LLM Citations: How AI Models Cite Sources, and How to Earn Them

LLM citations are the links an AI answer attaches to web pages it retrieved while writing that answer: Perplexity's numbered sources, ChatGPT's inline links, Claude's quoted spans, the supporting links in Google's AI Overviews. Search citations exist only when the model searched the web, and a citation shows where a sentence came from, not proof that the source shaped the answer.

This guide explains how each platform cites, using the vendors' own documentation (opened October 16, 2026), what the research says decides which pages get cited, and how to earn and track citations. It replaces our February 2026 "citation economy" post, which repeated market statistics we could not trace to a primary source.

What an LLM citation is, and what it is not

Three outcomes get lumped together as "AI visibility". They happen under different conditions, so measure them separately.

An answer from memory has no search citations, though the model may still write URLs it remembers, sometimes wrongly. OpenAI's wording is careful here: "Responses that use web search may include citations" (OpenAI Help Center). One more distinction: "AI citations" in academic style guides means citing ChatGPT in a paper, which is a different topic.

How LLM citations work on each platform

Every vendor cites a different unit, and the developer APIs show more than the apps do.

Our platform guides go deeper on each: Perplexity SEO, how to rank in ChatGPT and how to rank in AI Overviews.

Cited is not the same as used

The citation list is a subset of what the model read. OpenAI states it directly: "The number of sources is often greater than the number of citations." A page can be retrieved, read and shape the answer without ever being linked.

The reverse also happens. A 2026 mechanistic study of Llama-3.1-8B found that the decision to attach a citation comes from "a distributed, multi-stage 'attributional ensemble'" of model components. In a controlled single-document test on the PopQA benchmark, amplifying the critical components repaired over 90% of missed citations without changing the source (arXiv:2606.28358).

Gains on multi-document questions were modest, and the authors suggest that "inline citations can create a false sense of security."

Vendors say the same in plainer words. OpenAI warns ChatGPT users that "Search results and citations can be incomplete, outdated, or incorrect" and tells them to open a cited source and check that it supports the answer.

Treat citation counts as a noisy proxy for influence, not a measure of it.

Why you cannot add up citations across platforms

A Claude citation is a quoted span of up to 150 characters. A Perplexity citation is a whole page behind a number. An AI Overview link is a card beside a summary that often does not appear at all.

Adding them into one "total citations" figure adds unlike things, and each platform searches at a different rate, so the denominators differ too. Report citations per platform, with the questions asked and whether each answer searched. The cross-engine mechanics behind this are in our piece on LLM brand visibility.

What gets a page cited: what the evidence supports

Ranked from strongest evidence to weakest. Note how little of it is something you add to a page.

1. Being retrieved at all. Every operator documents that blocking its search crawler removes you from, or reduces you in, its results, and a page that was never retrieved cannot be cited. This is the one step with a documented mechanism and full publisher control.
2. Topical relevance to the question. A survey of 45 GEO studies found "topical relevance and context position are the most reproducible levers" (arXiv:2607.14035). Position in the candidate list is set by the search operator, not by you.
3. Brand prominence, as an association. The top 20 news outlets took 67.3% of OpenAI models' news citations in AI Search Arena data from March to May 2025 (arXiv:2507.05301). That concentration is consistent with size, but the study does not measure it.
4. Formatting, in a lab. In a 252,000-trial test where two anonymized pages competed inside the model's context, "formatting-only edits have little impact" (arXiv:2605.25517). It is a two-document synthetic setup by authors who ship a GEO checklist, so treat it as a hint, not proof.

And what the evidence says does not work. C-SEO Bench found most citation-oriented rewriting methods "not only largely ineffective but also frequently have a negative impact" (arXiv:2506.11097). Ahrefs' matched-control schema test concluded "Adding schema produced no major uplift in citations on any platform" (Ahrefs, May 11, 2026).

How to get cited by AI: six checks

1. Allow each vendor's search crawler. OAI-SearchBot (ChatGPT), Claude-SearchBot (Claude), PerplexityBot (Perplexity), Googlebot (AI Overviews and AI Mode), Meta-WebIndexer (Meta AI). Training crawlers such as GPTBot and ClaudeBot are a separate decision with no documented effect on citations.
2. Confirm the firewall agrees. Request a key page with each crawler's published user-agent string and check for a 200, not a 403 or a challenge page. OpenAI's eligibility rule requires both robots.txt permission and a host or CDN that "allows traffic from OpenAI's published searchbot IP addresses."
3. Put citable facts in the raw HTML. Vercel found ChatGPT's and Claude's crawlers fetch JavaScript but "don't execute them" (Vercel, December 2024). Open "View page source" and search for your price.
4. State facts as facts. "The Team plan costs $29 per user per month" can be quoted; "affordable plans for every team" cannot.
5. Keep one version of every fact. If your pricing page, an old blog post and a directory listing disagree, any of them can be the page that gets cited.
6. Check you are not suppressing yourself. Google says nosnippet prevents content "from being used as a direct input for AI Overviews and AI Mode"; max-snippet:0 and data-nosnippet limit what is shown.

One point of logic rather than evidence: a fact that exists only on your site, such as your own benchmark or your own pricing, can only be cited from your site. Publishing those facts clearly is the closest thing to a citation lever you fully own.

How to track LLM citations

99Visibility's AI visibility audit counts citation as 15% of its 0 to 100 score, alongside mention, accuracy and sentiment. Perplexity, AI Overviews and AI Mode search on every answer; ChatGPT, Claude and Gemini are asked with search on and decide per prompt whether to use it. Each answer is labelled as searched or from memory, and a "citation" on an answer from memory is a URL the model wrote from what it remembered.

The citation economy, minus the hype

The idea behind the phrase holds: an answer can satisfy a question without a click, so being cited and described correctly matters on its own. What does not hold is most of the data sold with it.

Even first-party claims deserve care. Google says that when people click from results pages with AI Overviews, "these clicks are higher quality", and the page publishes no data behind it (Google Search Central). Measure your own referrals and conversions before you plan around anyone's ratio.

Frequently asked questions

What are citations in an LLM?

They are links or references an AI answer attaches to sources it retrieved during that answer. Each platform cites a different unit: a quoted span on Claude, a numbered whole source on Perplexity, a URL tied to part of the answer on ChatGPT. An answer from the model's memory has no citations.

How do LLM citations work?

The model searches, reads a set of retrieved pages, writes an answer, and attaches citations to some of the pages it read. OpenAI documents that the pages consulted usually outnumber the pages cited. Research shows the decision to cite is partly separate from what actually shaped the answer.

Do LLMs struggle with citing sources?

They can. OpenAI warns that citations "can be incomplete, outdated, or incorrect" and tells users to open a cited source and check it. A lab study on one small model also showed the decision to cite can be steered inside the model without changing the source, in a single-document test.

How do I get cited by AI?

Let each vendor's search crawler fetch your pages through robots.txt and your firewall, put your facts in the page's HTML, state them plainly with numbers, and keep them identical across your site. No markup or rewriting tactic has shown a reliable effect.

To see which AI platforms cite and mention your brand, and which facts they get wrong, run a free 99Visibility audit.