Retrieval-augmented generation (RAG)

A RAG system has two halves. The retriever takes a query, searches an external corpus such as a web index or a vector database, and returns the passages it judges most relevant. The generator, a large language model, then writes an answer conditioned on those passages plus the original question. The 2020 paper that named the technique, from researchers at Facebook AI Research, University College London and New York University, framed this as combining parametric memory, the knowledge baked into model weights during training, with non-parametric memory, the documents fetched at inference time.

The practical benefits are the reason every major AI search product adopted the pattern. Retrieval gives the model access to information newer than its knowledge cutoff, lets operators update knowledge by updating the index instead of retraining the model, and supplies provenance: because the answer is grounded in specific retrieved passages, the system can show which pages it drew on. That provenance is what surfaces as the citation links in Perplexity, ChatGPT search, Gemini and Google AI Overviews.

In production AI search, the pipeline is more elaborate than the original paper. A user question is often decomposed into several background searches, a pattern Google calls query fan-out in AI Mode. Retrieved pages are split into chunks, re-ranked for relevance, deduplicated, and trimmed to fit the model's context window. The generator then synthesizes one answer from the surviving chunks. Every stage is a filter, and a brand can be eliminated at any of them.

Before RAG, a model could only recommend brands it memorized during training, which favored companies with years of accumulated coverage. With RAG, the candidate set is assembled fresh for each question from whatever the retriever finds. That changes the competition: the fight is now about being in the retrieved set and surviving the re-ranking, chunking and synthesis stages, and it is winnable by newer brands with the right content.

Retrievability comes first. If crawlers such as OpenAI's OAI-SearchBot cannot fetch your pages, or the underlying index has never seen them, the generator cannot cite you no matter how good the content is. After retrieval, chunk-level quality matters more than page-level authority. Models pull passages, so a page that states a clear claim with a number and a named source in one self-contained paragraph gives the generator something it can quote. Sprawling pages that bury the answer tend to lose out even when they rank well in classic search.

The evidence base behind AI answers also differs from organic rankings. An Ahrefs analysis of ChatGPT's most-cited pages found 28.3 percent of them had zero Google organic visibility, which means retrieval pipelines routinely surface pages that rank trackers would call invisible. Treating RAG visibility as a byproduct of SEO therefore misses a large share of what actually gets cited.

Because retrieval happens per query, presence is measured with prompts, and it must be measured repeatedly. Build a set of the real questions buyers ask in your category, run them on a schedule against the engines that matter, and record two things for every answer: whether your brand was mentioned, and which sources were cited. Mention rate tells you whether you appear in the synthesized text; citation share tells you whether your domain is part of the evidence the generator used.

Single checks mislead, because RAG output is volatile. Retrieval results shift as the index updates, and generation is probabilistic, so the same prompt can produce a different shortlist an hour later. SparkToro's 2025 research found under a 1 percent chance that two identical ChatGPT runs name the same set of brands. Sound measurement averages across many runs and prompts rather than reacting to any one answer. This is the approach Reachroller takes, tracking ChatGPT live through the official API with web search enabled and scoring visibility from stored evidence rather than single snapshots, with other engines rolling out.

Diagnosis follows from the pipeline stages. If your pages are never cited, check crawler access and indexation first. If you are cited but never named in the answer text, the problem is usually positioning: the retrieved passage supports a fact without presenting your brand as an option. If you are named inconsistently, that is normal variance, and the fix is broadening your retrievable footprint so more of the candidate passages mention you.