieves multiple candidate passages (typically 5-20), ranks them by relevance, and selects the top results for context injection.

Stage 3: Context Injection and Generation

The RAG system injects the retrieved passages into the LLM’s generation context along with the original query. The LLM generates a response based on the retrieved context, ideally citing sources by name, URL, or inline reference. Citation likelihood depends on passage clarity, source authority, and how directly the passage answers the query.

A critical distinction applies here: the LLM’s training data influences how it interprets and synthesises retrieved passages, but the retrieved passages determine which sources get mentioned. A company with effective RAG optimisation – clear chunks, strong entity signals, fresh timestamps – can achieve high citation rates even if the LLM’s training corpus barely mentioned them. Conversely, a well-known brand with poor RAG optimisation may be excluded from responses if its content is not retrievable or citable.

This three-stage process repeats for every AI-generated response involving external sources. For B2B companies, optimisation targets each stage: query understanding (entity disambiguation, semantic clarity), retrieval ranking (chunk quality, source authority, freshness), and citation (passage coherence, direct answers). Understanding these mechanics is the foundation of effective Answer Engine Optimisation (AEO).

How to Optimise Your Content for RAG Retrieval

Optimising for RAG requires addressing content structure, technical implementation, and entity signals. These tactics apply across B2B contexts: software companies, professional services firms, manufacturers, distributors, and service providers.

Structure Content as RAG-Ready Chunks

RAG systems retrieve passages, not entire pages. Your content must be chunked into logical, self-contained units that make sense when extracted and cited in isolation.

Every major section should have a clear H2 heading that functions as a standalone retrieval target. Headings like “What Is Predictive Maintenance?”, “How Does Multi-Currency Consolidation Work?”, or “When to Use Cold Chain Logistics” create retrievable chunks that directly answer queries. Generic headings like “Overview”, “Details”, or “More Information” produce chunks that lack context when retrieved, reducing retrieval likelihood.

Make each section self-contained. The first two to three sentences of each section should establish context without requiring readers – or RAG systems – to have read previous sections. Include the key concept or entity name in the opening sentence.

Research from Anthropic on Contextual Retrieval demonstrated that adding contextual prefixes to chunks – explaining what each chunk is about before the chunk content – reduced retrieval failures by up to 67% compared to naive chunking approaches when combined with reranking. This reinforces why self-contained, well-contextualised sections dramatically improve retrievability.

Implement Entity Disambiguation with Structured Data

RAG systems prioritise sources with explicit entity definitions. Use JSON-LD schema to identify your company, products, services, and key personnel as distinct entities.

Required schema types for B2B RAG optimisation include: Organisation schema defining your company entity (with name, url, logo, sameAs, and description); Product or SoftwareApplication schema for each product or service (with name, description, brand, offers); Person schema for executives and subject matter experts (with name, jobTitle, worksFor, sameAs linking to LinkedIn); and DefinedTerm schema for proprietary concepts, methodologies, or terminology.

When a RAG system retrieves a passage mentioning “Platform X”, it needs to verify that “Platform X” refers to your specific product, not a generic term or competitor offering. Explicit schema creates unambiguous entity references that increase retrieval confidence. This is a core component of effective AI data surface optimisation.

Maintain Freshness Signals

RAG systems weight recent content more heavily for queries requiring current information. Every page should include dateModified timestamps in both HTML meta tags and JSON-LD schema.

Update pricing and product specifications whenever prices or specs change, not on arbitrary schedules. Review documentation and help content quarterly, updating timestamps when revisions occur. For thought leadership and research, retain the original publication date but add dateModified when substantive updates occur. Update company information whenever organisational changes happen.

Stale timestamps signal low reliability. A technical specification page with a dateModified timestamp from 2021 signals to RAG systems that the information may be outdated, reducing retrieval likelihood even if content is still accurate. Update timestamps to reflect actual content review, not just to game freshness signals.

Create Direct Answer Formats

RAG systems prioritise passages that directly answer questions. FAQ pages, definition sections, and how-to content produce high-quality chunks that are more likely to be retrieved and cited.

Implement FAQ schema (FAQPage in JSON-LD) for common buyer questions. Each question-and-answer pair creates a retrievable chunk that directly matches common queries. For example, a question like “What industries does your company serve?” paired with a comprehensive answer containing specific industry names, certifications, and capabilities creates a highly citable passage.

Build Internal Link Graphs

RAG systems use link relationships to understand topical authority and content hierarchy. Link related concepts using consistent anchor text to reinforce entity relationships.

Link from feature pages to the main product page using the product name as anchor text. Link from use case pages to relevant feature pages using descriptive anchor text. Link from blog posts to documentation using terms defined in your DefinedTerm schema. These links help RAG systems understand that your content forms a coherent knowledge graph, not isolated pages.

Common RAG Optimisation Pitfalls to Avoid

Publishing long, unstructured pages. A 5,000-word page with no H2 headings produces poor RAG chunks. Break content into logical sections with clear headings that function as retrieval targets.

Burying key information in PDFs. Most RAG systems retrieve web content more reliably than PDF content. Publish critical specifications, pricing, and capabilities as HTML pages with schema, not just downloadable PDFs.

Using jargon without definition. If you use proprietary terms or industry acronyms, define them inline and use DefinedTerm schema. RAG systems cannot retrieve passages containing undefined terms confidently.

Blocking AI crawlers. Some B2B companies block user agents containing “bot” or “crawler” strings, inadvertently blocking AI retrieval systems. Use IP-based rate limiting instead of blanket user-agent blocks.

Inconsistent entity naming. If your company is referenced as “Acme”, “Acme Inc.”, “Acme Manufacturing”, and “Acme Corp” across different pages, RAG systems struggle to unify these references into a single entity. Use a consistent legal name and define alternate names in your Organisation schema’s alternateName property.

What Changed Recently in RAG

2024: Anthropic published research on Contextual Retrieval showing that adding contextual prefixes to chunks and combining with reranking reduced retrieval failures by up to 67% compared to naive chunking approaches.

2025 Q3-Q4: OpenAI upgraded ChatGPT search to use real-time RAG with sub-second retrieval latency, making freshness signals more critical for time-sensitive queries. Google began using multi-hop RAG – retrieving sources, then retrieving additional sources mentioned in initial results – for complex queries requiring synthesis across multiple domains.

2020: Meta AI Research published the foundational RAG paper (Lewis et al.) introducing the retrieval-augmented generation architecture and demonstrating significant improvements in factual accuracy on knowledge-intensive tasks. In the same year, Google Research published REALM (Guu et al.), demonstrating that retrieval-augmented pre-training outperformed prior methods by 4-16% absolute accuracy on open-domain question answering benchmarks.

How CiteCompass Approaches RAG Optimisation

CiteCompass approaches RAG optimisation as the technical foundation underlying all AI visibility work. While GEO (Generative Engine Optimisation) and AEO (Answer Engine Optimisation) address broader strategic concerns, RAG optimisation focuses specifically on the retrieval and citation mechanics that determine whether your content surfaces in AI responses.

Layer 1: Content Structure

We audit existing content for chunk quality, heading structure, and self-contained sections. Most B2B content is written for human reading flows – progressive narrative, linked concepts – rather than RAG retrieval, which demands standalone chunks with explicit entity references. Retrofitting content for RAG retrieval involves restructuring sections to function independently while maintaining coherent narrative for human readers.

Layer 2: Entity Architecture

We implement comprehensive schema markup that explicitly defines your company, products, services, people, and proprietary concepts as distinct entities. This creates the entity disambiguation signals that RAG systems require to cite sources confidently. Without entity architecture, RAG systems may retrieve your content but attribute it generically (“according to industry sources”) rather than citing your brand.

Layer 3: Verification Signals

We establish freshness timestamps, cross-surface consistency (synchronised information across web content, feeds, and live AI data surfaces), and third-party validation (reviews, certifications, mentions). These signals increase RAG retrieval confidence by allowing AI systems to verify claims across multiple sources.

What Makes This Approach Different

Most content optimisation focuses on human engagement metrics such as time on page, scroll depth, and conversion. RAG optimisation requires measuring AI engagement metrics: retrieval frequency, citation attribution rates, and passage coherence when extracted in isolation. The CiteCompass AI Visibility Suite provides monitoring tools that track which content gets retrieved by AI systems, whether retrieval results in citation, and where citation gaps exist relative to competitors.

We do not claim that RAG optimisation alone guarantees AI visibility. Citation Authority results from combined optimisation across content quality, technical implementation, entity signals, and trust markers. However, without RAG optimisation, even excellent content remains undiscoverable to AI systems. RAG is the necessary technical foundation, not the complete strategy.

References

Lewis, P., et al. (2020). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. Meta AI Research. https://arxiv.org/abs/2005.11401 – Foundational paper introducing RAG architecture, demonstrating that retrieval-augmented models produced more factual and specific language than purely generative models on knowledge-intensive tasks.

Guu, K., et al. (2020). REALM: Retrieval-Augmented Language Model Pre-Training. Google Research. https://arxiv.org/abs/2002.08909 – Introduces REALM architecture showing that retrieval-augmented pre-training outperformed prior methods by 4-16% absolute accuracy on open-domain question answering benchmarks, with qualitative benefits in interpretability and modularity.

Anthropic. (2024). Contextual Retrieval. Anthropic Research. https://www.anthropic.com/news/contextual-retrieval – Demonstrates that adding contextual prefixes to chunks and combining with reranking reduces RAG retrieval failures by up to 67% compared to naive chunking approaches.

What Is RAG? How it Determines Your AI Visibility