AI Hallucinations

Research has shown that language models can repeat common misconceptions rather than reliably correct them. Evidence for this claim TruthfulQA found that language models can reproduce common human misconceptions and that larger models were not automatically more truthful on its benchmark. Scope: TruthfulQA's benchmark, model set, and 2021 results; not a current cross-model error rate. Confidence: high · Verified: 2026-07-14 Lin et al.: TruthfulQA NIST calls confidently presented false or erroneous model content “confabulation.” Evidence for this claim NIST's Generative AI Profile identifies confabulation—the confident presentation of false or erroneous content—as a generative-AI risk. Scope: NIST risk-management terminology and guidance; not a product-specific incidence estimate. Confidence: high · Verified: 2026-07-14 NIST: Generative AI Profile

TL;DR — An AI hallucination is when a chatbot says something that sounds right but is just wrong — a made-up fact, a fake source, or a link to a page that doesn’t exist. In an ungrounded response, the model predicts likely-sounding words rather than checking a source. GroundingGrounding is anchoring an AI model's answer to source documents it retrieves at the moment you ask — not to the patterns frozen into its weights during training. Retrieval-Augmented Generation (RAG) is the most common way to do it. (giving the model retrieved sources) can help but doesn’t fully fix it. For your website, the two things to watch are AI pointing people at URLs you never made, and AI getting facts about your brand wrong.

What an AI hallucination is

An ungrounded language-model response predicts likely continuations from learned patterns rather than verifying each statement against a source. Search and retrieval tools can add current evidence, but they do not make synthesis error-proof. When a generated statement sounds confident but is false, that is commonly called a hallucination.

A useful way to picture it: the model is very good at writing the kind of sentence that usually comes next. It is not checking whether that sentence is true. When it doesn’t really “know” something — a niche detail, a recent change, an obscure brand — it still writes a smooth, confident-sounding answer, because that’s what it was trained to do.

Some researchers prefer the word confabulation (borrowed from memory research, where people sincerely fill in gaps with false details). It’s a better word because the AI isn’t lying or trying to trick you. It just generated the most plausible-looking continuation, and that continuation happened to be wrong.

The two kinds that hit your website

1. Made-up URLs. The AI recommends a page on your site that doesn’t exist. A user clicks and lands on a 404. AI assistants do this far more than Google does. On Ahrefs’ own data, 3.6% of the traffic from AI went to pages that don’t exist on the site.
2. Wrong facts about your brand. The AI tells a prospect the wrong founder, old pricing, a discontinued product, or something a competitor said about you — stated as fact, with no “I’m not sure” attached.

What you can actually do

• RedirectA redirect sends browsers and crawlers from a requested URL to a different one. An HTTP redirect specifically is a 3xx status code paired with a Location header; meta refresh and JavaScript redirects achieve a similar navigation without being a 3xx response themselves. Permanent redirects (301/308) are Google's signal the target should be canonical; temporary ones (302/303/307) aren't. the made-up URLs. If AI keeps sending people to a page you never built, set up a redirect from that fake URL to the closest real page. (You first have to find them — watch your analytics for traffic to 404s.)
• Check what AI says about you. Ask the big AI tools the kinds of questions your customers ask, and see if the answers are right. Do it regularly — answers change as models update.
• Publish clear, correct facts. The cleaner and more consistent your public information is (your “About” page, structured dataStructured data is a standardized way of labeling page content (using the schema.org vocabulary in JSON-LD, Microdata, or RDFa) so search engines can understand its meaning. It's not a direct ranking factor — its value is rich results and entity understanding., profiles), the easier it is for AI to ground on the truth instead of guessing.

One thing that surprises people: blocking AI crawlersAI crawlers are bots from AI companies that fetch web pages to train language models, build AI-search indexes, or answer live user questions. They come in three categories, each with its own user-agent tokens and its own robots.txt controls. does not fix hallucinations about your brand. If a model already learned something wrong about you, blocking it from re-crawlingCrawling is how search engines use automated bots (like Googlebot and Bingbot) to discover URLs and download pages. A page has to be crawlable to be indexed, but crawling on its own isn't a ranking factor. doesn’t un-teach it. Monitoring and clear public facts do more than blocking.

Want the mechanism — why bigger models can hallucinate more, what the research shows, and how groundingGrounding is anchoring an AI model's answer to source documents it retrieves at the moment you ask — not to the patterns frozen into its weights during training. Retrieval-Augmented Generation (RAG) is the most common way to do it. fails — switch to the Advanced tab.

TruthfulQA foundA 302 (\"Found\") is a temporary redirect: it forwards users to a new URL while telling search engines the original URL should stay in the index. It's a weak canonicalization signal, not the zero-equity dead end of SEO folklore. that scale alone did not make the tested models more truthful on its benchmark. Evidence for this claim TruthfulQA found that language models can reproduce common human misconceptions and that larger models were not automatically more truthful on its benchmark. Scope: TruthfulQA's benchmark, model set, and 2021 results; not a current cross-model error rate. Confidence: high · Verified: 2026-07-14 Lin et al.: TruthfulQA NIST treats confabulation as a risk to measure and manage, not as a product-specific fixed error rate. Evidence for this claim NIST's Generative AI Profile identifies confabulation—the confident presentation of false or erroneous content—as a generative-AI risk. Scope: NIST risk-management terminology and guidance; not a product-specific incidence estimate. Confidence: high · Verified: 2026-07-14 NIST: Generative AI Profile

TL;DR — Hallucination is a characteristic of next-tokenA token is the smallest unit of text (or image/audio/video) an LLM processes — roughly 4 characters, or about ¾ of an English word. A context window is the maximum number of tokens (input plus output) a model can hold at once, like its short-term memory. prediction, not a bug you can patch out. Models optimize for plausible continuations, not truth, so they confidently produce false statements when training signal is thin. groundingGrounding is anchoring an AI model's answer to source documents it retrieves at the moment you ask — not to the patterns frozen into its weights during training. Retrieval-Augmented Generation (RAG) is the most common way to do it. and RAGRAG is the retrieve-then-generate pattern behind AI search: the system retrieves relevant passages from an external index at query time, injects them into the model's context, and generates an answer grounded in those sources — without changing the model's weights. reduce it (73–86% in some benchmarks) but don’t eliminate it — the best grounded models still fail ~10–15% of the time, and TruthfulQA found larger models can produce more convincing falsehoods. For SEO it’s measurable: AI sends users to 404s ~2.87x more than Google, 3.6% of our AI traffic at Ahrefs went to non-existent pages, and citation error rates across AI search enginesAI search uses large language models and retrieval-augmented generation (RAG) to synthesize an answer from multiple sources rather than returning a ranked list of links. Examples include Google AI Overviews, ChatGPT Search, and Perplexity. run north of 60%. The fixes are tactical: redirectA redirect sends browsers and crawlers from a requested URL to a different one. An HTTP redirect specifically is a 3xx status code paired with a Location header; meta refresh and JavaScript redirects achieve a similar navigation without being a 3xx response themselves. Permanent redirects (301/308) are Google's signal the target should be canonical; temporary ones (302/303/307) aren't. hallucinated URLs, monitor brand answers, and publish verifiable, entity-clear content.

What hallucination actually is

An AI hallucination is output that’s confidently stated but factually incorrect, fabricated, or unsupported by any source. The cleaner technical term is confabulation — borrowed from neuropsychology, where it means producing false accounts with genuine confidence and no intent to deceive. That precision matters: the model isn’t lying. It has no notion of truth to lie about. It’s running the one task it was trained for — predicting the next token — and sometimes the most probable token is false.

This is the single most important framing on this page: hallucination is an inherent byproduct of language modeling that prioritizes plausibility over factual accuracy. It emerges from LLMsA large language model (LLM) is a deep-learning model trained on massive text corpora to predict the next token and generate human-like text. LLMs use the transformer architecture and power AI search features like Google's AI Overviews (Gemini) and Bing Copilot (GPT-4). functioning as designed, not from implementation failures. You don’t fix it the way you fix a bug. You reduce it.

Why it happens — the mechanism

LLMs are trained on a next-token objective: given a sequence, predict the most probable next token. That’s a language completion task, not a fact retrieval task. The model optimizes for what text sounds like it should follow, not what is true. When the training signal for a specific fact is thin — rare events, niche topics, anything that changed after the model’s knowledge cutoffA knowledge cutoff (or training data cutoff) is the date after which an LLM was no longer trained on new data. The model has no awareness of anything that happened later — unless the system bolts on live retrieval (RAG/web search) to fetch it at query time. — the model fills the gap with statistically probable language. If a sequence looks like a question, the probability mass overwhelmingly favors generating an “answer-shaped” response over admitting uncertainty, regardless of whether the model actually has the knowledge.

Microsoft Research’s Ece Kamar makes the empirical version of this point: models hallucinate more around facts that are less common in the training data. The long tail is where confabulation lives. Your obscure product SKU, a recent rename, a founder who left last quarter — exactly the facts least represented on the web are the ones a model is most likely to get confidently wrong.

The types worth keeping straight

• Factual hallucination — the stated fact is simply wrong.
• Citation / attribution hallucination — the model cites a source it didn’t use, misattributes a quote, fabricates a byline, or invents a URL. The most directly measurable SEO version.
• Entity / brand hallucination — wrong names, dates, stats, pricing, or company facts, stated with confidence.
• Intrinsic vs. extrinsic (the Ji et al. taxonomy): intrinsic output contradicts the source provided; extrinsic output cannot be verified from the source — neither confirmed nor contradicted, a fabricated addition.

Reading hallucination claims correctly

Five distinctions get flattened in most hallucination coverage, including earlier drafts of this page. They matter because they change what “fixed,” “grounded,” or “cited” actually means when you’re evaluating a tool or a claim.

• Name the reference. “Hallucination” isn’t one failure mode. It can mean contradiction of the supplied context, an unsupported addition the context neither confirms nor denies, or a plain world-factual error — and those three aren’t interchangeable (Ji et al.’s survey is the source taxonomy). A checker built to catch one of the three will quietly miss the other two.
• Factuality is not faithfulness. Factuality tests a claim against an external, real-world reference — is it true? Faithfulness tests a claim against the evidence you actually supplied — is it supported by that evidence? An answer can pass one and fail the other: faithfully summarizing a wrong source, or stating something true that the supplied evidence never backs. FActScore’s atomic-fact method — breaking a long answer into individually checkable claims — exists because a single holistic score hides that split.
• A citation existing isn’t groundingGrounding is anchoring an AI model's answer to source documents it retrieves at the moment you ask — not to the patterns frozen into its weights during training. Retrieval-Augmented Generation (RAG) is the most common way to do it.. Citation quality is really three separate checks: does the citation exist and resolve, does it actually entail (support) the specific claim sitting next to it, and does the answer cover every claim that needed a citation in the first place (the breakdown Gao et al.’s ALCE work uses). A visible, clickable footnote can pass the first check and fail the other two — which is exactly the gap behind the 60%+ citation-error numbers above.
• RAG failures have named stages, not one label. When a RAG-backed answer is wrong, the cause is one of: the evidence was never in the corpus, it was chunked so the needed fact got split across a boundary, it existed but wasn’t retrieved, it was retrieved but ranked too low to reach the generator, or it reached the generator and got misused anyway (RAGTruth’s taxonomy of RAG errors). “The AI hallucinated” collapses five different root causes — and different fixes — into one word.
• Benchmark percentages don’t travel. Every rate on this page — 73–86%, ~85–90%, 58% vs. 94% — is scoped to a specific task, dataset, reference policy, judge, prompt set, corpus, model, model version, and test date. Don’t quote “model X is Y% accurate” as a portable fact outside that benchmark’s own evaluation contract; a different judge, corpus, or model version can move the number substantially.

What the research shows

• TruthfulQA (Lin et al., 2021) established the inverse-scaling problem. Across 817 adversarial questions, the largest model of the day (GPT-3-175B) was only 58% truthful vs. a 94% human baseline — and crucially, “larger models produce more imitative falsehoods because they are better at learning the training distribution.” Bigger isn’t safer. A bigger model is better at reproducing the popular misconceptions baked into web text, so it can be more confidently wrong, not less. (This is directly relevant to your brand: if wrong information about you circulates online, more capable models are better at absorbing and replicating it.)
• Ji et al. survey (2022/2023) is the gold-standard academic taxonomy — intrinsic vs. extrinsic hallucination across summarization, dialogue, QA, translation, and LLMs.
• Lanham et al. (2023), from Anthropic, found chain-of-thought reasoning isn’t always faithful: “as models become larger and more capable, they produce less faithful reasoning on most tasks.” The step-by-step explanation a model shows you doesn’t necessarily reflect how it actually reached the answer — it can be post-hoc rationalization. So “the AI explained its reasoning” is not proof the answer is grounded.
• FACTS Grounding (Google DeepMind, Dec 2024) is the number that anchors the grounding debate: even given explicit source documents, the best models scored only ~85–90% grounded. That remaining 10–15% is misreading, misinterpreting, or failing to use the retrieved content — not confabulating from nothing.

Does grounding (RAG) solve it? No — it reduces it

This is the nuance most coverage botches. grounding — anchoring an answer to documents the system retrieves at inference time, usually via RAG — is the primary mitigation, and it works: web-search access reduced hallucination by 73–86% across tested models in grounding benchmarks. But it does not eliminate hallucination, and it fails in different ways than a bare LLM:

• it can misinterpret the retrieved passage,
• select the wrong passage,
• combine multiple sources incorrectly, or
• fail to recognize an adversarial or nonsensical query.

Google’s Elizabeth Reid drew exactly this line about AI OverviewsAI Overviews are the AI-generated summary box Google shows above or within its regular search results, written by Gemini models from pages retrieved out of Google's normal Search index. It's a Search feature, not a separate platform or index.: they “generally don’t ‘hallucinate’ or make things up in the ways that other LLM products might” because they’re “built to only show information that is backed up by top web results.” When they get it wrong, she said, it’s “usually for other reasons: misinterpreting queries, misinterpreting a nuance of language on the web, or not having a lot of great information available.” That’s a real distinction — grounded systems fail at retrieval and interpretation more than at pure confabulation — but it is not a guarantee of accuracy. And even RAG-heavy legal research tools have shown hallucination rates up to ~33%, contradicting vendor “hallucination-free” claims. Anthropic says the quiet part plainly in its own developer docs: “while these techniques significantly reduce hallucinations, they don’t eliminate them entirely. Always validate critical information.”

The SEO impact — measurable, and addressable

URL hallucination. This is the most concrete one, and it’s mine to quantify. Analyzing Ahrefs’ own AI referral traffic, I found (in my analysis of Ahrefs’ AI search traffic by page type) that 3.6% of the traffic went to hallucinated pages — pages that don’t exist on ahrefs.com. AI systems (primarily ChatGPT) recommended URLs that were never built. The broader Ahrefs study of 16 million AI-cited URLs (Ryan Law and Xibeijia Guan) puts it in context: AI assistants direct users to 404 pages 2.87x more often than Google Search — ChatGPT’s clicked-through 404 rate was 1.01% vs. Google’s 0.15% baseline. The tactical fix is straightforward: watch referral traffic to 404s, identify hallucinated URLs above a threshold (say, 10 visits/month), and 301-redirect them to the closest real page. You capture traffic the model is trying to send you anyway.

Attribution hallucination. The Columbia/Tow Center study (2025) tested 8 AI search engines and found incorrect answers to over 60% of queries when asked to identify article title, date, publisher, and URL. ChatGPT misidentified 134 articles but signaled uncertainty only 15 times — high-confidence wrong. Over half of some models’ responses cited fabricated URLs. The lesson for SEO: being cited in an AI answer does not mean you were accurately represented. A citation is not comprehension.

The analytics blind spot. This compounds the URL problem. AI traffic is poorly attributed today — some platforms don’t pass referrer data, and Google’s AI Mode used noreferrer, lumping clicks into “Direct”/“Unknown” (Google confirmed that as a bug after I reported it). If hallucinated-URL traffic doesn’t attribute cleanly, you may never see it — the 3.6% finding only surfaced because we had analytics access to look. Fix attribution first, or you’re monitoring blind.

Brand hallucination & reputation. An AI stating wrong facts about your pricing, leadership, or history to a prospect can damage trust before they ever reach your site. Worse, it’s now a negative-SEO vector: Ahrefs’ brand experiment (Makosiewicz) showed that a single seeded Medium article could push some models to propagate misinformation in 37–39% of answers about a brand. “Any growing brand can be knocked off course in AI searchAI search uses large language models and retrieval-augmented generation (RAG) to synthesize an answer from multiple sources rather than returning a ranked list of links. Examples include Google AI Overviews, ChatGPT Search, and Perplexity. results by an upset person with a Medium account.” No links to build or disavow — just false content on an indexed platform.

What to actually do

1. Redirect hallucinated URLs (the URL-hallucination fix above).
2. Monitor AI answers about your brand on a schedule — answers shift as models update and as new content gets indexed. One-time audits go stale.
3. Build entity signals — consistent NAP, schema, Wikidata, official profiles — so grounded systems have clean, verifiable facts to retrieve. This is the single best lever against future brand hallucination.
4. Verify AI-generated content before you publish it. Unedited hallucinations that reach the web get indexed and can feed the next model’s training — the model-collapse feedback loop. Don’t be a node in it.
5. Don’t expect crawlerA crawler — also called a spider or bot — is an automated program that fetches web pages, extracts their links, and queues new URLs to visit. Search engines use crawlers to discover and download content for their index. blocking to fix it. Blocking GPTBot/ClaudeBot stops future training on your content; it does nothing about facts already in the weights. (See the AI crawlersAI crawlers are bots from AI companies that fetch web pages to train language models, build AI-search indexes, or answer live user questions. They come in three categories, each with its own user-agent tokens and its own robots.txt controls. article for that distinction.)

The honest summary: hallucination is permanent infrastructure, not a passing bug. You can’t make AI stop confabulating. You can make your own facts the cheapest, cleanest thing for it to ground on — and you can catch and redirect the URLs it invents.

AI summary

A condensed take on the Advanced version:

• Hallucination = confident, false, or unsupported output. It’s a characteristic of next-tokenA token is the smallest unit of text (or image/audio/video) an LLM processes — roughly 4 characters, or about ¾ of an English word. A context window is the maximum number of tokens (input plus output) a model can hold at once, like its short-term memory. prediction (optimizing for plausible text, not truth), not a bug — so it can be reduced, never fully eliminated. “Confabulation” is the more precise term; the model isn’t lying.
• It clusters in the long tail. Models hallucinate more around facts that are rare in training data — niche products, recent changes, obscure brands.
• Naming the failure precisely matters. Contradiction-of-context, unsupported-addition, and world-factual-error aren’t the same thing; factuality (true?) and faithfulness (supported by the evidence given?) are different tests an answer can pass one and fail the other on; a citation existing isn’t the same as it entailing the claim next to it; and a RAGRAG is the retrieve-then-generate pattern behind AI search: the system retrieves relevant passages from an external index at query time, injects them into the model's context, and generates an answer grounded in those sources — without changing the model's weights. answer can fail because evidence was missing, chunked badly, unretrieved, under-ranked, or mishandled at generation — five distinct causes “it hallucinated” hides.
• Bigger isn’t safer. TruthfulQA: larger models produce more convincing imitative falsehoods (58% truthful vs. 94% human). Chain-of-thought reasoning can be unfaithful (Lanham et al.).
• GroundingGrounding is anchoring an AI model's answer to source documents it retrieves at the moment you ask — not to the patterns frozen into its weights during training. Retrieval-Augmented Generation (RAG) is the most common way to do it./RAG helps a lot but doesn’t fix it. Web access cut hallucination 73–86%, yet the best grounded models still fail ~10–15% (FACTS GroundingGrounding is anchoring an AI model's answer to source documents it retrieves at the moment you ask — not to the patterns frozen into its weights during training. Retrieval-Augmented Generation (RAG) is the most common way to do it.) — by misreading or mis-selecting sources. Reid: AI OverviewsAI Overviews are the AI-generated summary box Google shows above or within its regular search results, written by Gemini models from pages retrieved out of Google's normal Search index. It's a Search feature, not a separate platform or index. are grounded so they “generally don’t hallucinate… in the ways other LLMA large language model (LLM) is a deep-learning model trained on massive text corpora to predict the next token and generate human-like text. LLMs use the transformer architecture and power AI search features like Google's AI Overviews (Gemini) and Bing Copilot (GPT-4). products might,” but still err via misinterpretation. Anthropic: techniques “don’t eliminate them entirely.”
• SEO impact is measurable: AI sends users to 404s ~2.87x more than Google; 3.6% of Ahrefs’ AI traffic went to non-existent pages; AI citationAn AI citation is the visible source link an AI answer engine shows next to its generated text — the clickable reference that credits the web page it used. A citation's presence is a separate thing from whether the cited page actually supports the statement, and from being retrieved (read behind the scenes) or merely mentioned (named without a link); citation is driven more by brand mentions and being retrievable than by traditional ranking. error rates run 60%+ (Tow Center). Being cited ≠ being represented correctly.
• Attribution is broken too, so hallucinated-URL traffic can be invisible — fix tracking before you monitor.
• Brand risk is a negative-SEO vector: one seeded Medium article pushed some models to repeat misinformation in 37–39% of answers.
• Fixes: redirectA redirect sends browsers and crawlers from a requested URL to a different one. An HTTP redirect specifically is a 3xx status code paired with a Location header; meta refresh and JavaScript redirects achieve a similar navigation without being a 3xx response themselves. Permanent redirects (301/308) are Google's signal the target should be canonical; temporary ones (302/303/307) aren't. hallucinated URLs, monitor brand answers regularly, build entity signals, verify AI content before publishing, and don’t expect crawlerA crawler — also called a spider or bot — is an automated program that fetches web pages, extracts their links, and queues new URLs to visit. Search engines use crawlers to discover and download content for their index. blocking to undo facts already in the weights.

Official documentation

What the model makers and search engines say about hallucination in their own words.

Google

• What happened with AI Overviews and next steps — Elizabeth Reid (May 2024) on why grounded AI OverviewsAI Overviews are the AI-generated summary box Google shows above or within its regular search results, written by Gemini models from pages retrieved out of Google's normal Search index. It's a Search feature, not a separate platform or index. err differently than standalone LLMsA large language model (LLM) is a deep-learning model trained on massive text corpora to predict the next token and generate human-like text. LLMs use the transformer architecture and power AI search features like Google's AI Overviews (Gemini) and Bing Copilot (GPT-4)..
• GopherCite: supporting answers with verified quotes — DeepMind on models that hallucinate “convincing nonsense,” and citing evidence (and admitting “I don’t know”) to counter it.
• FACTS Grounding benchmark — DeepMind’s factuality benchmark; the ~85–90% grounded ceiling.
• Guidance on AI-generated content — Search Central: AI content is judged on quality and E-E-A-T, not production method.

Anthropic

• Reduce hallucinations — Claude developer docs: let the model say “I don’t know,” ground with direct quotes, verify with citations, and the explicit caveat that techniques don’t eliminate hallucination.

OpenAI

• Why language models hallucinate — OpenAI’s explainer of the mechanism.
• Does ChatGPT tell the truth? — Help Center: “Hallucinations are plausible but false statements generated by language models.”

Microsoft

• Why AI sometimes gets it wrong — Microsoft’s technical definition of “ungrounded” content and its mitigation tooling.

Quotes from the source

On-the-record statements from the people building these systems. Each link is a deep link that jumps to the quoted passage where the source supports it.

Google — Elizabeth Reid, VP/Head of Search (May 2024)

• “This means that AI OverviewsAI Overviews are the AI-generated summary box Google shows above or within its regular search results, written by Gemini models from pages retrieved out of Google's normal Search index. It's a Search feature, not a separate platform or index. generally don’t ‘hallucinate’ or make things up in the ways that other LLMA large language model (LLM) is a deep-learning model trained on massive text corpora to predict the next token and generate human-like text. LLMs use the transformer architecture and power AI search features like Google's AI Overviews (Gemini) and Bing Copilot (GPT-4). products might.” Read the post
• “When AI OverviewsAI Overviews are the AI-generated summary box Google shows above or within its regular search results, written by Gemini models from pages retrieved out of Google's normal Search index. It's a Search feature, not a separate platform or index. get it wrong, it’s usually for other reasons: misinterpreting queries, misinterpreting a nuance of language on the web, or not having a lot of great information available.” Jump to quote
• “Because accuracy is paramount in Search, AI Overviews are built to only show information that is backed up by top web results.” Read the post

Google DeepMind — GopherCite (March 2022)

• “Language models like Gopher can ‘hallucinate’ facts that appear plausible but are actually fake.” Read the post
• “Because language models can hallucinate convincing nonsense, GopherCite uses reinforcement learning from human preferences (RLHP) to train models that generate answers whilst also citing specific evidence for their claims.” Read the post

Anthropic — Claude developer documentation

• “Even the most advanced language models, like Claude, can sometimes generate text that is factually incorrect or inconsistent with the given context. This phenomenon, known as ‘hallucination,’ can undermine the reliability of your AI-driven solutions.” Read the docs
• “Remember, while these techniques significantly reduce hallucinations, they don’t eliminate them entirely. Always validate critical information, especially for high-stakes decisions.” Jump to quote

OpenAI — Help Center

• “Hallucinations are plausible but false statements generated by language models.” Read the article
• “Like any language model, ChatGPT can produce incorrect or misleading outputs, and sometimes it might sound confident—even when it’s wrong.” Read the article

Lin et al. — TruthfulQA (2021)

• “Larger models produce more imitative falsehoods because they are better at learning the training distribution.” Read the paper

Lanham et al. — Anthropic researchers (2023)

• “As models become larger and more capable, they produce less faithful reasoning on most tasks.” Read the paper

Mark Howard, COO, Time (via the Tow Center / CJR study)

• “It’s critically important how our brand is represented, when and where we show up.” Read the study

Hallucination types — cheat sheet

The four kinds you’ll meet

The numbers worth memorizing

• 3.6% — share of Ahrefs’ AI referral traffic that went to non-existent pages (Patrick Stox, Ahrefs).
• 2.87x — how much more often AI assistants send users to 404s vs. Google (Ahrefs, 16M URLs). ChatGPT clicked-404 rate 1.01% vs. Google 0.15%.
• 60%+ — AI searchAI search uses large language models and retrieval-augmented generation (RAG) to synthesize an answer from multiple sources rather than returning a ranked list of links. Examples include Google AI Overviews, ChatGPT Search, and Perplexity. citation error rate across 8 engines (Tow Center / CJR).
• ~85–90% — how grounded the best models are even when handed the sources (FACTS GroundingGrounding is anchoring an AI model's answer to source documents it retrieves at the moment you ask — not to the patterns frozen into its weights during training. Retrieval-Augmented Generation (RAG) is the most common way to do it.) — i.e. a 10–15% failure floor.
• 73–86% — hallucination reduction from giving models web-search/groundingGrounding is anchoring an AI model's answer to source documents it retrieves at the moment you ask — not to the patterns frozen into its weights during training. Retrieval-Augmented Generation (RAG) is the most common way to do it..
• 58% vs. 94% — best model vs. human truthfulness on TruthfulQA.
• 37–39% — answers in which some models repeated brand misinformation after a single seeded article (Ahrefs brand experiment).

Myth → reality

• “RAGRAG is the retrieve-then-generate pattern behind AI search: the system retrieves relevant passages from an external index at query time, injects them into the model's context, and generates an answer grounded in those sources — without changing the model's weights. eliminates hallucinations.” → It reduces them; grounded models still fail ~10–15%.
• “Newer/bigger models hallucinate less.” → TruthfulQA: bigger can mean more convincing falsehoods.
• “AI OverviewsAI Overviews are the AI-generated summary box Google shows above or within its regular search results, written by Gemini models from pages retrieved out of Google's normal Search index. It's a Search feature, not a separate platform or index. hallucinate just like ChatGPT.” → They’re grounded; errors skew toward misinterpretation, not pure confabulation (Reid).
• “Being cited means being represented correctly.” → 60%+ citation error rate says no.
• “Blocking AI crawlersAI crawlers are bots from AI companies that fetch web pages to train language models, build AI-search indexes, or answer live user questions. They come in three categories, each with its own user-agent tokens and its own robots.txt controls. stops brand hallucination.” → Facts already in the weights stay there.
• “You can tell it’s wrong because it sounds unsure.” → Confidence is the defining feature; ChatGPT flagged uncertainty in only 15 of 134 wrong answers.

How to reduce hallucination — the playbook

These are the levers, in roughly the order of leverage, drawn from what the model makers and the research actually support.

1. Ground the answer (and accept the ceiling). Give the model real sources at inference time — that’s RAGRAG is the retrieve-then-generate pattern behind AI search: the system retrieves relevant passages from an external index at query time, injects them into the model's context, and generates an answer grounded in those sources — without changing the model's weights. / groundingGrounding is anchoring an AI model's answer to source documents it retrieves at the moment you ask — not to the patterns frozen into its weights during training. Retrieval-Augmented Generation (RAG) is the most common way to do it.. It’s the single biggest lever (73–86% reduction). But internalize the ceiling: the best grounded models still fail ~10–15% (FACTS GroundingGrounding is anchoring an AI model's answer to source documents it retrieves at the moment you ask — not to the patterns frozen into its weights during training. Retrieval-Augmented Generation (RAG) is the most common way to do it.), and they fail by misreading or mis-selecting sources, not by inventing from nothing. Grounding changes the failure mode; it doesn’t remove it.

2. Let the model say “I don’t know.” Anthropic’s first recommendation: “explicitly give Claude permission to admit uncertainty… this simple technique can drastically reduce false information.” GopherCite improved sharply once it could abstain. An answer-shaped response to a question it can’t support is the core failure; permission to abstain attacks it directly.

3. Demand citations and verbatim quotes. Make the output auditable: have the model quote sources for each claim, and for long documents extract word-for-word quotes before reasoning. You’re not trusting the citation blindly (60%+ are wrong) — you’re making the claim checkable.

4. Don’t trust the reasoning trace as proof. Chain-of-thought can be post-hoc rationalization (Lanham et al.). “It explained its reasoning” is not “it grounded its answer.” Verify the conclusion, not the narration.

The SEO-side framework — protect your entity

Think of it as three jobs, because you can’t change the models:

1. Detect — monitor AI answers about your brand and watch analytics for traffic to 404s. (Fix attribution first, or hallucinated-URL traffic is invisible.)
2. RedirectA redirect sends browsers and crawlers from a requested URL to a different one. An HTTP redirect specifically is a 3xx status code paired with a Location header; meta refresh and JavaScript redirects achieve a similar navigation without being a 3xx response themselves. Permanent redirects (301/308) are Google's signal the target should be canonical; temporary ones (302/303/307) aren't. — 301 hallucinated URLs above a traffic threshold to the closest real page. Capture the intent the model is misdirecting.
3. Ground the truth for them — clean, consistent, verifiable public facts (schema, Wikidata, official profiles, an unambiguous About page) so grounded systems retrieve your facts instead of a competitor’s claim or a satirical article. This is the only durable defense against brand hallucination — blocking crawlersA crawler — also called a spider or bot — is an automated program that fetches web pages, extracts their links, and queues new URLs to visit. Search engines use crawlers to discover and download content for their index. isn’t, because the weights already hold what they learned.

Routine hallucination review

1. Define high-risk claims: pricing, legal or safety statements, product capabilities, executive names, locations, and official URLs.
2. Create stable prompts for those claims and record the platform, model, mode, date, locale, account state, and whether live search was enabled.
3. Save the complete response, citations, and source URLs. A paraphrased note is not enough evidence for later comparison.
4. Classify each statement as supported, contradicted, unverifiable, stale, or not applicable using an owned source of truth and cited sources.
5. Check fabricated URLs against logs, backlinks, and analytics before deciding whether to redirectA redirect sends browsers and crawlers from a requested URL to a different one. An HTTP redirect specifically is a 3xx status code paired with a Location header; meta refresh and JavaScript redirects achieve a similar navigation without being a 3xx response themselves. Permanent redirects (301/308) are Google's signal the target should be canonical; temporary ones (302/303/307) aren't., publish a page, or leave a true 404.
6. Assign the fix to owned content, technical routing, third-party correction, or platform feedback, then schedule a dated retest.

Incident playbook: an AI answer states a harmful falsehood

1. Capture evidence. Save the full answer, prompt, sources, product/model, date, account state, and a screen recording if the behavior is transient.
2. Verify impact and truth. Have the responsible subject-matter owner identify the exact false claim and the authoritative fact. Do not “correct” a disputed claim by publishing another unsupported statement.
3. Trace likely sources. Review cited pages, your own current and archived content, structured facts, feeds, profiles, and prominent third-party mentions.
4. Fix controlled evidence. Correct contradictory owned pages, consolidate entity facts, and update stale documentation without changing dates cosmetically.
5. Handle URLs deliberately. RedirectA redirect sends browsers and crawlers from a requested URL to a different one. An HTTP redirect specifically is a 3xx status code paired with a Location header; meta refresh and JavaScript redirects achieve a similar navigation without being a 3xx response themselves. Permanent redirects (301/308) are Google's signal the target should be canonical; temporary ones (302/303/307) aren't. a fabricated URL only when a real equivalent satisfies the intent. Otherwise preserve the correct missing-page response and avoid teaching users that every invented path resolves.
6. Report externally. Use the platform’s feedback channel and request corrections from third parties when evidence supports it.
7. Retest and monitor. Repeat the same prompt in documented conditions and keep the incident open until the answer is corrected or the residual risk is accepted.

Prompt: fact-check an AI answer against supplied evidence

Audit the AI answer below using only the supplied source-of-truth documents. Break the
answer into atomic factual claims. For each claim, label it supported, contradicted,
not found, stale, or opinion; quote the exact supporting or conflicting passage; and
state the smallest correction. Treat a citation as evidence only if its page supports
the adjacent claim. Do not fill gaps from memory or the open web.

AI answer:
[PASTE COMPLETE ANSWER AND CITATIONS]

Approved source-of-truth documents:
[PASTE DOCUMENT NAME, DATE, URL OR EXCERPT]

Prompt: triage fabricated URLs

Classify these AI-referred URLs using the supplied HTTP, backlink, analytics, and page
inventory evidence. For each URL, recommend one of: keep a true 404/410, redirect to a
genuinely equivalent page, restore a previously valid page, or investigate further.
Never recommend redirecting to the homepage just to remove a 404. Explain which user
intent and evidence make a destination equivalent.

Evidence:
[PASTE URL | STATUS | REFERRERS | BACKLINKS | VISITS | CLOSEST REAL PAGE]

DevTools Console: extract visible citations from an answer page

Run this on a saved or accessible answer page after checking the platform’s terms. The selectors are intentionally generic and may need adapting to the interface.

copy(JSON.stringify([...document.querySelectorAll('a[href]')]
  .map(a => ({text: a.textContent.trim(), url: a.href}))
  .filter(x => /^https?:/.test(x.url)), null, 2));

Shell: check a list of suspected fabricated URLs

Run against URLs you own or are authorized to test.

while IFS= read -r url; do
  curl -sS -o /dev/null -L --max-redirs 5 -w '%{http_code}\t%{url_effective}\t%{url}\n' "$url"
done < suspected-urls.txt

Regex: flag claim-shaped numbers for review

This review regex finds percentages, currency amounts, and multi-digit numbers. It does not decide whether a claim is wrong.

(?:[$€£]\s?\d[\d,.]*|\b\d+(?:\.\d+)?%|\b\d{2,}(?:[,.]\d+)*\b)

Tools for investigating hallucinations

• AI Brand Visibility captures a labeled model response and keeps mentions separate from citations for local comparison.
• Citation Gap Checker flags unsupported numerical and research-style claims in content you control. It is a review queue, not a truth engine.
• Redirect Checker shows the status and final destination of a fabricated or stale URL before you decide how to handle it.
• Server/CDN logs reveal requests to invented paths, while backlink and analytics tools show whether people or other sites are actually using them.
• Platform feedback controls are the direct reporting route, but keep your own evidence because correction timing and persistence are not guaranteed.

Validate a hallucination response

Hallucination metrics

Test yourself: AI Hallucinations

Resources worth your time

My related writing (Ahrefs)

• 80% of Our AI Search Traffic Goes to Our Homepage, Product Pages, and Free Tools — where the 3.6% hallucinated-pages finding comes from.
• AI Assistants Are Breaking Web Analytics and Hurting Their Future — why hallucinated traffic can be invisible.
• Google Made It So You Can’t Track Clicks From AI Mode (Partially Fixed) — the noreferrer attribution bug I reported.

From others

• How Often Do AI Assistants Hallucinate Links? (16 Million URLs Studied) — Ryan Law & Xibeijia Guan (Ahrefs); the 2.87x 404 figure.
• I Ran an AI Misinformation Experiment. Every Marketer Should See the Results — Makosiewicz (Ahrefs); the seeded-Medium-article brand attack.
• AI Search Has a Citation Problem — Tow Center / Columbia Journalism Review (2025); the 60%+ citation error study. Nieman Lab summary.
• Survey of Hallucination in Natural Language Generation — Ji et al.; the intrinsic vs. extrinsic taxonomy.
• TruthfulQA — Lin et al.; the inverse-scaling finding.
• Measuring Faithfulness in Chain-of-Thought Reasoning — Lanham et al. (Anthropic).
• Vectara Hallucination Leaderboard — ongoing grounded-summarization hallucination rates.
• What are AI hallucinations? — Google Cloud’s official explainer of the mechanism and mitigation strategies.
• How to identify and fix AI hallucinations about your brand — Search Engine Land’s practical guide to monitoring and correcting AI brand errors.
• The AI feedback loop: researchers warn of model collapse as AI trains on AI-generated content — VentureBeat on model collapse and how hallucinated content compounds across training cycles.

Stats worth citing

• 3.6% of AI referral traffic went to pages that don’t exist on ahrefs.com — hallucinated URLs, primarily from ChatGPT. My first-party finding. Source
• AI assistants send users to 404s ~2.87x more often than Google Search — across 16M AI-cited URLs. ChatGPT clicked-404 rate 1.01% vs. Google’s 0.15%. Source
• Over 60% citation error rate across 8 AI search enginesAI search uses large language models and retrieval-augmented generation (RAG) to synthesize an answer from multiple sources rather than returning a ranked list of links. Examples include Google AI Overviews, ChatGPT Search, and Perplexity.; ChatGPT flagged uncertainty in only 15 of 134 wrong answers (Tow Center / CJR, 2025). Source
• ~85–90% grounded is the ceiling for the best models given the sources — a 10–15% failure floor (FACTS GroundingGrounding is anchoring an AI model's answer to source documents it retrieves at the moment you ask — not to the patterns frozen into its weights during training. Retrieval-Augmented Generation (RAG) is the most common way to do it., DeepMind, 2024). Source
• 73–86% hallucination reduction from web-search/groundingGrounding is anchoring an AI model's answer to source documents it retrieves at the moment you ask — not to the patterns frozen into its weights during training. Retrieval-Augmented Generation (RAG) is the most common way to do it. access; 58% best-model truthfulness vs. 94% human on TruthfulQA. Source
• 37–39% of answers repeated brand misinformation after a single seeded source (Ahrefs brand experiment). Source

AI Hallucinations

An AI hallucination is when a large language model generates output that is confidently stated but factually wrong, made up, or unsupported by its source. It's a side effect of next-token prediction — not a bug that can be fully eliminated.

Related: Grounding, Retrieval-Augmented Generation (RAG), Knowledge Cutoff

AI Hallucinations

An AI hallucination is when a large language modelA large language model (LLM) is a deep-learning model trained on massive text corpora to predict the next token and generate human-like text. LLMs use the transformer architecture and power AI search features like Google's AI Overviews (Gemini) and Bing Copilot (GPT-4). produces text that sounds correct but isn’t — a confident, plausible-looking answer that is factually wrong, fabricated, or unsupported by any source. The term comes by loose analogy from psychiatry; some researchers prefer confabulation (from neuropsychology, where it means generating false accounts with genuine confidence and no intent to deceive), because it doesn’t imply the model has sensory experiences. The model isn’t lying — it’s predicting statistically likely tokensA token is the smallest unit of text (or image/audio/video) an LLM processes — roughly 4 characters, or about ¾ of an English word. A context window is the maximum number of tokens (input plus output) a model can hold at once, like its short-term memory., and sometimes the likely token is false.

The root cause is architectural. LLMs are trained to predict the most probable next token given everything before it — a language-completion objective, not a fact-retrieval one. When the model lacks a strong training signal for a specific fact (rare events, niche topics, recent changes), it fills the gap with statistically probable language that can sound completely authoritative while being wrong.

The taxonomy SEOs care about: factual (wrong facts), citation/attribution (invented sources, misattributed quotes, fabricated URLs), entity/brand (wrong names, dates, stats, or company facts), and the academic split of intrinsic (output contradicts the provided source) vs. extrinsic (output can’t be verified from the source). GroundingGrounding is anchoring an AI model's answer to source documents it retrieves at the moment you ask — not to the patterns frozen into its weights during training. Retrieval-Augmented Generation (RAG) is the most common way to do it. and RAGRAG is the retrieve-then-generate pattern behind AI search: the system retrieves relevant passages from an external index at query time, injects them into the model's context, and generates an answer grounded in those sources — without changing the model's weights. reduce hallucination substantially but don’t eliminate it — a grounded model can still misread, misattribute, or extrapolate from the documents it retrieves.

Related: Grounding, Retrieval-Augmented Generation (RAG), Knowledge Cutoff

Changelog