Writing for AI Citation

1. What this playbook is

This playbook turns the seven citability signals into practical writing recipes. Each recipe explains what to write, what to remove, and what to preserve so a passage can pass the chunk-extraction test in the Citability audit. When the audit marks a signal with ⚠️ or ❌, use the corresponding recipe in §4. The underlying signals are defined in Citability §4.

Microsoft made the structural case directly in February 2026: “Clear headings, tables, and FAQ sections help surface key information and make content easier for AI systems to reference accurately” (Bing AI Performance in Webmaster Tools). Aggarwal et al. 2024 found that substantive changes, including citations, statistics, and quotations, increased visibility in AI answers, while keyword stuffing did not. The finding is useful, but the reported magnitude is best treated as an upper estimate: competitive pressure reduces the gains available to any single participant (C-SEO Bench).

Before structure can help, an AI system must be able to retrieve the page (AI Crawlers; see Generative Engine Optimization for the full process) and consider the source trustworthy enough to use (E-E-A-T). Structure cannot compensate for thin content, even when the page contains many citations, and low-substance patterns remain detectable (AI Content Detection). Once retrieval and trust are in place, the seven signals improve the page’s suitability for grounding.

2. Four decisions to make before writing

Make four decisions before writing the first sentence. They determine how each recipe in §4 applies to the current page and follow the same principle used by the Citability audit and AI Citation Tracking: define the scope before evaluating the result.

If an audit has already identified a failed signal, go directly to the matching recipe in §4.N and run the revised passage through the pre-publish checklist in §6. For a new draft, follow §3, §4, §5, and §6 in order. If stale content prompted the rewrite because a fact changed, an engine changed its interface, or a competitor now ranks above the page for its target query, apply the new-draft sequence to the affected section. Content Freshness explains how to set the review cadence.

3. An overview of the seven recipes

Each row below corresponds to one recipe in §4. The common failure patterns come from the per-signal tables in the Citability audit, and the definitions come from Citability §4.

Signals 1, 2, and 7 provide the greatest leverage in practice because they determine whether a passage can be extracted and understood, so they apply to every page. Signals 3 and 4 depend on the page format, so do not add questions or steps where they do not fit. Signal 5 becomes more important as a page relies more heavily on tables. Signal 6 is quick to review and correct.

The same recipes apply across AI search products, although individual engines may place more weight on different signals; §8 compares those differences. Microsoft describes the underlying shift this way: “The unit of value shifts from documents to groundable information — discrete, supportable facts with clear provenance” (Bing — Evolving role of the index, May 2026). Write each passage so it can provide that kind of information on its own.

4. The seven rewrite recipes

Each recipe uses the same five-part format: the result, the mechanism, numbered steps, a before-and-after example, and a common pitfall. The consistent format makes the seven recipes easy to compare.

4.1 Self-contained chunks

What it produces. Paragraphs that remain clear when removed from their original context because the subject is named, each pronoun has an antecedent, and no reference depends on a neighboring passage.

Mechanism. Engines retrieve and select individual chunks rather than treating the page as an indivisible unit. A paragraph that depends on the paragraph above it loses the self-containment described in Citability §4.1 as soon as it is extracted.

Recipe.

1. Give every paragraph an explicit subject. Repeat the relevant noun instead of starting with a pronoun that refers to the previous paragraph.
2. Make sure every pronoun has an antecedent in the same paragraph. If the referent appears several sentences earlier, use the noun phrase again.
3. Replace phrases such as “as above,” “as we saw,” and “earlier” with the information they refer to.
4. When referring to another section, briefly restate the relevant point. Write “the precedence rule (the longer path wins) applies” instead of “the precedence rule from §2 applies.”

Before. “As above, it applies; but as we noted in §2, the precedence rules can override.”

After. “The robots.txt Disallow directive applies to every URL the user-agent block names. Where two rules conflict, the longer-path rule wins — see §2 for the full precedence walk-through.”

Pitfall. Do not split every paragraph into a single sentence in an attempt to satisfy Signal 1. Google’s May 2026 guidance is explicit: “There’s no requirement to break your content into tiny pieces for AI to better understand it” (AI Optimization Guide). Signal 1 requires coherent, self-contained ideas rather than fragmented prose. A series of isolated sentences can fail just as readily as a chain of paragraphs connected only by pronouns. The table in §7 shows this pattern in practice, and AI Content Detection explains why repeated over-chunking can resemble low-effort content at scale.

4.2 Inverted-pyramid sections

What it produces. Each H2 and H3 begins with its main claim, followed by context, justification, and qualifications.

Mechanism. Live-fetch engines, particularly ChatGPT search, favor a direct answer near the beginning of the fetched page or section. During grounding, the engine may see only the first window of a section before selecting a quotation, so an immediate answer is easier to use.

Recipe.

1. State the answer in the first sentence under every H2 and H3.
2. Place the setup and justification after the claim. If the claim seems to require two introductory paragraphs, clarify the claim before adding the background.
3. Rewrite formulaic openings that begin with “Let’s consider,” “There are several factors,” or “First, it is important to understand.”
4. Make the opening sentence quotable on its own. The same sentence can serve as the standalone claim described in §4.7.

Before. “There are several factors to consider when thinking about robots.txt. Many practitioners debate whether path-length or specificity should decide precedence. Eventually, after weighing both, the longer path wins.”

After. “In robots.txt, the longer-path rule wins when two Disallow directives conflict. The rest of this section explains the precedence model and the two edge cases in which it does not apply.”

Pitfall. Avoid writing the section twice. A summary sentence that merely repeats the H2 title, followed by the same material the section would contain anyway, adds no information. The opening sentence should make a claim rather than announce a topic. The layout already renders the page-level TL;DR from frontmatter, so do not repeat it in a body blockquote.

4.3 Question-shaped headings

What it produces. On genuine Q&A pages, H2 and H3 headings use questions that people actually ask, including FAQ and troubleshooting questions and queries that begin “when should I” or “why does X.”

Mechanism. Question headings can match the subqueries produced through query fan-out (see Answer Loop §3.1). Google AI Overviews and Gemini’s web-grounded mode place particular value on this alignment because headings are strong retrieval signals for index-based engines.

Recipe.

1. Use question-shaped H3s only when the page is genuinely organized as Q&A. Renaming every section on a definitional reference page as a question makes the page harder to scan.
2. Use real queries from Search Console, support tickets, sales-call transcripts, or the engine’s “people also ask” panel. §7 explains why invented questions are counterproductive.
3. Write the first sentence of the answer so it remains clear without the heading. An engine may quote the answer by itself.
4. Ask one question per heading. Combine two questions in one H3 only when a single claim genuinely answers both.

Before. ”### Considerations regarding crawler precedence rules”

After. ”### When two robots.txt rules conflict — which wins?”

Pitfall. Do not add a block of invented questions to every page simply to increase the number of FAQ headings. Engines can recognize this boilerplate and give less weight to low-effort content, while repeated manufactured questions can trigger the patterns described in AI Content Detection. The appropriate number of question headings is the number of real user questions the page answers, whether that number is five or zero.

4.4 Step lists

What it produces. Numbered procedures written in the imperative mood, with one action per step and enough context for each step to make sense on its own.

Mechanism. An ordered procedure can be extracted as a single coherent unit. When a query is procedural, such as “how do I…” or “what are the steps to…”, engines consistently prefer explicit numbered instructions over steps embedded in prose.

Recipe.

1. Present every procedure as a numbered list instead of embedding the steps in a paragraph.
2. Put one action in each step and use the imperative mood: write “Run X” instead of “You should run X” or “X needs to be run.”
3. Give each step enough context to stand alone. If a step depends on an earlier result, restate that result in the step.
4. Divide branching procedures into separate lists. Avoid nesting conditions such as “if X, go to step 5” inside one sequence.

Before. “First you should consider whether the file already exists, and then it may be worth fetching the current version, after which you would edit the relevant directive, and finally re-deploy and verify.”

After.

1. Fetch the current robots.txt:  curl https://example.com/robots.txt
2. Locate the User-agent: * block.
3. Add the line  Disallow: /private/  on its own line within that block.
4. Re-deploy the file at the site root.
5. Re-fetch the URL and confirm the new directive is present.

Pitfall. Do not number items that have no sequence, such as “1. Important context 2. Another consideration 3. A tip.” Numbers imply an order of execution, so using them only for emphasis makes the structure misleading. Use bullets or prose when the items are not sequential.

4.5 Self-labeling tables

What it produces. Captioned tables whose rows can be understood independently. The caption explains the comparison, column headers use full noun phrases, and cells use complete clauses.

Mechanism. Self-labeling columns allow an engine to quote one table row without relying on the surrounding paragraph. Microsoft states: “Clear headings, tables, and FAQ sections help surface key information and make content easier for AI systems to reference accurately” (Bing AI Performance).

Recipe.

1. Put a full-sentence caption above every table to explain the comparison, such as “When two robots.txt Disallow rules conflict, the longer path wins.”
2. Use full noun phrases in column headers, such as “Winning rule” rather than “Wins?”, so the relationship between headers and cells remains clear.
3. Write each cell as a complete clause rather than a one-word answer. “Yes” lacks context; “Longer path wins per RFC 9309 §2.2.2” provides it.
4. Test one row by pasting it into a new engine session and asking “What is this saying?” If the engine cannot explain the row clearly, revise it so it does not depend on neighboring rows.

Before.

| Rule | Wins? | Notes |
|---|---|---|
| Long | Yes | See above |

After.

When two robots.txt Disallow rules conflict, the longer matching path wins.

Pitfall. Do not use a table for decoration when a sentence would communicate the same information more clearly. Signal 5 concerns structured comparisons that can be extracted, not visual layout. A small table that says only “Option A: yes; Option B: no” is effectively one sentence; placing it in a grid weakens self-containment without improving Signal 5.

4.6 Heading discipline

What it produces. A page with exactly one H1, a clean H2 → H3 hierarchy, no skipped levels, and no headings used only for decoration.

Mechanism. A clear hierarchy gives each passage an identifiable section. Engines retrieve and quote content by section, and the heading acts as the passage’s strongest implicit label. Skipped levels and decorative headings make those boundaries less reliable.

Recipe.

1. Use one H1 per page. The layout renders the frontmatter title as the H1, so do not repeat it in the body.
2. Follow H2 headings with H3 headings. Do not skip to H4 simply to display smaller text.
3. Use each heading to name a distinct unit of content. If you cannot summarize that unit in one sentence, delete the heading or merge the material into a nearby section.
4. Make the first sentence fulfill the promise of the heading. It will usually serve both as the direct answer in Signal 2 and the quotable claim in Signal 7.

Before. H2 → H4 → H3 (skipped); decorative H3s like ”### Diving in!”, ”### Let’s explore”, ”### A note”.

After. Clean H2 → H3 → H3 nesting throughout; topic-titled headings (”### When two rules conflict”, ”### What the precedence model assumes”).

Pitfall. Avoid using H3 headings solely to create visual breaks. If an H3 contains only one sentence, merge that sentence into the parent section. A one-sentence subsection usually does not justify a separate heading.

4.7 Quotable claims

What it produces. One concise, standalone claim per H2 that keeps its attribution when extracted. This is usually the same opening sentence created by the inverted-pyramid recipe in §4.2.

Mechanism. A single claim that remains clear when extracted is the smallest useful unit of citability. Models favor concise, standalone claims during selection, while heavily hedged, multi-clause sentences are difficult to quote cleanly.

Recipe.

1. Write one self-contained sentence in each H2. Begin with the subject or verb, and do not bury the claim in a subordinate clause.
2. Remove unnecessary hedges. Replace “It could perhaps be argued that, in some cases, X may not always Y” with a direct claim such as “X does Y” when the evidence supports that wording.
3. Keep the attribution in the same sentence as the claim, using language such as “per RFC 9309 §2.2.2” or “per Aggarwal et al.” A separate footnote may be lost during retrieval.
4. Make sure the claim remains understandable without its H2 heading, because an engine may quote the sentence alone.

Before. “It could perhaps be argued that, in some cases, retrieval may not always lead to use.”

After. “Retrieval makes your page a candidate; grounding determines whether the engine uses it.”

Pitfall. Never invent a statistic such as “47% of marketers report…” to make a sentence look quotable. Unsourced numbers fail the trust checks described in E-E-A-T and can trigger the patterns covered by AI Content Detection. A fabricated number creates more risk than a cautious sentence. Put the source URL in the same sentence as every statistic so the claim remains both quotable and trustworthy.

5. A reusable section template

The following annotated MDX template combines Signals 1, 2, 5, 6, and 7 in one section. Copy it into a draft and replace each instruction with the relevant content.

## H2 — a heading that names the topic (Signal 6; also Signal 3 when phrased as a real question)

One-sentence quotable claim that resolves the H2's question or asserts the
section's main fact. (Signals 2 and 7 usually use the same sentence.)

Add two or three sentences of substantive context that support the claim.
Give the paragraph an explicit subject, and make sure each pronoun has an
antecedent within the paragraph. (Signal 1.)

Write a full-sentence caption above the table that explains what is being
compared. Do not use a short label in place of the caption.

| Column A (full noun phrase) | Column B (full noun phrase) | Column C (full noun phrase) |
|---|---|---|
| A clause that makes sense alone | A clause that makes sense alone | A clause that makes sense alone |
| A clause that makes sense alone | A clause that makes sense alone | A clause that makes sense alone |

(Signal 5.)

End with one sentence that adds new information rather than repeating the
section. When helpful, guide the reader to the next H2 or an inline link
that provides more detail. (Signal 1.)

Apply three rules when filling in the template. First, read the opening sentence without its heading and confirm that it still makes a clear claim. Second, test each paragraph by pasting it alone into ChatGPT search or Perplexity and asking “What is this passage saying?” Third, write the table caption as a full sentence rather than a label. “Comparison of robots.txt precedence rules” is only a label, while “When two robots.txt rules conflict, the longer matching path wins” states the comparison clearly.

Treat the template as a starting point rather than a rigid form. A section without a meaningful comparison should not add a table simply to fill the slot in §5. Apply only the signals that fit the page’s actual format, as the mistakes in §7 illustrate. Chinese and English also differ in paragraph length, punctuation rhythm, and section openings. The Chinese version adapts the recipes accordingly; Multilingual GEO explains the language differences.

6. Checklist before publishing

Use this checklist while drafting, before the page is published. It applies the chunk-extraction test from the Citability audit during writing and groups each check by signal so you can return to the matching recipe in §4 when something fails.

SIGNAL 1 — Self-contained chunks
[ ] Each paragraph opens with its own subject (no pronoun without a same-paragraph antecedent).
[ ] No "as above / as we saw / see §X" reference appears without a restated noun phrase.
[ ] Three randomly selected paragraphs pass the chunk-extraction test when read alone.

SIGNAL 2 — Inverted-pyramid sections
[ ] Every H2's first sentence states the section's claim.
[ ] No section opens with "Let's consider..." or "There are several..."

SIGNAL 3 — Question-shaped headings (only where page form is Q&A)
[ ] Each question heading matches a real user query from Search Console, support tickets, or sales calls.
[ ] No invented "frequently asked" questions have been added solely to satisfy Signal 3.

SIGNAL 4 — Step lists (only where the page contains a procedure)
[ ] Every procedure uses a numbered list, the imperative mood, and one action per step.
[ ] Non-sequential content does not use numbered lists.

SIGNAL 5 — Self-labeling tables
[ ] Each table has a full-sentence caption above it.
[ ] Each column header is a full noun phrase, and each cell is a complete clause.

SIGNAL 6 — Heading discipline
[ ] The page follows an H2 → H3 hierarchy, skips no levels, and uses no decorative headings.
[ ] One H1 per page (frontmatter title); no H1 in body.

SIGNAL 7 — Quotable claims
[ ] Each H2 contains one concise claim with necessary attribution and no unsupported hedging.
[ ] Every statistic has its source URL in the same sentence as the number.

TRUST (E-E-A-T overlap — not a citability signal, but still essential)
[ ] Author identity and credentials visible on the page.
[ ] Every number is supported by a source; no statistics are invented.

Passing the checklist is necessary but not sufficient. E-E-A-T still affects whether an engine trusts the source, and good structure cannot hide thin content (AI Content Detection). Review the checklist from top to bottom because it is also ordered by priority. On most pages, a Signal 1 failure matters more than a Signal 5 failure, and any failure in the TRUST block takes precedence over the structural checks.

7. Patterns to avoid

The following writing mistakes correspond to the broader anti-patterns in Citability §6 and the audit failures in Citability audit §7.

For AI-assisted drafting, start with a human-written core, use the model to expand it, and then have a person verify and tighten the prose. Aggarwal et al. tested substantive improvements rather than adherence to a template, and an empty template can still exhibit the low-quality patterns it was meant to avoid. AI Content Detection covers the risks of an LLM-first draft, while the Aggarwal paper entry explains why its reported magnitude should be treated as an upper estimate rather than a promise.

Google’s May 2026 guidance states: “There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary” (AI Features and Your Website). The recipes in §4 are ways to structure clear, well-sourced content, not special techniques required only for AI search.

8. How priorities vary by AI search product

The same signals apply across AI search products, but individual engines place more weight on some signals than others. If you need to focus your effort on one target engine, use the following priorities.

Language introduces another set of differences. Chinese and English differ in paragraph length, punctuation rhythm, and section-opening conventions, and these differences affect how chunks and direct-answer passages are written. The Chinese version adapts these recipes to its language rather than translating them mechanically. See Multilingual GEO for the practical differences that matter on Chinese-language engines.

Competition reduces the gains available to a single participant. These recipes remain directionally useful across products, but the headline results from single-participant benchmarks are not guaranteed. As competitors adopt similar methods, the average gain decreases (C-SEO Bench, NeurIPS ‘25 D&B). Evaluate whether a recipe makes the revised passage easier to extract than the previous version; do not assume the page will see a +40% lift.

9. What to do after publishing

After publishing the rewrite, complete the following three tasks.

Run the chunk-extraction test again on the live page. Select three passages at random: the TL;DR, the first paragraph under one H2, and one table row. Paste each passage by itself into a new ChatGPT search or Perplexity session and ask “What is this passage saying?” If the engine hedges, asks for context, or completes the passage incorrectly, record a citability finding. Use the Citability audit for the full diagnosis, then apply the corresponding recipe in §4.

Add the page to your tracking set. Include its target queries in the fixed prompt set defined in AI Citation Tracking §3, then wait two measurement cycles before interpreting the trend because a single week’s change is usually noise. Monitor Citation Rate and Average Position, as defined in GEO Metrics, rather than page views alone. The first cycle after a rewrite often shows a temporary decline while engines crawl and embed the page again, so evaluate the longer trend.

Separate citations from mentions in the report. A page may be quoted with a link, which counts as a citation, or paraphrased without a link, which counts as a mention. Both matter, but they indicate different problems; see Citation vs Mention. When a page is mentioned but not cited, its trust signals often pass while attribution fails, pointing to an E-E-A-T or attribution issue rather than a structural citability issue. Liu et al. found that “51.5% of generated sentences are fully supported by citations” (Liu et al., EMNLP ‘23). Structural citability work alone does not resolve that attribution gap.

Rewrite the affected section when one of the triggers from §2 occurs: a fact changes, an engine changes its interface, or a competitor begins to rank above the page for its target query. Use Content Freshness to set the review cadence, then repeat the sequence in §2, §4, and §6 for that section.

10. Further reading

• Core concepts: Citability defines the seven signals, E-E-A-T explains the trust conditions that structure cannot replace, and Answer Loop describes the four-step process in which these recipes support grounding.
• Related playbooks: Citability audit diagnoses the problems addressed by these recipes, Full GEO Audit includes them in its Layer 4 review, and AI Citation Tracking explains how to measure the page after publication.
• AI search products: Perplexity and ChatGPT search.
• Research: Aggarwal et al. 2024 — GEO: Generative Engine Optimization examines substantive content changes; C-SEO Bench provides a more limited estimate under competition; and Liu et al. 2023 measures citation support in generated answers.
• Quality and language: AI Content Detection explains the risks of over-optimization, and Multilingual GEO covers the differences between languages.

References

Academic:

1. Aggarwal, P., Murahari, V., Rajpurohit, T., Kalyan, A., Narasimhan, K. & Deshpande, A. (2024). GEO: Generative Engine Optimization. KDD ‘24. arXiv:2311.09735 · ACM DL · in-project paper entry
2. Puerto, H., Gubri, M., Green, C., Oh, S. J. & Yun, S. (2025). C-SEO Bench: Does Conversational SEO Work? NeurIPS ‘25 Datasets & Benchmarks. arXiv:2506.11097
3. Liu, N. F., Zhang, T. & Liang, P. (2023). Evaluating Verifiability in Generative Search Engines. Findings of EMNLP 2023. arXiv:2304.09848

Official platform documentation (verified 2026-05):

• Google Search Central — Google’s Guide to Optimizing for Generative AI Features on Google Search · A new resource for optimizing for generative AI in Google Search · AI Features and Your Website · Top ways to ensure your content performs well in Google’s AI experiences on Search
• Microsoft Bing — Evolving role of the index: From ranking pages to supporting answers · Introducing AI Performance in Bing Webmaster Tools (Public Preview)
• OpenAI — ChatGPT search Help Center
• Perplexity — What is an answer engine, and how does Perplexity work as one?