LLM Web Search Is Inaccurate More Often Than You Think

LLM web search is inaccurate when the model builds its answer from a handful of search results it skimmed seconds earlier and something goes wrong along the way. The search can surface weak pages, and the model can misread good ones. Both mistakes arrive in the same confident tone.

The error rates are high enough to change how you use these tools. In a 2025 test by the Tow Center for Digital Journalism, eight AI search tools answered more than 60% of 1,600 queries incorrectly when asked to identify the source of a news excerpt. A separate study led by the BBC and the European Broadcasting Union found at least one significant issue in 45% of news answers.

AI search still saves time, provided you treat each answer as a lead to check. Here is where each kind of error comes from, and how to catch it in about a minute.

How LLM web search builds an answer

An LLM, short for large language model, gathers facts very differently from a person at a keyboard. The steps below follow Anthropic's documentation for its Claude web search tool, and other assistants work on the same broad pattern:

1. The model decides whether the question needs a search at all.
2. It writes a short search query, and may repeat this several times in one answer.
3. The results are loaded into its context window, meaning the text the model can hold in view while it writes.
4. The model writes the answer and attaches citations to the sentences it believes came from a source.

Each step can fail on its own, and the final answer sounds equally sure of itself either way. Grounding is the term for tying a model's output to retrieved text, and every step in that list is a place where the grounding can slip.

The model reads the web the way most people read terms and conditions: a fast skim, followed by complete confidence about what it said.

The citation format shows how thin the link to the source can be. In Claude's tool, each citation carries a cited_text field holding up to 150 characters of the source. That is about two lines, enough to confirm a number and rarely enough to confirm its context.

Chat apps add their own layer on top of this, from hidden instructions to memory. Our explainer on what ChatGPT wraps around its LLM separates the model from the product around it.

Retrieval failures: data voids and LLM SEO

If the search returns a weak page, the model summarises a weak page. Many inaccurate answers are decided at this stage, before the model has written a word.

Google's own account of its AI Overviews errors, published in May 2024, gives the clearest example. For the query "How many rocks should I eat?", the only substantial content on the web was satire, republished on a geological software company's site, and the overview drew on it. Google calls this a data void: a question so rare that low-quality pages are the only ones available to answer it.

The system did its job there. It found the most relevant page on the internet about eating rocks, and the internet had never written a serious one.

Your own questions land in data voids more often than you might expect. Anything niche or very recent qualifies. Copies cause a quieter version of the same problem: the Tow Center found tools citing syndicated reprints in place of the original, including a Yahoo News copy of a USA Today story, and DeepSeek misattributed the source in 115 of its 200 answers.

LLM SEO adds pressure from the publishing side. Sites now write pages designed to be lifted into AI answers, with a direct statement under every heading. Sounding like an answer and being right are separate properties, and a retrieval system can measure only the first one directly. Our guide to brand visibility in AI search engines covers those mechanics from the publisher's chair.

A page can also carry text meant for the model alone, such as instructions hidden in white-on-white type. A reader never sees it, yet the model reads everything that lands in its context window. The attack is called prompt injection, and it is one more reason to open the source yourself.

AI search citations that point at the wrong thing

A citation proves only that the model looked at a page. Whether the page supports the sentence is a separate question.

Stanford researchers measured that gap in 2023. Auditing four generative search engines, including Bing Chat and Perplexity, they found that only 51.5% of generated sentences were fully supported by their citations, and 74.5% of citations supported the sentence they were attached to. The answers read as fluent and informative, which the authors warned gives a misleading impression of trustworthiness.

The mechanism is a blend. The model holds two kinds of knowledge at once: what it learned in training, called parametric memory, and the search results in its context window. When the results are thin, it can fill the gaps from memory, and the citation lands on whichever source sits closest to the topic.

The 2025 Tow Center test found the same pattern in sharper form. Grok 3 answered 94% of queries incorrectly, and 154 of its 200 citations led to error pages. ChatGPT misidentified 134 articles, yet it signalled low confidence only 15 times in 200 responses and never declined to answer.

Confidence is the cheapest thing a language model produces.

Asking for LLM news today: where freshness breaks

News is the hardest case for AI search, because the right answer changes by the hour and the model's own sense of "now" is fixed in the past.

Every LLM has a training cutoff, the date after which it saw no data. Ask about LLM news today and the model depends entirely on what the search returns, while its memory keeps offering an older version of events. When the retrieved pages are stale or undated, the older version can win.

The largest test to date bears this out. In October 2025, the European Broadcasting Union and the BBC published a review of more than 3,000 news answers from four assistants, including ChatGPT and Gemini, scored by journalists at public broadcasters in 18 countries. Of those answers, 45% had at least one significant issue.

Sourcing problems affected 31% of answers, making them the largest single cause. One in five contained major accuracy issues, including hallucinated details and outdated information. Gemini fared worst, with significant issues in 76% of its responses.

Dates on the page matter as much as dates in the model. Anthropic's tool returns a page_age field showing when a site was last updated, and a page refreshed yesterday can still carry a story from 2023. The model can easily miss that difference, and it will brief you, calmly and in full sentences, on last spring.

An LLM prompt that makes web search answers checkable

The aim is an answer whose failures are visible. Asking for quotes and dates does most of that work. Paste this ahead of any question where the facts matter:

Today is 17 October 2026. Search the web and answer: [your question]

For every factual claim:
- quote the exact sentence from the source that supports it
- give the source URL and its publication date
- mark the claim "unsupported" if no source states it directly

Prefer primary sources, such as official documentation or the
original report. If sources disagree, show both and note which
is more recent. If you cannot find an answer, say so instead of
filling the gap from memory.

The quote requirement does the heavy lifting. A model can attach a citation to anything, but a verbatim sentence forces it back to the retrieved text. A quote your browser's find command cannot locate on the cited page is a fabrication you caught in ten seconds.

Then check the answer in three moves:

1. Open the cited page and search it for the quoted sentence.
2. Compare the publication date with the date of the event you asked about.
3. Trace any number back to the organisation that produced it.

Tool-specific search prompts follow the same logic, as our guide to Grok AI search prompts shows for finding posts on X.

If you build on an API, you can enforce some of this in code. The Claude web search tool accepts an allowed_domains list that restricts results to sources you trust, plus a max_uses cap on searches per request. Citations are always on, so your application can compare each cited_text with the claim it supports before a user sees the answer. Designing that verification step around a model is ground covered in AI Engineering, which treats prompts and agent tools in one volume.

The one-minute rule for AI search answers

Treat every AI search answer as a tip from a well-read stranger: often right, and always worth a minute of checking before it reaches a report or a decision.

Spend that minute on the claims that carry weight. The more recent the facts, the more likely the answer is stale, and the more specific the number, the more it matters that you find it on the source page. If an answer recommends eating rocks, check that one twice.