When an AI draft cites a source you cannot find
Hallucinated references are the most common reason editors distrust generative drafts. Most of them can be sorted in two minutes if you know which symptom you are looking at.
Start with the symptom, not the tool
The model said a study exists, and you cannot find it. That is the whole problem. Most cases can be sorted in two minutes once you know which symptom you are looking at, and the next table is organised by symptom.
Language models predict text that looks like a reference, which is why a fabricated link often has a believable domain, a tidy path and a plausible title. A quick way to check DR domain by domain separates three cases. The domain is real and busy, real but tiny, or without any link footprint at all. Three cases only.
The table below maps each case to a next step. No score replaces opening the page, so the table assumes that a human reviewer is sitting at the desk with time to read the source before approving a sentence. For a whole list at once, use the bulk workflow. Fine.
Troubleshooting table
| Symptom | Likely cause | What to check | Editor action |
|---|---|---|---|
| URL returns 404 on a domain with many referring domains | The model guessed a path on a real site | Search the site for the topic by title | Replace with the real page or drop the claim |
| Domain scores 0 and has no referring domains | Invented domain, parked domain, or brand-new site | Open it, then look at the registration age in a WHOIS tool | Treat as unverified until a human confirms an author and date |
| Domain exists, DR in single digits, content looks scraped | Low-quality republisher | Compare it with the likely original source | Link the original, not the copy |
| Real DR but the quote is not on the page | Model paraphrased from memory | Search the page text for the key figure | Quote only what you can see, with the date |
| Redirects to an unrelated site | Domain expired and was bought | Check where the redirect ends | Remove the link and note why |
| Same source cited with different years in one draft | Model blended two documents | Find the original and its date | Correct the year, re-read the claim |
Two calls that settle most rows
The first call is free and needs no key. The second lists who links to a domain you doubt and costs 2 credits, which is $0.0002.
# 1. Free single lookup
curl "https://genaicontent.org/api.php?action=domain&q=suspicious-source.example"
# 2. Who links to it? (2 credits)
curl "https://genaicontent.org/api.php?action=refdomains&q=suspicious-source.example&per=50&key=sk_live_..."A young journal, a regional government page or a small clinic blog can all score low and be right. The check tells you where to spend your reading time. It never tells you what to believe.
Questions from reviewers
How do I check the DR of a domain an AI gave me?
Paste it into the free box on the home page or call the domain action with no key. You get a 0 to 100 score from our own index. Then open the actual URL, because a score cannot confirm that the cited page says what the model claims.
Why does a real site show DR 0?
Either it is very new, or the index has not yet seen its links. We refresh about once a month, so recent sites can lag. Treat DR 0 as unverified, and ask for a human check of the author and date.
Can I ask the model to fix its own citations?
You can, but it often produces a fresh set of equally plausible errors. Better to give it sources you have already verified and tell it to use only those, then review the output against the list.
Is it worth checking every link in every draft?
At $0.0001 for up to 2,000 domains per bulk call, cost is not the obstacle; reviewer attention is. Check them all automatically, then read only the ones the numbers flag.