What is an AI citation?

The term describes an observable fact. When a generative system answers a question, it sometimes accompanies the text with links to web pages, and each of those links is a citation. The unit is the URL inside an answer, and that is where the boundary with AI visibility runs, since AI visibility works with the brand as its unit and with a set of repeated questions.

No standard defines what counts as a citation. A visible link below the paragraph, a card in the sidebar, the name of a site mentioned without a link, and an icon that expands the sources when tapped are different formats coexisting in the same interface. Every product presents them its own way and changes them without notice. Anyone stating a number of citations is applying their own counting criterion, and that criterion is rarely published.

Two questions that often get mixed up belong apart. The first is whether a page appears as a source, which is what this concept covers. The second is how much presence a brand has across the whole set of a model's answers, which is an aggregate analysis with different methods and different limits. This glossary treats them separately because the data available to answer them is not the same.

In the generative features of Google Search, the answer does not come from a single query. The system breaks the question into several related searches across subtopics and different sources, and composes the text and the links it shows from that pool. That is why the list of cited pages can differ considerably from the classic results for the same query. The generative summary and the conversational mode also use different models and techniques, so they do not match each other either.

To be a candidate, a page has to be indexed and eligible to be shown with a snippet in Search. Google states that there are no additional requirements and no special optimizations for appearing in these features, and that the nosnippet, data-nosnippet, max-snippet and noindex directives still control what may be used. That is a statement from the provider, not a rule verifiable from outside.

The awkward part is measurement. Search Console gained a dedicated report for generative features that shows impressions by page, country, device and date. It includes no clicks, no CTR, none of the queries that triggered the appearance, and no position. Those impressions were already counted in the previous totals, so the report separates out a figure that already existed rather than adding a new one. Outside Search, conversational assistants publish no equivalent report for publishers. In practice this means you cannot count how many times a page has been cited; you can establish that appearances happened and which pages they concentrate on.

The first decision that depends on this is what you report and with what degree of certainty. A team promising a number of monthly citations is promising something no available source delivers today. A team reporting impressions in generative features, per page, and labelling them as such, is saying something the client can verify in their own account.

The second decision is where to invest in content. Because the system breaks the question into subtopics, a page that resolves one narrowly framed matter has more chances of entering one of those partial searches than a long page that touches many topics in passing. This gives content structure a direction without guaranteeing anything, since the final selection is made by a model whose behaviour is neither stable nor auditable from outside.

The third is about control. Blocking snippets reduces the chance of appearing as a source, and that block also affects classic results if applied carelessly. There is also a limited test of a setting that lets a site stay out of the generative features without losing positions in Search. Whoever activates it gives up impressions and any visits that might come from them, and that decision should be made with data at hand rather than out of generic caution.

• Review the Search Console generative features report page by page and cross-check it against the classic performance of those same URLs. If the report has not appeared in your property yet, say so in the report instead of replacing it with an estimate.
• Check that the pages you want as sources are indexed and carry neither nosnippet nor a restrictive max-snippet, because a page without a snippet is out before any other consideration applies.
• Build each page around one narrowly framed question, with the answer in the opening lines and the detail below, so that the usable excerpt does not depend on reading the whole text.
• Keep dated screenshots of the answers where your site appears, along with the exact question and the product used. Without that record there is no way to reconstruct later what was actually shown.
• Keep the verifiable details of the page current, such as dates, figures and authorship, because an outdated source is easy to discard and hard to win back.
• Note for each observation which product and which date it belongs to, since interfaces change without notice and a three month old screenshot may describe a format that no longer exists.

• Presenting a citation count as if it were measured data. No public source delivers it today, and the number always comes out of someone's own criterion.
• Confusing impressions in generative features with clicks. The report carries no clicks, and inferring them from the variation in total traffic mixes causes that cannot be separated afterwards.
• Reading one isolated answer as evidence of a trend. The same question asked again may cite other pages without anything having changed on the site.
• Comparing the source list of a generative summary with that of a conversational mode as if both measured the same thing, when they answer using different techniques.
• Applying nosnippet to whole templates to protect content, then discovering that the snippets in the classic results disappeared as well.