Will AI answers quote your page? A section-by-section rewrite you can test

What does it mean for an AI answer to quote your page?

It means an engine retrieved your page for a query it wrote itself, lifted a passage from it, and built part of its answer on that passage, usually with a link. The engine works with sections, not whole pages. A page gets quoted when one of its sections answers the engine's query cleanly and can stand alone once it has been cut out.

The query is often not the buyer's. Ahrefs, which sells AI visibility tools, tracked 66.6 million free, logged-out ChatGPT responses on GPT-5.6 Luna in August and September 2026 and found that 14.4% triggered a web search.¹ When ChatGPT did search, it ran about 1.6 fan-out queries per searching response.¹ Each fan-out is a search query the model writes for itself, so the passage it lifts has to match that rewritten query.

From a buyer's prompt to a quoted passage

1. Buyer's prompt
2. Fan-out queries the engine writes
3. Pages retrieved
4. Passage lifted from one section
5. Answer, with or without a link

Prior work, and what this method adds

Researchers and vendors have studied which content features get cited, and we draw on their work below. Our own review of schema, llms.txt and Reddit found that quotable pages had better support than markup. This piece turns that finding into a method for one page, with a test at the end so you can see whether the rewrite changed anything.

A disclosure first. Sigzen AI sells AI visibility audits and sprints at published prices, and answer-first rewrites are part of the sprint. Weigh our view with that in mind. Most of the studies cited here also come from companies that sell tools or services in this category.

Why most company pages are hard to quote

Most company pages were written to be read top to bottom by a person who already chose to visit. An engine reads them differently. It wants the one section that answers its query, and it needs that section to make sense without the rest of the page. Five faults break that, and most service pages carry at least three of them.

• Preamble before the answer. The section opens with context, history or a mission line, and the fact the buyer asked about arrives in the fourth sentence.
• An unnamed "we". A passage that says "we support every major accounting system" loses its subject once it is lifted. The engine has to guess who "we" is, and it may not credit you.
• Facts locked away. Prices sit in an image, features in a tab that scripts load later, terms in a PDF. Some crawlers fetch only the HTML the server sends.
• One fact in several versions. The pricing page, a blog post and a partner listing give three different starting prices. The engine picks one, and it may pick the oldest.
• Figures without a source or date. "Most customers save hours each week" gives an engine nothing to check and nothing a careful answer can repeat.

None of these is a ranking problem. A page can rank well and still be hard to quote, which is one reason ranking and citation overlap less than many teams expect.

Step 1: pick one page and the prompts it should answer

Start with one page that buyers' prompts should land on, and write down the prompts before you touch the copy. A rewrite without a fixed prompt set cannot be tested, and a rewrite that cannot be tested is an opinion.

Choose the page by commercial weight, not by traffic. A pricing page, a core service page or a comparison page usually matters more than a popular blog post. Then write ten to twenty prompts a buyer would type when that page is the right answer. Include the constraint buyers attach: company size, country, budget, the system they already run.

Prompts have grown more specific. Semrush, which sells AI visibility tools, found in US clickstream data that search-enabled ChatGPT prompts averaged 8.7 words in January–February 2026, up from 4.7 words a year earlier.² Google says the average AI Mode search is triple the length of a traditional query, a US company statement from May 2026.³ Your prompts should look like that: whole questions with the constraint attached.

If you need help building the list, our guide to building a prompt set from real buyer questions covers sales calls, support tickets and search data as sources.

Step 2: how do you find the passages an engine would lift?

Map each prompt to the section of the page that should answer it, then read only that section as if it were the whole page. If the answer is missing, buried or unclear without the sections around it, that section is the rewrite target. This inventory takes about an hour for one page.

Work in a table. One row per prompt, with the section that should answer it, the sentence that currently carries the answer, and which of the five faults applies. Rows with no matching section are gaps, and gaps usually matter more than weak sections because no rewrite can quote a fact the page never states.

Illustrative: a passage inventory for a fictional payroll vendor's service page.

Our view: the gap rows are where most of the value sits. Teams like to polish the sections they already have. The prompts with no answer on the page are the ones an engine will answer from someone else's site.

Step 3: rewrite each section answer-first

Each section should open with a sentence that answers its prompt on its own, in under 30 words, naming your company and the buyer's constraint. The rest of the section supports that sentence with detail, a source and a date. A section that passes this test can be lifted whole and still be right.

Six rules for a quotable section

1. Answer in the first sentence. Put the fact first and the context after it.
2. Name the company. Write "Contoso Growth files statutory returns in Mexico" rather than "we handle compliance".
3. State the constraint. Say who the answer is for: the company size, the country, the system it connects to.
4. One idea per section, in about 120 to 180 words. Long sections bury the answer, and very short ones lose the context that makes them safe to lift.
5. Source every figure. Give the number, where it comes from and the period it covers, in the same sentence.
6. Use the buyer's words in the heading. A question heading should be the question buyers ask, in their words, not a slogan.

One section, before and after

Before

• Heading: "Built for growth"
• Opens with the company's founding story
• "We work with businesses of every size"
• Countries listed in a map image
• "Customers save hours every week"

After

• Heading: "Who is Contoso Growth for?"
• Opens: "Contoso Growth runs payroll for firms of 10 to 500 staff in Mexico and Colombia."
• Company named in the first sentence
• Countries in the text
• Claims tied to a dated, named source, or cut

What not to change

Keep the page readable for people. An answer-first section reads well to a buyer too, because it respects their time. What the rewrite should not do is turn every heading into a question, repeat the company name in every sentence, or pad sections with keyword variants. Our view: if a sentence would embarrass you in front of a buyer, it does not belong on the page, whatever a tool scores it.

Does adding statistics and sources still help?

The evidence says yes in direction, though the size of the effect comes from older lab work. More recent field data points the same way: engines checking facts lean on official, sourced pages. The practical rule is to give engines a checkable fact with a named source rather than a claim.

What the lab study found

In a 2024 lab study, now dated, Aggarwal and colleagues found that adding citations, quotations and statistics to source pages were the best-performing methods, with relative gains of 30–40% on one of their visibility metrics.⁴ Keyword stuffing did little.⁴ That benchmark ran on generative engines of its time. Treat the size of the gain as dated and the direction as useful.

What current field data adds

Ahrefs' October 2026 test points the same way for factual questions. In 20,000 US prompts, 84.4% of factual questions about pricing, product features or company details cited at least one official source, against 68.3% for comparison questions.¹ When a buyer asks a checkable question, the engine looks for the company's own page and quotes it if the fact is there.

ChatGPT in 2026

Most answers do not search, but searching answers trust official pages

14.4%of 66.6M free, logged-out responses triggered a web searchAhrefs, Aug–Sep 2026

~1.6fan-out queries per searching responseAhrefs, Aug–Sep 2026

84.4%of factual questions cited an official source, 20,000 US promptsAhrefs, Oct 2026

Company pages are a large share of what gets cited overall. Profound, which sells an AI visibility platform, counted 11.84 billion citations on eight engines from 16 April to 16 July 2026, and about 57% pointed to company-owned websites: 47% on ChatGPT and 69% on Gemini.⁵ "Company-owned" means any company's site, including your competitors'. Their pages are quoted when yours are not.

Our view: this is the strongest case for rewriting your own pages before you spend on outside mentions. Factual prompts already send engines to official sources, so the cheapest gain is making sure the fact they look for is on your page, stated plainly.

Step 4: put every key fact in server-sent text

Every fact a buyer's prompt asks about should be in the HTML your server sends: plain text, not an image, not a tab loaded by scripts, not only in a PDF. Some crawlers read only that first response. Copy that appears later may never reach them.

Vercel's analysis of AI crawler traffic, published in December 2024 and now dated, reported that the major AI crawlers it observed did not render JavaScript.⁶ Crawlers change, and we would not rely on that study for any one engine today. The safe assumption is the same either way, and Google's guidance for its AI features asks for important content to be available in text form.⁷

A ten-minute check

• Open the page with JavaScript switched off. Every answer sentence from step 3 should still be there.
• Search the page source for your starting price, your countries and your main integrations. If a fact is only in an image's pixels, it is not in the source.
• Open closed accordions and tabs in the source. Their text should be in the HTML even when the panel is shut.
• Check robots rules and any bot settings at your CDN, so the crawlers that fetch pages for answers are allowed in.

Then fix the version problem. List every place a key fact appears: your page, old blog posts, partner listings, directories, press releases. Make them agree, and date the canonical version. If an engine already repeats a wrong figure, our guide to fixing wrong AI answers sets out how to trace it to the page it came from.

How do you test whether the rewrite worked?

Run the same prompt set before and after the rewrite, several times on each engine, and compare the share of answers that name you, cite the page and state the fact correctly. Change one page at a time, and keep the change only if the difference is larger than run-to-run noise.

Set the baseline first

Run the baseline in the week before you publish the rewrite. AI answers vary from run to run, so one run per prompt tells you little. MaxAEO, which sells AI visibility measurement, has written about how runs of the same prompt cluster, so repeated runs carry less independent information than their count suggests.⁸ Run each prompt several times on each engine and record three things per answer: named or not, page cited or not, fact right or wrong.

Wait, then run it again

Publish the rewrite and check that the engines can fetch it. Then wait two to four weeks before the second run, because engines do not re-read every page at once. Run the identical prompt set, the same number of times, on the same engines and tiers, and compare the three shares with a margin of error on each.

A one-page rewrite test

1. Day 1

   Pick the page and write ten to twenty buyer prompts with constraints.

2. Days 2–5

   Baseline: several runs per prompt per engine; record named, cited, correct.

3. Days 6–7

   Passage inventory, rewrite to the six rules, server-text check, publish.

4. Weeks 3–5

   Rerun the identical prompt set and compare shares with margins of error.

Be careful about what moved. If a model update lands between the runs, the change you see may be the engine, not your page. Our note on measuring around a model release covers how to separate the two. For what a vendor should report when it runs this for you, our measurement spec lists the questions to ask, including of us.

The payoff from being cited can be large where it applies. Seer Interactive, a search agency, tracked a panel of 53 brands from January 2025 to February 2026: pages cited in a Google AI Overview earned 120% more organic clicks per impression than when they were not cited.⁹ That is one agency's panel and one Google surface, so read it as a strong signal, not a forecast for your page.

What will a page rewrite not fix?

A rewrite changes what engines find when they search. It does little for answers drawn from memory, for engines that seldom link, or for categories where outside sources dominate. Know those limits before you promise anyone a result from it.

Most answers do not search. If 14.4% of free, logged-out ChatGPT responses in August–September 2026 triggered a web search, the rest drew on what the model already held.¹ A rewrite reaches those answers only slowly, through later training and through the third-party pages that repeat your facts. Our analysis of ChatGPT's search rate explains what that means for the prompts you track.

Links are rare in some sectors. Similarweb found that about 6.8% of US ChatGPT answers carried a citation in May 2026, with travel near 23% and professional services under 4%.¹⁰ In a low-link sector, judge the rewrite by whether the engine names you and states your facts correctly, not by referral clicks.

Engines also differ in how much room they leave. In Semrush's 2026 AI Visibility Index, covering 126 million US prompts from January to April 2026, ChatGPT cited an average of 15 sources per response and Gemini an average of 3.¹¹ Where an engine cites only a handful of sources, your page competes with the few authoritative pages it already trusts.

Ranking is a separate question too. In an Ahrefs study of 863,000 search results and 4 million cited URLs, published in March 2026 with the data period not stated, 38% of pages cited in AI Overviews ranked in Google's top 10 for that query.¹² A rewrite can help a page that ranks lower get quoted. It will not lift a page that engines never retrieve.

Which page should you rewrite first?

Rewrite the page whose prompts are factual and commercial: pricing, a core service, or the page that answers "does it work with my system". Those prompts send engines to official sources most often, and the facts on them are the ones buyers check before they call.

Our view: the pricing page is usually the best first test. Its facts are specific, its prompts are factual, and a wrong price in an AI answer costs you directly. After that, take the service page with the most gap rows in its inventory, then any comparison page that names your competitors.

Leave the blog until last. Posts can be quoted too, but a buyer comparing vendors wants your facts from your own pages, and those are the pages an engine checks first. If the rewrite works on one page, roll the six rules out page by page, testing each on its own prompts. Sigzen AI's AI visibility audit runs the baseline across five engines and names the three pages to fix first.