Definition
Relevance engineering is a framework for search and AI visibility, created by Mike King of the agency iPullRank, that treats being found as an engineering problem: measuring how relevant content is to a query with information retrieval methods and changing it until the numbers improve, rather than working through an SEO checklist. iPullRank describes it as the intersection of information retrieval, artificial intelligence, user experience, content strategy and digital PR, and abbreviates it r19g, a shorthand in the style of i18n for internationalization.
King introduced the framework in his opening keynote at iPullRank's first SEO Week, held in New York from April 28 to May 1, 2025. In the talk he called SEO poorly defined and said relevance engineering applies to every search surface, from Google and AI Overviews to conversational assistants and Amazon. iPullRank's introduction to the framework and its later AI Search Manual set out the method in more detail.
The framework rests on a few technical ideas:
- Relevance can be scored. Queries and documents are represented as vectors (embeddings), and the closer a passage's vector is to the query's, the more relevant a retrieval system considers it. iPullRank recommends checking drafts against target queries, for example with cosine similarity, before publishing.
- Retrieval happens at the passage level. AI Overviews and other answer engines lift short fragments, so content should be broken into clearly scoped units (iPullRank suggests paragraphs of about 50 to 150 words) with explicit subject, predicate and object statements. This ties directly to passage ranking.
- Engines expand queries. Query fan-out and semantic search mean a page competes on many related sub-queries, not one keyword.
- Visibility is off-site too. Reviews, forums and digital PR mentions are part of the system being engineered, and link or PR targets should be chosen for topical fit rather than authority scores alone.
Relevance engineering is not a rival name for the discipline. iPullRank's AI Search Manual uses generative engine optimization as the name for optimizing toward AI answers and presents relevance engineering as its method: the way of doing GEO with measurement rather than guesswork. In practice the framework overlaps heavily with technical SEO, content strategy and vector search concepts that information retrieval researchers have used for years; what it adds is a push for SEO teams to adopt those tools and to plan content, UX and PR together.
Two caveats apply. It is a proprietary agency framework, not a standard or a body of peer-reviewed research. And some of the claims attached to it, such as Google modeling site and author expertise by averaging vector embeddings, are iPullRank's interpretation rather than documented Google practice. The useful part for most teams is the habit it promotes: test whether a passage actually matches the questions you want to win, instead of assuming it does.
Examples of Relevance Engineering
- A content team embeds its draft pricing page and the query 'how much does project management software cost for 20 users', finds the pricing section scores far below a competitor's page, and rewrites it to state per-seat prices in the first two sentences.
- An insurer splits a 1,200-word FAQ article into 40 self-contained question and answer blocks of about 80 words each, so each block can be retrieved and quoted on its own by AI Overviews.
- A B2B SaaS company drops a guest-post target with high domain authority but unrelated topics and instead pitches a smaller trade publication whose articles are close to its own content in embedding space.
- A team runs the sub-queries an engine might generate for 'best CRM for law firms', maps each one to a passage on its site, and writes new sections for the three sub-queries with no matching passage.
