How AI Assistants Recommend Support Software (Study)

When AI assistants recommend support software, about half of what they cite is best-of lists: 704 of 1,338 citations in our October 2026 tracking of ChatGPT, Gemini and Perplexity went to listicles. Vendors' own product pages, help docs and pricing pages drew another 481 citations (36%). Being cited, though, is not the same as being recommended: the engines read lists written by vendors and then name the third-party tools on them, mostly not the author.

This study looks at 193 answers to 20 prompts a SaaS support lead, product manager or founder types when shortlisting tools, and asks two questions: which pages the engines cite, and which tools they actually name. The short answer for a SaaS company that wants to be recommended: the pages an assistant reads about your product are your help center and your pricing page, so those are the pages to keep factual and current, and independent lists matter more than your own. Disclosure: HappySupport makes this study and is one of the tools in the data; our own numbers are reported the same way as everyone else's.

The methodology section lists all 20 prompts verbatim, the models and dates, what counts as a citation and as a mention, and the rules we used to sort pages by type. Every number below can be recomputed from the tracker exports.

Best-of lists get half of all citations

Listicles are the single largest source AI assistants cite for support software questions. In the tracker export of 19 October 2026 (1,338 citations of 625 pages), best-of lists account for 52.6% of citations. The smaller export that matches our answer sample (945 citations, 14 to 17 October) gives almost the same split.

Perplexity leans on lists even harder. Its answers carry every source inline, and 74 of its 93 citations (80%) point to lists; 27 of its 31 answers cite at least one.

The most cited lists are mostly written by vendors that sell a tool in the category. Of the five most cited lists, four come from a vendor: a small documentation vendor's in-app widget list (25 citations), GitBook's documentation tools list (24), HappySupport's in-app widget list (16) and the same small vendor's automated documentation list (10). The fifth is Dupple's list of AI documentation generators (11), from a publisher.

What this means: lists are where assistants build a shortlist, and vendor pages are where they check it. A buyer asking ChatGPT for "the best in-app customer support widget" gets an answer assembled from other people's roundups, then corrected against the vendors' own docs and pricing.

Cited is not named

Being cited by an AI assistant does not mean being recommended by it. The domains in the table below collect similar numbers of citations for very different reasons: the incumbents are cited for their own help docs, product pages and pricing, the challengers mostly for lists they wrote.

Remaining citations on each domain are other articles and head-to-head comparisons. HappySupport's flat-price list and its documentation tools guide fall under "articles" because their titles carry no list words (see the rules below).

ChatGPT cites vendor-written lists but names the third-party tools on them. The small documentation vendor's lists were linked as a source in 8 ChatGPT answers. Those 8 answers named Intercom, Zendesk and Help Scout 7 times each, Crisp and Chatwoot 6 times each, and the vendor that wrote the list 0 times. Across all 80 ChatGPT answers and 31 Perplexity answers, the vendor was named in none.

Gemini does name the author. The same vendor was named in 14 of 80 Gemini answers. Gemini answers carry no source links in the export, so we cannot tie each of those answers to a page, but the tracker lists Gemini among the engines citing the vendor's lists. In 11 of the 14 answers, Gemini describes the vendor in the terms its own list uses for itself: a tool for teams that ship fast (fast-shipping, fast-moving or shipping weekly).

HappySupport is the second example, and the numbers are not flattering. ChatGPT linked a HappySupport list as a source in 11 of its 80 answers and named HappySupport in 1 of them. Across all 193 answers, HappySupport was named 3 times: once each by ChatGPT, Gemini and Perplexity. The one Gemini mention came in a documentation tools answer, the topic of the only HappySupport guide Gemini cited; our widget list, our most cited page, was cited by ChatGPT only. The tracker also checked nine of the most cited pages not written by us, four of them lists, for the brands on each page; none mentions HappySupport.

GitBook and Mintlify look like counterexamples: they are cited mostly for their own lists and named in 34 of 38 documentation answers each. They also appear on the independent lists, though. Dupple's list, for one, covers both. A list you write about your own category can put you in front of the engine; it does not, on its own, make the engine recommend you.

Which tools AI assistants name for each type of question

AI assistants name a small set of incumbents for broad "best X" questions. Intercom is named in 129 of 193 answers (67%) and Zendesk in 107 (55%). The table counts answers naming each tool, per topic.

Narrow prompts are where smaller tools get named. For "which in-app support widget offers interactive click-through guides", Stonly (7 of 10), UserGuiding (6) and Userpilot (5) lead and Intercom drops to 3. For "As a solo software founder, what is the best AI support chatbot for my website", Chatbase is named in 11 of 11. For "What AI help center software has flat monthly pricing instead of per-seat fees", Crisp (5 of 9), HelpCanvas (4), help.center (4) and Gorgias (4) lead; HelpCanvas and help.center are small vendors whose pricing pages are among the most cited.

The engines also differ in how many tools they name. ChatGPT and Gemini name 5.0 and 5.5 of our 64 tracked tools per answer on average; Perplexity, whose answers are much shorter (median 56 words against 595 for ChatGPT), names 2.6.

AI assistants go to primary sources

AI assistants go to the vendor's own pages for facts. Perplexity logged 7 web searches for these prompts, and three of them explicitly ask for the official source:

The last two are URLs of list pages rather than search terms, which is the shortlist step again.

Help docs carry the incumbents. Zendesk's help center articles collected 50 of zendesk.com's 77 citations and Intercom's collected 47 of intercom.com's 107. The most cited are plain explainers: "Comparing the Zendesk Web Widgets" (13 citations), "Overview of Zendesk AI offerings" (10), "Fin AI Agent explained" (9) and Intercom's "Pricing FAQs" (8).

Pricing pages are where prices come from. ChatGPT put a price figure in 53 of its 80 answers, against 17 of 80 for Gemini and 4 of 31 for Perplexity, and at least 25 of those 53 ChatGPT answers link a vendor pricing page in the text. The most cited pricing pages belong to Intercom (16 citations), Crisp (10), Chatbase (10), help.center (9), HelpCanvas (6) and Help Scout (6).

Three pages in the data were written for AI assistants outright: two Chatbase pages and one Tidio page at an /llm-info address, with 5 citations between them. That is a small number, but it shows the engines will read such a page when it exists.

What this means for a SaaS company's own help center

Your help center and your pricing page are what AI assistants read about your product. Lists largely decide whether you make the shortlist; your own pages decide what the assistant says about you once you are on it. For support leads and product managers, four things follow from the data:

1. State facts in plain text. Prices, plan limits, what each plan includes and what the product does not do. Zendesk's most cited page is a comparison of its own widgets; Intercom's most cited help articles explain what Fin does and how pricing works. A fact that is only in a screenshot, a sales call or a "contact us" button is a fact the assistant fills in from someone else's list.
2. Keep help docs current with the product. Help docs are 12.6% of all citations and the largest share of what Intercom and Zendesk are cited for. An article that describes last quarter's UI is now quoted to buyers, not only to customers.
3. Publish comparisons that are fair. Your own list gets you cited; it rarely gets you named, because the engines take the third-party facts from it and pass over the self-recommendation. A comparison that states every tool's limits, yours included, is the version worth being quoted from.
4. Get onto independent lists. The most named tools appear on lists their makers did not write: Dupple's list covers GitBook and Mintlify, InstantDM's chat widget list covers Intercom and Zendesk. Narrow, specific questions (click-through guides, flat pricing, solo founders) are where smaller tools break through, so a clear, specific positioning helps more than a broad one.

HappySupport is built around the second point: a help center whose articles get update proposals from code changes in GitHub, GitLab or Bitbucket, which your team reviews before anything is published, so the pages an assistant reads match the product you ship. How that keeps documentation usable for AI is on the documentation for AI page, and the AI help center page shows the chatbot that answers your customers from the same articles, with citations. Git Handoff is on the Professional and Scale plans; getting onto independent lists stays manual work.

Methodology

The tracker. An external answer-engine tracker ran 20 fixed English prompts against ChatGPT, Gemini and Perplexity and recorded each answer, the tools it detected and the sources cited. We used two citation exports (17 and 19 October 2026) and one answer export (17 October 2026). We did not edit the answers or citations.

The 20 prompts, verbatim, with the topic we assigned and the most named tools:

Models, dates and sample sizes. 193 answers collected 14 to 17 October 2026: ChatGPT 80 and Gemini 80 (one answer per prompt per day), Perplexity 31 (1 to 3 per prompt, mostly on 15 and 16 October), plus one Claude answer and one Google AI Overviews answer to prompt 8, which count in the totals but not in engine comparisons. The 19 October citation export covers answers collected 14 to 19 October.

What "cited" means. A citation is a source the engine attached to an answer, as recorded by the tracker; page and domain counts come from the citation exports. For answer-level statements ("linked in 8 answers") we used the source links inside the answer text. Perplexity's text carries all of its sources, ChatGPT's carries some of them, so ChatGPT figures are a floor, and Gemini's carries none.

What "named" means. A tool is named in an answer when its name appears in the answer text outside source links. We matched 64 tools, including product names (Fin for Intercom, Lyro for Tidio, Agentforce for Salesforce). A mention counts whether the tool is recommended or used as a reference point.

How pages were classified, first matching rule wins: Reddit, Medium, dev.to and YouTube are community; arxiv.org, ft.com, businessinsider.com and two academic sites are news and research; a help., support., docs. or developer subdomain, or a /help/, /docs/ or /hc/ path, is help docs; an /llm-info path is a page for AI assistants; a /pricing path is a pricing page; "vs" in the title or URL without "best", "top" or "alternatives" is a head-to-head comparison; "best", "top", "alternatives", "ranked", "compared" or a count of tools ("9 best", "10 AI chatbots") in the title or URL is a list; blog, guide, library and resource paths are other articles; everything else is a product or home page. Lists whose titles carry none of those words count as articles, so the list share is a floor.

Limitations. Four days of answers, one tracker vendor, English prompts only, and a small Perplexity sample. Answers vary from run to run. Tools outside our 64-name list are not counted. We report counts, not causes: the data shows what the engines cite and name, not how they rank internally. HappySupport is a vendor in this category and its own pages are in the data.

Frequently asked questions

How does ChatGPT decide which software to recommend?

In our study, ChatGPT's recommendations follow the sources it reads: best-of lists drew 53% of all citations across the engines, and vendors' own product pages, help docs and pricing pages another 36%. ChatGPT put a price in 53 of its 80 answers and mostly named the third-party tools on the lists it cited rather than the lists' authors. We observe citations and mentions, not ChatGPT's internal ranking.

Do AI assistants trust vendor listicles?

They use them, but ChatGPT does not repeat the author's self-recommendation. A small documentation vendor's own lists were linked in 8 ChatGPT answers and the vendor was named in none of them; HappySupport's lists were linked in 11 and HappySupport was named in 1. Gemini behaves differently: it named the same small vendor in 14 of 80 answers, usually with the label the vendor's own list gives it.

How can a SaaS company get recommended by ChatGPT?

The data points to two levers. Get onto lists you did not write, because the tools named most often are the ones independent roundups cover. And make your own help center and pricing page state the facts plainly, because those are the pages the engines check: Intercom's and Zendesk's help docs alone collected 97 citations. Specific positioning helps too, since narrow prompts are where smaller tools get named.

Which support tools do AI assistants recommend most?

Across 193 answers, Intercom was named in 129 and Zendesk in 107. For documentation tools, GitBook and Mintlify were each named in 34 of 38 answers. For turning screen recordings into help guides, Scribe was named in 10 of 10. Narrow questions change the list: Stonly, UserGuiding and Userpilot lead for click-through guides, and Chatbase was named in 11 of 11 answers for solo founders.