Write pages that give an engine something it cannot get anywhere else: your own facts and numbers, a direct answer to a specific question, and sources it can check. Google says unique, useful content will likely do more for a site's presence in its AI features, in the long run, than anything else in its guide. A study accepted at KDD 2024 found that adding sources, quotations and statistics made a page 30 to 40 percent more visible in a test engine's answers, but its prompts let the model invent that material, so treat it as a reason to test real evidence, not as a measured gain.
The rest of this guide turns that into habits, with the evidence behind each one.
What happens to a page before it is cited
An answer engine searches before it writes. Google describes its AI features as grounded in Search: its ranking systems retrieve pages from the index, and the model writes an answer from them. AI Overviews and AI Mode may also run several related searches on subtopics, a technique Google calls query fan-out.
ChatGPT works in a similar way. OpenAI says that when ChatGPT search uses outside search providers, it typically rewrites the question into one or more targeted queries for them.
So a page clears two hurdles. First it has to be found by one of those searches, which depends on crawler access and on the HTML the server sends; robots.txt for AI crawlers covers that. Then it has to be worth drawing on. That second hurdle is the writing.
Say something only you can say
Google's guide to its generative AI features draws a line between commodity content, built from common knowledge anyone could write, and non-commodity content that brings an expert or first-hand view. It gives a generic list of tips for first-time homebuyers as its commodity example.
For a business, non-commodity usually means one of three things:
- Your own data. Measurements, prices, benchmarks, survey results, usage figures.
- First-hand testing. What you tried, how, and what happened, with the method stated.
- A clear position. Which option you would pick for which buyer, and why.
Google's own checklist for helpful content asks the same question plainly: does the page provide original information, reporting, research or analysis?
Put numbers and named sources in the sentences
One published study tested this directly: the GEO study by Pranjal Aggarwal and five co-authors, accepted at KDD 2024. They built a test engine on GPT-3.5 Turbo that answers from the top five Google results, and ran it on the 1,000-query test set of a 10,000-query benchmark. They rewrote one source at a time and measured how much of the answer drew on it, weighted by position.
- Citing credible sources, adding quotations and adding statistics each raised that measure by 30 to 40 percent over the unedited page. The paper's published prompts let the rewriting model invent that material: quotes "even though fake and artificial", statistics "even if hypothetical", and sources it "may invent". Its test engine was built on GPT-3.5. So the gain shows what quote-, statistic- and citation-shaped text did to that engine, not that real evidence has the same effect on today's engines.
- Keyword stuffing gave little or no improvement. On Perplexity, the only live engine they tested, it did 10 percent worse than the unedited page, while added quotations did 22 percent better.
- The effects varied by subject. Statistics helped notably for law and government topics and for opinion questions.
Two limits matter. The study measured how much an answer used a page that had already been retrieved, not whether a page gets retrieved. And its results come from a test engine built on GPT-3.5 Turbo and from Perplexity as it was when the study ran.
Microsoft gives the same advice for Copilot and Bing's AI summaries. Its post introducing Bing's AI citation reporting tells publishers to support claims with examples, data and cited sources.
In practice, replace claims with figures a reader can check. "Setup is quick" gives an engine nothing to use. A sentence built like "median setup time was [N] minutes across [N] new accounts in [month and year]" gives it a fact to repeat and a date to weigh.
Answer the question in the first two sentences
In our scans of seven engines, a cituna.com guide headed with a buyer's literal question, "Which AI visibility tool includes Google Search Console?", was cited 21 times by ChatGPT and Gemini. That is one page, not a law. But a heading that repeats the question gives an engine answering that question an easy match.
Google's checklist asks whether the main heading gives a descriptive, helpful summary of the page. Put the answer in the first two sentences and the background after it. How to structure a page for AI answers covers the layout in detail.
Name things so they cannot be misread
Engines guess when a name is ambiguous. In our scans, Perplexity answered "best Profound alternative" by citing alternative-rock music sites 183 times, and Perplexity and Gemini answered "best Otterly AI alternative" with 96 citations to roundups about Otter.ai, a transcription app.
Put the category next to your product name in titles, headings and first sentences: "Acme, the invoicing app for freelancers" leaves less to guess than "Acme".
Say who wrote it and how you know
Google's helpful content guidance asks whether it is obvious who wrote a page, whether it carries a byline, and whether it explains how the content was produced, such as how many products were tested and how. Named people are also a source the engines read: LinkedIn drew 228 citations in our scans, and the LinkedIn pages cited include individuals' own articles.
Add a byline that links to an author page, a line on how the facts were checked, and the date they were checked. Our glossary entry on E-E-A-T explains the thinking behind this.
Keep every fact in text the crawler receives
Google advises making sure important content is available as text. That matters even more outside Google: Vercel and MERJ found that the crawlers from OpenAI, Anthropic and Perplexity do not run JavaScript, while Google's does. In our test of nine AI visibility pricing pages, two gave a crawler no readable price.
Write prices, specifications and comparisons as HTML text. Not in an image, and not only in a script.
What you can skip
Google's guide lists tactics you can ignore for Google Search:
- Breaking content into tiny chunks for AI.
- Rewriting pages in a special style for AI systems.
- llms.txt and other AI text files, which Google Search does not use. Google says keeping one for other services neither helps nor harms you in its Search.
- Chasing inauthentic mentions across the web.
- A separate page for every variation of a question. Google says doing that to manipulate its generative AI responses violates its scaled content abuse policy.
- A target word count. Google says it has no preferred one.
What to do first
- Rewrite the first two sentences of your five most important pages so they answer the page's question.
- Replace vague claims with figures, each with its source and date.
- Add a byline and a line on how the facts were checked.
- Put your category next to your product name in every title.
- Fetch each page with curl and confirm the key facts are in the HTML.
Correction
23 September 2026: added what the GEO paper's published prompts allowed (invented quotes, statistics and sources) and the GPT-3.5 test engine, so the 30 to 40 percent figure is not read as proof that real evidence works the same way.