Where ChatGPT Gets Its Recommendations for AI Tools
Amit Malvi
Founder, SearchAxe. Former Senior SEO Manager, Invideo AI
Published Updated 12 min read
Quick answer
Where ChatGPT gets its recommendations for AI tools comes down to two inputs: what the model learned from public web pages before its training cutoff, and the pages it retrieves when it searches. Published studies show the cited pages are led by best-of lists, review platforms, publishers and reference sites, alongside the vendor's own landing pages. Reddit is read often but rarely cited, and the mix shifts with the type of prompt.
On this page
- Where ChatGPT gets its recommendations for AI tools, in short
- Which channels feed ChatGPT search, and which ones get credited?
- Which kinds of pages does ChatGPT cite for software and AI tool prompts?
- How much of what ChatGPT cites can an AI startup influence?
- Does the source mix change with the type of prompt?
- What should an AI tool startup do about each source?
- What the published research cannot tell you yet
Key takeaways
- ChatGPT recommends from two inputs, its trained memory and live search, and in Chatoptic's consumer-category test, turning browsing on barely changed which brands appeared.
- Pages arriving through general web search were cited 88.46% of the time they were retrieved in Ahrefs' 1.4 million prompt study; Reddit's feed was cited 1.93% of the time.
- Best-of lists, review platforms, reference sites, publishers and vendor landing pages make up most of what ChatGPT cites for software.
- Only 32.3% of ChatGPT's 1,000 most-cited pages were ones a marketer could pitch or influence, according to Ahrefs.
- No published study isolates AI tool categories yet, so check your own prompts before acting on any dataset.
Where ChatGPT gets its recommendations for AI tools is a question about two inputs: what the model absorbed from the public web before its training cutoff, and the pages it retrieves when it searches. Published citation studies show the cited pages are led by best-of lists, review platforms, publishers and reference sites, alongside a handful of vendor pages. If you run growth or the company at a seed to Series B AI tool startup and want to know which pages to be on, this is the map, built from third-party research and OpenAI's own documentation. Perplexity, Claude, Gemini, Google AI Overviews, Google AI Mode and Copilot each draw on their own mix. This article covers one engine only.
Where ChatGPT gets its recommendations for AI tools, in short
ChatGPT draws on memory and on search, and for an AI tool shortlist the search side leans heavily on best-of lists and review platforms. OpenAI's help article on how its models are developed says training uses information that is publicly available on the internet, information from third-party partners, and information from users, trainers and researchers, and that it does not intentionally collect from sources known to be behind paywalls. So a tool that lives mainly in gated communities, private Slack groups or paywalled newsletters leaves little trace in the model.
Search adds pages on demand. OpenAI's search help page says ChatGPT ranks search results using multiple factors intended to find relevant, reliable information, that placement is not guaranteed, and that a site becomes eligible by allowing OAI-SearchBot and OpenAI's published search IP addresses. The two inputs mostly agree on which brands to name. Chatoptic's September 2025 study ran 1,000 queries across five consumer categories with browsing on and off. Brand overlap with Google moved only from 61% to 62%, and brand order in the two modes correlated at 0.463, which Chatoptic described as far more consistent than ChatGPT's agreement with Google. The order can still shift between modes.
OpenAI states the blend outright for shopping. Its shopping documentation, written for physical products, lists the inputs it considers: structured metadata from first-party and third-party providers, other third-party content, and model responses generated before it considers any new search results. We have not seen an equivalent OpenAI document for software, but the pattern in the studies below fits the same two-input picture.
Which channels feed ChatGPT search, and which ones get credited?
General web search does almost all the crediting. In Ahrefs' April 2026 study of 1.4 million ChatGPT prompts, URLs that came through the search channel were cited 88.46% of the time they were retrieved, against 12.01% for news, 1.93% for Reddit, 0.51% for YouTube and 0.40% for academic sources.
That split matters for where you spend effort. Reddit threads shape what ChatGPT reads about your category, but a thread rarely shows up as the link a buyer sees. The pages that get credited are the ones that surface through ordinary web search, which is why Bing and Google indexing still sit under every ChatGPT tactic. Ahrefs also found that cited pages had titles closer in meaning to ChatGPT's own sub-queries than pages it retrieved and dropped. How to take part on Reddit without wasting time is covered in why AI engines cite Reddit.
Find out which AI engines recommend you
A free consultation on where you appear today across AI engines and search.
Schedule a Free ConsultationWhich kinds of pages does ChatGPT cite for software and AI tool prompts?
Five kinds of page dominate: third-party best-of lists, review and listing platforms, reference sites, publishers, and the vendor's own landing pages. None of the studies below isolates AI tools as a category, so read them as the closest evidence available for software, not as AI-tool data. Semrush's July 2025 study gives the broadest count: 50% of links in ChatGPT 4o responses pointed to business or service websites.
| Page type | What the evidence shows | Study |
|---|---|---|
| Best-of lists, including vendors' own | "Best X" blog lists were 43.8% of cited page types across software, product and agency prompts | Ahrefs, 750 prompts, 26,283 source URLs, December 2025 |
| Review and listing platforms | G2 was 1.1% of all ChatGPT citations and 6.7% of citations among its top 10 sources; TechRadar was 0.9% of all citations | Profound, 680 million citations, August 2024 to June 2025 |
| Reference sites | Wikipedia was ChatGPT's most cited source in the same dataset, at 7.8% of all citations | Profound, as above |
| Publishers and wire services | PR Newswire, Forbes and Medium gained the most ChatGPT citations after a mid-September 2025 shift | Semrush, 230,000 prompts over 13 weeks, November 2025 |
| Vendor's own pages | For software, first-party citations showed general landing pages in 37.2% of prompts, the vendor's own best list in 34%, the homepage in 15.6% and documentation in 7.6% | Ahrefs, December 2025 |
| Business and service sites overall | 50% of links in ChatGPT 4o responses pointed to business or service websites | Semrush, 500+ digital marketing topics, July 2025 |
Two cautions sit under the table. First, the shares move: Semrush's November 2025 study found ChatGPT cited Reddit in close to 60% of responses in early August 2025 and Wikipedia in roughly 55%, falling to around 10% and under 20% after mid-September. Second, Profound's figures are shares of all citations across every topic, so a 1.1% share for G2 is large for one domain and says nothing yet about your sub-category. The Ahrefs best-lists study is the closest to software buying prompts, with examples such as "best CRM software for enterprise".
How much of what ChatGPT cites can an AI startup influence?
Less than a third of the most visible citations, on one count. Ahrefs' October 2025 analysis of the 1,000 pages ChatGPT cited most in September 2025 found only 32.3% were pitch-worthy or otherwise influenceable; the rest were reference pages, homepages, app store listings and similar pages outreach cannot reach. The same study found 28% of those pages had zero organic keywords, so a page's Google or Bing traffic is a poor filter for deciding which publishers to approach.
The share that comes from editorial coverage is contested. Muck Rack's May 2026 "What Is AI Reading?" report, covering more than 25 million links from ChatGPT, Claude and Gemini across 17 industries, put earned media at 84% of citations and journalism at 27%. PR Daily argued in August 2026 that the 84% bundles Wikipedia, academic, government and third-party corporate pages under one label, and that journalism alone is closer to a quarter. The journalism figure is the one to plan around. Muck Rack sells AI visibility monitoring, as do Ahrefs, Profound and Semrush, so treat every dataset here as directional.
Does the source mix change with the type of prompt?
Yes, and for an AI tool the prompt type decides which of your pages and placements matter. Shortlist prompts lean on lists and platforms, brand prompts lean on your own site, and trend questions lean on publishers. The evidence is uneven, so the table marks where it is thin. Seer Interactive's February 2025 study found Bing results for "top" and "best" questions skewed to affiliates, news and aggregators.
| Prompt type | Example | Sources the evidence points to | How strong the evidence is |
|---|---|---|---|
| Shortlist | "best AI video editor for YouTube creators" | Third-party best-of lists, review platforms, vendor landing pages | Strong: Ahrefs December 2025; Seer February 2025 found Bing's results for "top" and "best" queries skewed to affiliates, news and aggregators |
| Brand facts | "what does [your tool] do and how much is it" | Your own site; we have not seen a study of brand-fact prompts, though Ahrefs found that when ChatGPT recommends a brand on shortlist prompts, other pages on its site are more likely to be cited | Thin |
| Category trends | "how are AI agents changing customer support" | Journalism and press releases; Muck Rack found (across ChatGPT, Claude and Gemini) trend queries drew journalism citations at more than double the rate of how-to questions, and press releases 3.5 times more often than in best-of answers | Moderate: Muck Rack May 2026 |
| News | "latest AI music generator launches" | News channel, where Ahrefs found younger pages win when relevance is similar | Moderate: Ahrefs April 2026 |
| Comparison and alternatives | "[competitor] alternatives for small teams" | We have not seen a published source-mix study for these prompts | Thin |
Product categories behave differently again. In the Ahrefs best-lists study, landing pages showed up often for software and agency prompts but rarely for physical products, where a product page appeared in 87.2% of prompts that cited the brand's own site. If your AI tool is sold like an app, plan for the software pattern.
What should an AI tool startup do about each source?
Work through the sources in order of how much control you have, starting with your own pages and ending with editorial coverage. Each step targets one row of the tables above.
Getting onto the sources ChatGPT reads for your category
- Open the door for OAI-SearchBotOpenAI says a site becomes eligible for search answers only when OAI-SearchBot and its published IP ranges can reach it, so check robots.txt and your CDN rules first.
- Get indexed by Bing and GoogleCredited pages come through the general search channel, so product, pricing, use-case and docs pages need to be indexed; Bing Webmaster Tools and ChatGPT covers the Bing side.
- Build one landing page per use caseLanding pages were the most common first-party page ChatGPT cited for software vendors in Ahrefs' study (37.2% of prompts), so give each job your tool does its own page with a title that matches how buyers phrase it.
- Collect the lists ChatGPT already usesRun your shortlist prompts with search on, save every cited roundup and pitch the authors; the outreach is in getting an AI tool into best-of lists.
- Fill in the review and listing platformsG2 and similar sites are cited on their own and feed the lists; how review sites influence AI recommendations explains which profiles matter.
- Earn trade coverage for trend questionsTrend prompts lean on journalism, so a launch, dataset or opinion piece in the publications your buyers read feeds the answers that shape the category.
- Describe the product the same way everywhereThe model's memory is built from repeated public descriptions, so one category sentence on your site, profiles and press kit gives it a consistent picture to learn.
Steps one to three are in your hands within weeks. Steps four to seven depend on other people and take months. This sequence is the core of ChatGPT SEO work, and a team with time can run most of it in-house. If ChatGPT still leaves you out after the first three steps, why your brand does not show up in ChatGPT walks through the checks.
What the published research cannot tell you yet
We have not seen a published citation study limited to AI tool categories. The large datasets mix software with consumer products, agencies and general questions, and the categories in Chatoptic's study were car insurance, credit cards, hotels, online courses and web hosting. The datasets also come from companies that sell tracking tools, and the mix of sources shifted sharply at least once in the period they cover.
The fix is to check your own category. Run the prompts your buyers ask, save every cited page and tag it with the page types above. How to analyze which sources AI engines use to recommend competitors has the worksheet, and how ChatGPT decides which software to recommend explains what happens between retrieval and the final list.
Frequently asked questions
Does ChatGPT get its recommendations from training data or from the web?
Both. OpenAI says its models learn from publicly available internet content, partner data and information from users and trainers, and ChatGPT searches the web when a question benefits from current information. In Chatoptic's September 2025 study, turning browsing on barely changed which brands ChatGPT named, though their order shifted.
Which websites does ChatGPT cite most for software recommendations?
"Best X" blog lists, including vendors' own self-promotional lists, are the largest group, at 43.8% of cited page types in Ahrefs' December 2025 study of 750 prompts. Review platforms such as G2, reference sites such as Wikipedia, publishers and the vendor's own landing pages follow. Profound's dataset of 680 million citations found G2 at 1.1% of all ChatGPT citations. Shares shift over time, so check the sources cited for your own prompts.
Does Reddit influence ChatGPT recommendations if it is rarely cited?
It appears to. Ahrefs' April 2026 study of 1.4 million prompts found ChatGPT pulls Reddit through a dedicated feed at high volume, yet cited those pages only 1.93% of the time. Reddit shapes the context ChatGPT reads, while the credited link usually goes to a page from general web search. That makes Reddit useful for reputation and wording, but a weak route to a visible citation.
Can an AI startup pay to be recommended by ChatGPT?
Not in the answer itself. OpenAI's shopping documentation says product results are selected independently and are not ads, and that ads are separate from those results. For software, the routes in are organic: crawlable pages, Bing and Google indexing, inclusion in lists and review platforms, and coverage in publications. No published study we have seen measures whether paid list placements change ChatGPT's answers.
How do I find which sources ChatGPT uses for my AI tool category?
Write down the prompts your buyers ask, from shortlist questions to brand and trend questions, and run each one with search on, ideally logged out. Save every cited URL and tag it as a list, review platform, publisher, reference page, community thread or vendor page. Repeat over several weeks, because citations change, and prioritize the sources that recur and name competitors but not you.
