# AI Citations: How LLMs Choose Sources to Cite

> Source: https://rankxai.com/blog/ai-citations · Last updated: 2026-09-02

An AI citation is a source an AI assistant links or names when that assistant answers a question. Retrieval decides which pages enter the model's context, so citations follow relevance to the assistant's own sub-queries rather than search rankings. Being cited and being named are separate outcomes, and most citations never name the brand.

## What is an AI citation?

An AI citation is a source that ChatGPT, Claude, Gemini, Grok, Perplexity or Google's AI Overview links, footnotes or names when it answers a question. An AI citation is the generative-search equivalent of a search result, with one difference that decides everything else on this page. The citation is chosen after retrieval has already narrowed the web to a small set of candidate pages. An AI citation therefore reflects what a retrieval layer found relevant to its own reformulated question, not what ranked for the question a person typed.

Citations arrive in two forms that are easy to conflate and worth keeping apart. Assistant citations are the sources cited inside answers from AI platforms like ChatGPT, Claude, Gemini, Grok and Perplexity. [AI Overview citations](/glossary/ai-overview) are the sources Google's AI Overview lists above a results page. They come from different engines with different collection methods and different cadences, which is why RankX AI reports them on [two separate screens](/docs/concepts/citations-two-kinds) rather than as one number.

## How do LLMs decide which sources to cite?

LLMs decide which sources to cite in four stages, and only the last one is visible. Retrieval systems behind ChatGPT, Gemini and Perplexity expand the question into synthetic sub-queries, retrieve candidate sources for each, score passages from those candidates against the sub-queries, and generate an answer from the passages that scored highest. A citation is attached to whichever passages survived. The [generative engine optimization guide](/blog/generative-engine-optimization) covers those retrieval mechanics in full; what matters here is which signals predict surviving citation selection.

That pipeline has a name: [retrieval-augmented generation](/glossary/rag), usually shortened to RAG. Large language models answering from RAG are handed retrieved passages at question time rather than relying on what they memorised in training, and the citation is the receipt for which passages the model was handed. This is the single biggest difference from traditional search and organic search ranking, where the ranking IS the answer. In an [answer engine](/glossary/answer-engine), ranking only decides which candidate sources reach the model.

The largest published measurement is Ahrefs' analysis of 1.4 million ChatGPT prompts from February 2025, run with data scientist Xibeijia Guan. Retrieval channel decides most of it. URLs from ChatGPT's own search index were cited 88.46 percent of the time, against 1.93 percent for Reddit URLs and 0.51 percent for YouTube. Those are citation rates per channel rather than shares of a total, and they do not sum to 100.

Two smaller findings from the same study point at what to write. Semantic similarity between a page title and the [fan-out sub-query](/blog/query-fan-out) scored higher than similarity to the prompt itself, 0.656 against 0.602. And descriptive natural-language URL slugs appeared on 89.78 percent of cited pages against 81.11 percent of uncited ones.

One result in that study is worth more than the other three because it is a null. Pages that were not cited were more likely to carry a visible publication date than pages that were, 49 percent against 33.79 percent, and Ahrefs drew no strong conclusion from it. Any advice promising AI citations in return for date stamping should be read against that number.

## Why an AI citation is not the same as a brand mention

A citation and a mention are different outcomes, and most writing on this topic quietly assumes they are one thing. A citation is a source link. A mention is the brand name appearing in the answer text a person actually reads. Semrush, working with Kevin Indig, logged 3,981 domain appearances across 115 prompts, 14 countries and four AI engines, published on 9 June 2026. The split was 61.7 percent citation with no mention, 25.1 percent mention with no citation, and only 13.2 percent both.

The per-engine numbers reverse the intuition most readers arrive with. In the Semrush data ChatGPT cited a domain 87 percent of the times it appeared but named the brand only 20.7 percent of the time. Gemini did close to the opposite, naming the brand in 83.7 percent of appearances while citing it as a source in 21.4 percent. A brand optimising only for source links is close to invisible on Gemini, and a brand counting only name-drops is close to invisible on ChatGPT.

The distinction is commercial, not academic, and it is the one most marketers get wrong. Roughly 1 percent of users click a source link at all, so for the other 99 percent the answer text is the whole impression. A citation carrying no [brand mention](/glossary/ai-mention) is an [unlinked brand mention](/glossary/unlinked-brand-mention) inverted: the credit arrives without the recognition.

## Are AI citations accurate?

Often not, and the distance between a confident AI citation and a correct one is the least discussed fact in AI search. The Tow Center for Digital Journalism at Columbia University tested eight AI search engines across 1,600 queries, feeding each one a direct excerpt from a real news article and asking it to name the headline, publisher, date and URL. The engines answered incorrectly on more than 60 percent of queries.

The spread means citation accuracy has to be read per platform, exactly as citation rate does. Perplexity was the most accurate of the eight with a 37 percent error rate. ChatGPT Search returned incorrect information on 134 of its 200 queries, 67 percent. Grok 3 was wrong on 94 percent. Across all eight engines the systems rarely signalled uncertainty, which is what turns a plain error into a confidently wrong answer a reader has no reason to question.

Two consequences follow for anyone doing citation tracking. A citation pointing at your domain is not proof the answer described your page correctly, so read the answer text rather than only the source list. And misattribution runs both ways: a competitor's claim can arrive attached to your URL, and a [hallucinated URL](/glossary/hallucinated-url) on your domain can send a reader to a page you never published.

## What 218 AI answers showed about our own AI visibility

RankX AI published its own 30-day baseline rather than only citing other people's, and the citation half of it is blunt. Across 218 analysed assistant answers in the 30 days to 21 August 2026, RankX AI was named in 44, an overall mention rate of 20.2 percent, with a spread from 11.4 percent on Perplexity to 23.3 percent on Grok. The [full baseline study](/blog/our-ai-visibility-baseline) carries the method and the per-platform table.

The AI Overview half is the part relevant here. Across 60 AI Overview checks on five tracked keywords, an Overview rendered on 59 of them, 98 percent of the time. RankX AI was named inside an Overview 12 times. RankX AI was cited as a source zero times. Every one of those 12 mentions came from the brand-name keyword, and on all four category keywords the count was zero.

Read those 12 as noise rather than visibility. The Overview on that brand-name keyword is about the RANKX function in Power BI, so a brand-mention count on a homonym measures the homonym rather than the brand.

That is the mention-without-citation case from the previous section, measured on our own domain rather than described from someone else's dataset. It is also the reason RankX AI does not report a single AI visibility number. A brand reading citations alone would have recorded a flat zero for August 2026 while being named a dozen times. A brand reading mentions alone would have missed that RankX AI pages were never once used as a source.

## Do AI citations follow Google rankings and traditional SEO?

Partly, and the relationship between traditional SEO and AI citations is loosening fast enough that any single percentage has a shelf life of months. The honest way to hold it is as a trend rather than a snapshot:

| Measurement | Surface | Source |
| --- | --- | --- |
| Top-10 share of AI Overview citations fell from 76 to 38 percent in seven months | Google AI Overviews | Ahrefs, 863,000 SERPs |
| 88 percent of citations sat outside the organic top 10 | Google AI Mode | Moz, about 40,000 queries |
| 12 percent of assistant-cited URLs ranked in Google's top 10 | ChatGPT, Gemini, Copilot, Perplexity | Ahrefs, 15,000 queries |
| About 11 percent of domains were cited by both engines | ChatGPT against Perplexity | Profound, 100,000 prompts each |

The reconciliation is that Google AI Overviews sit on Google's ranking stack, Google AI Mode cites pages that rank for sub-queries rather than for the visible query, and ChatGPT retrieves through an index of its own. Three things follow. Being retrievable somewhere an engine looks is the floor. A top-10 position for the head term is neither required nor sufficient. And treating AI search as one channel is a category error when ChatGPT and Perplexity share roughly a ninth of their cited domains.

## Do AI citations actually drive traffic?

Citations pay even though almost nobody clicks them, and the two halves of that claim were measured separately by Seer Interactive across more than 3,100 queries between June 2024 and September 2025. Take the discouraging half first: organic click-through rate on queries carrying a Google AI Overview fell 61 percent, from 1.76 percent to 0.61 percent, and paid click-through rate fell 68 percent.

The second half is what pays for the work. On those same queries, brands cited inside the AI Overview earned an organic click-through rate of 0.70 percent against 0.52 percent for brands that were not cited, a 35 percent difference. On paid the gap was wider: 7.89 percent against 4.14 percent, a 91 percent difference. Being cited does not restore the clicks a Google AI Overview removes. Being cited moves a brand to the better side of a pool that is shrinking either way.

RankX AI therefore treats an AI citation as an impression rather than a visit. The click is the smaller prize and the brand named in the answer text is the larger one, which is why a team measuring only analytics sessions will conclude AI search does not matter while a competitor is being described to every buyer who asks.

## What has been tested and does not get you cited

Three widely sold tactics have been tested by someone who added them to real pages and measured the difference against controls, and all three failed. This section exists because almost every competing page on this topic still recommends at least one of them.

- **Schema markup as a citation lever.** Ahrefs tracked 1,885 pages that added JSON-LD between August 2025 and March 2026 against about 4,000 matched controls. Google AI Mode moved 2.4 percent and ChatGPT 2.2 percent, both statistically indistinguishable from zero, while AI Overviews fell 4.6 percent. Four separate statistical approaches agreed. searchVIU tested the same question from the other end and found no engine extracted a price present only in JSON-LD, because retrieval pipelines strip script tags before embedding. Schema still earns rich results and helps entity resolution. It does not buy citations.
- [**llms.txt**](/blog/llms-txt)**.** Ahrefs' server-log study across 137,000 domains found 97 percent of llms.txt files received zero requests, and no engine documents consuming the file. Ship one if a machine-readable surface is part of your product story, never as a visibility lever.
- **Generic GEO rewrites.** The original KDD 2024 paper reporting up to a 40 percent visibility gain from adding statistics and quotes measured citation share in a simulator with documents pre-placed in the context window, which removes retrieval from the experiment entirely. C-SEO Bench later found significant positive effects in 3 of 54 method and domain combinations, and SAGEO Arena found body-only rewrites reduced retrieval presence. The 40 percent figure is still repeated across this search result; it did not replicate.

## What actually makes a page citable

What survives replication is narrower and duller than the tactics above, and almost all of it operates on the passage rather than the page. Retrieval scores chunks, so the unit of work is a section that answers one question completely and still reads correctly with the rest of the page removed.

1. **Put the answer first, in each section and on the page.** Kevin Indig analysed 3 million ChatGPT responses and 30 million citations, isolating 18,012 verified citations and matching them to source sentences with sentence-transformer embeddings. 44.2 percent came from the first 30 percent of the page, in a ski-ramp pattern that held across randomised validation batches, and cited passages used definitional phrasing about twice as often as uncited ones.
2. **Name the entity instead of using a pronoun.** Entity clarity is what survives chunking: a passage is extracted with no surrounding context, so a sentence opening with RankX AI survives the trip and one opening with it does not. Cited passages run around 20 percent proper nouns against 5 to 8 percent for ordinary prose.
3. **Write densely rather than at length.** Dan Petrovic's instrumentation of Google's grounding API found extraction from any single page plateaus around 540 words, and pages over 3,000 words had roughly 13 percent of their content used. Length past that point is weight, not coverage.
4. **Cover the adjacent questions.** Citations are won against sub-queries nobody typed, so a page answering the neighbouring questions collects retrievals that a narrower page hands to somebody else. This is also where third-party sources matter: for commercial questions, AI models reach authoritative sources off your domain more often than pages on it.
5. **Put every quotable fact in visible body text.** Real numbers, real dates, real prices. Anything living only in markup is not read by the systems doing the citing.

None of that is exotic, which is the point. The two levers that held up across every replication are relevance to the sub-query and position within the document. Everything sold beyond those two has either failed a controlled test or has never had one run against it.

## Why is my page not being cited?

Work the causes in the order the evidence ranks them, because the common instinct, rewriting the page, belongs fourth on that list rather than first. LLM citation problems are usually mechanical before they are editorial. The three failures above it are mechanical, and each one makes a page unciteable no matter how good the writing is.

1. **The page is not server-rendered.** No AI crawler executes JavaScript, so anything injected after hydration does not exist for GPTBot, ClaudeBot or PerplexityBot. Check view-source rather than DevTools, and [how AI crawlers read your site](/blog/how-ai-crawlers-read-your-site) covers the mechanics.
2. **A search-purpose crawler is blocked.** Blocking GPTBot does not remove a site from ChatGPT Search, and blocking OAI-SearchBot does. A CDN can also be refusing those bots without anyone having decided to. Confirm 200s for OAI-SearchBot, Claude-SearchBot and PerplexityBot in your own server logs, and the [AI Crawler Access Checker](/tools/ai-crawler-access-checker) reads the robots.txt half in seconds.
3. **The page is not retrievable for the sub-queries.** Citations are won against reformulated questions rather than the visible one, so a page answering only its head term is absent from most of a [fan-out](/blog/query-fan-out).
4. **The answer is buried.** 44.2 percent of ChatGPT citations come from the first 30 percent of the page, so an answer saved for the conclusion is an answer the retrieval layer scored low.
5. **The brand is being named without being cited, or cited without being named.** Check both before concluding there is a content problem at all. In the Semrush data those two outcomes accounted for 86.8 percent of every brand appearance between them.

## How to measure AI citations: citation tracking that holds up

Measure citations as a rate across a fixed prompt panel over time, never as a single check. SparkToro and Gumshoe ran 2,961 repeats and found under a 1 in 100 chance that two runs of the same prompt return the same brand list, and roughly 1 in 1,000 that they return it in the same order. A screenshot of one good answer is not a measurement, and neither is a screenshot of one bad one.

Three numbers hold up. The first is your [citation rate](/glossary/citation-rate) across a defined panel of prompts, run repeatedly and tracked against itself rather than against anyone else. The second is your server logs, which say plainly whether the search-purpose crawlers are fetching your pages at all, and [how AI crawlers read your site](/blog/how-ai-crawlers-read-your-site) covers which user agents to look for. The third is your AI Overview citations on keywords you already track, the closest available view of what Google repeats to someone who never scrolls.

RankX AI measures the first and third of those, and the disclosure belongs on a page making this argument: [AI Visibility](/features/ai-visibility) tracks prompts across ChatGPT, Claude, Gemini, Grok and Perplexity and reports citations as an aggregate rather than from a single run. Two checks are free and need no account. The [AI Overview Checker](/tools/ai-overview-checker) shows whether a keyword triggers an Overview and which sources it cites, and the [AI Readiness Score](/tools/ai-readiness-score) checks a page against the extraction mechanics above.

## What is the difference between an AI citation and an AI mention?

A citation is a source link the assistant attaches to its answer. A mention is your brand name appearing in the answer text. They come apart more often than they line up: in a Semrush study run with Kevin Indig across 3,981 domain appearances, 61.7 percent were citations with no brand name in the text, and a further 25.1 percent were mentions with no citation. Only 13.2 percent were both.

## Do I need to rank in Google to get cited by AI?

Not for the query a person actually typed. Ahrefs found only 12 percent of assistant-cited URLs rank in Google's top 10 for the prompt, and Moz found 88 percent of Google AI Mode citations sit outside the organic top 10. Ranking somewhere an engine retrieves from is the floor. A top-10 position for the head term is neither required nor sufficient, because retrieval scores pages against synthetic sub-queries rather than the visible one.

## Does adding schema markup get a page cited more often?

No, on the only controlled test published. Ahrefs tracked 1,885 pages that added JSON-LD against roughly 4,000 matched controls: Google AI Mode moved 2.4 percent and ChatGPT 2.2 percent, both indistinguishable from zero, while AI Overviews fell 4.6 percent. searchVIU's testing found no engine extracted a price that existed only in JSON-LD. Keep schema for rich results and entity resolution, not as a citation lever.

## How many AI citations should a brand expect?

There is no honest external benchmark, because the number depends entirely on which prompts you track. Single checks cannot answer it either: SparkToro found under a 1 in 100 chance that two runs of the same prompt return the same brand list. The defensible measure is your own citation rate across a fixed prompt panel, run repeatedly, compared against your own previous month.