How to write content AI assistants quote accurately

TL;DR

• AI assistants rarely read your whole page. They retrieve a passage, then rewrite it. Errors creep in when that passage is vague, incomplete or contradicted elsewhere.
• Write passages that make sense on their own: name the product, state the full fact (number, unit, period, condition, date) and put the answer first.
• Keep one page as the source of truth for each fact, and fix or remove old pages that disagree with it.
• Put facts in plain HTML text. Many AI crawlers do not run JavaScript, and content in images, tabs or PDFs is easy to miss.

Why assistants misquote websites

When someone asks ChatGPT, Gemini, Copilot or Perplexity about your company, the assistant usually searches, retrieves a few passages from different pages and writes an answer from them. Microsoft's Bing team puts it plainly in its guidance on AI search answers: "AI assistants don't read a page top to bottom like a person would." They break pages into pieces, and "these modular pieces are what get ranked and assembled into answers."

That process creates predictable failure points. Each one below is something you control.

• The passage loses its context. A paragraph that says "It costs $49 and includes everything in Basic" makes sense under the right heading. Pulled out alone, nobody knows what "it" is, whether $49 is monthly or yearly, or what "Basic" contains.
• The fact is incomplete. "From $49" becomes "costs $49" in an answer, and the conditions disappear.
• Two pages disagree. Your pricing page was updated, but a 2024 blog post still mentions the old price. The assistant may retrieve either one.
• The claim is vague. "Fast setup" or "flexible pricing" gives the model nothing concrete to repeat, so it paraphrases, and paraphrase invites invention. A made-up specific is the classic hallucination.
• The fact is hard to reach. Prices in an image, specs in a PDF, details inside a tab that only opens with JavaScript.

The rest of this guide is about removing those failure points, one habit at a time.

Write passages that stand alone

The single most useful test is what we call the paragraph test: copy any paragraph from your page into a blank document. Does it still say who, what and how much? If not, it will not survive retrieval either. Microsoft recommends exactly this: "Sentences that make sense even when pulled out of context."

Name things instead of pointing at them

Replace "it," "this," "our solution" and "the platform" with the product or plan name, at least once per section. It feels repetitive to write and reads fine.

Before:

It integrates with your CRM and syncs every few minutes.

After:

Example CRM Sync connects to HubSpot and Salesforce and syncs contacts every 5 minutes.

Put the answer first

Under each heading, make the first sentence answer the question the heading asks. Explanation comes after. If a section is titled "Does Example support SSO?", the first words should be "Yes. Example supports SAML single sign-on on the Business plan." An assistant that retrieves only the top of the section still gets it right.

Use the buyer's words in headings

Headings define the slice of content that gets retrieved. Write them as the question or topic a buyer would type: "How much does Example cost for a team of 20?" works better than "Investment." Microsoft describes headings as working, for AI, "like chapter titles that define clear content slices."

State facts completely

A fact an assistant can repeat correctly has five parts: the number, the unit, the period, the condition and the date.

All the examples above are fictional. The point is the shape: no unit, no period and no condition left for the model to guess.

Anchor claims in something measurable

Microsoft warns that "terms like innovative or eco mean little without specifics." If you want to say setup is quick, say how long it takes and for whom. If you cannot measure it, consider leaving the claim out. An honest "setup time depends on how many data sources you connect" is quotable; "effortless setup" is not.

Say what you don't do

Assistants fill gaps with assumptions, often optimistic ones. A sentence like "Example does not offer on-premise installation" or "We don't ship outside the EU" prevents a wrong answer that no amount of positive copy can. Negative facts are some of the most valuable sentences on a B2B site.

Date anything that changes

Prices, plan limits, integration lists and compliance status change. Add "as of October 2026" to statements that will age, and show a visible "last updated" date on pages like pricing and security. A dated fact tells both the reader and the assistant how much to trust it.

Keep one source of truth

Many misquotes of B2B sites are not model errors. They are faithful quotes of the wrong page.

Make facts reachable

A fact that a crawler cannot see cannot be quoted.

• Text, not images. Pricing tables, spec sheets and comparison charts should be HTML text. If you use an image, repeat the key numbers in text nearby. Microsoft advises to "always provide alt text or present critical details in HTML."
• Not only in PDFs. Datasheets are useful downloads, but put the critical specs on an HTML page as well.
• Not hidden behind tabs or scripts. Microsoft notes that "AI systems may not render hidden content, so key details can be skipped." Vercel's study of AI crawler traffic found that "none of the major AI crawlers currently render JavaScript," with Google's Gemini and Applebot as exceptions. Server-render the facts that matter.
• Tables with headers. Comparison and spec tables need a header row and units in the header ("Price per user per month, USD"). A table without headers is a grid of numbers with no meaning.

Structured data, honestly

Structured data such as Product, Offer, FAQPage or Organization markup can make meaning explicit. Microsoft recommends it. Google says it "isn't required for generative AI search," while still recommending it for SEO in general. Either way, follow Google's structured data guidelines: "Don't mark up content that is not visible to readers of the page," and the markup "must be a true representation of the page content." Markup that disagrees with your visible text is another way to create two sources of truth.

Back up claims about the outside world

When you make claims about your market, standards or regulations, cite the source and link to it. The GEO research paper found that adding citations, quotations and statistics to content increased its visibility in AI-generated answers in controlled tests. Sourced claims are also easier for an assistant to repeat accurately, because the facts and their origin travel together.

Don't over-apply this. Inventing statistics to look quotable is worse than having none, and a single wrong number can be repeated thousands of times.

A before-and-after page section

Here is a typical pricing FAQ answer, rewritten.

Before:

Is there a discount? Absolutely! We offer great savings for teams who commit. Contact us to learn more about our flexible options.

After:

Is there a discount for annual billing? Yes. Example's annual plans cost 20% less than paying monthly. For example, the Pro plan is $48 per user per month billed annually, or $60 billed monthly (prices as of October 2026). Teams over 100 users can request volume pricing on the contact page.

The rewrite answers first, names the product and plan, gives both prices with periods, dates them and says what happens at the edges. Every sentence passes the paragraph test.

How to check your work

• Ask the assistants. Run the questions your buyers ask in ChatGPT, Gemini, Copilot and Perplexity. Compare each answer with your canonical page and trace any error to the passage it came from. Our guide to how AI assistants describe your company has a full audit routine.
• Read your pages without styles and scripts. View the raw HTML of key pages and confirm the facts are there.
• Test retrieval on your own assistant. If you run an AI chat on your site, ask it the same questions. It draws on the same pages, so a misquote there usually points to the same weak passage. Accuracy matters on your own site in a legal sense too: in Moffatt v. Air Canada, a tribunal held the airline responsible for what its website chatbot told a customer, even though another page on the site had the correct policy. Our guide to grounded answers covers the controls for that case.

Where sources disagree

Google's guide to its generative AI features says "you don't need to write in a specific way just for generative AI search" and "there's no requirement to break your content into tiny pieces." Microsoft recommends modular sections, Q&A blocks and short self-contained sentences. The two positions overlap more than they conflict. Neither asks you to fragment pages or write for machines instead of people. Both reward sections that answer one question clearly, facts stated in full and content that people find useful. Write that way and you satisfy both, along with the chunking step inside every retrieval system.

AI assistants will keep quoting your website whether you plan for it or not. Complete facts, one source of truth and plain HTML are what turn those quotes from approximate to accurate, and they make the page better for the people reading it, too.