LLM Citations: What They Are and How to Earn Them

An LLM citation is a linked reference to your page inside an answer written by an AI assistant such as ChatGPT, Perplexity or Google’s AI Mode. To earn one, your page has to be found when the assistant searches, judged useful enough to support part of the answer, and written so a passage can be lifted cleanly. The research published this autumn is surprisingly clear on which of those you can influence, and which you can’t.

This guide covers what a citation is (and isn’t), how assistants pick their sources, what the evidence says helps, and how to track citations without kidding yourself. Fair warning: nobody can promise you a citation, because answers change from prompt to prompt and even from run to run. What you can do is clear away the things that stop good pages being cited, then build the signals that make it more likely.

Want to know whether AI assistants cite your business at all? We’ll run your buyers’ questions through the main assistants and show you who gets the link.

Get an AI visibility check

TL;DR

• A citation is a linked source in an AI answer. A mention is your brand name in the text, with or without a link. Track them separately, because they move separately.
• Being retrieved and being cited are two different events. In a study of ChatGPT, retrieval explained 32% of the variation in citations within websites; every page feature measured, combined, explained 2.8%.
• Pages whose titles and paragraphs match the questions people ask get cited more. FAQ blocks, schema, summary boxes, author boxes and page speed showed no significant effect within websites.
• In the UK, brand websites supply a smaller share of citations than in any other large market GEO Compass benchmarks, so off-site mentions matter as much as on-page work.
• Measure with a fixed prompt set, repeated runs and a monthly log. One-off spot checks produce anecdotes, not trends.

Table of Contents

• What is an LLM citation?
• How AI assistants choose which sources to cite
• What the evidence says earns citations
• Make sure AI crawlers can reach you
• Ecommerce and lead generation: what changes
• How to track LLM citations
• Citations aren’t always accurate
• A realistic 90-day plan
• Our perspective
• Talk to us about AI search visibility
• FAQ
• Sources

What is an LLM citation?

Ahrefs puts it simply: a citation is when the AI attributes information to your content, including a link to your site. Semrush describes AI citations as linked references to specific sources or webpages that a system used to generate or support its response, and adds a point that’s easy to miss: assistants don’t cite anything when they answer purely from pre-trained knowledge.

That’s the whole game in one sentence. If the assistant answers from memory, there’s nothing to cite. If it goes and searches the live web, it gathers candidate pages, and that’s the moment your content can earn a link.

Citation, mention and ranking are three different signals

Search Engine Journal’s Matt G. Southern argues that mentions and citations are separate events, and that blending them into one score can make it unclear which one actually changed. You can be named without being linked. You can be linked without being named. Here’s how the three signals differ.

Citation

• What it is: your URL appears as a linked source in the answer.
• What it tells you: your content informed the answer, and the link can send a visitor.

Mention

• What it is: your brand name appears in the answer text.
• What it tells you: the assistant knows who you are and thinks you’re relevant.

Ranking

• What it is: your position in a traditional Google results page.
• What it tells you: classic search visibility, which also feeds what assistants retrieve.

A single “AI visibility score” can hide which of these moved. SEJ also points out that a visibility score can rise while clicks fall. If a report can’t tell you whether you gained links or only name-checks, it isn’t ready to make decisions with.

The same discipline applies to marketing reporting generally. Now that impact measurement has become auditable, vague visibility claims carry less weight, and a clear definition for each metric is where honest reporting starts. If the terminology is new, our AEO, GEO and AI SEO glossary defines the lot.

How AI assistants choose which sources to cite

Selection happens in stages, and you can drop out at each one. The assistant rewrites the user’s question into one or more search queries, retrieves a pool of candidate pages, then picks a handful to cite.

An October 2026 arXiv study of ChatGPT’s production web search, Understanding ChatGPT Citations, found that only a subset of retrieved pages end up shown as citations, and that pages retrieved more often across repeated runs are more likely to be cited. It also found that outside proxies (Google Search and the OpenAI Responses API combined) recovered only about a quarter of the URLs ChatGPT actually cited. The authors’ advice is to treat “being retrieved” and “being cited” as separate goals, and to plan for rewritten queries and run-to-run variation.

In practice, that means two things:

1. Being findable for the rewritten query, which is usually longer and more specific than what the user typed, matters as much as ranking for the head keyword.
2. One test prompt proves almost nothing. Run it several times and look at the pattern.

Rank still counts, but it doesn’t explain everything

A second October 2026 paper, What Drives AI Citations? by Mahe Chen and Nick Koudas at the University of Toronto, measured this directly. Within websites, retrieval explained 32% of the variation in ChatGPT citations, while all the page features they measured explained 2.8%. Looking at which of the pages a single answer consulted got picked, rank alone explained 26.8% of the choice, and adding question match, the engine’s own queries and every page feature lifted that only to 32.5%.

Two lessons follow. First, traditional SEO still earns its keep, because ranking is a big part of what gets a page retrieved. Second, most of the outcome isn’t explained by anything on the page, so expect some unpredictability and plan to measure over months rather than fix once and forget.

It’s also worth keeping the prize in proportion. Across roughly 60,000 websites in Ahrefs’ analysis, all large language models combined sent less than 1% of traffic, against 41.35% from Google Search. That’s why we treat AI search as an extension of search work rather than a replacement for it. Our comparison of SEO and GEO sets out where the two overlap.

What the evidence says earns citations

The most useful finding in the Toronto study is also the least glamorous: pages whose titles and paragraphs match the questions people ask get cited more. Within websites, a closer title match was linked to 14.1% more citations per standard deviation, and a closer best-paragraph match to 15.5% more. Write pages that answer specific questions directly. That’s it. That’s the trick.

Match the question in the heading and the first paragraph

• Write headings the way a buyer would phrase the question, then answer it in the first sentence or two underneath.
• Keep each section self-contained, so an assistant can lift one passage without needing the rest of the page.
• Cover the follow-up questions too. Rewritten queries tend to be more specific than the original, and a page that answers “how much does it cost for a three-bed semi in Chester” beats one that only answers “how much does it cost”.

Our explainer on answer engine optimisation goes deeper on writing for direct answers.

Don’t over-invest in on-page tricks

Here’s the bit some agencies won’t enjoy. The same study found that within websites, FAQ blocks, schema, summary boxes, author boxes and page speed showed no significant citation effect. Across websites they all looked positive, but the authors put that down to Simpson’s paradox: sites that already earn lots of citations also tend to use these features.

That doesn’t make them worthless. Structured data still helps machines read prices, opening hours and reviews accurately, and fast pages still matter to people. It does mean none of them is a shortcut. Treat them as hygiene, and spend the real effort on content that deserves to be retrieved. We cover more of the oversold tactics in the GEO tactics most agencies miss.

Build your brand’s presence beyond your own site

For UK businesses this is the big one. GEO Compass’s UK GEO benchmark says competition for a place in the answer is decided mostly off your own site. Brand and corporate sources make up 21% of UK citations against 31% across all the markets it tracks, and UK answers draw on an average of 9.9 sources each, more than any other market (the all-market average is 8.1).

More sources per answer means more seats at the table, but most of those seats go to someone else’s website. So the plan has to look outward:

• Earn coverage and mentions on publications, trade bodies and directories that your customers and the assistants both trust.
• Publish expertise-led content that other people want to reference.
• Keep your business name, address, services and prices consistent everywhere, so assistants describe you correctly.

This is the work behind digital PR and good regional link building, and it’s why a narrow “fix the page” project so often disappoints.

Make sure AI crawlers can reach you

Start with access, because nothing else matters if the bot can’t get in. OpenAI’s publishers and developers FAQ says any public website can appear in ChatGPT search, and that to be included in summaries and snippets you should make sure you’re not blocking OAI-SearchBot in robots.txt. It also confirms that ChatGPT adds utm_source=chatgpt.com to referral URLs, so you can see those visits in analytics.

A quick technical checklist:

1. Check robots.txt and confirm the AI search crawlers you want aren’t blocked.
2. Check your firewall, CDN and security plugins as well. We’ve seen bot-blocking rules lock out AI crawlers without anyone at the business knowing.
3. Make sure key content is in the raw HTML, not only rendered by JavaScript.
4. Set up a segment for utm_source=chatgpt.com in analytics so you can see what citations actually deliver.
5. Keep important pages indexable, quick to load and free of duplicate content.

A technical SEO audit is the quickest way to find blockers like these, and it’s the sensible first step before any content programme.

Ecommerce and lead generation: what changes

Product and category pages

Salesforce UK’s AI-readiness SEO checklist gives online retailers a practical list, and it’s candid that following it doesn’t guarantee rankings or visibility. The highlights:

• One clear H1 per page, with supporting H2 and H3 headings.
• Meta descriptions of 150 to 165 characters that include the product name, brand, a selling point or two and a call to action.
• No reused manufacturer copy, because search engines and AI may treat duplicates as low-value.
• Short, descriptive, hyphenated URLs without query strings.
• High-resolution images of 1,200 pixels or more, with alt text.
• AI crawlers such as GPTBot explicitly allowed in robots.txt, and firewall rules checked so they aren’t blocked by default.
• Server-side rendering, or key content in raw HTML, because most AI crawlers don’t render JavaScript.

Unique product copy, clear specifications and honest comparisons give an assistant something worth quoting. Structured product data helps it read prices, stock and reviews correctly, but as the Toronto study found, it supports accuracy rather than winning the citation on its own.

Service and lead-generation pages

For service businesses, the pages most likely to be cited answer what a buyer asks before they pick up the phone: what the service involves, who it suits, how long it takes, what affects the price and how to choose a provider. Location pages that describe real local work, rather than a template with the town name swapped in, give assistants a reason to surface you for local questions.

Our complete guide to AI SEO and GEO shows how these pieces fit together, and our guide to ChatGPT SEO covers the ChatGPT-specific steps.

How to track LLM citations

Tracking is where most teams go wrong. A few ad hoc searches on a Friday afternoon produce anecdotes. A repeatable process produces trends you can act on. It has three parts:

1. A fixed prompt set. Write prompts that mirror how buyers actually ask, including unbranded category questions, comparisons and local questions. Aether AI’s UK guide to measuring AI search visibility suggests 15 to 20 distinct buyer-style prompts, each repeated three to five times across sessions before you draw conclusions. Keep the set stable so months are comparable.
2. A reliable log. For each prompt and assistant, record whether you were mentioned, whether your URL was cited, which other sources were cited, and whether the description of your business was accurate.
3. A trend view. Turn the log into monthly figures, with mentions and citations kept apart.

On cadence, the same guide suggests weekly spot checks of five to ten priority prompts, a monthly citation report, quarterly competitor benchmarking, and a targeted re-check after a product launch, rebrand or PR push. That’s a sensible starting rhythm, and you can loosen or tighten it for your sector.

What to measure

Five numbers cover most of what you need. Report them monthly, side by side.

Mention rate

• How to calculate it: the share of test prompts where your brand name appears.
• Why it matters: it shows brand recall in AI answers.

Citation rate

• How to calculate it: the share of test prompts where your URL is linked.
• Why it matters: it shows credit and potential referral traffic.

Source share

• How to calculate it: which domains get cited alongside you, or instead of you.
• Why it matters: it tells you where to earn coverage next.

Accuracy

• How to calculate it: whether the assistant describes your services, prices and locations correctly.
• Why it matters: a wrong description costs enquiries and reputation.

Referral visits and enquiries

• How to calculate it: sessions from chatgpt.com and similar sources, followed through to enquiries or sales.
• Why it matters: it connects visibility to money.

Keep retrieval and citation apart where you can. If a page keeps getting retrieved but rarely cited, the content is the problem. If it never gets retrieved, findability is the problem. And keep the commercial maths in view. As an illustrative example: if ChatGPT sends a plumbing firm 40 visits a month, 10% of them enquire and a quarter of those enquiries become £1,200 jobs, that’s one job a month, or about £1,200 in revenue. Worth having, and worth knowing before you spend £2,000 a month chasing it. Ahrefs also found that only about 10% of its own 1,000 most-cited pages were among its top pages for ChatGPT traffic, so a citation and a click are not the same thing. If you want to compare lead quality, we looked at whether GEO leads convert better than Google leads.

Citations aren’t always accurate

Being cited is not the same as being cited correctly. A 2026 preprint, Verified Misguidance, analysed 112,000 responses from ten models across five providers and found that 30.6% of citations distorted their sources and 27.1% came from sources unsuited to the subject. That’s one study’s evaluation pool rather than a universal rate, but it’s a good reason to check how you’re described, not just whether you appear.

Where an answer gets your services or prices wrong, make the correct facts plain and easy to find on your own site and your third-party profiles. Clear, consistent statements leave less room for an assistant to improvise.

A realistic 90-day plan

1. Weeks 1 to 2: baseline. Build the prompt set, run it across the main assistants several times, and log mentions, citations and accuracy. Run the technical access checks.
2. Weeks 3 to 6: fix and rewrite. Clear crawl and rendering blockers. Rewrite your highest-value pages so headings match buyer questions and each section answers directly.
3. Weeks 7 to 10: build presence. Publish expertise-led pieces, pitch for relevant coverage and tidy up business profiles so they match.
4. Weeks 11 to 12: re-measure. Re-run the prompt set, compare against the baseline and decide where to focus next.

Given how much answers vary, judge progress by direction over several months, not by a single good or bad week.

Our perspective

We see LLM citations as the visible edge of a bigger shift. More buyers now ask an assistant first, and the businesses it recommends get the enquiry before a results page is ever opened.

That’s why we don’t treat AI search as a separate product. Technical foundations, expertise-led content, digital PR and measurement all feed each other. A page that ranks is more likely to be retrieved. A page that answers a question plainly is more likely to be cited. A brand that people talk about on trusted sites is more likely to be named. When one in-house team handles all of it, nothing falls between the gaps.

We also think reporting should be honest about what can and can’t be measured. Mentions and citations get tracked as separate signals, and both get tied back to enquiries and revenue. Any agency promising you a guaranteed spot in ChatGPT’s answers is selling something nobody controls, us included. Most UK businesses haven’t yet built a structured programme for this, which leaves real room for the ones that start now.

Talk to us about AI search visibility

If you’d rather have specialists run this for you, we combine traditional SEO with generative engine optimisation and answer engine optimisation, delivered by one in-house team: technical audits, expertise-led content, service and location pages, digital PR and analytics tied to enquiries. You can see how our AI SEO service works, or browse the full range on our services page. If you’d like an honest view of where you stand, get in touch.

Book an appointment

FAQ

What is the difference between an LLM citation and a brand mention?

A citation is a linked source in the AI’s answer, so a reader can click through to your page. A mention is your brand name appearing in the text, with or without a link. They can move independently, so measure them separately.

Can anyone guarantee that ChatGPT will cite my site?

No. Research on ChatGPT shows that retrieval, rewritten queries and run-to-run variation all affect which pages get cited. You can improve your odds by letting crawlers in, ranking well, matching real questions and building brand presence, then measuring over time.

How do I track LLM citations?

Use a fixed set of buyer-style prompts, run each several times across the main assistants, and log whether you were mentioned, linked and described accurately. Add an analytics segment for utm_source=chatgpt.com. Review monthly against your baseline.

Does schema markup get me cited?

Not on its own. A 2026 University of Toronto study found no significant citation effect from schema within websites. It still helps machines read your pages accurately, so keep it correct, but put most of your effort into answer quality and question match.

Why are UK businesses cited less from their own websites?

According to GEO Compass, UK answers draw on more sources per answer than any other market it tracks, and brand websites make up a smaller share of them. That pushes more of the competition onto third-party sites, which is why digital PR and consistent listings matter so much here.

Sources

• Ahrefs: How to Earn LLM Citations to Build Traffic & Authority (updated May 2026)
• Semrush: What Are AI Citations & How Do I Get Them? (July 2025)
• Search Engine Journal: Brands Are Tracking AI Visibility, But Are They Measuring The Right Things? (September 2026)
• arXiv: Understanding ChatGPT Citations: Evidence from Production Web Search Across Model Generations (October 2026)
• arXiv: What Drives AI Citations? Content, Retrieval and Model Knowledge in GEO, Chen and Koudas (October 2026)
• GEO Compass: AI search in the United Kingdom, GEO benchmark
• OpenAI Help Centre: Publishers and developers FAQ
• Salesforce UK: AI-Readiness SEO Checklist
• Aether AI: How to Measure AI Search Visibility in 2026 (UK Guide), October 2026
• alphaXiv: Verified Misguidance: Measuring Structural Citation Failures in Search-Augmented LLMs
• When impact became auditable and branding had to grow up (dante.media)

Recommended

• ChatGPT business ranking requires these 10 essential tactics
• What is ChatGPT SEO? What it involves and how to measure it
• The GEO tactics most agencies miss (and the ones they oversell)