How to Write Content That AI Assistants Actually Cite

Why AI assistants ignore most business websites

AI assistants do not rank your website the way Google does. They retrieve passages and repeat them. When a searcher asks ChatGPT "who should I hire for X" or "what is the best Y in Z," the assistant looks for a sentence somewhere on the open web that answers the question in a clean, quotable form — and cites the source that produced it. The domain that owns that sentence gets the citation. Everyone else gets skipped.

That is why so much high-effort marketing content never earns an AI citation. A well-designed page can be beautiful for a human reader and still be structurally impossible to quote. Marketing headlines like "Your Trusted Partner in Complex Matters" do not answer any real question. Practice-area paragraphs that lump ten services into one block of prose do not isolate a passage the model can lift. PDFs behind a form gate do not exist as far as most retrieval layers are concerned.

Content that AI assistants actually cite tends to share four features. It states the buyer's question in the buyer's own words as a heading. It follows the heading with a short, direct, self-contained answer paragraph. It carries FAQPage JSON-LD structured data so the question-answer pair is legible as a discrete unit. And it lives in HTML on a page an anonymous crawler can reach. Get those four right and you have written content ChatGPT can cite. Miss any one and you have written content that ranks in Google but not in AI.

What does "AI-friendly content" actually look like?

AI-friendly content is content that can be quoted in isolation. Read a page out loud one paragraph at a time. If a paragraph makes sense on its own — if it names the situation, gives an answer, and does not rely on the paragraph before it for context — an assistant can quote it. If a paragraph only makes sense inside the flow of the surrounding page, an assistant will skip it.

Concretely, AI-friendly content pairs a scenario-shaped heading with a two-to-three-sentence answer. The heading uses the buyer's vocabulary, not the industry's. "Divorce involving hidden assets" is a scenario. "Complex matrimonial dissolution" is a marketing description of the same work — and only the first phrase matches what a buyer actually types into an assistant.

The best pages are dense with these units. Not one big FAQ at the bottom of the page. Multiple scenario-answer blocks, each about one specific buyer situation, each capable of standing alone. A site with fifteen tightly scoped scenario blocks covering fifteen buyer questions will earn citations across fifteen different prompt buckets. A site with one long "About Our Practice" narrative will earn citations across none of them, no matter how well written the narrative is.

Viclaro's deep-dive breakdown of this pattern lives in our guide to how to rank in ChatGPT — read it after this piece if you want the fuller playbook.

The one-quotable-sentence pattern that gets cited

The specific pattern that shows up on top-cited sites across every category is the same: one H2 or FAQ heading per buyer scenario, followed by a single quotable sentence that answers the scenario directly, followed by two-to-three sentences of supporting evidence. The quotable sentence is the payload. Everything else is context an assistant can use if it wants, but the answer sentence has to work on its own.

To write one, start by writing the exact question a buyer would type into ChatGPT before they know your industry's vocabulary. Not "matrimonial services for high-net-worth individuals." Instead: "how do I get divorced when my spouse is hiding assets in a business?" That question becomes your H2. Then answer it in one sentence that a stranger could read out of context and understand. "We represent New York spouses in divorces involving hidden business assets, forensic accounting, and undisclosed income." That sentence is what ChatGPT lifts.

The three-part test for whether a sentence is quotable: does it name the situation? Does it name who is being served? Does it name the specific service being provided? If the answer to all three is yes, the sentence works. If the sentence is about the company's values, its history, or how much it cares, it is prose — not a citation candidate.

This is the pattern top-cited firms share across every vertical Viclaro measures. It is not the only signal AI assistants use, but it is the one that most reliably separates a site that gets cited from a site that does not.

FAQPage schema: what it does and doesn't fix

FAQPage JSON-LD is structured data you embed in the head or body of a page that tells search engines and retrieval systems "this specific text is a question, and this specific text is the answer." When an assistant's retrieval layer sees FAQPage schema, it treats the marked-up question and answer as a discrete, quotable unit. That is a significant citation advantage in the decision and validation prompt buckets — the "who should I hire" and "is this firm a good choice" moments where recommendations actually happen.

What FAQPage schema does: isolates a question-answer pair from surrounding page noise, signals that the passage is designed to stand alone, and improves the odds that an assistant lifts the pair verbatim into a recommendation. Sites that add FAQPage schema to existing scenario content typically see measurable citation lift in downstream AI panels within a snapshot cycle or two.

What FAQPage schema does not do: rescue bad content. Schema-wrapping a marketing paragraph does not turn it into a quotable answer. Schema-wrapping questions the firm wishes the buyer would ask, rather than questions the buyer actually types, produces a technically valid FAQPage that no assistant surfaces. And schema alone will not overcome a page that is not indexable by crawlers, hidden behind a login, or served only inside a PDF.

The pattern that works: write the scenario-answer content first, in the buyer's language, on an indexable HTML page, then wrap it in FAQPage JSON-LD. The schema amplifies content that is already citable. It does not create citability where none exists.

Content anti-patterns that lose AI citations

The first citation-killer is burying the answer in a PDF. Case results, credentials, methodology descriptions, and scenario coverage sitting inside a linked PDF are effectively invisible to most assistant retrieval layers. If it matters for citation, it belongs in HTML, above the fold, in plain text, on a page an anonymous crawler can reach without a form submission.

The second is the undifferentiated practice-area page. A single page titled "Family Law" that lists ten sub-services in one paragraph loses to ten short pages that each name one scenario in the buyer's vocabulary. The information is often on both versions of the site. Only one version has it in a form the assistant can quote.

The third is peer language instead of buyer language. Content written for other lawyers, other doctors, or other professionals uses vocabulary the buyer does not have yet. "Advanced reproductive technology for advanced maternal age" is what a physician says. "IVF after 40" is what the buyer types. Assistants match buyer phrasing to page phrasing — same information, invisible on the physician-vocabulary page.

The fourth is unsupported superlatives. "Premier," "best," "aggressive," "top-rated," "trusted" — assistants ignore these entirely. They add no retrievable signal. Specific proof — years of practice, case types handled, published work, named certifications, exact numbers — is what an assistant will quote when justifying a recommendation. If the copy would still be true if a competitor swapped their name onto it, it is invisible content by design.

The fifth is content that never asks a question. Pages that describe capabilities without ever stating the questions those capabilities answer offer nothing for a Q&A retrieval system to match on. If the page has no headings shaped like questions, it will not surface on question-shaped prompts.

How to test whether your content is AI-citable

The cheapest first test is manual. Open ChatGPT, Claude, Gemini, and Perplexity in four browser tabs. Ask each of them the exact question your ideal buyer would ask before they know your industry's vocabulary. If your firm is not named in any of the four answers, the site is not yet AI-citable for that prompt. If the firm is named in one but not the other three, the content has partial reach — usually because one model's training or retrieval favors what is already published, and the other three do not have enough to work with.

The manual test is not a measurement. It is a sample of one, and any single response from a chat interface is noisy. It tells you whether your content is missing entirely, not whether it is on the margin. For real measurement, you need a panel — multiple assistants, multiple samples per prompt, aggregate share of voice. Viclaro's free AI-visibility scan runs a small version of that panel against your specific domain in about a minute.

The second test is the "quote it out loud" test. Pick a paragraph you want AI to cite. Read it aloud to someone with no context about your business. If they can restate what you do, who you serve, and in what situation, an assistant can too. If they cannot, the paragraph is not yet quotable — and no schema markup will fix that.

The third test is the leaderboard test. Go to the Atlas leaderboards and look at the firms cited most often in categories similar to yours. Their pages will show the pattern in practice: scenario H2s, one-sentence answers, FAQPage schema, buyer vocabulary, no PDFs, no undifferentiated practice pages. The winners are not louder than everyone else. Their content is shaped for citation.

Related reading

How to Rank in ChatGPT: the full playbook for site structure, prompt buckets, and content patterns that earn AI citations.

AI Search Rankings Guide: what a defensible AI visibility program measures and how the metrics differ from Google SEO.

How to Read an AI Recommendation Ranking Without Fooling Yourself: understand share of voice, Wilson intervals, and per-model variance before you interpret any leaderboard.