What is an AI citation?
An AI citation is an explicit reference to a web page as a source inside an answer generated by an AI assistant or AI search engine, usually a link to a URL next to the sentence that page supports. It differs from a mention: an AI citation points to a specific page, while a mention names a brand or product without linking to anything.
Key takeaways
- An AI citation ties a piece of the answer to the URL it came from: it is a link to a page, not just a name.
- An assistant reading your page does not guarantee it cites it: the results it retrieves and the sources it shows are two different lists.
- A brand mention without a link cannot be clicked and does not show up as a visit in your analytics; a citation can bring the visit.
- Counting citations tells you how often you are used as a source, not where in the answer or with what weight.
When an assistant answers "which backpack should I get for my laptop?", your store can show up in two very different ways: as a link to your product page next to the sentence that uses it, or as a bare name in the middle of the text. The first is a citation; the second is a mention. Here "citation" has nothing to do with an academic reference or a court summons: it is the source the answer shows. A store should care about the difference because only a citation leads to a specific page, can be clicked and can be measured; it is the part of a generated answer that looks most like a search result.
How it works
An assistant with search follows three steps: it decides whether to search, retrieves results and writes the answer, indicating where each claim comes from. The vendors' technical documentation shows that last step in detail. In the Gemini API, the model analyzes the request and decides whether a Google search could improve the answer; the response includes the queries it ran and an annotated text in which each citation ties a specific span of the text to the source URL, marked by the position where it starts and ends.[1]Google AI for Developers (Gemini API). Grounding with Google Search Google gives the reason: showing the sources behind the model's claims builds trust with whoever uses it.[1]Google AI for Developers (Gemini API). Grounding with Google Search
Claude's search tool works in a similar way. Citations are always on when it searches, and each one carries the source URL, its title and up to 150 characters of the cited text.[2]Anthropic, Claude documentation. Web search tool Anthropic also requires that when an application shows those answers directly to its users, it includes the citations to the original source.[2]Anthropic, Claude documentation. Web search tool
One consequence of that structure is worth keeping in mind: retrieving a page is not citing it. In Claude's response, search results arrive as a list of their own, with the URL, title and date of each page, and the citations come separately, attached to specific sentences in the text.[2]Anthropic, Claude documentation. Web search tool A page can be among what was read and not appear among what was cited.
Search engines work much the same way. Google says AI Overviews and AI Mode show a broader and more varied set of links next to the answer than a classic search does.[3]Google Search Central. AI features and your website Bing, which already measures citations, defines them as the content shown as a source in AI-generated answers.[4]Microsoft, Webmaster Tools blog. Introducing AI Performance in Bing Webmaster Tools Public Preview And Perplexity explains that when someone asks it a question, its Perplexity-User agent may visit a page to answer and include a link to it in the response.[5]Perplexity. Perplexity Crawlers
Citation versus mention. A citation points to a URL of yours: it can be clicked, it can bring the visit, and it is recorded in the reports that count sources, such as Bing's, which gives citations per page.[4]Microsoft, Webmaster Tools blog. Introducing AI Performance in Bing Webmaster Tools Public Preview A mention names your brand or product without a link: it can weigh on the reader's decision, but it leads to no page, leaves no trace in your analytics and can only be counted by reading the answers one by one. Both matter; what you should avoid is counting them together as if they were the same thing.
In a store
One possible answer to the question above, with three products, each present in a different way. It is a fictional example: the store, the brands and the URLs are made up to classify the three cases.
| What appears in the answer (fictional example) | Type | Leads to your site |
|---|---|---|
| Urban 20 Backpack, linking to example-store.com | Citation of your page | Yes |
| "The Northfield brand also has a good reputation" | Mention without a link | No |
| Anti-Theft Backpack, linking to a review site | Citation of a third-party page | Only indirectly |
If the answer comes from an API like Claude's, the first row has this shape, with the span of text, the URL and the cited part of your page:[2]Anthropic, Claude documentation. Web search tool
{
"text": "The Urban 20 Backpack has a padded laptop sleeve",
"citations": [
{
"type": "web_search_result_location",
"url": "https://example-store.com/urban-20-backpack",
"title": "Urban 20 L waterproof backpack",
"cited_text": "Padded sleeve for laptops up to 15 inches..."
}
]
}
The second row has no equivalent: a mention carries no structure, which is why only someone reading the answer will find it. The third is the easiest to overlook, because it names you through someone else: the visit, if there is one, goes to the review site first.
What gets confused
Being cited does not tell you how much you weigh in the answer. Bing, which gives publishers a count of their citations, warns that its figures reflect how often your content is used as a reference, with no indication of its placement or presentation in the answer, and that they do not measure ranking, authority or a page's role within a particular answer.[4]Microsoft, Webmaster Tools blog. Introducing AI Performance in Bing Webmaster Tools Public Preview A citation at the end, among ten other sources, counts the same as the one that holds up the whole recommendation.
How to check
- Write a buying question your customer would ask, without your brand, and save it as is so you can repeat it the same way every time.
- Run it in each assistant in a new chat and note two separate lists: the URLs cited as a source and the brands named without a link. Repeat on several days: one answer is a sample of one.
- If your site is in Bing Webmaster Tools, open AI Performance: it gives citations per page and the grounding queries the AI used to retrieve your content.[4]Microsoft, Webmaster Tools blog. Introducing AI Performance in Bing Webmaster Tools Public Preview
- On Google, look at the generative AI performance report in Search Console, knowing that it counts impressions of links in AI Overviews and AI Mode, not citations.[6]Search Console Help. Generative AI performance report (Search)
- Work with a fixed set of queries. Repeated over the same questions every time, that count gives you a citation rate in a panel, and compared with your competitors' it becomes what is called AI share of voice.
Related concept: grounding
Grounding is the step before: connecting the model to search results so it bases its answer on them. In the Gemini API that is what it is called, grounding with Google Search, and the response records the queries the model ran.[1]Google AI for Developers (Gemini API). Grounding with Google Search Bing uses the same word for the phrases the AI used when retrieving the content that ended up cited.[4]Microsoft, Webmaster Tools blog. Introducing AI Performance in Bing Webmaster Tools Public Preview The practical difference is the one between what was read and what was shown: grounding is what the model consults; the citation is what it shows the person asking. A page can make it into the first and stay out of the second.
To keep going, the measurement and verification reading path lays out the rest of the terms, and the entry on GEO (Generative Engine Optimization) explains what you can do to give a page a better chance of being cited.
Synonyms
- English
- AI citation, LLM citation
- Español
- cita en IA, cita de fuente en IA
Sources
- [1] Google AI for Developers (Gemini API). Grounding with Google Search. Accessed . ↑ ↑ ↑
- [2] Anthropic, Claude documentation. Web search tool. Accessed . ↑ ↑ ↑ ↑
- [3] Google Search Central. AI features and your website. Accessed . ↑
- [4] Microsoft, Webmaster Tools blog. Introducing AI Performance in Bing Webmaster Tools Public Preview. Accessed . ↑ ↑ ↑ ↑ ↑
- [5] Perplexity. Perplexity Crawlers. Accessed . ↑
- [6] Search Console Help. Generative AI performance report (Search). Accessed . ↑
Change history
- First English version.
This definition is published under CC BY 4.0. Reuse this definition freely, crediting Clione with a link to this page.
Related terms
AI share of voice
What AI share of voice is, why it has no standard formula and what to ask before trusting the number from any tool or agency.
LLM attribution
What LLM attribution is, how Google Analytics detects visits from ChatGPT or Gemini, what ends up as direct or organic, and how to read it in your store.
Generative AI performance report
What Search Console's generative AI report measures: your site's impressions in AI Overviews and AI Mode, with no clicks or queries, and how to read it.
Grounding
What grounding is in AI, how it connects a model to a search engine's index, what Google-Extended controls and how to see if your store is used as a source.
Citation rate
What citation rate is in a query panel, how to build a repeatable panel for ChatGPT or Gemini, and what that number cannot tell you.
GEO
What GEO is in SEO, where the term comes from, what Google says about it and how to check whether ChatGPT, Perplexity or AI Mode cite your store.
Make your catalog easy for bots to read
Clione fills every product page with what bots read best: the JSON-LD graph with its structured data, the meta description and the FAQs. You review and press Push; then we check your live store so it all stays published.