Passage indexing is the ability of search and AI retrieval systems to index and rank individual passages or sections within a page, rather than treating the entire page as a single unit. Google introduced passage indexing in 2021; the same concept underlies how AI retrieval systems decide which excerpt of your page to surface for a given query, even if the rest of the page is less relevant.
Why passage indexing matters for AEO
A page may comprehensively cover a topic but have one section that is uniquely relevant to a specific query. Passage indexing allows retrieval systems to:
- Score and retrieve that specific section independently from the rest of the page
- Cite that excerpt in an AI response even if the overall page rank wouldn’t have surfaced it
- Serve a query that only partially overlaps with the page’s main topic
Example: A long guide on “content marketing strategy” may include a section on “content optimization for AI engines.” Passage indexing allows that specific section to be retrieved and cited in response to “how do I optimize content for AI?”, even though the broader page is about marketing strategy.
Passage indexing vs. chunking
The concepts are closely related but distinct:
| Passage indexing | Chunking | |
|---|---|---|
| Who does it | Search engine / retrieval system | RAG pipeline (pre-LLM generation) |
| Unit | Semantically coherent passage | Fixed or adaptive token segments |
| Purpose | Relevance ranking of specific content sections | Fitting content into LLM context windows |
| What you can influence | HTML structure, heading hierarchy, paragraph clarity | Same, plus token-aware content organization |
In practice, for AI retrieval systems the two merge: chunks are often aligned to passage boundaries, and re-rankers then score those chunks as passages.
Writing for passage-level retrieval
The practical implication: every section of your page is a potential entry point for retrieval, not just the page as a whole. This means:
- Each section should answer a specific question: write section headings as questions, or immediately follow a heading with a direct answer to the implied question
- Passages should be self-contained: a retrieved passage is shown without surrounding context; it needs to be intelligible and credible alone
- Include entity context in each passage: mention your brand name and the relevant category/topic in each major section so a retrieved passage always conveys the full context
- Don’t bury your best content: if your most citation-worthy claim is on page 3 of a long document, it may be indexed but never retrieved; distribute key claims across early sections
Identifying which pages are being cited
Retrieval happens at passage level, but what an engine hands back (and therefore what LLM Metrix can record) is a page. In the Citation Intelligence view you see the specific URLs and domains each engine cited, and the prompt that produced each one. The page is the finest unit; there is no anchor, fragment or excerpt to click through to, because engines don’t report which section of your page they drew on.
That still tells you most of what you need, as long as you do the last step yourself:
- A URL cited repeatedly, across engines or across prompts, is a page that is being retrieved from. Open it and read it against the prompts it won on: the section that answers those questions is almost certainly the passage doing the work. Strengthen that section, and add its siblings for the adjacent questions.
- A page you expected to be cited and never is may be too diffuse for any single section to surface with confidence. Splitting one sprawling page into several focused ones, or tightening its headings so each section answers one question outright, is the usual fix.
What you cannot do here is confirm the passage: nothing in the product identifies a section within a page, counts how often one is retrieved, or distinguishes “cited a different part of the page this week” from “cited the same part”. Mapping a cited URL back to a section is a judgement you make, and it is worth making, but don’t treat it as measured.