We Asked 4 AI Models 800 Buyer Questions. Here's What They Cite.

Which sources an AI assistant cites depends mostly on which assistant you ask. On the same set of about 800 buyer questions, with the same instructions and web search on, Grok linked a Reddit thread in 97% of its answers, Gemini in 26%, GPT-5 mini in 15% and Claude Haiku in 2% (SlopMog's High-Value discovery runs, March to September 2026).

The threads they did cite were young and small, a median of 5 months old with 18 comments, and the four models rarely picked the same one. So check each assistant your buyers use on its own, and don't generalize from one.

How we asked 4 models 800 questions

SlopMog's High-Value discovery writes 40 buyer-style questions for a site, without the brand name, and asks each one to 4 AI models with live web search. Between Mar 10, 2026 and Sep 29, 2026, those runs asked about 800 questions and stored about 2,700 answers. The models, as their APIs name them: Gemini 3 Flash (preview), GPT-5 mini, Claude Haiku 4.5 and Grok 4.1 Fast. Grok stopped answering our calls after April, so its numbers come from about 280 answers to the March and April questions.

Every model got the same instructions, and they matter here. They ask for “links to sources” and to “cite as many relevant sources as possible including Reddit threads, YouTube videos, blog posts, and review sites.” So the rates below show what each model does when it's asked for those sources. Comparing the models with each other is fair, because they all got the same nudge.

An answer counts as citing Reddit here when its text links at least one Reddit thread (a link containing reddit.com/r/ and /comments/), and YouTube when it links a video. Separately, SlopMog saved about 1,260 Reddit threads and YouTube videos from the models' search citations after checking each one still existed.

Reddit and YouTube links, by model

Grok links everything. Its average answer held about 36 links, and about 1 in 7 of those pointed at a Reddit thread. Gemini 3 Flash linked a thread in 26% of its answers, and in about 130 more it linked a Reddit page that wasn't a thread, such as a subreddit. GPT-5 mini wrote the longest link lists, about 19 an answer, but only about 1 link in 70 went to a Reddit thread.

Claude Haiku 4.5 is the outlier the other way. It averaged about 12 links an answer, so it wasn't short of sources, yet it linked a Reddit thread in about 20 of 800 answers, all of them in March and April. From June to September it linked none in roughly 520 answers, on the same questions where Gemini linked one in 32%. That matches Otterly.AI's count of zero Reddit citations among 379,321 Claude citations in June 2026¹.

The order changed over the months

In March and April, GPT-5 mini linked Reddit threads more often than Gemini did. From June on, Gemini linked them more than three times as often as GPT-5 mini. The two periods hold different questions from different sites, so the swing can't be pinned on the models alone. What it does show is that a single month of checks would have ranked these two models the other way round.

Which questions pull in Reddit threads

Comparison questions pulled in Reddit threads most often. Gemini linked one in 35% of them and GPT-5 mini in 31%, against 20% and 9% for how-to questions. How-to questions went to YouTube instead, more often than any other type (14% of answers against 9% to 10%). The review row is the smallest sample in this post, fewer than 20 questions, so don't lean on it.

How often the models agree

Leave Grok out and agreement gets thinner. Across all the questions, Gemini, GPT-5 mini and Claude Haiku all linked a Reddit thread on under 1% of them, and at least one of them did on about 300 (37%). Thread by thread, 95% of the Reddit threads SlopMog saved came from one model's citations only, most of them Grok's, and almost none (0.1%) came from all four.

Same question, different sources

Each question here was asked once, so this measures disagreement between models. How much one model disagrees with itself from week to week takes repeat runs, so ask the same question several times per engine before you trust a result. The free way to do it is in tracking brand mentions in ChatGPT, Gemini and Perplexity.

What the cited threads look like

Small threads get cited. About half the threads had 10 to 49 comments, 31% had fewer than 10 and 2% had 200 or more, so a thread doesn't need to go viral to end up in an answer. Semrush saw the same in 248,000 Reddit posts cited by Google AI Mode, Perplexity and ChatGPT Search, 70% of which had fewer than 20 comments². In our data it was 54%.

Age is where our data parts ways with that. The cited threads were young: 20% were under 30 days old when we checked them and 74% were under a year. Semrush's cited posts averaged about 900 days², ours 336. Different engines, different questions and a different year, so the gap says more about setup than about any rule. It does show that an answer built from a live web search can cite a thread from last month.

The most common title was a plain question, with 43% carrying a question mark and 13% the word “best”. “Vs” (3%) and “alternative” (2%) were rare, even though comparison questions pulled in the most threads. The threads came from about 310 subreddits. The busiest 10 held 32% of them, and none held more than 7%.

The roughly 290 YouTube videos skewed small too. Of those with a view count, 62% had under 10,000 views and 3% had over a million. 27% had “how to” in the title and 19% a year. Getting mentioned under that kind of video is its own job, covered in YouTube comment marketing.

What this data can't tell you

• What the apps show. These are API models with web search on. The ChatGPT, Gemini, Claude and Grok apps your buyers open run their own search and settings. ChatGPT is the clearest case: in Ahrefs' data, about 1 in 5 of its answers cited Reddit in March 2026 and about 1 in 43 by late September³, which does ChatGPT still cite Reddit covers engine by engine.
• How a neutral question would do. Our instructions asked for Reddit threads and YouTube videos, which lifts both rates for every model. Wording moves these numbers a lot, as one practitioner's test of 748 answers found.
• What every category looks like. The questions were written for the sites added to SlopMog, our own test sites included, and three sites' runs supplied 70% of them. 28% mention software, a tool, an app or a platform. A mattress brand's questions could come out differently.
• What the newest versions do. SlopMog has since moved to newer Claude and Grok versions, so read those two columns as history.
• Whether ChatGPT's August drop reached the API. GPT-5 mini linked a Reddit thread in about 25 of its roughly 120 answers from August and September (21%). That's too few answers to say.
• Whether the thread traits hold beyond Grok. 95% of the saved threads date from March and April, when Grok supplied most of them. The roughly 190 Reddit threads that at least one other model cited looked much the same: 69% under a year old and 41% titled as a question.

What to do with these numbers

Check each assistant your buyers use on its own, with several runs per question. A Grok screenshot says nothing about Claude. For Reddit, go after the threads that answer comparison and problem questions in your category, and don't skip small ones: the median cited thread here had 18 comments. How to find the Reddit threads Google and ChatGPT already use has the free, step-by-step version, and mentioning your product in a comment that stays up covers what to write once you're there.

For assistants that barely cite Reddit, like Claude Haiku here, Otterly.AI found Claude leaning on brand and product pages instead¹. Being named on the pages those searches land on is the other half of the job, set out in how to get recommended by ChatGPT, Gemini and Perplexity.

Running those questions every week is the job of High-Value discovery. It asks your buyers' questions to Gemini, Claude, GPT and Grok with live web search, ranks every Reddit and YouTube thread they cite, and drafts comments in each thread's tone for you to edit, approve or skip. A High-Value comment is 60 credits ($6), and plans start at $99 a month.