Which external sites shape what AI cites about a brand?

Direct answer

Engines show only part of the sources they retrieve, and they publish controls to limit what content enters an answer. In the Observatory, one cut's answers rested on 3,033 distinct domains, and the first ten held 19.3% of the appearances. An external site's influence is measured by counting which answers it appears in, with the window and the model stated.

What the site's publisher controls

Google documents that, to appear in AI Overviews and AI Mode, a page has to be indexed and eligible to be shown with a snippet in Google Search¹. On top of that there are two explicit levers: the nosnippet directive prevents content from being used as a direct input for those features, and max-snippet limits how much may be used, with the exception of a publisher having separately granted permission².

The consequence for a brand is indirect and worth holding clearly. Those levers belong to whoever publishes each site, so a publication or a directory that restricts its snippet stops feeding the answer even while writing about the brand. An external site's influence rests on decisions that site makes, and falls outside the reach of the brand it covers.

What shows and what stays inside

OpenAI documents that its web search tool returns a field with every URL retrieved, wider than the citations that end up displayed³. The visible list is a subset: a domain can enter the answer's context and stay out of the citations.

A 2026 work separates those two things by name: a document's contribution to the answer, and its citation, which is what sends traffic back to the creator⁴. Measuring only what is cited understates influence; measuring only what is retrieved overstates the visible result.

On the measurement side, a first party already publishes it. Bing Webmaster Tools reports to a site owner the total citations, the daily average of cited pages, the grounding queries the AI used to retrieve the content, and citation activity per URL⁵.

How the citations split

The shape of the split matters as much as who leads it. In the Observatory, across 6,228 answers in the September 2026 cut, 6,064 arrived with at least one source. Those answers cited 3,033 distinct domains and added up to 19,184 appearances, counting one per answer.

Concentration is strong at the top and thins out fast. The most cited domain is gub.uy, the Uruguayan state's, which appears in 1,179 answers, 6.1% of the total. The first three hold 11.7%, the first ten 19.3% and the first fifty 36.0%.

And there is a very long tail: 1,229 domains appeared in a single answer. They are 40.5% of the cited domains and barely 6.4% of the appearances. A site entering one answer once is the typical case, and it moves nothing.

Behind gub.uy come reddit.com in 599 answers, mercadolibre.com.uy in 458, instagram.com in 355 and youtube.com in 324. The head mixes a government domain, a forum, a marketplace and two social platforms, surfaces where a brand rarely runs the page.

What the public measurements say

Third-party measurements come from companies selling AI visibility monitoring, so each figure travels with its origin. According to a study by Profound, a company that sells AI visibility analytics, across 680 million citations between August 2024 and June 2025, Wikipedia held 7.8% of ChatGPT's citations and Reddit 2.2% in AI Overviews⁶. According to a study by Peec AI, a company in the same business, across 30 million sources in five engines, the three most-cited domains were Reddit, YouTube and LinkedIn⁷. According to a study by Semrush, a company that sells SEO software, across more than 230,000 prompts between July and October 2025, those rates moved sharply week to week⁸.

The three panels are their own and incomparable between them. What they hold together, and what matches the Observatory, is the shape: a few surfaces concentrate a large part of the citations.

A 2026 position paper names the risk in that shape: concentrated influence from low contestability, and undisclosed commercial influence embedded in evidence and reasoning⁹.

What the cited site gains

Little traffic, and that deserves saying. The Pew Research Center, a research centre with no SEO product, measured across 68,879 searches by 900 US adults in March 2025 that a click on a source cited inside an AI summary happened on 1% of visits¹⁰.

On the infrastructure side, Cloudflare, a company that sells AI crawler blocking and a pay-per-crawl model, defines the crawl-to-refer ratio as the requests from a platform's agents divided by those arriving with that platform in the Referer header, and for the week of 19 to 26 June 2025 reported 70,900 to 1 for Anthropic¹¹.

How I measure it

I run a bank of unbranded prompts over the category, repeated across several models, and I store the cited domains per answer. From that I build two things: the list of domains ranked by how many answers they appear in, and the concentration curve, which is what says whether the head carries the category or whether it rests on many different sources.

I count one appearance per answer, so a domain cited three times inside the same answer weighs as one. And I drop the engines' own domains from the list: they appear through the shape of the answer, and say nothing about influence over the category.

The field's critical survey helps set expectations for what this figure allows: topical relevance and position within the context are the most reproducible findings, and generic heuristics transfer poorly across contexts¹². In the September cut I measured how much the head repeats across categories: 66 domains hold some top three, and 59 of them do it in a single category. The exception carries weight. The domain gub.uy enters the top three of 18 of the 35 categories, and the seven domains that repeat take 46 of the 105 places.

How to improve what other sites say about your brand

1. Start by learning which domains travel with your category's answers, before deciding where to invest. The list almost always mixes government, forums, marketplaces and social networks, and that mix tells you what ground your category is played on.

2. Look at the shape of the split before deciding: if ten domains gather 20% of the appearances, working many fronts suits you; if they gather 60%, concentrating on those does.

3. Review and correct your entry on the head domains that allow editing. It is the work with the best ratio of effort to result, because those pages already enter the answers without you having to earn the place.

4. When an outside page says something out of date about your brand, ask for the correction with the figure and the source at hand. It tends to work better than asking without evidence, and it leaves you a record of what you claimed.

5. Keep in mind that each site's publisher decides whether its content feeds the answers. If a publication limits its snippet, it stops feeding them even while speaking well of you, and that sits outside your reach.

6. Count one appearance per answer and leave individual citations aside. It is worth it because a domain cited three times inside one answer inflates the count and makes you believe it weighs more than it does.

Which AuraMetrics modules address this

See which sources AI cites when it answers about the category. Prompt Tracking shows, for each prompt and model, the sources the answer cited, and AI Market Leaders groups them by brand.

Audit the site's trust signals. Trust Auditor audits a site's trust signals: authors, policies, certificates and inbound links.

Review the site's technical signals with the GEO Score and the GEO Framework. The GEO Score scores the site from 0 to 100 on the four pillars of the GEO Framework.

Frequently asked questions

¿Cómo se analiza la influencia de los sitios externos en lo que la IA cita sobre una marca?

What can be measured is which of the category's answers each domain appears in, with the window and the models stated. What informs is the curve: a category where ten domains hold 20% of the appearances is worked differently from one where they hold 60%.

¿Los enlaces entrantes influyen en las citas de la IA?

Inbound links and AI citations are two different things and get measured separately. A 2026 work is a reminder that what sends traffic back to the creator is the citation, and not the influence over the answer's text⁴.

How I measured it

Observatory of brands in AI: a fixed bank of unbranded prompts across 35 categories in Uruguay, three repetitions per model, querying the real ChatGPT, Gemini and Google AI Mode products. September 2026 cut: 6,228 answers, 6,064 with at least one source, 3,033 distinct domains and 19,184 appearances, counting one per answer.

Declared scope: domains are normalised to the registrable domain, so a subdomain counts with its site, and the engines' own domains are excluded. The Observatory covers Uruguay, so the head of the list reflects that market.