In-Context Ranking (ICR): The Next Frontier in AI Search Algorithms

In-context ranking (ICR) is an AI-driven approach to information retrieval where large language models (LLMs) evaluate queries and candidate documents simultaneously within a single prompt to determine document relevance. Unlike traditional keyword-based search algorithms, ICR leverages deep semantic understanding to assess meaning and intent.

For SEO and content professionals operating in New Zealand, the UK, and beyond, understanding how ICR works is no longer optional; it’s the foundation of future-proof search strategy. This post explains how AI systems read and rank content, how ICR differs from legacy relevance models, and what technical steps you can take today to prepare your content for AI-driven search quality standards.

Introduction

Before we get too deep into this topic, it should be noted that this is a pretty technical post. Like any content covering algorithms and rankings, there is a lot of information to digest, and some of it will be relevant to you, and some sections may go into too much depth for your specific job role. Hopefully, you can gather enough insights to have a better understanding of ICR and understand AI-driven search.

Search has always evolved in waves. The shift from directory listings to keyword matching, then from keyword matching to semantic relevance, each wave redefined what it meant for content to be “found.” The next wave is already breaking, and it’s built on a fundamentally different principle: rather than matching words or even meanings in isolation, AI systems are now reading your content the way a subject expert reads a research paper; holistically, contextually, and with a judgment about whether your document truly answers the question in front of it.

This is in-context ranking (ICR), and it is reshaping what search quality actually means for technical SEO specialists, search strategists, and content professionals in New Zealand, the UK, and globally.

At Digital Hothouse, we have been tracking the mechanics behind AI algorithms since the earliest integration of machine learning into Google’s ranking stack. ICR is not a minor adjustment; it’s a structural rethinking of how search works at the document level. Here is what you need to know.

How AI Systems Read and Rank Content

To understand ICR, you first need to understand how modern AI systems process a document. Traditional search algorithms worked primarily through inverted index lookups: a document was scored based on the presence, frequency, and positioning of query-relevant terms. This was efficient and scalable, but fundamentally limited to surface-level signals.

Large language models (LLMs), by contrast, encode contextual relationships across entire passages. When an LLM processes your content, it is not tallying keyword instances; it is building a representation of what your content means, who it’s for, and how fully it addresses a given topic. This shift from word counting to meaning representation is at the heart of how AI algorithms now determine document relevance.

Research from Google DeepMind and Google Research has formalised this as In-context Ranking (ICR): a paradigm in which the task description, candidate documents, and the query are all placed together inside the model’s input prompt, and the model is tasked to identify the most relevant document or documents. This means the AI is not assessing your page in a vacuum; it is assessing it against competing documents, simultaneously, within a shared context window.

The implications for content creators and SEOs are significant. Being “good enough” no longer means meeting a minimum threshold; it means being demonstrably more relevant than the alternatives the AI is reading at the same time.

In-Context Ranking vs Traditional Relevance

Traditional machine learning ranking systems, such as RankBrain, Neural Matching, and early BERT integrations, were significant advances, but they still largely operated on pre-computed representations. A document was embedded into a vector space at indexing time, and queries were matched against those stored vectors. Relevance was a calculation made outside of and prior to the actual retrieval moment.

ICR works differently. According to the foundational research published on arXiv, ICR leverages the contextual understanding of LLMs by formatting the query and a list of candidate documents directly into the model’s input prompt, so that relevance judgements happen in real time, with full awareness of the competing content. The model is not recalling a pre-computed similarity score; it is reasoning about relevance in the moment.

Two structural properties identified in the research are particularly important for SEOs to understand:

• Inter-document block sparsity: document tokens attend primarily to content within their own block, rather than across all documents simultaneously. This means each document is largely assessed on its own terms, but the query signal threads through all of them.
• Query-document block relevance: specific query tokens develop strong attention weights towards relevant document tokens, particularly in middle processing layers. In plain terms: the AI is looking for the sections of your content that directly speak to the intent of the query.

For SEO practitioners, this is the technical grounding behind a shift many have observed empirically: traditional ranking position correlation with AI Overview inclusion has dropped to r=0.18, and 47% of AI Overview citations now come from pages ranking below position 5. The AI is not following the organic results; it’s making its own document-level judgements.

Query-Document Similarity at Scale

One of the longstanding challenges with ICR is computational efficiency. When a model needs to evaluate hundreds of candidate documents against a query simultaneously, attention computation scales quadratically with context length. Put simply: it gets very expensive, very quickly.

The Google DeepMind BlockRank approach addresses this directly. By structuring the attention mechanism to reflect the naturally sparse cross-document attention pattern, BlockRank reduces complexity from quadratic to linear without sacrificing ranking performance. In testing with Mistral-7B across the BEIR benchmark, MSMarco, and NQ datasets, the method processed up to 500 candidate documents within a single context window of approximately 100,000 tokens in under a second; a 4.7x efficiency gain over prior approaches.

In simple terms: the computational barriers that have limited widespread ICR deployment are being solved. As these efficiencies reach production systems, the coverage and frequency of AI-driven ranking will increase. Content that is not structured for deep read processing will be progressively disadvantaged.

For search strategists managing large content estates in New Zealand and the UK, the scalability of ICR is a signal to audit content architecture now, before AI-driven selection becomes the dominant path to visibility.

Content Density and Semantic Saturation

A common misconception in AI-era SEO is that longer content automatically performs better. ICR research and practical AI Overview data tell a more nuanced story: what matters is not length, but semantic completeness and entity density.

According to research analysing over 15,000 AI Overview results, content scoring above 8.5/10 on semantic completeness is 4.2 times more likely to be cited, while pages with 15 or more connected entities show 4.8 times higher selection probability. These are not marginal differences; they represent a structural advantage for content that has been deliberately engineered for semantic saturation: the practice of covering a topic with sufficient depth, breadth, and entity interconnection that an AI system can extract a complete, self-contained answer from your content without requiring additional context.

This aligns with the information retrieval principles that underpin how AI systems are trained and evaluated. Relevant documents are not just topically related; they are informationally complete relative to the query’s intent. For content teams, this reframes the brief: the question is not “how long should this article be” but “does this article, standing alone, answer the query completely and credibly?”

Critically, cosine similarity scores between content and target queries above 0.88 show 7.3 times higher AI selection rates than content scoring below 0.75. Closing that gap requires not just adding relevant keywords, but building genuine topical architecture: structured sections, clearly labelled sub-topics, explicit entity relationships, and answer-first formatting.

Diagnostic Checks for ICR Readiness

Before moving to technical implementation, it’s worth conducting a content-level audit against ICR criteria. At Digital Hothouse, our AI Optimisation service includes a structured ICR readiness assessment. The core questions are as follows:

• Answer completeness: Does each priority page answer its target query fully, without requiring the user to visit another page to complete the answer?
• Entity coverage: Does the content reference the key entities (people, places, concepts, organisations) that the AI would expect to see in a credible treatment of this topic?
• Passage extractability: Are there clear, self-contained passages of 100 to 200 words that could be extracted verbatim and still make complete sense as a standalone answer?
• Source authority: Does the content cite verifiable, authoritative external sources? AI systems now show 89% higher selection rates for content with cross-referenced, verifiable citations.
• Structural clarity: Is the content structured so that the answer to the primary query appears near the top, with supporting depth available below?
• Author expertise signals: Is there clear attribution to a named, credentialed author? 96% of AI Overview content now comes from verified authoritative sources.

If any of these checks reveal gaps, they represent direct opportunities to improve AI search quality scoring and increase the probability of citation in AI-generated results.

Technical Implementation: Semantic HTML, Structured Data, and Indexability

Content quality is necessary but not sufficient. Even the most semantically complete content will underperform if the technical infrastructure does not support AI crawling and comprehension. Here are the implementation priorities:

Semantic HTML

Use HTML elements that communicate document structure to both crawlers and AI systems. This means proper heading hierarchy (H1 through H3 reflecting actual content structure, not stylistic choices), paragraph elements for prose, and list elements for enumerable content. Avoid div-soup layouts where important content is buried in non-semantic containers.

Structured Data

Schema markup is a direct line of communication to AI ranking systems. Research consistently shows 73 to 89% higher selection rates for content with properly implemented structured data. Priority schema types for ICR-readiness include: Article and BlogPosting (with author, datePublished, and publisher), FAQPage, HowTo, and BreadcrumbList. For New Zealand and UK businesses, LocalBusiness and Organisation schema with region-specific identifiers support the localisation signals that AI systems use to serve geographically relevant results.

Indexability and Crawl Efficiency

An AI cannot rank what it cannot read. Core indexability requirements include: clean canonical signals, no conflicting noindex directives on pages intended for AI inclusion, fast server response times (AI crawlers apply similar thresholds to Googlebot), and logical internal linking that surfaces topically related content. XML sitemaps should be current and submitted. For large content estates, log file analysis remains the most reliable way to verify that priority pages are being crawled at the expected frequency.

Core Web Vitals and Page Experience

While ICR is a content-level process, the pages that host that content still need to meet baseline page experience standards. Google’s own guidance confirms that pages with poor Core Web Vitals face a compounding disadvantage: lower crawl priority, weaker traditional ranking signals, and reduced likelihood of appearing in AI-curated surfaces. Both our New Zealand and UK client data show that technical health and AI citation frequency are positively correlated.

What Most Businesses Are Missing

The majority of SEO content produced today is still optimised for traditional keyword ranking. It’s written around phrase density, internal linking targets, and DA-weighted backlink profiles. Those signals still matter, but they are increasingly insufficient for AI search visibility.

The businesses pulling ahead in AI-driven search results are those treating content as a knowledge asset rather than a keyword vehicle. They are mapping topics to entities, writing for answer extraction rather than click-through, structuring content so that passages stand alone as authoritative responses, and building author credibility signals into the content infrastructure itself.

This is the work that our SEO services and AI Optimisation programme are built around: not chasing algorithm updates reactively, but understanding the underlying mechanics well enough to build content strategies that remain effective as AI search quality standards continue to evolve.

The AI search engine market, valued at $43.6 billion in 2024, is projected to capture 62.2% of total global search volume by 2030. Organic click-through rates have already dropped 61% on queries triggering AI Overviews. The window to build ICR-ready content architecture while competitors are still operating on legacy models is open now, but it will not remain open indefinitely.

Related Reading from the Digital Hothouse Blog

• Understanding AI Overviews: What the Data Tells Us About Content Selection
• Semantic SEO in 2025: Moving Beyond Keywords
• Structured Data for AI Search: A Practical Implementation Guide
• E-E-A-T in the Age of Generative AI: What Authority Really Means Now

Frequently Asked Questions

What is in-context ranking (ICR) in simple terms?

In-context ranking is a method used by AI systems to evaluate multiple documents and a search query at the same time, within a single processing window. Rather than scoring documents individually against a pre-computed index, the AI reads the query alongside candidate documents simultaneously and determines which content most fully and accurately addresses the search intent. Think of it as having a subject expert read several articles side by side and decide which one genuinely answers the question best.

How is ICR different from traditional SEO ranking?

Traditional search ranking relied primarily on signals like keyword density, backlink profiles, and pre-computed semantic embeddings. ICR uses the contextual reasoning capability of large language models to make relevance judgements in real time. The practical difference is that traditional ranking is largely about matching signals; ICR is about genuine answer quality. A page can rank well organically through authority signals while still being overlooked by AI systems if it does not provide a complete, extractable answer to the query.

Does in-context ranking affect traditional organic search results?

ICR directly influences AI-generated surfaces such as Google AI Overviews, AI Mode results, ChatGPT Search citations, and Perplexity answers. Traditional organic rankings (the blue links) still operate on a partially separate set of signals, though the overlap is increasing. Importantly, 47% of AI Overview citations come from pages below position 5 in organic results, which confirms that AI selection and traditional ranking are not the same process. Optimising for ICR can improve AI visibility independently of organic rank.

What content changes make the biggest difference for ICR readiness?

The highest-impact changes are: writing self-contained answer passages of 100 to 200 words near the top of key pages; increasing entity density by explicitly naming and connecting the key concepts, people, and organisations your topic involves; adding verifiable citations to authoritative external sources; and implementing clear heading structure that signals what each section covers. Schema markup, particularly FAQ and Article schema, also shows consistent correlation with higher AI selection rates.

Is ICR relevant for businesses targeting both New Zealand and UK audiences?

Yes, and the signals differ slightly by market. UK AI Overview data mirrors US patterns closely in terms of query type distribution and selection criteria. New Zealand search behaviours tend to weight local authority signals more heavily, partly due to the smaller ecosystem of high-authority local sources. For businesses operating across both markets, content should be structured with universal semantic clarity (entity richness, answer completeness) while using region-specific schema where applicable; particularly LocalBusiness and geo-targeted structured data fields.

How do I know if my site is currently being cited in AI search results?

Several tools now track AI Overview citations, including SE Ranking’s AI Results Tracker, Semrush’s AI Toolkit, and manual SERP monitoring for brand and topic mentions. Google Search Console is beginning to surface some AI Overview impression data. For a structured audit of your current AI citation footprint and ICR readiness, contact the Digital Hothouse team for an AI Optimisation assessment tailored to your market and content estate.

What is BlockRank and should I know about it?

BlockRank is a method developed by Google DeepMind researchers that makes in-context ranking far more computationally efficient; processing up to 500 candidate documents simultaneously with a 4.7x speed improvement over prior approaches. It’s a research method at present, not a confirmed live Google ranking signal. However, it represents the direction of travel for AI-driven retrieval systems and signals that the computational barriers to wider ICR deployment are being actively removed. Staying current with developments like BlockRank is part of what separates proactive search strategy from reactive algorithm chasing.