Answer in the first two sentences, write 40 to 60 word passages under question headings, and back claims with sources and statistics. Rules with examples.
LLaunchScaler·Published ·11 min read
To write content that AI answers cite, put the full answer to the question in the first one or two sentences, then build the page from short, self-contained passages under question headings, each one backed by a specific number, a quotation or a named source. In the GEO study (Aggarwal and colleagues, KDD 2024), adding citations, quotations and statistics were the three edits that raised a source's visibility in generative engine answers the most, by roughly 30 to 40%.
Google is a special case: its guide to generative AI search says its AI features run on its normal ranking systems and that you don't need to write in a special way, chunk content or add AI files for them. The rules below are ordinary good writing, which is what Google asks for, and what the research found other generative engines reward. Each section says which engine a point applies to.
How do AI search engines pick which passages to cite?
AI search engines retrieve pages first and write second. A page can only be cited if the engine's crawler is allowed to fetch it and its search index holds it. From there, the engine pulls passages that answer the question or one of its sub-questions, so a passage that answers cleanly on its own is the easiest thing to cite.
For Google, a page must be "indexed and eligible to be shown in Google Search with a snippet," and since the Search generative AI control rolled out to all sites (as of August 31, 2026), the site must also be included under Settings, then Search generative AI in Search Console. Inclusion is the default. Google's guide says its AI features retrieve pages through its core ranking systems and use query fan-out, "a set of concurrent, related queries generated by the model," to fetch more results. One question becomes several, and each sub-question is a chance for a different section of your page to be picked. The same guide says Google Search ignores llms.txt and other AI text files and needs no special schema.org markup for these features.
ChatGPT's search works on its own crawler. OpenAI says OAI-SearchBot "is used to surface websites in search results in ChatGPT's search features" and recommends allowing it in robots.txt. GPTBot is a separate crawler for model training, so blocking GPTBot does not remove you from ChatGPT search, while blocking OAI-SearchBot does.
So writing for AI search starts with checks that have nothing to do with prose: the page is indexed in Google with a snippet (no nosnippet), the site is not excluded in Search Console's Search generative AI setting, and robots.txt does not block the search crawlers. After that, the words decide it.
Questions, answered
What people ask about this
01
How do you write content for AI search?
Answer the question in the first one or two sentences, then break the page into short passages under question headings that each make sense on their own. Support claims with specific numbers, quotations and links to the original source, because those were the edits that raised visibility most in the GEO study.
What did the GEO study find about writing for AI search?
The GEO paper tested nine ways of rewriting a web source and measured how much of the generative engine's answer drew on that source. Citing sources, adding quotations and adding statistics came out on top, with relative gains of 30 to 40% on its main visibility metric. Keyword stuffing, the classic SEO move, gave little to no improvement.
The researchers built a benchmark called GEO-bench of 10,000 queries drawn from nine datasets across different domains. They measured visibility two ways: Position-Adjusted Word Count, the share of the answer's words tied to a source, weighted so that citations earlier in the answer count for more, and Subjective Impression, a score across seven aspects such as relevance and influence of the cited sentence, graded by GPT-3.5 in the style of G-Eval.
Method in the paper
What the rewrite changes
What the paper reports
Cite Sources
Adds citations to credible sources
Among the top three: 30 to 40% relative gain on Position-Adjusted Word Count, 15 to 30% on Subjective Impression
Quotation Addition
Adds quotations from credible sources
Among the top three; best method on Perplexity.ai, with a 22% gain on Position-Adjusted Word Count
Make the text more persuasive, more fluent or simpler
Smaller gains than the top three
Unique Words, Technical Terms
Add unusual words or jargon
Modest gains over the unedited source
Two findings matter for a small site. First, sources lower in the search results gained the most: the paper reports that Cite Sources produced a 115.1% visibility increase for websites ranked fifth. Second, the effect differed by domain, which is why the authors call for "domain-specific optimization methods." Statistics Addition helped most on Law & Government and opinion questions, Cite Sources on factual questions, and Quotation Addition on People & Society, explanation and history queries. Test on your own queries rather than trusting any single percentage.
The broader lesson is that engines reward evidence. A sentence with a number, a source and a date gives the model something it can repeat and attribute. A sentence of general opinion gives it nothing it could not have written itself.
How do you write an answer-first opening?
Write the first two sentences as if they were the only thing a reader will see. The first sentence answers the question as asked, using its main words. The second adds the condition, number or exception that makes the answer complete. Anything that sets the scene goes after that, or nowhere.
A method that works on almost any topic:
Copy the exact question from your target query, for example "how long does Google take to index a new page".
Write the answer in one sentence that contains the question's key noun and verb, and gives a specific value or range if one exists.
Add a second sentence with the main condition or the next step.
Delete any sentence before those two that does not answer the question.
Read the two sentences out of context. If a stranger could act on them, the opening is done.
Compare two openings for the same page. "Indexing speed is a topic many site owners wonder about, and the answer depends on many factors" gives an engine nothing to quote. "Crawling a new page can take Google anywhere from a few days to a few weeks, according to its own documentation, and requesting a recrawl more than once for the same URL does not make it faster" can be lifted word for word into an answer and credited to you.
No engine requires a passage length, and Google says outright that there is "no requirement to break your content into tiny pieces" and "no ideal page length." Treat the length as a writing discipline, not a ranking rule: open each section with about 40 to 60 words that answer its heading's question in full, then go into detail. At that length the answer is complete and can be quoted whole, by a reader skimming or by an engine, without the paragraph before it.
Self-contained is the harder half, and it matters most for engines outside Google that quote a single section. An engine that quotes your third section will not quote your first, so each section's opening has to carry its own subject. Four habits make a passage stand alone:
Name the subject in the first sentence instead of starting with "it", "this" or "they".
Avoid references backward or forward, such as "as mentioned above" or "see below".
Put the answer first and the explanation after, so a quote cut at 50 words still answers.
Keep one claim per passage; if a paragraph answers two questions, split it under two headings.
Headings help readers find the answer they came for. Phrase each H2 as the question people type, in their words. Google says its systems understand synonyms, so exact-match keywords are unnecessary. The point of a question heading is that a reader, or a fan-out query, lands on the section that answers them. "Pricing" tells a reader little. "How much does a content audit cost?" tells them they are in the right place.
Detail belongs below the 40 to 60 word answer, not in it. Steps, caveats, edge cases and examples can run as long as the topic needs. Structure each section as a short answer first, then the depth.
How do you add statistics, quotations and sources an engine can trust?
Treat every claim as something an engine may repeat under your name. A statistic needs a number, a unit, a date or period, and a link to the page it came from. A quotation needs the speaker's name and role and a link to where they said it. A source should be the original: the documentation, the study or the dataset, not an article about it.
The rules that keep evidence usable:
Put the number in the sentence that makes the claim, not in a footnote. "Google reports a relative gain of 30 to 40%" beats "significant gains" with a link.
Link the primary source. For a Google behaviour, link Google's documentation. For a study, link the paper.
Quote short and exact. One sentence in quotation marks, copied from the source, with the speaker named.
Date anything that changes. Prices, limits and rules go stale, so write "checked in September 2026" next to them.
Leave out what you cannot verify. An invented statistic is worse than none, because it can be repeated with your name attached.
Your own data counts as a source, and often the best one. Search Console numbers, support ticket counts or benchmark results from your product are original information that no other page has, which is exactly what Google's helpful content questions ask for: "Does the content provide original information, reporting, research, or analysis?"
A content brief template is the place to collect these before writing starts, so the writer has the numbers and links in hand instead of reaching for vague claims.
Why do tables and visible dates matter?
Tables give an engine structured comparisons it can quote row by row, and visible dates tell both readers and engines how current a page is. Use an HTML table whenever you compare options on the same fields, and show a published or updated date that matches your structured data.
For tables, three rules:
Use real HTML table markup, not an image. Google's AI features guidance asks that "important content is available in textual form."
Give every row the same columns, so a row can be read out of context.
Write "Not published" in a cell when a vendor does not state a figure, instead of guessing.
For dates, Google's byline date guidance asks for a user-visible date labelled like "Posted" or "Last updated", and for datePublished and dateModified in Article or BlogPosting structured data. It says the visible and structured values must match. Its helpful content guidance also warns against "changing the date of pages to make them seem fresh when the content has not substantially changed," so change the updated date only when you change the facts.
How do you know whether AI answers use your pages?
For Google, Search Console now has a Generative AI performance report, rolled out to all sites as of August 31, 2026. It counts impressions, meaning how often links to your site were shown in AI Overviews and AI Mode, and groups them by page, country, device and date. Compare a page's AI impressions before and after you rewrite it to these rules. For ChatGPT, Perplexity and other engines there is no equivalent report, so ask each engine your target questions and note which pages it cites.
What should you check before publishing?
Run this list on every draft before it goes live:
The first two sentences answer the target question completely and name its subject.
Every H2 is a question in the searcher's own words, in sentence case.
The first 40 to 60 words under each H2 answer that heading's question alone.
No passage opens with "it", "this" or "as mentioned".
Every number has a unit, a period and a link to its original source.
Quotations are exact, short and attributed by name.
Comparisons are in an HTML table with the same fields in every row.
The page shows a published date, and an updated date only if facts changed.
The page is indexed with a snippet, the site is included under Search generative AI in Search Console, and robots.txt allows Googlebot and OAI-SearchBot.
Nothing on the page is a claim you could not source.
Writing to these rules takes time per article, and most of it goes into choosing the question and finding the sources. LaunchScaler's content engine plans a month of articles from your own Search Console queries (the searches you already rank for just off page one), with your readiness audit and real keyword data filling in the rest. It writes one article a day, thirty a month, from what your product actually does, in your voice. Every draft lands in your workspace for you to rewrite, trim or scrap, and by default it waits for your review before it publishes to WordPress, Webflow, Ghost, Shopify, Notion, Medium or a signed webhook.
It costs $99/mo per website, with 3 days free before the first charge. Start the engine and your first article arrives on day one.
02
Do you need special markup or an AI text file to appear in Google's AI Overviews?
No. Google's guide to generative AI search says Google Search ignores llms.txt and other AI text files and needs no special schema.org markup. A page has to be indexed and eligible to show with a snippet, and the site must be included in Search Console's Search generative AI setting, which is the default.
03
Does keyword stuffing work for AI search engines?
No. The GEO paper tested adding more query keywords and found it offered little to no improvement in generative engine responses, while citing sources, adding quotations and adding statistics each improved visibility by roughly 30 to 40%.
04
How long should a paragraph be for AI answers?
No engine requires one, and Google says there is no need to chunk content for its AI features. Opening each section with about 40 to 60 words that fully answer its heading is a practical habit, because the answer can then be read or quoted without the sentences around it.
05
Should I show dates on articles written for AI search?
Yes. Google recommends a visible published or updated date and matching datePublished and dateModified values in Article structured data, and asks you not to change dates when the content has not substantially changed.
Competitor alternatives and X vs Y pages that hold up: every claim sourced and dated, one table with the same fields, who each suits, checked quarterly.
