The Newsroom · guide · 20 August 2026
RAG Analysis
RAG analysis audits content the way AI retrieval actually reads it: query fan-out, embedded passages and citation assembly. The approach, the insight, and three case studies.
By Piers Butler
RAG analysis audits content by simulating the retrieval pipeline that AI assistants actually run. Retrieval-augmented generation systems do not read pages the way people do: they decompose a question into many parallel sub-queries, retrieve the best-matching passages from an embedded corpus, and assemble a cited answer from the union. RAG analysis reproduces that pipeline against your own content, building the fan-out query cluster for each URL, embedding your passages beside the pages the corpus already holds, and measuring whether you would be retrieved, and cited, at each step. The output is a diagnosis of exactly where in the pipeline your content drops out.
How does AI retrieval actually choose content?
Every major AI answer surface, Google's AI Overviews and AI Mode, ChatGPT search, Perplexity, Claude, runs some version of the same three-stage pipeline. First, query fan-out: the user's single question is silently decomposed into many concurrent sub-searches. A query like "best system scaffold for residential builds" fans out internally into variants covering cost comparisons, regulations, rental options and alternatives, each run in parallel. Second, retrieval: for each sub-query the system pulls the best-matching candidates, and the match is made on meaning, embedded passages compared in vector space, not on exact keyword strings. Third, assembly: the answer is synthesised from the union of retrieved passages, and the citations go to the pages those passages came from.
Each stage changes what "optimised" means. Fan-out means the unit of competition is no longer a keyword but a semantic neighbourhood of twenty to fifty related queries per URL. Embedding-based retrieval means a passage must be self-contained and semantically dense to be matched at all. And citation assembly means the winning content is the passage that adds something the other retrieved passages do not, because a synthesiser has no reason to cite a paraphrase of what it already has. Upstream of all three sits a brutal gate: the corpus only contains what was crawled and indexed, so a page outside the index is not losing the retrieval competition, it was never entered.
What are the steps of a RAG analysis?
The approach, developed by Laurelin Labs across its audit engine and content tooling and used in Newsroom's own editorial pipeline, mirrors the pipeline stage for stage.
| Step | What is done | Pipeline stage it tests |
|---|---|---|
| 1. Gate check | Confirm every audited URL is indexed and snippet-eligible | Corpus membership |
| 2. Cluster build | Construct each URL's fan-out cluster: head, mid and tail queries from ranking data, keyword expansion and embedding similarity | Query fan-out |
| 3. Passage embedding | Embed the page's passages and the top competing pages for each cluster query into the same vector space | Retrieval matching |
| 4. Information gain | Score what each passage adds beyond the retrieved competition; judge first-hand evidence with quoted spans | Citation assembly |
| 5. Dual-surface tracking | Track classic rankings and AI citations for every cluster query, in parallel, over time | Outcome verification |
Two design principles keep it honest. The comparison set is always the real competition: information gain is measured against the pages actually retrieved for the cluster's queries, not against an abstract quality bar. And judgments come with evidence: where a language model grades something qualitative, such as first-hand experience, it must return verbatim quoted spans from the text, so a human can confirm the finding in seconds rather than take a score on faith.
What insight does RAG analysis provide that a normal audit does not?
Four findings recur, and each reverses a habit of classic SEO. First, the cluster replaces the keyword: pages briefed against two or three exact-match terms systematically under-cover the fan-out, and the fix is briefing against the neighbourhood. Second, extraction beats persuasion: passages written to be lifted whole, self-contained, front-loaded, one claim per passage, get retrieved; elegant prose that only works in context does not. Third, information gain is the citation currency: where a page merely restates what the corpus already holds, the analysis shows it being retrieved and then passed over in assembly, which no rank tracker can see. Fourth, the gate dominates everything: the most common catastrophic finding is not weak content but content missing from the corpus entirely, through indexation failures no one was watching. The companion piece on the AEO gate checks covers that layer in full.
What does RAG analysis look like in practice?
Method note: the three case studies below are drawn from real engagements analysed with this methodology by Laurelin Labs. Client identities are withheld and figures are rounded; each is presented as a generic case study to illustrate the analysis, not to identify a client.
Case study 1: the publisher whose content left the corpus
A specialist industrial-energy publisher migrated its site and reported that nothing seemed lost. RAG analysis started, as always, at the gate, and found the opposite: more than 600 legacy URLs, together holding over 320,000 impressions and still ranking in positions four to eleven, had been left out of the migration inventory and were returning errors or redirecting to the homepage. Indexed pages were decaying month on month, and weekly impressions had fallen roughly 65 percent. No amount of content work could matter until re-entry into the corpus was fixed, so the recovery plan was hundreds of exact redirect rules before a single word was rewritten. The retrieval lesson: a corpus does not warn you when your content leaves it.
Case study 2: the hub pages nobody could retrieve
The same engagement surfaced a subtler retrieval failure. The publisher's seven topic hub pages were structurally correct, every article nested under the right topic, yet earned close to zero impressions between them, and cluster analysis showed why: only about seven percent of the topic's validated search demand was phrased at the broad hub-title level, while over ninety percent sat in narrower fan-out queries about specific technologies and applications. The hubs were the right vessel with nothing retrievable inside: no topic sections, no self-contained passages matching the narrow queries, no inbound links with topical anchors. The fix was content inside the existing structure, not a new structure, which is precisely the distinction a retrieval-level analysis exists to make.
Case study 3: the navigation rebuilt around demand clusters
A B2B music-industry publication wanted its information architecture to reflect how its market actually searches. The analysis embedded and clustered roughly 13,700 real queries, sized the top clusters at about 41,000 monthly searches, and found that a single theme, AI in the industry, carried around seventy percent of top-of-funnel demand the editorial structure barely acknowledged. Four candidate architectures were then tested against the clusters: the best housed effectively all of the measured demand, while a purely editorial structure housed under one percent of it. The retrieval lesson: an information architecture is a bet on which clusters exist, and the bet can be measured before it is made.
Frequently asked questions
What is RAG analysis in plain terms?
It is auditing your content by simulating the pipeline AI assistants actually run: decompose the target question into its fan-out queries, embed your passages alongside the pages the retrieval corpus already holds, and measure whether your content would be retrieved and cited. It replaces guessing about AI visibility with a measurement of it.
How is RAG analysis different from classic keyword rank tracking?
Rank tracking measures one page against one query on one surface. RAG analysis measures a page against its whole query cluster on two surfaces at once: classic search rankings and AI citations. A page can be invisible in classic results yet cited heavily by AI assistants, or the reverse, and only tracking both reveals which.
Are the case studies in this article real?
Yes. They are drawn from real client engagements analysed by Laurelin Labs, with identities withheld and figures rounded. They are presented as generic case studies to illustrate the method, not to identify any client.