Nearly half of ChatGPT answers cite ten or more domains: what that changes
Promptwatch's October 2026 data shows 47.6% of cited ChatGPT answers name ten or more domains. What it means for how you read a citation, and how to check your own prompts.
Promptwatch's data report on unique domains per response says 47.6% of ChatGPT responses that include at least one citation cite ten or more different domains. At least eight domains appear in 66.1% of them. Only 6% cite three domains or fewer, and 1% cite a single one.
For a client who asks "how do we get cited?", this number changes the question. A citation is rarely a win in a two-horse race. It's one seat at a crowded table.
What the report covers
The window is September 28 to October 4, 2026, in ChatGPT, limited to responses with at least one citation. That last filter matters: answers with no citations at all are outside the count. The report breaks the figure down by prompt type, and the differences are the useful part.
| Prompt type | Cite 10+ domains | Other figures |
|---|---|---|
| Organic | 52.5% | 71.8% cite at least 8; 4.5% cite three or fewer |
| Competitor comparison | 49.1% | 66.5% cite at least 8; 0.35% cite one domain |
| Brand-specific | 21.3% | 40.7% cite five or fewer; 13.8% cite three or fewer |
For brand-specific prompts, the most common count is five domains, at 14.9% of responses. So a question about you by name tends to draw on a smaller set of sources than a question about your category.
What it does not mean
It doesn't mean ten domains share the credit equally. The report counts distinct domains per response, not how much each one shaped the answer. A response citing twelve domains could rest mostly on two. Nor does it tell you how prominent a citation is, where it sits, or whether the reader clicks it.
It's also one week, one engine, and one vendor's dataset. We wouldn't assume Perplexity or Google's AI features behave the same, and we wouldn't assume other languages or markets look like this one.
And the report doesn't show causes. It describes how many sources answers use. It doesn't explain why.
How we read it for a client
Three readings, with the caveats attached.
Being cited is a shared outcome. If almost half of answers draw on ten or more domains, a response that cites you almost certainly cites your competitors, review sites and a forum as well. The report's own advice is to plan on sharing every answer. We'd add that "share of citations" is a better measure than "were we cited yes or no", because yes/no hides how crowded the answer was.
Brand prompts are a different game. With a typical count of five, a branded question has fewer slots. That puts more weight on which five. If two of them are stale third-party pages and one is your product page, you want to know that. Our note on why a competitor page owns the AI citation covers the version of this problem where the other source is a rival's.
Single-run checks mislead. With this many sources per answer, the set of cited domains will vary between runs. One screenshot of one answer shows one draw. The report recommends measuring across many organic prompts, and we agree. See our short piece on whether a single run of a prompt is enough.
What to do without a tool
Take ten prompts a buyer might ask in your category and run each a few times. For each answer, write down the domains cited and mark which ones are yours, a competitor's, a review site, a forum or a publisher. Then count.
A spreadsheet is enough. After a day or two you'll see patterns: the same three or four domains turning up in most answers, perhaps one competitor everywhere, perhaps a listicle you've never heard of. The report's suggestions follow from that list. Earn mentions on the third-party reviews, listicles and forums that already appear. Publish your own comparison pages where a comparison prompt is the target. Check which competitors share your answers, since they are the ones you're being weighed against.
Write down how the answer describes you, too. Being one of ten domains is a lesser problem than being described with a wrong price.
Measuring it properly
The spreadsheet stops being practical once you track more than a handful of prompts across several engines. That's where we move clients onto Promptwatch. It runs prompt panels daily per model, country and language, so the "many prompts, repeated runs" requirement is covered without anyone copying answers by hand.
Several features map to the report's advice:
- Citation analytics at domain and page level show which domains sit alongside you in responses, and how that mix moves over time.
- Position in Answer and Share of Voice turn "were we mentioned" into a number you can compare with competitors.
- Competitor visibility and heatmaps show which brands appear in the same responses as yours, and across which topics or models.
- Offsite mentions show your brand named on third-party pages without a link, which is the group the report's advice about reviews, listicles and forums points at.
- Sentiment and Brand Narrative cover the report's other suggestion, tracking how ChatGPT describes you. Brand Narrative is available on every plan, per the changelog.
The prompt types in Promptwatch (organic, brand, competitor comparison) line up with the three groups in the report, which makes it simple to check whether your own numbers look like the report's or differ.
A warning on targets
It's tempting to take 47.6% and build a goal around it. Don't. The figure is a description of how crowded answers are in one week of data. It can't tell you what your share should be, or whether getting from five cited domains to ten would help.
What it can do is reset expectations. If you're measuring success as "ChatGPT cites us and nobody else", the data says that's rare. Measure your presence relative to the others in the answer, run enough prompts that one odd response doesn't mislead you, and keep notes on the source mix. That's the whole job.