Where ChatGPT Gets Its Citations: What Domain-Level Data Shows

ChatGPT citations do not come from one fixed index of approved publishers. The source mix changes with the prompt, search behavior, available pages, and product updates. Domain-level data offers a way to observe that mix at scale, but it needs careful interpretation.

The useful question is not simply which domain is number one. A GEO team should ask which types of sites supply answers in its category, which domains are gaining or losing share, and whether the company's owned pages appear beside independent sources.

What the public domain report measures

Promptwatch publishes a live report called Which Websites Does ChatGPT Cite the Most?. We reviewed the page on August 30, 2026. It says the report draws on more than 100 million citations tracked across more than 3,000 websites, ranks 24 domains, and displays 40-day trends.

The underlying collection comes from actual AI product interfaces, according to the methodology on the page. The report aggregates non-identifiable citation patterns and updates continuously. That means its displayed percentages can change after this article is published. The source page, not a copied figure here, should be used for the latest reading.

This is broad, cross-topic data. It can reveal source behavior across a large response set, but it cannot tell a niche software company that the same domain mix appears for its ten purchase prompts. A market-specific prompt panel remains necessary.

Social sources are concentrated

The report separates social and community platforms from other domain types. In its displayed analysis, Reddit is the only social platform cited at material scale, while LinkedIn appears as a much smaller second source and several other social domains barely register in the aggregate view.

That does not mean every company should start a Reddit campaign. Domain share across all topics says nothing about whether the relevant subreddits contain accurate, current conversations about a particular market. It also says nothing about whether participating would fit the community.

What it does show is that "social content" is too broad a category for GEO reporting. ChatGPT may treat a detailed community thread differently from a short feed post or a video page. Buyers evaluating citation software should check whether the product preserves the exact URL and source type. A row labeled "social" discards the part an editor needs.

Social dependence also carries volatility. Promptwatch's separate Reddit citation decline study found an abrupt change in August 2026. Reddit averaged 3.83% of ChatGPT citations from July 18 through August 7, then averaged 0.52% from August 14 through August 17. The study describes that as an 86.4% relative drop and explicitly says the cause was not established. A data-collection issue could not be ruled out.

The lesson is methodological: a domain can look structurally important until the retrieval mix changes. Owned content and several credible offsite sources are safer than dependence on one community domain.

Review and reference sites play different roles

The live report also groups B2B software review sites, knowledge bases, and news publishers. Its category analysis points to review platforms such as G2 and Trustpilot, reference sources such as Wikipedia and arXiv, and product-oriented publishers such as TechRadar.

These sources do not perform the same job in an answer. A reference page may support a definition. A review platform may supply customer context for a vendor comparison. A buying guide can provide a ready-made shortlist. Looking only at the root domain misses the content format and the claim the answer borrowed.

For GEO work, inspect cited URLs in context:

1. Read the prompt that caused the citation.
2. Open the exact cited page rather than the domain homepage.
3. Identify the paragraph, data point, or comparison the answer used.
4. Note whether the answer named your brand, a competitor, or neither.
5. Check whether your own site has a better page for the same intent.

This turns domain research into an editorial brief. Without the response and page, a domain ranking is little more than a media list.

A citation domain is not an endorsed brand

ChatGPT may cite a company's article while never mentioning that company's product. It may recommend a company while citing an independent publisher. Those are separate events.

This is why domain citation share should not be labeled brand share of voice. The cited domain supplied a source. The answer text determines which brands received attention. If the report needs both, maintain separate measures and add an overlap view showing answers where the company was named and its page was cited.

The distinction also prevents a misleading competitive claim. A publisher with a large citation share is not necessarily competing for the same customer. It may be an intermediary that shapes the answer. That makes the publisher strategically relevant, but in a different way from a product named in the recommendation.

Source inventory changes the difficulty of winning

The number of available citations varies by engine. Promptwatch's live average sources per response report says ChatGPT answers with web search typically contain around five sources. Only ChatGPT responses where web search triggers are included in that figure. Google AI Overviews and Perplexity sit around ten sources on the same report, while Copilot is described as much more volatile.

Five average slots create a different contest from ten. A domain can have strong content and still miss a narrow source list. This is another reason to track rates over repeated runs rather than treating one uncited answer as a verdict on page quality.

For buyers, ask whether a tool records the number of sources in each response. A citation count without the source inventory cannot distinguish one slot won out of five from one won out of ten.

Turn the market view into a brand view

Start with a stable group of high-intent prompts. Track them by engine and locale. For each response, retain every cited page, the source domain, answer position, and whether the company was named.

Then compare three views:

• The broad public domain trend, which can flag shifts in source types
• Your category's domain set, which shows the publishers and communities relevant to buyers
• Your owned-domain performance, which shows whether company pages enter the answers

Review changes over a rolling period. A one-day jump can come from ordinary answer variation or collection volume. A sustained move across related prompts deserves page-level investigation.

Promptwatch provides domain and page citation analytics alongside prompt responses, citation trends, crawler logs, and visitor analytics. See our Promptwatch review for the product overview. If your team needs to move from a public market chart to the exact sources shaping its own buyer prompts, Promptwatch is our recommendation.

The most productive output is not a static list of top domains. It is a recurring source review that shows which outside pages influence your category, which owned pages get used, and where a retrieval change altered the mix.