How Do LLMs Decide What to Cite? UK Guide 2026

Last updated: 25 October 2026

LLMs decide what to cite using two distinct processes: retrieving live web content through search tools (as ChatGPT, Perplexity and Copilot do) or generating references from memorised training data, which risks fabrication. Citation accuracy varies hugely depending on the method used and whether outputs are cross-checked.

Key Takeaways

• Citation accuracy for popular generative search engines sits at only around 74%, according to CiteFix (Amazon), 2025, meaning roughly one in four citations misattributes or misrepresents a source.
• Hallucination rates in LLM-generated citations have varied widely across experimental studies run between 2023 and 2025, according to LLM hallucinations in the wild (arXiv, 2026).
• GPT-3.5 fabricated 55% of citations and GPT-4 fabricated 18% in a study spanning 42 multidisciplinary topics, according to arXiv, 2026.
• Checking whether three or more LLMs agree on the same citation substantially raises accuracy, according to M.Z. Naser, Clemson University, 2026.
• Aether AI's measured baseline across 700 live AI engine scans (July–August 2026) found four UK brands named in just 3.3% of answers, published at aether-ai.co.uk/insights/ai-visibility-baseline-2026 — a reminder that even being retrieved at all is far from guaranteed, let alone cited accurately.

What does it mean for an LLM to 'cite' a source?

An LLM citation is a reference — a link, footnote or named source — that a large language model (LLM, a system trained on vast text data to predict and generate human-like responses) attaches to a claim in its output. Unlike a human citing a journal article after reading it, an LLM's citation can come from one of two very different places: a live document it has just retrieved, or a pattern it remembers from training.

This distinction matters enormously for UK professionals using tools like ChatGPT, Gemini or Copilot in client work. A citation that looks identical on screen — a tidy footnote with a URL — might be a verified reference to a real, checked page, or it might be a plausible-sounding fabrication the model generated because similar citations appeared often in its training data.

Research from Clemson University generated a large number of citation instances across 10 commercially deployed LLMs in four academic domains, verified against CrossRef, OpenAlex and Semantic Scholar — three recognised bibliographic databases used to check whether an academic reference actually exists, according to M.Z. Naser, 2026. The scale of that study alone shows how systematically citation behaviour now needs independent verification rather than visual trust.

Do LLMs retrieve sources in real time, or rely on memorised training data?

LLMs cite sources through two mechanisms: real-time retrieval, where the model searches the live web and grounds its answer in documents it has just read, or parametric generation, where it produces a citation from patterns memorised during training without checking anything. The second route carries a far higher hallucination risk.

Retrieval-augmented generation (RAG — a technique that pairs an LLM with a live search or document-lookup step before it writes an answer) is the mechanism behind tools like Perplexity, Microsoft Copilot and ChatGPT's browsing mode. When RAG works correctly, the model drafts its answer using passages it has just pulled from an indexed source, then attaches a citation pointing back to that exact passage.

A standard chatbot without browsing enabled has no such safety net. It generates text, including any citation-like references, purely from statistical patterns learned during training — which may be a year or more out of date and contains no live fact-check step. This is precisely why hallucination rates differ so sharply between tools: a 2024 study in the Journal of Medical Internet Research found hallucination rates of 39.6% for ChatGPT (GPT-3.5), 28.6% for GPT-4, and 91.4% for Bard when generating references for systematic reviews — a dramatic range entirely explained by differences in retrieval grounding and model training.

How do ChatGPT, Perplexity and Copilot decide which websites to cite?

ChatGPT, Perplexity and Copilot select citation sources through a retrieval and ranking pipeline that scores candidate web pages on relevance, authority and recency before passing the top results to the model for summarisation. None of these tools cites the entire web equally — each applies its own ranking layer first.

The general pipeline runs in three stages:

1. Query expansion — the tool turns the user's prompt into one or more search queries, often rephrasing or splitting it.
2. Retrieval and ranking — a search index (Bing for Copilot, a proprietary crawler for Perplexity) returns candidate pages, ranked by signals including domain authority, content freshness and topical match.
3. Synthesis and attribution — the LLM reads the top-ranked passages and writes an answer, attaching citations to the specific sentences it drew from each source.

Research on generative engines found that primary-source citation rates vary by topic and region: generative engines cited primary sources in 60–65% of Japanese political answers but only 25–45% of equivalent US ones, while low-barrier sources accounted for roughly 30% of citations overall, according to Exposing Citation Vulnerabilities in Generative Engines (arXiv, 2025). "Low-barrier sources" means pages that are easy to crawl and rank well on structure rather than necessarily being the most authoritative — a finding with direct implications for any UK business trying to be cited rather than merely indexed.

Why do LLMs cite sources that don't support the claim, and how common is hallucinated citation?

A hallucinated citation is a reference an LLM generates that either does not exist, misrepresents the source it points to, or supports a different claim than the one attached to it. This happens because the model is optimising for plausible-sounding text, not verified truth, particularly when retrieval fails or is absent.

Three structural reasons drive this:

• No retrieval step. If the model isn't actually searching, any citation it produces is a statistical guess dressed up as a reference.
• Retrieval-generation mismatch. Even with RAG, the model may retrieve a relevant document but then generate a sentence that drifts from what the source actually says.
• Training data saturation. Citation formats are so common in training data that models learn to produce citation-shaped text even when no real source backs it.

The scale of the problem is well documented. RAG systems — including popular generative search engines — show citation accuracy of only around 74%, according to CiteFix (Amazon), 2025. Separately, across experimental studies run from 2023 to 2025, hallucination rates in LLM-generated citations have varied widely, according to LLM hallucinations in the wild (arXiv, 2026). And in a 42-topic multidisciplinary study, 55% of GPT-3.5 citations and 18% of GPT-4 citations were entirely fabricated, according to arXiv, 2026. For a UK compliance officer, solicitor or auditor, that range should settle any assumption that a citation on screen means a citation has been checked.

How does retrieval-augmented generation change citation selection compared to a standard chatbot?

Retrieval-augmented generation changes citation selection by forcing the model to ground its answer in specific, retrieved passages before writing, rather than generating text from memory alone. This single architectural difference is the main reason RAG-based tools outperform standard chatbots on citation reliability, though neither is fully reliable.

The table shows a clear progression: the more verification steps sit between retrieval and the final answer, the more trustworthy the citation becomes. This is also why Aether AI's Head of Content is cautious about drawing conclusions from small, unverified samples of AI output.

"Reporting share-of-voice from a handful of hand-picked prompts. Tiny prompt sets produce whatever story you want. Run a stable, sizeable prompt set, keep the misses in the data, and let the trend be boring — boring trends are the trustworthy ones." — Lauren Dawkins, Head of Content, Aether AI

That same discipline applies to citation verification: one spot-check proves nothing, but a repeated, consistent pattern across models and time is evidence worth acting on.

Can UK businesses influence whether AI tools cite their content?

UK businesses can influence their chances of being cited by AI tools through content structure, clarity and technical accessibility — a practice known as generative engine optimisation (GEO), the discipline of structuring content so generative AI systems are more likely to retrieve, quote and cite it. GEO methods have been shown to boost content visibility in generative engine responses by up to 40%, according to Aggarwal et al., KDD 2024, the foundational peer-reviewed study on the topic from Princeton, Georgia Tech, IIT Delhi and the Allen Institute for AI.

Practical GEO levers include:

• Clear definitional openings. Sections that define a term in their first sentence are cited roughly twice as often, because retrieval systems extract standalone claims more easily than buried ones.
• Clean heading hierarchy. A page with proper H1/H2/H3 structure signals clearly where one passage ends and another begins — important because citation systems quote passages, not entire pages.
• Named, checkable facts. Specific regulators, standards, dates and figures are more extractable than vague claims.
• Crawlable technical foundations. If a search index or crawler cannot reliably access a page, it cannot be retrieved for RAG in the first place.

As of August 2026, Aether AI's own measured baseline shows how far most UK brands still have to go on this: across 700 live AI engine scans of four UK brands' target keywords run in July–August 2026, the brands were named in only 23 answers — 3.3% of scans overall, while the longest-established brand (a security services firm) reached 6.3% across its 270 scans. These figures are published at aether-ai.co.uk/insights/ai-visibility-baseline-2026 and illustrate that even structured, well-optimised content competes in a low-visibility environment — being cited at all is the first hurdle, before accuracy of that citation even becomes a question.

How do ChatGPT, Gemini, Copilot and Perplexity compare on citation accuracy?

ChatGPT, Gemini, Copilot and Perplexity differ meaningfully in citation transparency because each uses a different retrieval architecture and surfaces sources with varying consistency. None of the published research reviewed here ranks these four tools head-to-head by name on a single accuracy scale, but the underlying mechanisms explain why professionals notice real differences in practice.

Tools built around persistent, mandatory retrieval — where every answer is grounded in a live search before generation — tend to show source links more consistently than conversational assistants where browsing is optional or user-triggered. Microsoft Copilot draws on the Bing index; Perplexity operates its own retrieval and ranking layer; Google's AI Overviews and Gemini draw on Google's Knowledge Graph and search index; ChatGPT's citation behaviour depends on whether browsing mode is active for a given query.

The practical takeaway for a business professional in London, Manchester or anywhere else in the UK is the same regardless of tool: the presence of a citation is not proof of accuracy. Given that popular generative search engines show only around 74% citation accuracy as a category, according to CiteFix, 2025, every tool in this category warrants the same verification habit, not selective trust in one over another.

What should UK professionals check before trusting an AI-generated citation?

UK professionals should verify an AI-generated citation by opening the linked source directly, confirming the cited page actually contains the claim attributed to it, and checking the publication date before relying on it in client work, regulated reporting or published content. This single habit closes most of the gap between a plausible-looking citation and a genuinely reliable one.

A practical verification sequence:

1. Click through, don't just read the footnote. A citation that displays a real-looking URL can still misrepresent what that page says.
2. Check the date. Training-data-only answers can surface outdated statistics, superseded guidance or withdrawn regulatory positions.
3. Cross-check across tools. Running the same query through two or three different AI assistants and comparing citations mirrors the multi-model consensus method that has been shown to substantially raise accuracy in controlled research, according to Naser, Clemson University, 2026.
4. Repeat the query. Within-model repetition — asking the same tool the same question again — raised accuracy from 28.6% to 88.9% in the same study, a simple, low-cost check available to anyone.
5. Treat primary sources as the baseline. Where the citation points to a secondary summary rather than the original regulator, standard or dataset, trace it back further.

For regulated sectors — financial services overseen by the Financial Conduct Authority (FCA), legal work subject to Solicitors Regulation Authority (SRA) standards, or data handling governed by the Information Commissioner's Office (ICO) under UK GDPR — this verification step is not optional diligence; it is the minimum bar before an AI-sourced fact enters client-facing or regulatory material.

Are there legal or compliance risks from relying on AI citations in UK professional work?

Relying on unverified AI citations in UK professional or regulated work carries real legal and reputational risk, because a fabricated reference presented as fact in client advice, a regulatory submission or published content can constitute professional negligence or a misleading claim under UK consumer protection law. The risk sits with the professional who publishes the claim, not with the AI tool that generated it.

Specific exposure points for UK businesses include:

• Professional standards bodies. The SRA, FCA and other regulators expect verified, accurate advice; an AI hallucination does not excuse an inaccurate statement to a client.
• Advertising and consumer protection rules. Aether AI notes that the Advertising Standards Authority (ASA) and the Consumer Protection from Unfair Trading Regulations 2008 require claims made to the public to be substantiated — a fabricated statistic lifted from an AI tool is not substantiation.
• Data protection. Where AI tools process client data to generate citations or summaries, UK GDPR and ICO guidance on data handling still apply regardless of which AI platform is used.
• Defamation and misrepresentation. Publishing a false claim about a third party because an AI tool invented a supporting citation does not remove liability from the publisher.

Given that hallucination rates for citations have reached as high as 91.4% in some systematic contexts, according to the Journal of Medical Internet Research, 2024, UK firms operating in regulated sectors should treat AI-generated citations as a draft lead, never a final, verified source.

Your AI citation verification checklist

• Identify whether the tool is retrieving live or working from memory — check if browsing or search mode is active.
• Click every citation link before quoting it, rather than trusting the footnote text alone.
• Confirm the source date to rule out outdated or superseded guidance.
• Cross-check high-stakes claims across at least two different AI tools.
• Repeat the same query once to see whether the citation holds up consistently.
• Trace secondary summaries back to the primary regulator, standard or dataset wherever possible.
• Log verification steps for any AI-sourced fact used in regulated or client-facing work, in line with SRA, FCA or ICO expectations.
• Treat structural quality of your own published content as a visibility factor, not just an accuracy one — clear headings and definitions affect whether your content gets retrieved at all.

FAQ

How do LLMs decide what to cite?

LLMs decide what to cite either by retrieving live web content through a search layer and grounding their answer in it, or by generating a citation-shaped reference from memorised training patterns with no live verification. The first method is substantially more reliable, though neither guarantees full accuracy.

What is the difference between a real-time citation and a hallucinated one?

A real-time citation comes from a document the model has just retrieved and read via search or RAG; a hallucinated citation is generated from statistical patterns in training data and may point to a source that doesn't exist or doesn't support the claim. Hallucination rates have ranged from 19.9% to 91.4% across studies run between 2023 and 2025, according to arXiv, 2026.

Can I trust citations from ChatGPT or Perplexity?

You should treat AI citations as a starting point requiring verification, not a finished fact-check, because even popular RAG-based generative search engines show citation accuracy of only around 74%, according to CiteFix (Amazon), 2025. Always click through and confirm the source actually supports the claim.

How can a UK business get cited more often by AI tools?

A UK business can improve its chances of being cited through generative engine optimisation (GEO) — structuring content with clear definitions, clean heading hierarchy and specific, named facts. GEO methods have been shown to boost visibility in generative engine responses by up to 40%, according to Aggarwal et al., KDD 2024.

Why does GPT-4 fabricate fewer citations than GPT-3.5?

GPT-4 fabricated 18% of citations compared with 55% for GPT-3.5 in a 42-topic multidisciplinary study, according to arXiv, 2026, likely reflecting improvements in training methodology and, in some deployments, stronger retrieval grounding. Model version alone does not guarantee accuracy, however — retrieval setup matters just as much.

Is it a compliance risk to use AI citations in regulated UK industries?

Yes — relying on an unverified AI citation in financial services, legal advice or any FCA, SRA or ICO-regulated context can expose a business to professional negligence or misleading-claims risk, because liability sits with the publisher, not the AI tool. Given hallucination rates reaching up to 91.4% in some studies, according to the Journal of Medical Internet Research, 2024, verification before publication or client use is essential.

What is the single best way to verify an AI-generated citation?

The single best check is to open the linked source directly and confirm it actually contains the claim attributed to it, combined with cross-checking the same query across multiple AI tools. Multi-model consensus has been shown to substantially raise citation accuracy, according to Naser, Clemson University, 2026.

Tracking your own citation accuracy with Aether AI

Understanding how LLMs decide what to cite is only useful if a business can see whether it is actually being cited — and how accurately — across the tools its customers use. Aether AI addresses this directly through citation tracking across six AI engines (ChatGPT, Perplexity, Google AI Overviews, Claude, Gemini and Copilot), giving UK businesses visibility into where and how often their content is being pulled into AI-generated answers rather than guessing from anecdotal spot-checks.

As of August 2026, Aether AI's own measured baseline — 700 live AI engine scans of four UK brands' target keywords — found those brands named in just 3.3% of answers overall, published in full at aether-ai.co.uk/insights/ai-visibility-baseline-2026. That same rigour, applied consistently rather than to a handful of hand-picked prompts, is what the platform's tracking is built to deliver for any business trying to understand its real AI visibility rather than an illusion of it.

If your business wants to see where it currently stands, Aether AI offers a free AI-visibility audit at /audit, alongside self-service tools for keyword and competitor tracking, Google Search Console integration, and automated GEO-optimised content generation — with transparent pricing available at /pricing.

Related Reading

• GEO vs SEO: What Is the Difference? (2026 Guide)
• Does GEO Actually Work? The UK Evidence in 2026
• Do AI Answers Differ Between Users? UK Business Guide