Writing for AI citation is a craft problem before it is a technical one. The test is simple: pull one sentence out of your paragraph, drop it into a stranger’s answer, and see whether it still means something, still says something specific, and still points back to you. Sentences that survive that extraction get quoted. Sentences that need three paragraphs of setup do not.
What follows is a writer’s guide to making passages worth pulling: the citable sentence, specificity over hedging, structure a machine can lift from, primary sourcing, and what schema actually does. One limit runs through all of it. Google’s own documentation says you do not need to write differently for its AI features at all, and we take that seriously rather than writing around it.
The anatomy of a citable sentence
A citable sentence has four properties, and you can test any sentence against them in about five seconds.
- Self-contained: it carries its own subject. “This makes migration harder” fails, because “this” lives in the previous sentence. “Drupal 10 reaches end of life on December 9, 2026” passes.
- Specific: it contains a hard element. A number, a date, a version, a named product or organization. Sentences made entirely of adjectives cannot be verified, so they are not worth repeating.
- Factual rather than promotional: “Our approach delivers world-class results” is unquotable in a neutral answer. “The tool caps free accounts at 500 API calls per day” is quotable.
- Attributable: it is clear whose claim it is. Either you observed it, or you name who did.
None of this is new. It is what good reference writing has always demanded. What changed is the penalty for vagueness, because a system choosing among five sources reaches for the one that gave it something concrete to repeat. A useful habit: copy the sentence you would most want quoted into a blank document. If it reads as a complete thought, keep it. If it collapses, rewrite it so it stands alone.
Specificity is the whole argument
Vague writing is usually a sourcing failure wearing a style disguise. Writers hedge when they do not actually know the number.
Compare two versions of one claim. "Many organizations see significant gains from AI search visibility" says nothing checkable. "In the KDD 2024 paper that introduced the term, adding citations, quotations, and statistics produced relative improvements of 27 to 41 percent on the authors’ position-adjusted word count metric" says something a reader can go verify. The second is longer by a dozen words and worth several times as much.
The rules are dull and they work. Name the company, not "leading vendors." Give the date, not "recently." Use the version number, not "the latest release." Give the figure with its unit and its source, or cut the claim. Google’s guide to optimizing for generative AI features lands in the same place: it contrasts commodity content ("7 Tips for First-Time Homebuyers" is its example) with non-commodity content built on real experience, and says the second kind will likely influence visibility more than anything else in the guide. Specificity is how experience shows up on the page.
For the wider strategic frame, our explainer on generative engine optimization covers the discipline this craft guidance sits inside.
Structure a machine can lift from
Structure is not decoration. It is how a retrieval system decides which part of your page answers a given question. Four patterns do most of the work.
- Headings that state the answer, not the topic. “Drupal 10 support ends December 9, 2026” beats “Timeline considerations.” The heading labels the passage underneath it.
- Definitions in the first sentence under the heading. If a section explains what something is, define it immediately. Do not build to it.
- Direct question and answer pairs. A heading phrased as a real question, answered in full in the next sentence, is the cleanest structure a passage can have. It is why FAQ sections punch above their weight.
- Lists and tables for parallel data. Three or more comparable items belong in a list or a table. Prose that enumerates five things in one paragraph is harder for everyone.
Google’s documentation backs the readability half of this, noting that people appreciate pages organized by paragraphs, sections, and headings. What it does not do is treat structure as a special ritual for AI.
Passage-level clarity, and the objection to it
Here practitioner advice and platform documentation diverge, so it is worth being precise about who says what.
Retrieval-augmented generation, the technique behind most cited AI answers, fetches relevant sources and generates a response grounded in them. Google confirms it uses RAG (which it also calls grounding) in its generative AI features, plus query fan-out, where the system issues related queries to gather more sources. Our primer on retrieval-augmented generation covers that loop.
From the mechanism, many practitioners infer a writing rule: because retrieval operates on passages, write passages that stand alone. The inference is reasonable. It is also, for Google, explicitly not required. The same guide states there is "no requirement to break your content into tiny pieces for AI to better understand it," that its systems can handle multiple topics on one page and surface the relevant part, and that there is no ideal page length. It adds that you do not need to rewrite content for AI systems, because they understand synonyms and general meaning.
So which is it? Both. Self-contained passages are good writing regardless of the reader, and they help a human skimming on a phone. Chopping a coherent article into fragments to feed an imagined chunker is what Google warns against. Write for clarity and the passage-level benefit arrives as a side effect.
Primary sourcing is a citation strategy
The strongest evidence in this area concerns attribution, and it comes from research rather than platform documentation.
The 2024 paper that introduced the term, GEO: Generative Engine Optimization, tested nine content edits against a baseline. Its top three performers were adding citations, adding quotations from credible sources, and adding statistics, which delivered relative improvements of 27.3 percent, 40.9 percent and 30.6 percent respectively on the authors’ position-adjusted word count metric. Improving fluency scored 27.8 percent on the same metric but only 9.3 percent on their subjective impression measure, which is a useful reminder that the two metrics do not move together. Keyword stuffing, borrowed from old-school SEO, scored 8.8 percent below the unoptimized baseline on position-adjusted word count, and 9.1 percent below when tested against Perplexity, though it did edge above baseline on subjective impression.
Two findings deserve emphasis. The gains skewed toward sites that were not already dominant: the citation treatment raised visibility 115.1 percent for a source ranked fifth, while the top-ranked source lost visibility on average. And effects varied by domain, with citations helping most on factual queries.
Read that as an argument for reporting rather than asserting. Go to the primary source, quote it accurately, link it. Which is also why the crawler layer matters: OpenAI’s crawler documentation states that sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers, Perplexity documents PerplexityBot as the agent that surfaces and links sites in its results, and Anthropic documents that disabling Claude-SearchBot may reduce a site’s visibility in search responses. That is not optimization. It is permission, and it is a prerequisite.
Structured data: useful, not a shortcut
Schema markup gets oversold here, so the documented position is worth stating plainly. Google says structured data is not required for generative AI search and there is no special schema.org markup you need to add, while still recommending it for rich-result eligibility. It says the same of llms.txt files: Google Search ignores them, and creating one neither helps nor hurts.
The principle that survives scrutiny is consistency. Google’s guidance on AI features and your website asks that structured data match the visible text on the page. Markup describing content a reader cannot find is a quality problem, not an optimization. Mark up what is genuinely there: Article, FAQPage where you have real question and answer pairs, Organization details. Our comparison of where GEO and SEO diverge covers how much of the technical layer the two disciplines share.
What you cannot control
Three limits belong in any honest treatment of this.
You cannot control citation. To appear as a supporting link in Google’s AI features, a page must be indexed and eligible to appear with a snippet, and Google states outright that meeting every requirement does not mean it will crawl, index, or serve the content. Everything above changes the odds. Nothing sets them.
The platforms move. Google’s optimization guide was last updated in July 2026, and its position on AEO and GEO tactics has already hardened into an explicit mythbusting section. The GEO paper’s authors note their methods may need to adapt as engines evolve.
And much published GEO advice is correlation dressed as causation. Studies observing which pages get cited most cannot separate the writing from the domain authority, the backlinks, or the brand behind it. The GEO paper is stronger than most because it ran controlled before-and-after edits, but on a benchmark, not across the live web. Google adds its own caution: no third-party tool has access to its internal ranking or AI systems. Measure with the approach we outlined for tracking both channels, expect noise, and hold conclusions loosely.
The part you fully control is the sentence. Make it stand on its own, make it specific, and say where you got it.
Frequently Asked Questions
What is a citable sentence?
A citable sentence still makes sense when removed from its paragraph. It carries its own subject rather than leaning on “this” or “that,” contains a specific element such as a number or named entity, states a fact rather than a promotional claim, and makes clear whose claim it is. Those properties let an answer engine lift it into a response without distorting it.
Does Google require special formatting to appear in AI Overviews?
No. Google’s documentation states there are no additional requirements to appear in AI Overviews or AI Mode and no special optimizations necessary. A page must be indexed and eligible to appear in Search with a snippet. Google also notes that meeting every requirement does not guarantee it will crawl, index, or serve the page.
Should I break my articles into small chunks for AI retrieval?
Not mechanically. Google’s guide says there is no requirement to break content into tiny pieces and no ideal page length, because its systems can understand multiple topics on a page and surface the relevant part. Self-contained passages remain good practice for readers who skim, but fragmenting a coherent article to satisfy an imagined chunking algorithm is the failure mode.
Does adding citations and statistics actually increase AI citation?
In controlled research, yes, with caveats. The KDD 2024 paper that introduced generative engine optimization found that adding citations, quotations, and statistics produced relative improvements of 27.3, 40.9 and 30.6 percent on its position-adjusted word count metric. That came from a benchmark, not from platform documentation. Strong evidence, not a documented ranking factor.
Does keyword stuffing help with AI answer engines?
No, and the evidence suggests it hurts. In the GEO paper’s testing, keyword stuffing scored 8.8 percent below the unoptimized baseline on position-adjusted word count and 9.1 percent below against Perplexity, although it edged slightly above baseline on the paper’s subjective impression metric. Google separately warns that producing content variations primarily to manipulate rankings or generative AI responses violates its scaled content abuse spam policy.
Do I need schema markup or an llms.txt file to get cited?
Google says no to both. Structured data is not required for its generative AI features and there is no special schema.org markup to add, though it remains worth using for rich-result eligibility. Google Search ignores llms.txt files entirely, so creating one neither helps nor harms visibility there.
How do I check that AI engines can read my site?
Audit robots.txt and any CDN or firewall rules against the crawlers each platform documents. OpenAI uses OAI-SearchBot for ChatGPT search appearance and states that opted-out sites will not be shown in ChatGPT search answers. Perplexity uses PerplexityBot, and Anthropic uses Claude-SearchBot. Both OpenAI and Perplexity say robots.txt changes can take around 24 hours to take effect.
Does this replace normal SEO?
No, and treating it as a replacement is the common error. Search ranking still drives the majority of clicks for most sites, and the practices that make a passage quotable are largely the practices that made it a good page anyway: a clear claim, stated plainly, under a heading that matches a real question. Our look at how ranking and AI citation have decoupled covers what the measurement actually shows and where the widely quoted numbers come from.
Update (2026-08-03): this article was revised on publication day.
The version that went out at 07:00 rounded the GEO paper’s keyword-stuffing result to "roughly 10 percent"; the paper’s own figure is 8.8 percent below the unoptimized baseline on the position-adjusted word count metric, and the article now quotes it exactly, alongside similar precision fixes to how the paper’s findings are cited. A FAQ on whether this discipline replaces normal SEO was also added. No claim was reversed; the direction and substance of every finding are unchanged.
