Catch broken and hallucinated links in AI output
Language models confidently cite URLs that look real but do not exist. This tool pulls every link out of a block of LLM output, removes duplicates, checks each one for valid format, and flags the patterns that usually signal a hallucinated or unusable link — so you can verify before you publish or click.
LLM hallucinated links follow recognizable patterns. Models tend to invent:
- Plausible but non-existent journal articles — a DOI or URL that resolves to a 404 because the paper doesn’t exist.
- Wikipedia pages that were never written — a valid-looking Wikipedia path for a topic that has no article.
- Invented API endpoints — a real base domain but a path the provider never created, common in code-generation and documentation tasks.
- Placeholder domains —
example.com/somethingoryoursite.com/pagecopied from training data.
How it works
A regular expression scans the text for both http/https URLs and bare www.
domains, trims trailing punctuation, and deduplicates the results. Each link is then
parsed with the browser’s URL API and checked against several heuristics: uncommon
or missing top-level domains, example.* placeholder domains, localhost and
loopback hosts, non-ASCII characters that can hide homograph attacks, unusually long
hostnames, and opaque path segments LLMs tend to invent. Valid, unremarkable links
are marked clean; anything questionable is flagged with the reason. Everything runs
client-side.
What the flags mean
| Flag | What it means |
|---|---|
| Uncommon TLD | The top-level domain is outside the common list — may be real but worth checking |
| Placeholder domain | Matches example.* or similar template domains from training data |
| Loopback/localhost | Points to a local address — never a valid public citation |
| Non-ASCII host | Contains characters that can disguise homograph phishing domains |
| Malformed URL | The browser URL parser rejected it — the link is broken regardless |
Tips and notes
A green check means the URL is well-formed, not that the page exists — the tool cannot fetch links, so always open flagged ones (and ideally the clean ones too) before trusting a model’s citations. The uncommon-TLD flag is intentionally conservative and will sometimes flag legitimate but rare domains; treat it as a prompt to look closer. Use the copy button to pull a clean, deduplicated list of links into a checker or your notes. Because nothing leaves your browser, it is safe to run on confidential drafts.
Checking links in bulk
After extraction, the deduplicated list of clean URLs is ready to copy into an
external link checker (Ahrefs Broken Link Checker, Screaming Frog, or a simple
curl loop) to verify reachability. The extractor handles the tedious step of
finding and deduplicating the URLs; reachability checking requires a network
request the browser tool deliberately avoids for privacy and security reasons.