Where AI answers get their sources: 36 answers to 12 B2B buying questions, analysed (November 2026).

Key findings.

Where do ChatGPT, Gemini and Google’s AI Overviews get their sources when a buyer asks for a supplier? In our sample, mostly from company websites: about 91% of all citations pointed to a company or vendor site. Directories and review platforms were about 2%, and forums and communities about 1.5%.

The practical reading: your own site is the source engines lean on most, so its pages must answer the buying question directly. Which off-site sources matter depends on the engine and the sector. The sample is small, so treat these figures as a dated snapshot. We run the same kind of source check for clients of our GEO and AEO services.

How we ran the study.

We put 12 buying questions, one from each of 12 B2B and local-service sectors, to three AI engines on 10 November 2026, from the United States, in English. That gave 36 answers. We recorded the domains each answer cited and sorted them with fixed rules.

The engines

We used Google AI Overviews (the AI summary at the top of a normal Google results page), ChatGPT and Gemini, as their consumer interfaces return them. The answers were collected through the DataForSEO API: its SERP API for Google results, including the AI Overview and its reference links, and its LLM Scraper, which DataForSEO describes as extracting “structured data from AI answers as they appear in supported interfaces”, for ChatGPT and Gemini. Google AI Mode was not part of this test; we cover it in our Google AI Mode guide for B2B.

The questions

Each question is phrased the way a buyer shortlisting suppliers would type it: some national, some local, some for regulated or accredited services.

What we counted as a citation

A citation here means one distinct domain cited by one answer. If an answer linked to three pages on the same website, that website counts once for that answer. For AI Overviews we used the reference links attached to the overview; for ChatGPT and Gemini, the source list returned with each answer. Google’s own documentation says AI Overviews give “a jumping off point to explore links to learn more” and may use “query fan-out”, running several related searches to gather supporting pages.

How categories were assigned

Every domain went into one of five categories by a fixed list of domain rules:

• Directories and review platforms: named review and listing sites, such as Yelp, software-review sites and adviser directories (NAPFA, the fee-only adviser network, sits here).
• Forums and communities: Reddit, Quora, YouTube, LinkedIn and similar social or community sites.
• Publishers and media: named news, finance and technology publishers, such as NerdWallet, WSJ and PCMag.
• Associations, standards and government: .gov and .edu sites, accreditation bodies such as A2LA and ANAB, and most .org domains, such as the Exit Planning Institute and CPA Academy.
• Company or vendor sites: everything that matched no rule, in practice mostly businesses’ own websites and blogs.

A second rule flagged a company-site page as a list or comparison page when its web address contained “best”, “top-”, “vs”, “versus”, “alternative”, “compare”, “comparison” or “review”.

What this study cannot tell you.

This is a small, dated snapshot: what three engines cited for 12 questions on one day, not how they behave in general.

The study can still show patterns large enough to survive these limits: the weight of company websites, the gap in answer size and how little the engines’ sources overlap. The raw data, including every cited URL, is available on request for anyone who wants to check or extend the work; we do not publish the full file because it lists individual businesses.

Company websites dominate the sources.

All three engines cited company and vendor websites far more than anything else: 86.5% of citations in AI Overviews, 85.7% in ChatGPT and 97.8% in Gemini. Across all 36 answers, the figure was 91.1%.

Much AI visibility advice starts with off-site sources: reviews, directories, Reddit, the press. In these answers, the engines mostly sent buyers to businesses themselves, or to other businesses writing about the category.

What kind of company pages get cited

Three page types come up again and again in the cited addresses:

• Service pages that answer the question directly. For the dental IT question, many cited pages were IT providers’ pages built specifically for dental practices, rather than general home pages.
• City and location pages. For the Chicago courier and Houston calibration questions, many cited pages were city-specific: a Chicago courier company’s own Chicago service-area page, or a calibration firm’s Houston lab page.
• Vendors’ own “best” and comparison articles. Software companies comparing their own category (see chapter 08).

Gemini’s 97.8% needs a caution: it cited several articles from sites our rules could not place, so part of its gap over the other engines is a product of our rules.

The other 9%: who else gets cited.

The non-company sources were few but not random. ChatGPT leaned on associations and regulators in regulated sectors, AI Overviews alone cited forums and video, and Gemini cited almost nothing outside company sites.

ChatGPT: the credential checkers

ChatGPT cited an association, standards or government source in 4 of its 12 answers (9.1% of its citations), in sectors where buyers check credentials:

• For the fee-only financial adviser question, it cited the SEC’s adviser search (adviserinfo.sec.gov), NAPFA, a certification board’s planner search, NerdWallet’s adviser finder and a WSJ guide to adviser disclosure forms.
• For the Houston calibration question, it cited the A2LA accreditation directory and ANAB, the accreditation body.
• For exit planning, it cited the Exit Planning Institute’s “find a CEPA” directory.
• For CPE courses, it cited CPA Academy, which AI Overviews also cited.

For a regulated or accredited business, the body that certifies you may be one of the few non-company sources an engine checks. Keep that listing complete and consistent with your site.

AI Overviews: the only engine citing communities

Forums and communities made up 8.1% of AI Overview citations, in 3 of the 11 overviews. Reddit was cited twice: a remote-work subreddit thread for the team-building question and a tax professionals’ thread for the CPE question. A YouTube video was cited for the fee-only adviser question. Neither ChatGPT nor Gemini cited a forum, community or video site in any answer.

Directories and publishers: rare everywhere

Directories and review platforms were 2.0% of all citations: Yelp once in AI Overviews (Chicago courier), a directory in 2 ChatGPT answers and 1 Gemini answer. Publishers were 1.5%: ChatGPT’s NerdWallet and WSJ citations, and a PCMag CRM roundup in Gemini.

Our guide to the off-site sources AI engines use to recommend B2B companies explains how to earn a place in reviews, directories and lists. This study adds a caution: for many questions, those sources were not cited at all.

How many sources each engine cites.

Gemini cited the most sources per answer, at 7.4 on average, followed by ChatGPT at 6.4 and Google AI Overviews at 3.4. An AI Overview gives each cited source a bigger share of a smaller answer.

Every ChatGPT and Gemini answer returned content. Google showed an AI Overview on 11 of the 12 questions; the exception was the local physical therapy question, where Google showed normal results without an overview.

Why the count matters

Fewer sources means fewer places. With about three per overview, a good answer can still miss out. With seven there is more room, but each source carries less weight.

The engines rarely agree on sources.

For the same question, the three engines mostly cited different websites. Across the 12 questions, 179 distinct domains were cited, and only 21 of them (11.7%) were cited by two or more engines for the same question.

Only three questions had any domain cited by all three engines: team building, dental IT and the calibration lab question, with one shared domain each. The CPE courses question had the most agreement, with 5 of its 16 domains cited by two or more engines. The business intelligence software and physical therapy questions had none.

What low overlap means

• Each engine is its own audience. A citation in one does not carry over, so a one-engine check misses most of the picture.
• The source pool is wide. For most questions there is no fixed set of sources to break into.
• Measure several runs. Answers also vary between runs. Our guide on how ChatGPT decides which companies to recommend explains why.

List and comparison pages carry real weight.

A sizeable share of company-site citations were pages with “best”, “top”, “vs”, “alternative” or “comparison” in their address. Gemini leaned on them most: 33 of its 87 company-site citations. In AI Overviews it was 9 of 32, and in ChatGPT 10 of 66.

Most were written by businesses in the same category: a software vendor’s “best business intelligence tools” article, a CPE provider’s “best online CPE courses” guide, a dental IT firm’s “best dental IT support companies” list.

Which questions drew the most list pages

Gemini’s citations for CPE courses and small-business CRMs were mostly list pages. List pages were rarer for the Chicago courier and Houston calibration questions, where engines mostly cited the businesses’ own service and location pages.

A word of caution

This is not a case for lists that rank yourself first with no evidence. A comparison that is fair, dated and specific to one buyer’s situation helps a reader; one written only to name its author the winner does not. Say who each option suits, show the criteria and include cases where another option fits better. The play on honest comparison pages covers how.

What this means for a B2B company.

Start with your own website, the source type engines cited most. Then add the off-site sources your sector’s engines actually use, and test in more than one engine.

A checklist for the next 30 days

• Write down the 10 to 20 questions your buyers type before they shortlist suppliers.
• Run each in AI Overviews, ChatGPT and Gemini, and save the cited sources with the date.
• For each question, check whether you have a page that answers it directly. If not, that is the first page to write.
• List the association, regulator, accreditor and directory profiles you hold, and correct any that disagree with your site.
• Re-run the same questions in a month and compare several runs. If that becomes routine, our comparison of AI visibility tools covers the options.

If engines know your company but leave it off shortlists, our diagnosis of why brands go missing from AI shortlists covers the usual causes. For the wider method, see our guide to generative engine optimisation (GEO). If you want help with the whole programme, our AI search visibility service starts with exactly this kind of source baseline.

The source self-check.

Tick each source type your company already has. The check shows how many of this study’s source types you cover, overall and per engine. Nothing you tick is sent anywhere.

The engine tiles count only the source types each engine cited in our 36 answers. A low score is not a verdict, just a list of places to look. If you plan to hire help for the gaps, see what AI search optimisation costs first.

Cite this study.

You are welcome to quote and chart these findings with a link to this page. Please keep the date and sample size next to any figure.

Raw data: the full table of answers and cited URLs, the question list and the categorisation rules are available on request through our contact page. We share it for research and checking, but do not publish it in full because it lists individual businesses.

Questions about the study.

Where does ChatGPT get its sources for buying questions?

In our November 2026 sample, mostly from company and vendor websites: 85.7% of ChatGPT’s citations. Associations, standards bodies and government sites were 9.1%, mainly for regulated services. It cited 6.4 sources per answer on average.

How do AI engines choose which sources to cite?

The engines do not publish detailed selection rules. Google says AI Overviews may run several related searches (query fan-out). Our data shows the outcome: engines mostly cited pages that address the question directly, often service, location or comparison pages.

Do Reddit and review sites matter for AI visibility?

Less than expected in this sample. Reddit appeared only in Google AI Overviews (2 of 11), and directories and review platforms were 2.0% of citations. Check your own questions before investing.

Why do Google, ChatGPT and Gemini cite different sources?

Each engine runs its own searches and selection, and their answers differ in size: 3.4 sources per AI Overview, 6.4 for ChatGPT and 7.4 for Gemini. Only 11.7% of cited domains were shared by two or more engines for the same question, so visibility in one engine says little about the others.

How reliable are these figures?

They are a dated snapshot: 12 questions, one run per engine, US English, rule-based categories. Use them as direction, not benchmarks; some figures would move on a repeat.

Sources: Prolinkage study, data collected 10 November 2026 via the DataForSEO API (SERP API; LLM Scraper), 12 questions × 3 engines, US English, rule-based categorisation. External: Google Search Central, AI features and your website. All accessed 10 November 2026.