In-Context Ranking
Re-ranking candidate documents for a query by feeding them all into an LLM's context and letting it pick the most relevant.
In-context ranking (ICR) is a way of re-ordering candidate documents for a query by placing the query, the document list and a task description directly in a language model's context window and letting the model identify the most relevant items. It leans on the LLM's contextual understanding instead of a separate scoring model.
The catch is cost. The attention mechanism scales quadratically with input length, so doubling the number of documents can quadruple the compute — impractical for large candidate lists. This is the bottleneck that BlockRank tackles by exploiting the block-sparse structure of attention.
ICR matters for AI SEO because it's increasingly how relevance gets decided inside generative systems: not classic keyword scoring, but a model reading candidates and choosing. It's a core part of modern relevance engineering.
Paper: Scalable In-context Ranking with Generative Models
Related concepts
Mentioned in
-
New Feature: Grounding Snippet Ablation
The new Ablation tab in DEJAN's Content Optimizer measures how many rank places each sentence of an AI-cited source is worth.
-
Gemini 4 Argon's personality: no hype, hard facts, fewer tables
Release-day tests on Gemini 4 Argon: it punishes hype and still wants hard facts like prices, but tables, a favourite of every Flash model, fall from 93.6% to 31.5%.
-
Impact of Parametric Memory and Grounding on Ordinal Positioning of Brands in Generative Responses
Mathematical formulas for evaluating entity association, semantic embedding similarity, ordinal rank positioning, and generation performance in AI models.
-
AI Influence With Confidence Threshold Manipulation
A study reveals that language model abstention is governed by a single internal confidence signal that outranks question difficulty and fact retrievability.
-
AI Models Encode Brand Data but Fail to Recall a Third of It
Google Research found frontier AI models encode 95-98% of brand facts in parametric memory but fail to recall 26-34% of them. The bottleneck is retrieval, not storage.
-
Gemini picks whichever web page it reads first 92% of the time. Only pages rewritten to its measured preferences ever beat that.
Shown two web pages, Gemini picks the first one 92% of the time. Pages rewritten to 31 measured text preferences win 15% from the second slot. Ordinary pages: 0%.
-
The Paradox of Average and AI Visibility
A study by DEJAN AI examines how AI rankers prioritize prototypical content over brand authority, suggesting a dual strategy for SEO and human conversion.
-
BlockRank: A Faster, Smarter Way to Rank Documents with LLMs
BlockRank is a novel method for in-context ranking that uses structured sparse attention and contrastive training to improve LLM efficiency and accuracy.
