Citation signals

This page is the writing half of GEO. Turning GEO on adds a TL;DR block, a date and an author to your pages; nothing in a toggle decides whether a sentence is liftable. That is decided by how the section around it is written, and the rules below are the ones with a mechanism behind them.

Every rule is stated the same way: the rule, why the retrieval and generation stack behaves that way, and a source you can open. Where the evidence is thin, the rule says so.

The rules at a glance#

Why passages, not pages#

Google describes its passage ranking system as one that identifies "individual sections or 'passages' of a web page to better understand how relevant a page is to a search" (ranking systems guide). Retrieval-augmented assistants do the same thing more aggressively: the document is split into chunks, each chunk is embedded and scored on its own, and only the winners reach the model.

Anthropic states the failure mode plainly. A chunk reading "The company's revenue grew by 3% over the previous quarter" "doesn't specify which company it's referring to or the relevant time period", so it cannot be retrieved for a question that names the company. Re-attaching that context before embedding "reduced the top-20-chunk retrieval failure rate by 35%", and combined with contextual BM25, "by 49%" (Anthropic, contextual retrieval).

That is a vendor working around prose that does not name its own subject. Prose that does name its subject needs no such repair. Which gives the two rules that matter more than all the rest:

• Name the subject in full inside every section. "To rotate it, call the endpoint" is unretrievable — nothing in it says which product, or that the thing being rotated is an API key.
• Answer in the first sentence after the heading, then elaborate. "Before we get into rotation, it is worth understanding…" is the answer arriving too late to be the passage that gets picked.

The quote test. Paste a section into an empty file. If it no longer says what it is about, rewrite it. That is exactly the operation a retriever performs on your page.

What Docsbook does with these rules mechanically#

Three parts of Docsbook read the shapes above directly, so following the rules changes real output rather than only your odds:

• Your first paragraph becomes your machine summary. If a page has no tldr: in frontmatter, the GEO TL;DR block is built from the document's first real paragraph — headings, blockquotes, lists, code fences and image-only lines are skipped — capped at 280 characters. A lede shorter than 40 characters produces no block at all.
• A question heading becomes structured data. With AEO on, any ### heading ending in a question mark is emitted as a Question in FAQPage JSON-LD, with the paragraphs beneath it as the answer, up to 20 questions per page and 1,000 characters per answer. Phrasing a heading as the reader's question is therefore not a style preference — it is the input to a detector.
• audit_geo checks the mechanical preconditions, including whether at least 200 words of body prose are present in the raw HTML with no JavaScript executed, and whether the page states any date at all.

Rules with a measured effect size#

The GEO paper (KDD 2024) built GEO-bench, which "consists of 10K queries", and tested nine content transformations against a generative-engine setup using "only the top 5 sources fetched from the Google search engine for every query". Its visibility metric, Position-Adjusted Word Count, weights how much of the generated answer is attributed to your source by where the citation appears.

These are the figures the paper itself prints. It reports bands and a headline, not a per-method percentage table, so neither do we:

Read those bands as "which sentences a model reuses once your page is already in front of it", not as a traffic forecast — see the limits below. The paper also reports that effects vary by domain: its top-performing tags for Statistics Addition are Law & Government, Debate and Opinion, and for Quotation Addition People & Society, Explanation and History.

Practically: give every page something extractable — a limit, a timeout, a version, a one-sentence definition, or a number with its source named. Concrete facts are what a model lifts verbatim. And never invent one: a wrong figure repeated by an assistant is worse than no answer, and it will outlive your correction.

Rules about what not to do#

Do not keyword-stuff. It is the only transformation in the benchmark that scored below leaving the page alone.

Do not strip your prose down to make it "quotable". The 2026 critical survey of 45 GEO studies reports a controlled result (Kim et al., 2026; 171,003 documents, 2,700 queries) in which optimising only the body of a page "reduces average top-20 presence by approximately 9%, top-10 presence after reranking by 16%, and final citation by 6%" (arXiv 2607.14035). The survey's explanation is the one to remember: a rewrite can raise the chance of being cited given retrieval while lowering the chance of being retrieved at all, and the total effect goes negative. Prose carries the synonym variety that matches varied questions. Add the definition and the number alongside it, never instead of it.

Do not block the agents that do the citing. Vendors run different crawlers for different purposes, and blocking one does not block the others:

Sources: OpenAI, Anthropic, Perplexity. Blocking training while allowing search is a coherent business decision. Blocking the search agent while expecting citations is the incoherent one, and it is almost always accidental — audit_geo reports it as critical.

Do not expect a special file or schema to substitute for the writing. Google states there are "no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary", and "You don't need to create new machine readable files, AI text files, or markup" (Google AI features). See llms.txt for the same question asked of that file specifically.

Limits and open questions#

• The measured gains are conditional on already being retrieved. The 2026 survey concludes the foundational results are "valid within its experimental setting but conditional on a source already being present in a fixed context", and that "no reviewed technique shows a stable, longitudinal, cross-platform causal effect on organic discoverability or downstream behavior". Everything above improves your odds once you are in the context window. Getting into it is SEO work.
• Under question: does freshness raise citation rate? Google documents "query deserves freshness" systems for Search ranking, and a page with no date gives a model nothing to attribute currency to. No primary source measures a citation-rate lift from a visible date. Treat freshness as hygiene until someone publishes a measurement.
• One run proves nothing. Audits summarised in the survey: Schulte et al. (2026) "observe daily source-level Jaccard scores of approximately 0.34–0.42" across four engines and 45 days, and propose "seven to eight repetitions per prompt as a starting point"; Kirsten et al. (2026) report "page overlap across two months is 18% for AI Overviews, compared with 45% for organic Google". Repeat any before-and-after check several times before believing it.
• The numbers above are from one benchmark and one era. GEO-bench was evaluated in 2024 on a fixed engine configuration. The direction of the findings has held up in review; the magnitudes should not be quoted as your expected result.
• Docsbook measures none of this for you. What your analytics can show is the trace: crawler hits from assistant user agents and referral traffic from assistant domains. A citation you did not observe is not a metric.

Related#

• GEO — what Docsbook injects into the page, and how to check it.
• llms.txt — the site-level machine index and its evidence.
• AEO — question headings, FAQPage and HowTo markup.
• SEO — indexing and crawlability, the stage before citation.
• Analytics — where assistant crawler hits and referrals show up.