How to Create Content AI Can Quote

Most advice on "quotable content" tells you to rewrite your pages for AI. Add statistics. Add expert quotes. Break everything into short answer blocks under question-shaped headings. It sounds reasonable. When researchers tested it, it did little or made things worse.

The problem customers bring to a fixing tool is usually a different one. AI has read their pages, and it repeats the wrong number from one of them. The pricing page says $79 a month. A launch post from two years ago says $59. The engine cannot tell which one is current, so it uses whichever one it retrieved.

This lesson replaces the rewriting playbook with something you can defend: state each fact plainly, once, in the same words everywhere, in HTML a crawler can read without running JavaScript. Then go and find the pages that disagree.

What the evidence says about writing for AI

The tactics sold as "GEO writing" have been tested more carefully than almost anything else in this field, and the results are not kind to them.

Three other sources point the same way. A 252,000-trial factorial experiment (arXiv:2605.25517) found that "formatting-only edits have little impact." A July 2026 critical survey of 45 GEO studies (arXiv:2607.14035) warns that "citation-oriented rewrites can impair retrieval." And Google's AI optimization guide, updated July 10, 2026, says it directly: "You don't need to write in a specific way just for generative AI search." The same guide tells you not to worry about covering every long-tail variant of a question.

Two caveats keep this honest. First, nearly all of these experiments placed documents directly into the model's context. They measure what happens after a page has been retrieved, not whether writing style helps a page get retrieved at all. Second, they do not prove that content quality is irrelevant. They show that rewriting for citation is not a lever. Writing clearly for the people who read your site is still worth doing. It is not a GEO tactic.

The 252,000-trial study did find one thing that helped consistently: "Including explicit price information and a recent timestamp." That result comes from a synthetic corpus with two candidate documents per trial, in a paper from a company that ships its own GEO checklist, so treat it as unresolved. It is still a reason to put your real price on the page in text rather than behind a "Contact sales" form. It is not a reason to change the date on a page whose facts did not change.

Quotable means stated plainly, once, the same way

Drop the idea that quotability is a writing style. Treat it as a property of your facts. A fact is quotable when three things are true:

1. It exists in the raw HTML. A crawler that does not run JavaScript can read it.
2. It stands on its own. One plain sentence carries the whole fact, without "as mentioned above" or a pronoun pointing back at an earlier paragraph.
3. It is consistent. Every page that states the fact states the same value.

The first condition has the strongest evidence in this whole course. Vercel's crawler study (December 2024) found that ChatGPT's and Claude's crawlers fetch JavaScript files but "don't execute them. They can't read client-side rendered content." Independent re-testing in June 2026 found the same for GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, Claude-SearchBot and PerplexityBot. Googlebot renders JavaScript; most AI crawlers do not. A price that loads into a pricing toggle after the page renders does not exist for those crawlers. Rendering, noindex and Snippet Controls walks through the check, and the short version takes two minutes: view the page source, not the browser inspector, and search it for the fact.

The second condition is about definitional sentences. A true definitional sentence names what you are, who it is for and what it does, with nothing a reader could not verify:

• Checkable: "Ledgerly is invoicing software for B2B companies with 10 to 500 employees. It sends invoices, tracks payments and chases late payers automatically."
• Not checkable: "Ledgerly is the leading AI-powered financial operations platform for modern businesses."

The second version contains no fact. Leading by what measure? Which operations? An engine that repeats it repeats a slogan, and an engine that paraphrases it can drift anywhere. Write the first kind, put it near the top of your homepage and About page, and use the same sentence in your directory profiles so the wording matches everywhere.

The third condition is where most brands fail, and it gets its own section.

Contradictions: the plausible reason AI gets your facts wrong

When an engine searches before answering, it does not read your whole site. It turns the user's question into search queries of its own, retrieves a handful of pages and works from those. OpenAI's help center says ChatGPT search "typically rewrites your query into one or more targeted queries." Which of your pages comes back for a given query is not something you control. Research on retrieval pipelines also shows that a document's position in the model's context strongly affects whether it gets cited, and no publisher can set that position.

Put those together. If your pricing page says $79 and a two-year-old announcement post says $59, some queries will retrieve the post. The engine then states $59 with full confidence, because as far as it can see, you said so. This is not a proven causal chain; nobody has published a before-and-after study of it. It is the plausible mechanism behind the most common complaint we hear, and it is entirely under your control.

Contradictions hide in predictable places:

• Old blog posts announcing a price, a feature or a launch date
• Comparison pages written against a competitor's old plans and your old plans
• Help center articles that describe a retired plan or feature
• The footer, the careers page and the press kit, which nobody updates after an office move, a new CEO or a round of hiring
• JSON-LD blocks hardcoded into a template years ago, covered in Schema Markup for AI
• Third-party pages you do not own, covered in Building Off-Site Corroboration

Answers generated from training data alone, with no search, are frozen at the model's training cutoff. Nothing you fix this quarter changes them, which is why this work targets what engines retrieve now. A 99Visibility audit tells the two apart: each answer is marked "Searched the web" or "Answered from memory." A wrong fact on an answer from memory is a stale belief from training. On an answer that searched, it is something a page said, and that page is what this module helps you fix.

99Visibility's accuracy check compares what each engine says against the brand facts you enter: pricing, features (including features you do not have), founding date, location, CEO and leadership, ownership, acquisitions, headcount and outdated facts. Those fields are also a good template for what your own site must state consistently. Find What AI Gets Wrong About You covers how the check works.

Build a canonical facts block

A canonical facts block is one short section, usually on your About page, that states each core fact once. It is plain visible text in the server-rendered HTML: not an image, not an accordion that loads on click and not a PDF. Every other page either links to it or repeats its wording exactly.

Cover the facts engines get asked about and the facts the accuracy check verifies:

• Your name and the one-sentence definition
• Founding year and headquarters location
• CEO and other named leadership
• Ownership: independent, or owned by whom, and any acquisitions
• Headcount, as a range you can keep current
• Plans and prices, with a link to the pricing page as the detailed source
• What you do not offer, stated plainly

The last item matters more than it looks. Engines can credit a product with a competitor's features, which is why the accuracy check flags features a brand does not have. A sentence like "Ledgerly does not run payroll" gives a reader a straight answer and gives your own audit something to compare against. Nobody has shown that it stops an engine from guessing, so do not oversell it to your team.

Give the block one owner. When a fact changes, the owner changes it there first and then works through every page that repeats it. That routine is the whole system, and it costs nothing but attention.

Comparison tables and definitions in plain HTML

Buyers ask engines to compare products. If you do not publish your own comparison, the page that gets retrieved for "Ledgerly vs. X" is someone else's, written with their framing and their version of your price. Publishing one gives the engine at least one source where your facts are right.

Build it as a real HTML <table> in the server-rendered page. Not an image of a table, not a widget that loads its rows from an API, and not a tab that fetches its content when clicked. Keep it fair: state the competitor's facts as of a date you name, link to their pricing page, and update the table when either side changes. A comparison page that gets a competitor wrong is a contradiction with their site, and engines can retrieve both.

None of this is a formatting trick. The 252,000-trial study found formatting-only edits have little impact, so do not expect a table to get you cited because it is a table. It is there because a table is the clearest way to state facts that are clumsy in prose, and clear facts leave less room for an engine, or a buyer, to get them wrong.

What to remove, not add

Some tactics sold as GEO writing are worse than useless. Treat these as risk flags when you review a page:

• Injected statistics. Numbers added to look authoritative, with no source, or with a source that does not say what the sentence claims. Each one is an unverifiable claim attached to your brand.
• "According to experts" boilerplate. Phrases like "according to authoritative sources" or "experts agree" with no named expert. They give a reader nothing to check.
• Keyword stuffing. In the original GEO paper (KDD 2024), keyword stuffing lowered the authors' own visibility metric, 17.8 against a baseline of 19.5.
• Whole-page AI rewrites. The study usually cited for "write prose that models find predictable" (arXiv:2509.14436) produced its effect by having an LLM rewrite entire pages, which changed tone, structure, hedging and length at once. Nothing in it isolates a style you could copy.

Common mistakes

• Fixing the pricing page and stopping. The pricing page was rarely the problem. The older pages that disagree with it are.
• Putting facts where crawlers cannot read them. Prices in images, in client-rendered toggles or behind a "Contact sales" form are not on the page as far as most AI crawlers can tell.
• Writing slogans as definitions. "The leading platform" gives an engine nothing true to repeat.
• Chunking pages into tiny blocks "for LLMs". Google says there is no ideal page length, and the chunking research measures how a retrieval system splits pages inside its own index, which you do not control.
• Adding FAQ markup instead of answers. If a question matters, answer it in visible text. Schema Markup for AI explains why the markup earns nothing new.
• Bumping dates. Changing a "last updated" date on a page whose facts did not change games a signal nobody has shown engines weight, and it makes your own records less trustworthy.