Writing pages that AI assistants cite

Andy Bright

Director of Product and Technology

An AI assistant writes its answer from the pages it reads, as the first chapter explains. This chapter is about your own pages, and what makes them useful to that process.

Advice about writing "for AI" is often ordinary advice about clear writing under a new name, and some of it isn't supported by any evidence. This chapter sets out what Google and Microsoft say, what research has tested, and what you can skip.

What an assistant looks for in a page

When someone asks ChatGPT "how much does a coffee subscription cost in the UK?", it runs some searches, gets back a list of pages and reads parts of them. It's looking for passages that answer the question, or part of it, in a form it can repeat with confidence.

Most pages on a topic say much the same thing. Google calls this commodity content, content "often based on common knowledge, which could originate from anyone". Its guide to optimising for generative AI⁠(opens in a new tab) contrasts it with content that offers "unique expert or experienced takes that go beyond common knowledge". Google's own example of commodity content is an article called "7 Tips for First-Time Homebuyers". Its example of the other kind is "Why We Waived the Inspection & Saved Money: A Look Inside the Sewer Line". For Fernhill, "10 tips for better coffee at home" would be commodity content, and "We tested 12 grind sizes in a cafetière: here's what tasted best" wouldn't.

If ten pages say the same thing, an assistant can use any of them, or none, because its model already knows it. A page with something the others don't have gives the assistant a reason to use it and to link to it. That might be a measured result, a price, a test or a first-hand account. Google says that content people find "unique, compelling, and useful will likely influence your website's presence in generative AI search in the long run more than any of the other suggestions in this guide".

Ahrefs saw the same pattern in ChatGPT. Its October 2026 study⁠(opens in a new tab) of millions of US ChatGPT answers found that ChatGPT "seems to be prioritizing original sources over sites that summarize them". Between March and September 2026, sites that summarise other people's work lost citations, and research bodies, regulators, brands and people who test products themselves gained them.

Start with what only you can say

Every business knows things nobody else can publish. For Fernhill Coffee, that includes:

• the date each coffee was roasted, and how long it stays fresh
• the farms it buys from, and what it pays for the beans
• the results of its own brewing tests, such as which grind works best in a cafetière
• how long deliveries to Ireland and Germany actually take, now that it has started sending them
• what its customers say about the coffee, from its own survey
• who its subscription isn't right for, such as people who want supermarket prices

Most of this is already known inside the business. It only needs writing down, with numbers and dates.

Google's guide to creating helpful content⁠(opens in a new tab) has a list of questions that work as a check. Two of them are whether a page provides "original information, reporting, research, or analysis", and whether it provides "insightful analysis or interesting information that is beyond the obvious".

Put your facts in plain text

When someone asks an assistant a buying question, it needs specific facts, such as prices, plans, sizes, delivery areas, limits and compatibility. If your pages don't state them clearly, the assistant takes them from somewhere else, such as a review site with last year's prices, or it leaves you out.

ChatGPT often searches a business's own website for these facts. Ahrefs found that 6.8% of ChatGPT's searches used the site: instruction, which limits a search to one website. Most of the site: searches that came up often were aimed at brands' own sites. Its advice⁠(opens in a new tab) is that prices, specifications and support details "need to be on your site, in plain text, on a page it can find".

Microsoft's guidance on content for AI answers⁠(opens in a new tab) gives an example. It suggests "42 dB dishwasher designed for open-concept kitchens" rather than "quiet dishwasher", and says "terms like innovative or eco mean little without specifics". It also lists ways that facts become harder for AI systems to find or use:

• inside tabs or sections that open when clicked, because "AI systems may not render hidden content". That means they may not build the page the way a browser does, so they miss the parts that open on a click
• only in PDF files
• only in images
• in long, unbroken blocks of text

Fernhill's subscription page used to open like this:

Fresh coffee, your way.
Flexible plans to suit every coffee lover, delivered to your door.
Find out more about our plans.

Nothing in it answers a question anyone would ask an assistant. The prices were in a table filled in by JavaScript, a programming language that runs in the visitor's browser after the page arrives. Most AI crawlers couldn't run JavaScript when they were last studied, so they saw an empty table, as the technical SEO guide's JavaScript chapter explains. Fernhill rewrote the opening, and its developer moved the table into the page's HTML, the code the server sends that holds the page's text. That second change is a job for a developer or your website platform's settings.

Fernhill's coffee subscription sends a 250g bag of freshly roasted coffee
every week, fortnight or month, to UK, Irish and German addresses. Prices
start at £9.50 a bag, including UK delivery. You can pause or cancel at
any time.

An assistant asked about prices, delivery to Ireland or how often bags arrive can now find each answer in a sentence or a table row.

Answer the question near the top

Kevin Indig, who writes the Growth Memo newsletter, matched 18,012 ChatGPT citations to the part of the page they came from. He found⁠(opens in a new tab) that 44.2% came from the first 30% of the page. Opening passages that ChatGPT cited were also almost twice as likely as those it skipped (36% against 20%) to use defining language, such as "is defined as" or "refers to". His data came from a commercial tracking tool, so treat it as one analysis, not a rule.

Microsoft makes a related point. Its guidance⁠(opens in a new tab) says "AI assistants don't read a page top to bottom like a person would". They break pages into smaller pieces, so each section should answer its question on its own. Microsoft recommends headings that describe what a section answers, such as "What Makes This Dishwasher Quieter Than Most Models?" rather than "Learn More", questions followed by direct answers, and tables for comparisons.

Google is more relaxed. Its AI optimisation guide⁠(opens in a new tab) says people "appreciate it when web pages are organized by paragraphs and sections, along with headings". But it also says "There's no requirement to break your content into tiny pieces for AI", and that there's no ideal length. "You don't need to write in a specific way just for generative AI search."

Both companies describe pages that a busy person can read easily, which suits AI assistants too. In practice:

1. Put the answer to each section's question in its first one or two sentences, in words that make sense if they're quoted on their own. Add the detail after.
2. Write headings that say what the section answers. "How much does the subscription cost?" is better than "Pricing made simple".
3. Keep to one idea per paragraph. Microsoft specifically warns against "overloaded sentences" that pack several claims together.
4. Use a table to compare options, prices or features, and a numbered list for steps.
5. Define terms the first time you use them, in a single sentence.

Back up what you say

What the research found

The original GEO study⁠(opens in a new tab), published in 2023, tested nine ways of rewriting a web page and measured how much of each page an AI-written answer used. Three changes worked best: adding quotations from relevant sources, adding statistics, and citing sources. Each raised the page's share of the answer by up to about 40% on the study's main measure. That measure counted how many words of the answer came from the page, and how early they appeared. Keywords are the words people type into a search engine. Adding extra keywords to a page, a tactic from old-fashioned SEO, made little or no improvement. On the main measure it did slightly worse than leaving the page alone.

The study has three limits:

• It used a research system the authors built, which read the top Google results and wrote answers with GPT-3.5, an older OpenAI model. In a smaller test of 200 questions on Perplexity, with the pages uploaded as files, the best gain on the same measure was 22%.
• The rewritten page was always one of the pages the system read. The study shows how a page is used once an assistant has found it, not how to get found.
• Pages that were already the first result lost ground when they were rewritten in these ways. The gains went to pages lower down the list.

Newer research has been more cautious. In May 2026, three researchers from Sprinklr, a company that sells customer experience software, ran 252,000 tests with six AI models (Vishwakarma and others⁠(opens in a new tab)). Each test showed a model two versions of a product review that differed in one detail, and recorded which one it cited first. How closely a version matched the question, and where it appeared in the list of sources the model was given, mattered most. Stating prices helped, and a 2026 date beat a 2019 one. Changes to formatting alone made little difference.

In July 2026 a review of 45 studies⁠(opens in a new tab) concluded that "no reviewed technique shows a stable, longitudinal, cross-platform causal effect on organic discoverability or downstream behavior". In other words, rewriting a page can change how an assistant uses it once it has found it. But no tactic has yet been shown to reliably help a page get found, or bring more visits, over time and across assistants. It found that relevance to the question, and where a page sits among the sources an assistant reads, were the effects that showed up most consistently.

What to do with it

The research supports writing that is relevant, specific and backed by evidence. Bing gives the same advice in its AI Performance announcement⁠(opens in a new tab): "Support claims with evidence. Examples, data, and cited sources help build trust when content is reused in AI-generated answers."

• Replace vague words with numbers wherever you have them. "Most deliveries arrive in two working days" is more useful than "fast delivery", as long as it's true.
• Link to your sources, especially for statistics. Link to the original research, not to a blog post about it.
• Quote named people, and say why they know what they're talking about. Fernhill quotes its head roaster, Sam Okafor, on how long each roast stays fresh.
• Give your data a date and a source. "In our September 2026 survey of 412 customers" can be cited. "Customers love it" can't.

Keep important pages up to date

Ahrefs studied 17 million citations in 2025 and found⁠(opens in a new tab) that the pages AI assistants cited were, on average, 2.9 years old, against 3.9 years for pages in Google's ordinary results. They had also been updated slightly more recently. ChatGPT showed the strongest preference for recent pages, and Google's AI Overviews the weakest. Ahrefs sells SEO tools, so treat this as one company's study.

Fresh means the content has changed, with new prices, a new plan or updated test results. Changing the date at the top of a page without changing anything else is a different matter. It may work for a while, but Google treats it as a warning sign. Its guidance on helpful content⁠(opens in a new tab) asks "Are you changing the date of pages to make them seem fresh when the content has not substantially changed?" The SEO consultant Lily Ray also warns against⁠(opens in a new tab) "artificial refreshing".

When you update a page:

1. Change what readers need changed, such as prices, availability or results.
2. Show a "last updated" date on the page, and change it only when the content changes.
3. Ask whoever looks after your website to keep two other dates the same as the one on the page. One is in your structured data, the code that labels facts for machines. The other is in your sitemap, the list of your pages that search engines read.
4. Tell Bing about the change with IndexNow, a way of notifying Bing and some other search engines that a page has changed. The technical SEO guide's sitemaps chapter explains steps 3 and 4.

Pages that help with buying questions

These types of page answer the questions people ask assistants before they buy.

Comparison pages done fairly

For questions that ask for the best of something, lists of recommendations are among the pages ChatGPT cites most. Ahrefs looked at the answers to 750 "best X" prompts in late 2025, such as "best home treadmill", and found⁠(opens in a new tab) that "best X" lists made up 43.8% of the pages ChatGPT cited. Getting onto other websites' lists is covered in the chapter on other websites.

Your own comparison pages can help too, if they're fair. A good one:

• says clearly who wrote it, such as "we make Fernhill Coffee, and this page compares our plans with others"
• uses the criteria a buyer would use, with dates next to prices
• describes competitors accurately, and says where they're the better choice
• is updated when products or prices change

A list on your own site that ranks your own business first, presented as if it were independent, is a different thing. It carries risks from Google and from consumer protection law, which the chapter on what to avoid explains.

Show who wrote it

Google's helpful content guidance asks whether it's "self-evident to your visitors who authored your content" (Google⁠(opens in a new tab)). Name the person who wrote each guide or article, and link to a page about them that shows their experience. Fernhill's brewing guides name Sam Okafor and link to his profile. The technical SEO guide's structured data chapter shows how to describe authors in code as well. No AI company says its assistant checks authors directly, but named, qualified authors make a page more trustworthy to readers and to Google.

Images and video

Video can help as well as text. Ahrefs found YouTube was the most-cited website in Google's AI Overviews⁠(opens in a new tab) in its March 2026 study, and the chapter on other websites covers video in more detail. Whatever you show in an image or video, say the important parts in text on the page too, because a crawler that reads only text can't see them.

Using AI to help write

Google has no objection to content written with AI help. Its guidance on AI-generated content⁠(opens in a new tab) says AI "can be particularly useful when researching a topic, and to add structure to original content". It also warns that generating "many pages without adding value for users may violate Google's spam policy on scaled content abuse", which is Google's name for mass-producing pages to manipulate its results.

An AI tool also can't supply facts it doesn't have. A page written only from what's already on the web is commodity content, and ChatGPT appears to be citing that kind of page less. Use AI tools to draft, shorten and organise. Supply the facts, numbers and experience yourself, and have someone who knows the subject check every claim.

What doesn't help

• Adding keywords. The GEO study found it made little or no improvement.
• Formatting for its own sake. The Sprinklr tests found formatting-only changes had little effect. Structure helps when it makes the content clearer.
• Chopping pages into small fragments. Google says there's no requirement to, for its own AI features.
• Writing to a word count. Google says there's no ideal length for its AI features.
• Changing dates without changing content. It may help briefly, but Google treats it as a warning sign.
• A separate page for every possible variation of a question. Google says creating pages for the searches behind a question "primarily to manipulate rankings or generative AI responses" breaks its scaled content abuse policy (AI optimisation guide⁠(opens in a new tab)). Answer the related questions on the pages where they belong.

A shape for a page you want cited

This is a suggested shape, not a rule from any AI company. Fernhill used it for its subscription page.

1. A heading that says what the page is about in the words customers use: "Coffee subscription: fresh beans every week, fortnight or month".
2. Two or three sentences under the heading that answer the main question.
3. The facts, in a table if there's more than one option.
4. Sections with headings that ask the follow-up questions customers ask, such as "Can I pause my subscription?" and "Do you deliver to Ireland?", each answered in its first sentence.
5. Evidence, such as test results, survey figures, or quotes from named staff or customers, with dates.
6. Who wrote the page and when it was last updated.
7. What to do next, such as how to subscribe.

What to do next

1. List the ten questions customers ask most before they buy. Sales calls, support emails and reviews are good places to find them.
2. For each question, find the page on your site that answers it. If there isn't one, or the answer is vague or buried, note it.
3. Check that your prices and key facts are in the page source, not only in images, PDFs or content added by JavaScript.
4. Rewrite the opening of your most important pages so the first two sentences answer the main question.
5. Write down three things your business knows that competitors can't publish, and plan a page around one of them.

Further reading

• Optimizing your website for generative AI features on Google Search⁠(opens in a new tab), Google's guide, including commodity and non-commodity content.
• Optimizing your content for inclusion in AI search answers⁠(opens in a new tab), Microsoft's practical advice, with examples.
• Creating helpful, reliable, people-first content⁠(opens in a new tab), Google's self-assessment questions.
• GEO: Generative Engine Optimization⁠(opens in a new tab), the 2023 study that named the field.