What ChatGPT cites when buyers ask about AI visibility: 45 runs, 324 citations
When a buyer asks ChatGPT about your category, the answer is built from a handful of web pages the model chose to read. If you want to be in the answer, you need to know which pages those are. So we asked.
On 22 October 2026 we sent 15 questions a B2B marketer might ask about AI search visibility to OpenAI's Responses API with web search turned on, three times each. That gave us 45 answers, 324 citations and 73 distinct domains. Here is what came back.
How we ran it
- Model and tool: gpt-5.5 through the OpenAI Responses API, with the built-in web_search tool and the user location set to the United States.
- Questions: 15, in four groups: choosing a tool (5), choosing an agency or service (2), how-to questions (5) and concept or budget questions (3). The full list is at the end.
- Repeats: each question ran three times, independently, on the same day.
- What we counted: every URL the model attached to its answer as a citation. We grouped domains by type by hand: AI visibility vendors (including Semrush and Ahrefs, which sell AI visibility tools), platform documentation from OpenAI and Google, agencies, media, academic papers and community sites.
One caveat matters more than the rest. The API is not the ChatGPT app. The app can use a different model, the user's memory and history, and its own search settings. Published analyses of the consumer app often list Reddit and review sites among the most-cited domains, and we saw none. Treat our numbers as what OpenAI's search stack does with a clean, logged-out question, not as a screenshot of what your buyer sees.
The headline numbers
| Question type | Runs | Citations per answer | Where citations went |
|---|---|---|---|
| Choosing a tool | 15 | 9.5 | 88% vendors' own sites |
| Choosing an agency | 6 | 7.0 | 48% agency sites, 33% Google docs |
| How-to | 15 | 6.7 | 90% OpenAI and Google documentation |
| Concepts and budget | 9 | 4.2 | 47% platform docs, the rest spread thin |
Across all 45 answers, vendors' sites took 41% of citations, OpenAI and Google documentation 38%, agencies 8%, media 3% and academic papers 2%. The most-cited domains by number of answers were developers.google.com (16 of 45), help.openai.com (14), semrush.com (14), peec.ai (14), developers.openai.com (11), otterly.ai (9) and tryprofound.com (9).
Finding 1: for tool questions, the model reads your pricing page
When the question was "which tool should I use", the model went straight to the vendors. Of the vendor pages it cited, 47 were product or home pages, 32 were pricing pages and 32 were help or documentation pages. Vendor blog posts were cited three times.
The search queries explain why. The model issued 468 search queries across the 45 runs, and 37% of them used a site: operator, for example site:peec.ai pricing or site:tryprofound.com. It decided which vendors to consider first, then went to their sites to check plans, prices and features. The answers quote those details back: plan names, monthly prices, prompt limits and which engines each tool covers.
The brands named most often across the 15 tool answers were Profound and Otterly (15 of 15), Peec AI (14), Semrush and Ahrefs (12 each), Scrunch (11) and AthenaHQ (7).
What this means for a B2B SaaS company: your pricing page and help center are citation targets, not just conversion pages. If prices sit behind "contact sales" and features live in a PDF, the model has nothing to quote, and it will quote a competitor who published theirs.
Finding 2: for how-to questions, your blog competes with OpenAI and Google
For questions like "how does ChatGPT decide which websites to cite" or "how do I get cited in Google AI Overviews", 90% of citations went to help.openai.com, developers.openai.com and developers.google.com. These answers also cited fewer pages, often one to three.
That is a hard place for a vendor blog. A generic "how to rank in AI search" post is competing with the platform's own documentation, and the model prefers the source. Posts that can win here add something the docs don't have: original data, a specific test, a number nobody else published.
Finding 3: agencies get cited for saying exactly what they do
For "best GEO agencies for B2B SaaS", the model built its shortlist mostly from the agencies' own service pages. The ones it picked said it plainly: an AI-search or GEO service page, written for SaaS or B2B tech, often with a published starting price. Several answers quoted those prices directly.
Positioning that reads "we help ambitious brands grow" gives the model nothing to match. A page titled "Generative engine optimization for B2B SaaS" with a price on it does.
Finding 4: no Reddit, no G2, no YouTube
Zero of 324 citations went to Reddit, G2, Capterra, YouTube, LinkedIn or Quora, and none of the 468 search queries targeted Reddit. Eight answers did tell the reader to work on Reddit presence, but none cited a thread.
We don't read this as "community sites don't matter". The consumer ChatGPT app behaves differently, and Perplexity and Google AI Overviews are often reported to cite community content far more. It does mean that in OpenAI's API search for this category, the official and vendor web did the talking.
Finding 5: one check is noise
The same question asked three times shared on average 42% of its cited domains and only 19% of its exact URLs. For "what is the difference between AEO, GEO and SEO", the three answers had almost nothing in common. Only the head-to-head comparison, "Profound vs Peec AI vs Otterly", returned the same three domains every time.
If you checked ChatGPT once, saw a competitor and panicked, or saw yourself and relaxed, check again. Visibility is a rate across repeated runs, not a yes or no.
What to do with this
- Publish pricing and plan limits as plain HTML. It is the page tool answers quote most after the homepage.
- Write help-center pages that answer feature questions in one paragraph each. The model checks them to confirm what you do.
- Name your category the way buyers do, on the page itself. "AI visibility tracking for B2B SaaS" beats a slogan.
- Skip generic how-to posts. Publish something the documentation doesn't have, such as your own measurements.
- Measure in repeats. Run each important question several times and report the share of answers that mention you.
- Watch for utm_source=openai in analytics. 67% of the URLs the model cited carried that tag, so clicks from ChatGPT-style answers are visible in your referral data.
What we did on brandseen.app
Our pricing is on the homepage in plain text: a $299 per month pilot for 90 days. Our FAQ is server-rendered and marked up with FAQ schema, and robots.txt explicitly allows OAI-SearchBot and the other AI crawlers. We will rerun this exact question set each month and add the results here.
The 15 questions
- What are the best tools to track how my brand appears in ChatGPT?
- Best AI visibility tracking software for B2B SaaS companies in 2026
- Which tools monitor brand mentions in Perplexity and Google AI Overviews?
- What are the best alternatives to Profound for AI search visibility?
- Profound vs Peec AI vs Otterly: which AI visibility tool should I choose?
- Best generative engine optimization (GEO) agencies for B2B SaaS
- Should I hire an agency or use a tool for AI search optimization?
- How do I get my SaaS product recommended by ChatGPT?
- How does ChatGPT decide which websites to cite in its answers?
- How to get cited in Google AI Overviews for B2B software queries
- Does llms.txt help with AI search visibility?
- How long does it take for ChatGPT to start citing a new web page?
- What is the difference between AEO, GEO and SEO?
- Is generative engine optimization worth it for an early-stage B2B SaaS startup?
- How much does AI search optimization cost for a SaaS company?
Questions people ask
Does the ChatGPT app cite the same sources as the API?
Not exactly. The app can use a different model, the user's memory and its own search settings. The API shows how OpenAI's search behaves on a clean question, which makes it repeatable, but it is not identical to what a signed-in user sees.
Why did vendor sites get so many citations when other studies say third-party sites dominate?
Because of the question type. When a buyer asks which tool to choose, the model checks each vendor's own pricing and feature pages to confirm details. Broader studies that mix many question types report lower vendor shares.
How many times should I run a prompt to measure AI visibility?
At least three times per question per engine. In our data, repeated answers to the same question shared only 42% of cited domains.
Does ChatGPT cite Reddit?
Not in this test: zero of 324 citations. Published analyses of the consumer ChatGPT app and of Perplexity report Reddit much more often, so check each engine you care about separately.
Sources
- Web search tool guideOpenAI
- Overview of OpenAI crawlersOpenAI
- Does llms.txt work? What the logs say, and why we still ship oneBrandseen, 22 October 2026
- Why ChatGPT recommends your competitor: 7 causes we check in every auditBrandseen, 22 October 2026