Passage Ranking and GEO: Optimal Content Structure for AI Citation

Why passage ranking matters for GEO

Search engines and AI systems stopped evaluating content only at the page level a long time ago. They split pages into individual passages and evaluate each passage independently. For visibility in AI answers, this shift is fundamental: the page as a whole no longer competes for a spot. Each individual, self-contained block of text does.

Google announced what it called passage indexing in October 2020 and began rolling it out in February 2021. According to Google, the technique affects 7 percent of all search queries across all languages once fully rolled out globally. The distinction matters: Google still indexes entire pages, but it additionally evaluates the content and meaning of individual passages. The more precise term is therefore passage ranking, not passage indexing.

With the rise of generative search, this principle has gained a new dimension. AI Overviews, ChatGPT, and Perplexity don't extract entire pages. They extract atomic units of information: individual sentences, list items, or table cells. If you want to be cited in these answers, you have to structure your content so that these units are fully understandable without surrounding context. This is exactly where Generative Engine Optimization (GEO) comes in.

In the DACH region, this development has reached Google Search itself: in March 2025, Google rolled out AI Overviews in Germany, Austria, and Switzerland, at launch only for signed-in users aged 18 and older. Since then, German-language pages have been competing for citations in AI answers directly within the search results.

How passage ranking works and turns into GEO citations

The mechanism runs in two steps. First, the search engine crawls the complete page (full-page crawling). Then an NLP model segments the content into thematic passages, guided by headings and semantic shifts in topic. Each passage receives its own relevance score for the query. This way, an otherwise broad page can rank for a very specific question if a single section answers that question precisely.

Generative systems go one step further. When they build an answer, large language models don't just grab any paragraph. They extract the smallest meaningful unit of information. A clearly worded sentence, a precise list item, or a table row can go straight into a generated answer and be attributed to a source. Long, convoluted paragraphs without a clear statement are unsuitable for this.

Research shows that this kind of structuring has a measurable effect. The Princeton paper on Generative Engine Optimization (KDD 2024) demonstrates that targeted GEO methods can increase visibility in generative answers by up to 40 percent. The study also points out that the effectiveness of individual tactics varies by subject area, which makes domain-specific optimization necessary.

A practical structural pattern amplifies this effect further. A data analysis of around 25,000 URLs and nearly 400 million AI citations across six AI engines found that 63 percent of all citations referred to listicles. Depending on the model, the share ranged from 40 to 65 percent, with Copilot at the low end and Gemini at the high end. AI systems clearly prefer to extract structured content that is broken down into points.

Best practices for a citable content structure

The following principles make content optimally usable for passage ranking and AI citation.

• Answer-ready blocks directly after the heading: Place a self-contained answer of about 40 to 60 words immediately below every subheading. This block must fully answer the question implied by the heading without needing the rest of the text. This is exactly the unit an LLM draws on as a citation.
• Consistent heading hierarchy: Use a clean H2-over-H3 structure without skipped levels. Clear headings give the NLP model the cut lines for passage segmentation. A chaotic hierarchy makes it harder to separate topics cleanly.
• Question-style subheadings: Phrase headings as concrete questions or precise statements that match search intent. That way, they directly mirror the prompts users type into AI systems.
• Answer-focused section length: Keep thematic sections compact, usually in the range of about 120 to 180 words per unit of meaning. A passage should cover exactly one topic completely before the next one begins.
• Use bullet lists and tables: Turn enumerations, comparisons, and step-by-step instructions into lists and tables. These formats deliver the atomic units that AI systems prefer to extract, as the high share of listicles among citations shows.
• Statistics and authoritative language: Back up key claims with concrete, linked data and phrase them precisely. Princeton's GEO research counts adding solid evidence among the effective levers for increasing the likelihood of being cited, but it also points out that the effects depend on the topic.

Structured data as a grounding signal

Schema.org markup in JSON-LD format helps AI systems clearly identify the entities and content types on a page. For editorial content, Article, FAQPage, and HowTo are the most relevant types. This markup provides a machine-readable grounding signal that supports the thematic classification of a passage.

This is where a concrete competitive advantage lies. According to the HTTP Archive Web Almanac, JSON-LD was present on 41 percent of all pages in 2024, compared with 34 percent in 2022. Content-specific types, by contrast, are rare: schema.org/article reached only 1.77 percent and schema.org/product 1.50 percent. If you consistently mark up articles, FAQs, and how-to guides semantically, you stand out from the vast majority and support the AI discovery of your content.

Technical foundation: crawlability and performance

No passage can be cited if it isn't reliably captured. The technical foundation is therefore a prerequisite, not an optional extra.

Crawlability comes first. Content must be delivered server-side or prerendered so that crawlers and AI bots can read it without executing JavaScript. In a Next.js architecture, this means consistently using server-side rendering or static generation for all citable content. Text that is loaded purely on the client side remains invisible to many AI crawlers.

Delivery speed matters too. Fast pages served via edge and CDN make frequent, complete crawling easier. Core Web Vitals still leave room for improvement: according to HTTP Archive, 48 percent of mobile pages passed all three Core Web Vitals in 2025. In detail, 62 percent of mobile pages achieved a good LCP, 77 percent a good INP, and 81 percent a good CLS. The weakest single metric is therefore Largest Contentful Paint, the load time of the largest visible element. That is where the first optimization step pays off.

AI crawler management: weighing visibility against protection

If you want to appear in AI answers, you need to manage AI crawlers deliberately. In robots.txt, you can allow or block bots such as GPTBot, ClaudeBot, and PerplexityBot individually. This is a strategic trade-off between visibility and content protection.

The data shows how unbalanced the ratio of effort to return currently is. According to Cloudflare, the ratio of crawls to referrals in July 2025 was around 38,000 to 1 for Anthropic, about 1,091 to 1 for OpenAI, and around 5.4 to 1 for Google. In addition, 79 percent of AI crawling was for model training, not for live search. If you optimize exclusively for direct referral traffic, you should know these ratios. For most B2B providers, the interest in being present in the answers still outweighs this, which is why blocking AI crawlers completely rarely makes sense.

IndexNow also speeds up the discovery of new and changed content by actively notifying participating search engines of updates. Combined with edge and CDN control, this lets you manage which bots see which content and how quickly.

Common mistakes

• Convoluted paragraphs without a clear core message: Long blocks of text that mix several topics can't be segmented cleanly and therefore can't be cited. Every passage needs a clear, standalone statement.
• Answer buried far down the page: If the actual answer only appears after a long lead-in, the citable block directly below the heading is missing. AI systems then find no compact unit to extract.
• Keyword stuffing instead of natural language: Over-optimized, unnaturally repetitive text hurts readability for humans and machines alike. Precise, authoritative wording gets you further than keyword density.
• Content rendered only on the client side: Text that is only generated in the browser via JavaScript remains invisible to many AI crawlers. Citable content belongs in the server-rendered HTML.
• Structured data ignored: If you skip Article, FAQPage, or HowTo markup, you give away an effective grounding signal that offers a clear lead, given how rarely content-specific schemas are used.

Metrics and measurement in the DACH region

Success measurement is shifting from the click to the citation. Traditional click counts are becoming less meaningful because AI answers intercept clicks. According to an analysis of 300,000 keywords, the click-through rate for position 1 drops by around 34.5 percent when an AI Overview appears. This study refers to informational search queries.

At the same time, AI Overviews are volatile. According to Semrush, in 2025 they appeared for 6.49 percent of search queries in January, rose to 24.61 percent by July, and fell to 15.69 percent by November, measured across more than 10 million keywords. If you manage visibility, factor in these fluctuations instead of relying on a single reference date.

Useful KPIs for GEO are therefore citation tracking in ChatGPT, Perplexity, and AI Overviews, the number of mentions per platform, and visibility within generated answers. The question is whether and how often your own brand shows up as a source, not just how many sessions you get.

Then there is the legal side. Where tracking AI referrals and user interactions involves personal data, the GDPR applies in Austria and Germany. If the tracking stores or reads cookies or similar identifiers on the user's device, you additionally need consent under § 165 para. 3 TKG 2021 in Austria or § 25 TDDDG in Germany. The exception is anything strictly necessary for the service the user has explicitly requested. In Switzerland, the revised Federal Act on Data Protection (revDSG) has applied since September 1, 2023, and the GDPR may also apply to Swiss companies. If you analyze referral sources from ChatGPT or Perplexity, build privacy-compliant data collection into your setup from the start.

Further reading

Passage ranking and GEO go hand in hand: only a clean content structure broken down into atomic units makes a page citable, and only a technical foundation of crawlability, rendering, and performance makes that structure visible in the first place. For implementation, combine three building blocks: answer-focused content with answer-ready blocks and lists, consistent JSON-LD markup for articles, FAQs, and how-to guides, and fast, server-rendered delivery with targeted AI crawler management. It is also worth setting up citation tracking to make the shift from clicks to citations measurable.