How to Check an AI Answer's Sources
The traps of search-connected AI
① Search-connected AI attaches links and footnotes, but a link being present is not the same as the link supporting the claim. In one measurement, 45% of AI news answers had a significant issue, and sourcing (31%) was the biggest cause.
② Sources fail in four ways: a dead link, a link that doesn't back the claim, a copy instead of the original, and stale or undated evidence. The first two are caught by clicking; the last two take one more look.
③ The fix isn't verifying every link — it's the habit of picking only numbers, quotes, dates and causal claims, and checking them against the source in 60 seconds.
① Getting started — a footnote doesn't mean it's verified
Our AI hallucination guide made this point in a single paragraph: "a source being present is the start of verification, not the end." This piece unpacks that paragraph. AI with a search feature now ends answers with numbers like [1], [2] and links, and seeing them, almost anyone relaxes — "there's evidence." The problem is that someone has actually measured how justified that relaxation is, and it's lower than you'd expect.
② How does search-connected AI attach sources?
You need the structure to see where it slips. Most search-connected AI works in the order search → read documents → summarize → footnote the documents it read (the same skeleton as the four stages in RAG explained). The footnote points to "a document pulled in while building this answer."
Fetching the document is search, but reading it and turning it into sentences is still next-token prediction. So there is no structural guarantee that the sentence carrying a footnote number actually says that in the document. A number can change, two documents' contents can blend into one sentence, or an interpretation that isn't in the document can slip in — and the footnote stays attached anyway.
The trap is that a footnote makes you trust it more. You'd doubt an unsourced answer, but the same content with a [1] on it easily feels already checked. The number tells you where to verify, not that verification happened.
③ How often is it wrong? Two measurements
There are numbers behind "sourcing is a problem." Here are two studies with different methods, side by side.
| Study | Method | Result |
|---|---|---|
| EBU–BBC (Oct 2025) | Journalists at 22 public-service media organizations (18 countries, 14 languages) rated 2,709 news answers from the free versions of ChatGPT, Copilot, Perplexity and Gemini on five criteria including accuracy and sourcing | At least one significant issue in 45% (81% counting minor ones). The biggest cause of significant issues was sourcing, 31%. Gemini stood out at 72% for sourcing issues; the rest were under 25% |
| Tow Center (Mar 2025) | Eight AI search tools were given a passage from a news article and asked to find the headline, date, publisher and URL (200 queries per tool, 1,600 total) | Wrong more than 60% of the time. Results ranged from 37% (Perplexity) to 94% (Grok-3), and tools often stayed confident even when wrong |
In the EBU–BBC study, the sourcing criterion asked "are the claims in the response supported by the source the assistant provides?" It wasn't checking whether links were fake, but whether a link was present and still didn't match the claim. The report's foreword also cites separate BBC research in which just over a third of UK adults said they completely trust AI-generated summaries to be accurate, rising to almost half of under-35s. With that many people trusting without checking, the gap easily turns into real harm (sources: EBU–BBC, News Integrity in AI Assistants report; Nieman Lab's write-up of the Tow Center study).
Both studies covered news questions and measured services as of the first half of 2025. The Tow Center task ("trace a quote back to its source") is a harder, special-purpose task than a typical question, so it must not be read as "60% of AI answers are wrong." Services keep changing, so the point is less the exact rates than the direction: sources can be wrong, and the AI will say so confidently anyway.
④ Four ways a source fails
Grouping the failures from the measurements and from everyday use into four also organizes how to catch them.
| Type | What happens | How to catch it |
|---|---|---|
| ① A dead link | It attaches a broken or invented address. Coverage of the Tow Center study says some tools frequently produced such addresses | One click. The easiest to catch |
| ② A link that doesn't back the claim | The page opens, but the claim isn't on it, or the number or conclusion differs (the core of what the EBU study counted as a "sourcing problem") | Open it and find the sentence yourself |
| ③ A copy, not the original | It cites a republished version, a summarizing blog or a roundup instead of the original article. When it isn't the original publisher, you can't tell whether corrections or updates were made | Ask "where did this first appear?" and climb to the primary source |
| ④ Stale or undated evidence | Something that was true no longer is, or there's no date to tell what it's as of. Pages written to rank in search (SEO pages) sometimes surface as evidence too | Check the page's publication and update dates, and if the question is "as of now," that the material is current |
① and ② are caught by opening the link, but ③ and ④ have a working link and plausible content, so they're easy to miss. ③ in particular ties straight back to the principle in researching with AI — "AI is a signpost to sources, not the source" — you have to ask once more whether the signpost points to the destination (the primary source) or a waypoint (a copy).
⑤ The 60-second check
You don't need to open every link. Spend your time on claims that would change your decision if wrong — numbers, quotes, dates and "A because of B" causal claims. Apply the five steps below to those four only.
- Open the link. If it doesn't open or it's a completely different page, you're done (type ①).
- Search the page for the claim's key words with Ctrl+F (⌘F on Mac). Put the number or proper noun in the search box and see whether that sentence is really on the page. If not, it's type ②.
- Compare numbers, dates and units by eye. "12%" becoming "1.2%", or "quarterly" becoming "annual," is common.
- Judge whether it's the primary source. Is this page the original announcer or document, or a piece that repeats it? If it repeats, follow it one step further to the original (type ③).
- Look at the publication date. If the question is about "now," the material must be current (type ④).
A prompt can shave off some of this work.
"For each claim, quote word for word the sentence from the source page that supports it, and give the URL alongside. For any claim you couldn't find verbatim in the source, mark it separately as 'supporting sentence not found.'"
This prompt narrows where to check; it doesn't do the checking. A sentence the AI puts in quotation marks may not exist in the source either, so the job isn't finished until you copy that sentence and search for it on the page, as in step 2. If it turns up, the quote is real; if it doesn't, don't trust that quote.
⑥ How far up should investing information go?
For information that feeds an investment decision, a single number can change the call, so the rule is to climb the source ladder all the way up. Don't stop at an AI summary or an explainer piece; go to where it was announced, as below.
| What you're checking | Where to climb to (primary source) | If you trust only an AI summary |
|---|---|---|
| Company earnings and guidance | The company's investor-relations press release and quarterly filing (SEC filings for U.S. companies) | Revenue and profit figures or the "versus expectations" baseline are easily relayed wrong |
| Rates, the dot plot, Fed remarks | The Fed's official statement, Summary of Economic Projections (SEP) and press conference transcript | The nuance of remarks or the vote breakdown shifts in commentary |
| Jobs and inflation data | The issuing agency's release (e.g. the U.S. Bureau of Labor Statistics) | Preliminary and revised figures, month-on-month and year-on-year, get mixed up |
| Analyst price targets and ratings | The original report, or consensus data with dates shown | A price target from months ago can be relayed as if recent |
Among these, analyst views suffer most often from the date problem (type ④). How consensus differs from individual reports and why stale targets are dangerous is covered in how to read analyst reports, and reading release data from the source is covered in the economic indicators calendar, fully explained.
⑦ Practical checklist
□ Did I pick the claims that would change my decision if wrong (numbers, quotes, dates, causal claims)?
□ Did I actually open those claims' links?
□ Did I search the page for the claim's key words and find the sentence?
□ Do the numbers, dates and units match the source exactly?
□ Is that page the original announcer (primary source), or a piece repeating it?
□ Does the publication date fit the reference point of the question?
□ Did I tag anything I couldn't check as "AI summary" and keep it out of my reasoning?
※ The research figures here are measurements from 2025 under specific conditions (news-related questions, free/consumer versions, services as of that time) and don't directly describe current services. Investing examples explain how to check information and are not investment advice. Investment decisions and responsibility rest with the investor.
※ This guide is provided for general educational purposes and simplifies technical details for readability.
New guides, when they land
We publish AI literacy guides twice a week. Subscribe and the next one comes to you — free, unsubscribe anytime.