An AI engine cites your page when it can fetch it and has reason to treat it as a source, and the engines differ on both counts. On October 28, 2026 we asked five of them how a B2B software company can get cited in ChatGPT and Google’s AI Overviews. Perplexity answered from Google’s and OpenAI’s own documentation and from nothing else, and ChatGPT cited the same two companies’ documentation beside two original studies. Gemini and Claude cited vendors’ posts, most of them how-to guides. Google’s AI Overview drew on a LinkedIn article, YouTube, a Medium post and several agency and vendor blogs, and did not cite Google at all, though the question was about Google’s own product.
One question, then, produced five reading lists, yet the advice on offer usually comes as a single list of tactics, which cannot be right for all five. On November 6, 2026 we read 20 pages that answer the question, the 8 ranking on Google for how to rank in AI Overviews and 12 from web searches. Of the 20, 12 name a single engine in their headings and 7 link to anything an engine’s owner has published. The owners, for their part, document what each engine searches and which fetchers it sends, and two of the five engines answered from exactly that documentation.
How the engines split in our run
Across ten questions the five engines had few sources in common. Their 50 answers carried 440 citations to 246 sites, and of the 246, 200 were cited by one engine only, 36 by two, 6 by three and 4 by four. No site was cited by all five.
The how-to question was one of ten that a buyer at a B2B software company might ask, all listed in our guide to generative engine optimization. We put the ten to ChatGPT, Perplexity, Gemini, Claude and Google’s AI Overviews on October 28, 2026, through DataForSEO’s answer endpoints, from the US. It was one run on one day, with questions we wrote and ten answers per engine. An endpoint need not return what a signed-in person sees in the app, and the exports do not record which model answered.
The engines differed in the kind of site they cited as well. Sorting the sites into kinds by hand, and counting a site once for each answer that cited it, we found an engine owner’s documentation only in ChatGPT’s answers and Perplexity’s, in 3 and 4 of their 10. Social platforms were cited in all 10 of the AI Overviews’ answers and in 3 of the other 40, and Perplexity went to GitHub in 5 of its 10. Gemini and Claude kept almost wholly to commercial sites (agencies, consultancies, software vendors, investors): 38 cases of 40 and 57 of 62.
None of this is where the usual advice points. Of the 20 pages we read, 10 name Reddit and 9 name Wikipedia, yet Reddit appeared in 3 of our 50 answers and Wikipedia in 1. Ahrefs, in a study of ChatGPT’s searches updated on October 22, 2026, reports that “ChatGPT’s Reddit citations dropped 88% in six months”, so the advice may describe last year’s engines. Two of this year’s, asked the question outright, cited documentation and original studies, which are sources and not summaries of them.
The table sets each owner’s account of what its engine searches beside what our run found, with a check for each.
| Engine | What it searches, by its owner’s account | What it cited in our run | The check |
|---|---|---|---|
| ChatGPT | Sites surfaced by OAI-SearchBot; partner search providers, sent rewritten queries | 58 citations; none in 3 of 10 answers | Ask with search on and open Sources; search Bing for the short query |
| Perplexity | Sites surfaced by PerplexityBot; Perplexity-User fetches a page on request | 193 citations; 104 sites, 75 cited by no other engine | Read the sources it lists; if absent, check the firewall |
| Gemini | Google’s Search index, for grounding, governed by the Google-Extended token | 42 citations; none in 4 of 10 answers | Confirm Google-Extended isn’t disallowed; see whether the answer cited anyone |
| Claude | Brave Search, TurboPuffer and Parallel Web Systems, by Anthropic’s subprocessor list; fetchers Claude-SearchBot and Claude-User | 63 citations; 60 sites, 38 cited by no other engine | Search Brave; ask Claude with web search on; confirm neither fetcher is blocked |
| Google AI Overviews and AI Mode | Google’s Search index; a page must be indexed and eligible for a snippet | AI Overviews only: 84 citations; LinkedIn in 7 of 10 answers, YouTube in 6 | URL Inspection; the page’s snippet controls; the Generative AI performance report |
ChatGPT
ChatGPT gave no citation on 3 of our 10 questions, and when it does search, the search may not be for your words. OpenAI’s help page on web search, as it stood on November 6, 2026, says ChatGPT search “sometimes partners with other search providers” and, when it does, “typically rewrites your query into one or more targeted queries”. Ahrefs calls these fan-out queries. Its study of 66.6 million free, logged-out responses, updated on October 22, 2026, found them in about 14% of responses, at about 1.6 each.
The fetcher behind ChatGPT’s search is OAI-SearchBot, which our guide to generative engine optimization lists beside the other engines’. OpenAI’s crawler overview, as it stood on November 6, 2026, says a change to robots.txt can take “~24 hours” to reach search results. Which crawlers to admit, and why, is a subject for a later guide.
On the how-to question ChatGPT gave five sources: two Google Search Central pages, an OpenAI help article and two original studies whose findings we have not checked. On the question of which agencies do this work it gave 26, lists of agencies among them. Ahrefs’ study of 1.4 million ChatGPT prompts from February 2025, updated on August 27, 2026, found that ChatGPT cites about half of the URLs it retrieves. Drawing on outside research into the pipeline, it adds that ChatGPT weighs a result’s title, snippet and URL before it opens the page.
To check, put your buyer’s question to ChatGPT with search turned on and open Sources. Then search Bing for the short query ChatGPT would be likely to write, and read your own title and snippet as it does, with nothing else to go on. Bing is only a proxy. OpenAI’s page names two search providers, Microsoft and Shopify, by linking their privacy policies, so Bing shows part of what ChatGPT may be handed and not all of it.
Our five citations fit a line in Ahrefs’ fan-out study: “ChatGPT seems to be moving from sites that summarize information to the sources that produce it”. That stands to reason, since a summary is something ChatGPT can write for itself. What it is short of is sources, and a page that sets down what a company knows and can show is one.
Perplexity
Perplexity cited more than any of the other four engines and ranged further. Each of its ten answers carried at least 15 sources, often several pages from one site, and its citations reached 104 sites, 75 of which no other engine cited. It cited GitHub in 5 of the 10 and research or reference sources in 3; ChatGPT did each once and the other three engines not at all. On the how-to question all 15 of its citations were Google’s and OpenAI’s own pages: eight pages of Google’s developer documentation, OpenAI’s crawler overview, and one OpenAI help article on search under six addresses. No vendor’s post was among them.
PerplexityBot crawls for Perplexity’s search results, and Perplexity-User fetches a page when a person’s question calls for it. Perplexity’s crawler page, last modified on October 29, 2026, warns that a site behind a firewall “may need to explicitly allow Perplexity’s bots”, and says robots.txt changes take “up to 24 hours” to register. Because Perplexity showed its sources on all ten of our questions, checking is quick: ask your buyer’s question and read the list. If your pages are absent, look at the firewall before the content, since a rule that turns away unfamiliar bots can keep both fetchers out even where robots.txt lets them in.
A README in a public repository or a page of documentation is as much a candidate here as a blog post, though neither is usually written to win a citation. Each records what a product does, in the words of someone who knows, and Perplexity treated pages of that kind as sources.
Gemini
Gemini was the likeliest of the five to cite nobody. Four of our ten questions drew no citation from it, more than from any other engine in the run, and in the six answers that carried sources it cited 40 sites, 38 of them commercial. Those on file for the how-to question are vendors’ how-to posts. The silences are not peculiar to our run: Ahrefs, in a study of citation stability published on October 29, 2026, found that only 4% of the questions it put to Gemini had citations on all 20 days it tracked.
Gemini’s grounding comes from Google’s Search index, by Google’s own account. Its page on common crawlers, updated on October 30, 2026, says the Google-Extended token governs model training and “grounding (providing content from the Google Search index to the model at prompt time to improve factuality and relevancy) in Gemini Apps”. The same token, it adds, “does not impact a site’s inclusion in Google Search”. On that description a site can rank on Google and still be withheld from Gemini by one line of robots.txt, perhaps a line somebody added to keep pages out of training.
Start the check in robots.txt, and confirm that it doesn’t disallow Google-Extended if you want your pages available for grounding. Before counting an absence against a page, look at whether the answer cited anyone at all, because often nobody is cited. Past that, Gemini grounds on the Search index, so a page that is in Google’s index is within Gemini’s reach.
Claude
None of the 20 pages we read links to Anthropic’s documentation, and none names Brave Search. Yet Anthropic’s list of subprocessors, as it stood on November 6, 2026, names three companies for “Web Search”: Brave Search, for all products, and TurboPuffer and Parallel Web Systems, each for all but Claude for Government. Claude’s fetchers, described on Anthropic’s help-center page dated November 4, 2026, are Claude-SearchBot, which indexes pages for search, and Claude-User, which retrieves a page when a person asks.
Claude was the steadiest of the five in our run. It cited between 4 and 7 sources on each question, 63 citations spread over 60 sites, 38 of which no other engine cited, and of the 62 times a site appeared in its answers, 57 were commercial sites. On the how-to question it cited seven vendors’ pages, and most of the six we have on file are how-to posts with the question’s own words in the title. We infer, from one run only, that a title stating the question a page answers is worth having. The rest of a cited page’s shape is described in our guide to generative engine optimization.
Begin with Brave, the one of the three whose entry on Anthropic’s list links to a search page, search.brave.com. Search it for your buyer’s question and see whether your page is in the results. Brave is a proxy for what Claude is handed, as Bing is for ChatGPT, and we have not measured how closely the two match. Then ask Claude the question with web search on and read its sources, and confirm that robots.txt blocks neither fetcher.
Google AI Overviews and AI Mode
To rank in AI Overviews a page needs nothing that ordinary search does not already ask of it. Google’s page on AI features, dated December 10, 2025, says a page “must be indexed and eligible to be shown in Google Search with a snippet”, the rule that our guide to generative engine optimization sets out. AI Overviews and AI Mode, the same page adds, “may use a ‘query fan-out’ technique”, so a page can be pulled in for a sub-question it handles well.
In our run a social platform appeared in all 10 of the Overviews’ answers, LinkedIn in 7 and YouTube in 6, and the LinkedIn pages on file are ordinary posts and Pulse articles. Google’s own documentation did not appear once, even on the question about AI Overviews. How far the Overviews’ citations overlap Google’s top ten is measured in our comparison of AEO and SEO.
Ahrefs’ list of the sites AI Mode cites most, refreshed on November 3, 2026 for all topics in the US, puts google.com first, then Reddit and YouTube. Our run measured AI Overviews on B2B questions on one day, and in it LinkedIn held the place among social platforms that Reddit holds on Ahrefs’ list.
Most of the check is in Search Console. URL Inspection confirms that the page is indexed, and the page’s snippet controls (nosnippet, max-snippet, data-nosnippet) limit what Google may show, so look for one set by accident. Google announced a Generative AI performance report on June 3, 2026 and has offered it to all sites since August 31, with impressions by page, country and device. The last check lies outside Search Console: whether a named person has posted the page’s argument on LinkedIn.
That last check is the one we would press. The LinkedIn pages the Overviews cited came under people’s names and under companies’, and ten answers cannot show that Google prefers either. We recommend the person’s post all the same, because a source is a page somebody stands behind: a named person who checked each claim against the product before it shipped.
Copilot, which we did not measure
Copilot was not among the five engines, so we have no figures for it. Bing’s Webmaster Guidelines, as they stood on November 6, 2026, say Bing and Copilot search experiences “rely on the same core crawling, indexing, and ranking foundation as traditional search”. Bing Webmaster Tools opened an AI Performance report in public preview on February 10, 2026, showing citations, the pages cited and the grounding queries behind them, and IndexNow tells Bing when a page has changed. Open that report, and confirm that your pages are in Bing’s index.
Other people’s pages, by engine
Mahe Chen, Xiaoxuan Wang, Kaiwen Chen and Nick Koudas reported in September 2025 that AI search strongly favors third-party, authoritative sources over brand-owned and social content, a finding our guide to generative engine optimization discusses. Our run adds that the third-party pages cited differ by engine. It also found Google’s Overviews citing LinkedIn and YouTube all the same, so the social rows below are our run’s result and not the study’s.
| Engine | Outside pages it cited in our run | The page to have there |
|---|---|---|
| Google AI Overviews | LinkedIn posts and Pulse articles (7 of 10 answers); YouTube (6 of 10) | A named person’s LinkedIn post; a video |
| Perplexity | Public repositories on GitHub (5 of 10 answers); research and reference sources (3 of 10) | A README and public documentation; research that shows its method |
| ChatGPT | Two original studies on the how-to question; lists on the agencies question | A study of your own, with an author and a method |
| Gemini and Claude | Commercial sites (38 of 40 and 57 of 62), seldom the same one twice | No single platform; the lists in the next row |
| All five, on the agencies question | Lists of agencies; the question drew 77 citations, the most of any | An entry on the lists the answers cite |
Start with the lists, since all five engines cited them when the question was which agencies to consider. Find the ones an engine cites for your category’s version of that question and pitch inclusion with facts you can stand behind. Buying a mention is among the things the same guide advises against, along with llms.txt and FAQ markup, though of the 20 pages we read, 12 still bring up FAQ markup and 5 llms.txt.
Why a single check proves little
A citation seen today may be gone tomorrow. Ahrefs, in a study published on October 29, 2026, followed about 37,500 citations across five AI tools for 20 days and found that “50–83% of citations are gone the next day, but about 75% of those come back within 20 days”. Our own figures are a single day’s. A page missing from today’s answer has not been dropped, and a page present in it has not arrived.
The same study supplies the remedy: “Count how many of the last seven days you were cited.” Put the same questions to each engine on seven days and keep one number per engine, the days on which your page was cited. On the days it was not, write down who was cited instead, because that list tells you what to do next. A spreadsheet will do, and none of it needs a framework.
Past a handful of questions the counting wants a tool. Which to use, and how far to trust a tool’s score, is the subject of a guide we are writing on measuring AI search visibility.
Does each engine need its own page?
No. Five engines read five different sets of pages, so a buyer’s picture of a company is assembled from its site, a LinkedIn post, a README, a list entry and a video. The pieces agree only if they were written from one source: positioning, product facts and terms set down once, in a repository the company owns, so that the list entry says what the pricing page says. No vendor can supply that, and it is the one piece that carries across all five engines.
Who signs each piece matters for the same reason. The two studies ChatGPT cited each carry an author’s name, and a page with a name on it has somebody answerable for its claims. We would have a person who knows the product check each claim before a page ships, and have the same person post the LinkedIn version.
On the question of how to get cited, the two engines that passed over the how-to posts landed on documentation and on original studies. Both are written by people who know the thing they describe, before any model is asked to write about it. A company’s part is to write down what it knows and can show, under a person’s name, and leave the summarizing to the engines.