Share of answer is the operating metric for this practice.

How we count it
Request an audit

Insight · 15 min

What gets cited

ChatGPT, Perplexity, Claude, Gemini, and Copilot cite claims they can reuse, not content in the abstract. To be cited, a page has to be reachable by the search system that engine uses, and it has to hold a claim that is specific, attributable, stable, and either corroborated by other sources or original to you. As of November 2026, no engine offers a way to opt in, pay for organic citations, or guarantee one. Generative engine optimization is the work of making your claims the ones that survive the selection. It sits on the same retrieval work as SEO and AEO, as SEO vs AEO vs GEO explains.

In short

  • Engines cite reusable claims: a number, a definition, a condition, a boundary, with the source and date stated next to it. Restated advice is close to invisible, because the model already holds a hundred copies.
  • Each engine retrieves from a different search system. ChatGPT search uses OpenAI’s crawler plus outside providers, Perplexity runs its own crawler, Claude’s web search has been matched to Brave, Copilot uses Bing, and Gemini grounds in Google Search. Being retrievable is per engine.
  • For B2B buying queries, comparison pages carry weight. Siege Media found versus pages were the strongest content predictor of AI referral traffic across 116 B2B sites.
  • Write sections that stand on their own: a question as the heading, the answer in the first sentence, the evidence in the next. Do not split pages or engineer chunk sizes.
  • Citations churn. In Ahrefs’ October 2026 tracking, 50% to 83% of cited sites were gone the next day, depending on the engine. Measure presence over repeated runs and report mention, citation, and link separately.

What makes a claim citable?

A citable claim has four properties.

  • It is specific. A number, a definition, a condition, or a boundary. “Setup takes about two weeks for teams under 50 people” can be reused. “Fast, flexible onboarding” cannot.
  • It is attributable. The page says who measured it and when, and who stands behind the page.
  • It is stable. The URL still resolves, and the paragraph still says the same thing next month. A claim that moves to a new URL every quarter loses the links and citations it earned.
  • It is corroborated, or it is original. Either an independent source could confirm it, or you are the only place it exists.

The last property covers two kinds of claim, and they need different work. A pricing index you measure is original: engines cite you because nobody else has it. A definition of a category is common: engines look for agreement across sources, and the page that states it most clearly and consistently with everyone else tends to win. Label each claim before you write the page.

Engines also lean on the subject’s own site for facts about the subject. In Ahrefs’ October 2026 study of 20,000 US prompts, 84.4% of factual prompts cited at least one official source, against 68.3% of comparison prompts. Official sites took 63.4% of citations on factual prompts and 46.4% on comparison prompts. Your site is the default source for what you sell, what it costs, and what it does. Others are the default source for how you compare.

Where does each engine get its sources?

Each engine retrieves from its own search system, and a page has to be reachable by that system before any writing matters. The table lists what each operator documents as of November 2026. Where the operator is silent, the table says what independent testing found and labels it.

EngineWhere cited pages come fromBot or control that decides accessWhat the operator documents
ChatGPT searchOpenAI’s own crawl plus outside search providers. OpenAI does not publish a list of providersOAI-SearchBot. GPTBot is training onlyOpted-out sites “will not be shown in ChatGPT search answers,” except as navigational links. About 24 hours for robots.txt changes (OpenAI)
PerplexityPerplexity’s own index, plus live fetches when a user asksPerplexityBot for the index. Perplexity-User for live fetchesPerplexityBot surfaces and links sites in results and is not used for training. Perplexity-User “generally ignores robots.txt rules” (Perplexity)
ClaudeA web search tool Claude calls when it decides current information helps. Anthropic does not name the text search provider. In March 2025, Simon Willison matched Claude’s results to Brave Search, which Anthropic had added to its subprocessor listClaude-SearchBot for indexing. Claude-User for fetches. ClaudeBot is training onlyClaude decides when to search, and “every response includes citations.” Images come from Bing (Anthropic)
Microsoft CopilotThe Bing search serviceBingbot and Bing’s indexCopilot generates a search query of “a few words informed by the user’s prompt” and sends it to Bing (Microsoft)
GeminiGoogle Search, through groundingGooglebot for crawling. Google-Extended governs use in Gemini Apps groundingThe model “automatically generates one or multiple search queries and executes them,” then cites URLs for text segments (Google)
Google AI Overviews and AI ModeGoogle’s core ranking systems and the Search indexGooglebotCovered in how AI Overviews choose a source

Three things follow from the table.

Ranking on Google is not the same as being retrievable in ChatGPT. In an August 2025 Ahrefs analysis, only 6.82% of the URLs ChatGPT retrieved for its own search queries were in Google’s top 10 for those queries, and 83.39% did not appear in Google’s results at all. Ahrefs concluded that ChatGPT combines several sources and re-ranks them. Bing matters for Copilot by design. It is one input among several for ChatGPT.

Every engine searches with its own queries, not yours. Copilot sends Bing a few words derived from the prompt. Gemini and Claude write their own search queries. ChatGPT rewrites the prompt. The page that gets cited is the one that answers the query the engine actually ran. That mechanism, query fan-out, has its own guide: what query fan-out is and how to cover sub-queries.

You can see some of it. Bing Webmaster Tools added an AI Performance report in public preview on February 10, 2026. It shows citation counts, cited pages, and a sample of grounding queries across Copilot, Bing’s AI summaries, and “select partner integrations.” It does not show clicks, and it does not cover ChatGPT. No other engine in the table offers a site owner a comparable report as of November 2026.

Both, and they need different work. An answer written from the model’s training data reflects what was on the web, and in the training set, before the model’s cutoff. It usually carries no link. An answer grounded in live search reflects what the engine retrieves today, and it carries citations.

Engines search less often than most people assume, and the share depends on the product and the prompt:

  • Ahrefs found that 14.4% of 66.6 million free, logged-out ChatGPT responses in August and September 2026 contained fan-out queries, its measure of a response that ran a search. Ahrefs notes that logged-out use searches less than paid or reasoning use.
  • DataForSEO found that 47.5% of 100,000 ChatGPT prompts from its own database triggered fan-out queries, in a study published in May 2026.
  • Anthropic’s documentation says Claude searches for recent events, current prices, and “specific organizations, people, or products that might have changed,” and answers stable knowledge directly.

The practical split: questions about your company, product, price, and current offer tend to trigger a search, so live retrieval is where you win or lose them. Category definitions and settled concepts are often answered from training data, so what you published in the past, and what others wrote about you, has already shaped the answer. Blocking a training crawler affects future models only. Blocking a search crawler affects today’s citations. The robots.txt guide lists every token.

Which pages get cited for B2B buying queries?

Comparison pages, more than any other content type in the best evidence available. Siege Media analyzed 116 B2B Google Analytics properties, 1,112 comparison pages, and 739,492 AI-referred sessions over the 90 days ending May 24, 2026. Versus pages had a Spearman correlation of 0.65 with AI search traffic, which Siege calls “double the next-best content type.” Sites with 21 or more comparison pages had 900% more median AI sessions than sites with one to five.

Read it with its limits. The study measures referral sessions, not citations. It is a correlation across sites, and sites that build many versus pages also tend to have more brand demand. Siege says content explains roughly a third of the variation, and that brand pull and category fit drive most of the rest. Four of the top 20 sites by AI traffic had fewer than five comparison pages.

The reason the pattern holds is the claim, not the template. A buyer asking “X vs Y for a 200-person team” needs specific, checkable differences: price model, limits, integrations, who each tool is not for. A versus page that states those with dates is a page of reusable claims. A versus page that says your product wins on every row is not, and engines have the rival’s page and three review sites to check it against.

What to put on a B2B comparison page:

  • The comparison criteria a buyer uses, as a table, with each cell a fact rather than a checkmark where possible.
  • Where the other product is the better choice. This is the sentence that makes the page credible to a reader and to a model.
  • Prices, plan limits, and the date you last checked them.
  • A link to the source for each claim about the competitor.

How should you write sections so an engine can lift them?

Write each section so it answers its heading without the paragraph above it. That means a heading phrased as the question, the answer in the first sentence with the number or condition in it, and the evidence in the next sentence. The page stays one page, written for a person.

This is a live debate, and the evidence supports a narrow position. Google says “there’s no requirement to break your content into tiny pieces,” and Danny Sullivan said in January 2026 that Google does not want people crafting content for Search specifically. Ahrefs’ Despina Gavoyannis argued in August 2025 that “chunk optimization isn’t actually an SEO tactic,” because “you can’t control how Google, ChatGPT, or Perplexity chunk your content,” and recommended “self-contained sections that answer a query clearly and can stand on their own.” Mike King of iPullRank replied in January 2026 that retrieval works on passages, and showed that splitting one paragraph that mixed two topics raised its similarity score for one topic from 0.541 to 0.645 in his own tool.

Our position, and the reasoning behind it:

  • Both sides recommend the same page. Ahrefs and iPullRank both end at self-contained sections with clear headings. The disagreement is whether to engineer the boundaries further.
  • The pro-chunking evidence measures similarity, not citations. A higher cosine score on one paragraph in one tool is a reason to keep each passage on one topic. It is not evidence that a given chunk size gets cited more. We have not found a published controlled test that ties chunk size to citation rate.
  • You do not control the chunker. Each engine splits pages its own way and changes it with models. Formatting for one engine’s token window is a bet on an implementation you cannot see.

So: one topic per section, the subject named in the section rather than “it” or “this tool,” the answer first, the evidence next to it, and normal paragraphs. Do not split a guide into many thin pages, do not repeat the question in every sentence, and do not write sentence fragments for a machine.

Why does original data get cited?

Original data gets cited because the engine cannot find the claim anywhere else. When ten pages restate the same advice, the engine can cite any of them, or none. When one page holds the only measurement of something, that page is the source.

What counts as original:

  • A benchmark or survey with a stated method and sample.
  • Your own pricing, specs, limits, and policies, stated in HTML text.
  • A public dataset or changelog with dates.
  • A method note: how you measured, what you excluded, what the result does not mean.

Two rules make original data citable. Put the number and its method in the same passage, so a lifted sentence still carries who measured it and when. And publish it in HTML. If the only copy of a claim is a PDF or an image, expect the distributor who pasted the table into a web page to become the cited source.

Does it matter what other sites say about you?

Yes, often more than your own pages for comparison and recommendation queries. Engines check common claims against several sources, and they cite review sites, forums, publications, and reference sites heavily when the question is which vendor to choose. Ahrefs’ factual-versus-comparison split above is that effect in numbers: your own site’s share of citations drops when the buyer asks how you compare. Earned mentions, review profiles, and consistent facts on the platforms engines already trust are a separate workstream, with its own evidence and its own volatility. It is covered in why AI cites what others say about you.

Should you let AI crawlers in?

Let in the crawlers for the engines you want to be cited by, and decide the rest on purpose. Access is a decision, not the default that shipped with the template.

Training crawlers and search crawlers are separate tokens with separate effects. Blocking GPTBot keeps your pages out of future OpenAI training sets and does not remove you from ChatGPT search. Blocking OAI-SearchBot does remove you. The same split exists for Anthropic (ClaudeBot and Claude-SearchBot) and Perplexity. Google-Extended does not control AI Overviews. A CDN setting can block a search bot before robots.txt is read. Write down which behavior you intend for each token, then check your server logs. The robots.txt guide has the full table.

Two other access problems cost citations quietly. Most AI crawlers do not run JavaScript, so content rendered only in the browser is invisible to them; see JavaScript and AI crawlers. And llms.txt is a map, not a ranking factor: it can point an agent at canonical pages, but it does not compensate for a thin page. See what llms.txt does.

Count all three, separately, and add accuracy.

  • Mention rate: the brand appears in the answer.
  • Citation rate: the brand’s page is attached as a source.
  • Link rate: the citation includes a working URL to your page.
  • Accuracy: what the answer says about you matches what is true today.

A program that celebrates mentions and never earns links is doing brand monitoring. That is useful, and it is a different engagement. Third parties move the link rate: a standards page, a documentation URL, a publication that cites a number you published.

Mentions and citations also behave differently over time. In Ahrefs’ October 2026 tracking of about 37,500 citations across five engines, 50% to 70% of cited websites appeared on only one of 20 days, compared with about one in five brand names. Ahrefs concluded that mentions make a steadier KPI. They are steadier. They are also not the same outcome, because a mention without a link sends nobody, and a citation can point to a rival’s page that describes you.

Being cited for a claim you do not stand behind is worse than absence. Log the generated sentence next to the current offer and flag mismatches. Share of answer is the metric that holds these together, and measuring AI visibility is the method.

What can’t you control?

Most of the outcome. You control whether a page is retrievable and what it says. You do not control which queries the engine runs, which retrieved pages it reads, which passage it lifts, or whether it cites anything at all.

The variance is large and measured:

EngineCited sites gone the next dayAnswers with no citations
Google AI Overviews50%0.9%
Google AI Mode58%0.5%
ChatGPT59%0.7%
Gemini81%64%
Copilot83%53%

Source: Ahrefs, October 29, 2026, the same questions asked daily for 20 days. Ahrefs notes that after excluding answers with no citations, roughly 50% to 60% of citations dropped out by the next cited answer on every engine.

AirOps and Kevin Indig ran 16,851 queries through ChatGPT three times each in April 2026 and found that “only 2.3% of page-query combinations are cited in all 3 runs.” About 58% of retrieved pages were never cited.

Three consequences:

  • One check proves nothing. A screenshot of a citation is one sample from a distribution. Run each query several times and report a rate.
  • Model changes move results without any change on your site. OpenAI, Anthropic, Google, and Microsoft change retrieval and models often. Date every measurement.
  • Nobody can guarantee a citation. Google says no third-party tool has access to its AI systems, and no engine sells organic citations. Treat a vendor guarantee as a reason to look elsewhere.

Worked example: rewriting a comparison page

This example is hypothetical. It describes a made-up B2B vendor, not a client, and it reports no results. It shows the edits, not an outcome.

A compliance-automation vendor has a page titled “Us vs. Competitor.” The page has a hero banner, a 12-row table of checkmarks in which the vendor wins every row, and a pricing note that says “contact sales.” It is rendered client-side.

The audit, claim by claim:

ProblemWhy it failsEdit
Table rendered in JavaScriptMost AI crawlers do not execute it, so the table does not exist for themRender the table in the HTML response
Checkmarks instead of factsA checkmark is not a claim a model can quoteReplace with values: “SOC 2 Type II and ISO 27001; HIPAA on the Enterprise plan”
Vendor wins every rowContradicts the rival’s page and review sites, so it reads as uncorroboratedAdd a section: “When the competitor is the better choice,” with two specific cases
“Contact sales” pricingThe buyer’s most common follow-up has no answer on the pageState the price model and starting price, with “as of November 2026”
No sources for competitor claimsUnattributable, and possibly out of dateLink each competitor claim to its public source and date it
No owner or review dateNot attributableAdd who maintains the page and when it was last checked

After the edits, the page holds about a dozen claims an engine can lift and a reader can check. Whether it gets cited is a separate question, answered by running the target queries on a fixed panel of engines, several times each, before and after, and logging mention, citation, link, and accuracy.

Common mistakes

  • Treating “AI” as one engine. ChatGPT, Perplexity, Claude, Copilot, and Gemini retrieve from different systems. Check access and presence per engine.
  • Blocking the search crawler to stop training. Blocking OAI-SearchBot or Claude-SearchBot removes you from cited answers. Block the training token instead, if that is the goal.
  • Writing the category definition again. Restated advice has a hundred copies. Publish the measurement, the price, the limit, or the method nobody else has.
  • Claims without a source or date. A number with no “who measured it, when” next to it is hard to reuse. Put the attribution in the same passage.
  • Facts only in PDFs, images, or scripts. If it is not in the HTML text, an AI crawler cannot quote it.
  • Splitting a guide into many thin pages. One page per intent, with self-contained sections, beats a URL per question. Google treats the second as scaled content abuse.
  • A comparison page that never concedes. Say where the other product fits better. It is the most credible sentence on the page.
  • Reporting a single screenshot. Citations churn daily. Report rates over repeated runs, with dates.

Engines cite the clearest checkable claim they can retrieve. The work is making sure each engine can retrieve yours, and that yours is the one worth quoting. If you want that measured on your own query set, that is what our GEO work does.

Sources

Bring us the domain.

We will tell you where you are absent, which page is in the way, and what has to be true before a citation is possible.

Request an audit