The Sources an AI Gave You — Do They Actually Check Out? We Clicked 753

Ask an AI something today and the answer usually arrives with source links underneath. News articles have them; blog posts have them. But have you ever actually clicked one? Most people don’t. The mere presence of a link reads as “this is backed up,” and we move on.

So we clicked them. All of them. We reopened every source our blog had ever cited — 753 links in total, across 427 domains — and spent a day checking two things. One: is the link still alive? Two: if it’s alive, does the page actually contain what we wrote?

Here’s the short version. Almost all the links were alive. But when we opened a sample of the live ones, about 1 in 6 didn’t quite match the number next to it. “The link opens” and “the link backs up the claim” turned out to be different questions.

Key takeaways

• We reopened 753 citations (427 domains). 87.4% loaded normally, and only 1.1% were fully dead. Link rot was lower than we expected.
• But 11.6% were alive yet refused automated access (bot-blocked). A link that looks fine to a person can be invisible to a crawler or an AI.
• In a 12-link sample of the live ones, 10 backed the claim exactly and 2 didn’t. An alive link is not the same as a supporting link.

A note before we start — this isn’t an exposé of any tool or outlet It’s a record of us auditing our own sources, plus a source-checking method anyone can use. The sample isn’t exhaustive, so treat the rates as estimates.

Why we audited our own citations

As AI answers started carrying sources, “has a source” quietly became shorthand for “trustworthy.” The trouble is that producing convincing form — a number, an institution, a year, a URL — is exactly what language models are good at. A blue underlined link is the most convincing form of all, so most readers assume it’s been checked.

The way to find out is to count. And we had a good dataset to count: our own 753 citations, each one added while fact-checking a post. Before pointing at anyone else’s sources, auditing the ones we know best felt like the honest place to start.

The experiment — what we measured, and how

The method is simple. We’re publishing the criteria up front so it can be reproduced.

Item	Detail
Scope	Every citation URL in the reference lists of posts published by 2026-07-10
Sample	753 unique URLs · 427 domains
Survival test	Requested each link with a browser User-Agent and recorded the HTTP status (all 753)
Evidence test	Opened 12 links from distinct domains and checked whether our cited figure appears on the page
Verdicts	Full match / Partial (figure or timing off) / No match (absent or contradicted)

Survival was checked by machine across all 753; evidence-match needs a human to read each page, so we capped it at a 12-link sample. Treat the survival rate as firm and the evidence-match rate as an estimate.

Result 1 — how many links were still alive

First, survival. Of 753 links, 87.4% loaded normally. Fully dead links — page gone or unreachable — came to just 8, or 1.1%.

Link rot in the low single digits is on the low end. For comparison: a 2024 Pew Research Center study found that 38% of webpages that existed in 2013 were no longer reachable by 2023 — more than one in three vanishing over a decade. Our survival rate is high for a boring reason: most of our links are recent, and we click each one at publish time. Put differently, the longer a source sits unchecked, the more likely it is to be gone.

Result 2 — alive, but invisible to machines

The more interesting group was the second one. Of 753 links, 87 — 11.6% — loaded fine for a person in a browser but returned a refusal (mostly a 403) to automated access. Bot-blocking. Plenty of news sites, government pages, and outlets fence off crawlers and automated tools.

Why does that matter? Search engines, AI summaries, and source-checking tools read pages with programs, not human eyes. If a page blocks bots, the tool simply can’t read that source. So an AI may skip the summary, or attach the link while filling the substance in incorrectly. A source that’s perfectly readable to you can be a wall to the machine — the flip side of the whole “will AI find and read my page” question.

Result 3 — an alive link isn’t automatically evidence

Now the real question: does an open link actually contain what we wrote? We opened a 12-link sample from distinct domains — academic institutions, think tanks, government statistics, news outlets, and asset managers — and compared. Most held up exactly: Stanford’s youth-employment figure and Indeed’s “5+ years of experience” data, for instance, matched our citations word for word. The two that didn’t are below.

Nothing was flat-out wrong — after fact-checking, most sources held up. But about 1 in 6 came back “partial”: the link is alive and points at a relevant page, yet our number and the page’s number don’t line up cleanly. Both cases came from our own citations, so we’ll be candid about them.

Type of drift	What we wrote	What the page actually said	Why it drifted
Intraday vs close	Pinned a stock index’s “−9.99% crash” to one wire story	That story (filed mid-session) reported the intraday drop of about −8% at the circuit-breaker moment	The number moves within a day; the −9.99% close was in a different article
Over-attribution	Attached three figures (a 43.8% rate, a “25 months straight” streak, a +13.6% jump) to one news link	Only the 43.8% was on that page; the other two came from the statistics office	Several numbers seen across sources got piled onto one link

Neither is a lie. The index really did close down 9.99%, and the 43.8% figure is real. But saying “this link is the source for this number” isn’t accurate. The link is alive; as evidence it’s half a step off — and that gap is the whole point of this experiment.

Why links and evidence drift apart

The drift isn’t laziness; it comes from how information flows. Four forces overlap.

First, the distance between the original and the retelling. A statistics agency publishes raw data, a news outlet rewrites it, another post cites that — and the number shifts a little at each hop. A live link often points to a re-citation of a re-citation.

Second, timing. For anything that changes within a day — markets, exchange rates, prices — the same event reads differently depending on whether the story was filed mid-session or at the close. The link is right; the moment is wrong.

Third, rounding and ranges. An original “16%” becomes “about 15%” in one retelling, then “mid-teens” in the next. Precision erodes with every step.

Fourth, pages change quietly. A news page may run an early version and then overwrite it once the close is in. Yesterday’s number on that URL can be a different number today — same link, new contents.

So “the link opens” only guarantees that the page exists. Whether it contains the claim is something you can only learn by opening it and looking.

The 30-second source check

Whether it’s someone’s article or an AI’s answer, here’s a check that takes about half a minute and needs no special tools.

Step	Check	How
1. Click it	Is the link alive?	Actually open it. A 404 or timeout means no evidence
2. Read the domain	Is it the original?	Is it the institution/outlet’s own page, or a re-citation?
3. Find the number	Is the figure there?	Use Find-in-page (Ctrl/Cmd+F) to search the page for that exact number
4. Check timing & unit	Is it the right number?	Do the date, period, and unit (percentage points vs percent, intraday vs close) match the claim?

Step 3 is the one that matters. Open the page and search for the number yourself. That single move separates an “alive link” from a “supporting link” — and it works just as well on the figures an AI drops into an answer. Search for the number; if it isn’t there, you’ve caught convincing form standing in for real evidence.

What we’re fixing

This audit left us homework too. The two partial citations — an intraday figure written as a close, and several numbers stacked on one link — were corrected right after we ran the check, so each figure now points to a source that actually carries it. Going forward, the “many numbers, one link” habit becomes “one number, one source.” And one more admission: while double-checking this very experiment, we caught an error in our own verification — we first filed one citation as “off,” then reopened the page and found it matched the original exactly, so we corrected it. Checking, it turns out, isn’t a one-and-done. We relearned that writing this piece.

FAQ

If a source link works, can I trust it?

Not by itself. We reopened all 753 of our own citations and 87.4% loaded fine — but when we checked a sample of the live ones, 1 in 6 didn’t match the number we’d written. Usually it was an intraday figure quoted as a closing figure, or several claims piled onto one link. “The link opens” and “the page backs up the claim” are two different things.

Why be extra careful with links an AI puts in its answer?

Because AI is good at producing convincing form — numbers, institution names, tidy URLs. Even when the link is real and loads, whether the page actually contains what the AI said is a separate question. Sometimes the link points to a real page that says something else. Treat a citation as a claim to verify, not proof.

How common is “link rot,” where a link dies?

More common than you’d think. A 2024 Pew Research Center study found that 38% of webpages that existed in 2013 were no longer accessible by 2023. In our own citations only 1.1% were fully dead — because most were added recently and we recheck sources at publish time. Older links are far likelier to be gone.

What does “alive but blocked” mean?

Some sites load fine for a person in a browser but refuse automated programs — crawlers, search bots, some AI fetchers. About 11.6% of our links were like this. So an AI or checking tool may fail to read that source and either skip it or fill the gap incorrectly.

Is there a fast way to check a source link?

Four steps: click it and see if it loads; check whether the domain is the original source; use Find-in-page to search for the exact number; confirm the date and unit match. The “30-second source check” table above lays it out.

Conclusion

A source link looks like a badge of trust, but it doesn’t guarantee one. Reopening our own 753 citations, we found the links were alive more than 87% of the time — yet alive wasn’t the same as accurate. About 1 in 6 of the live links we sampled sat half a step off the number beside it, and roughly 1 in 10 was visible to people but closed to machines.

Next time you see a source link — under an AI answer or in an article — click it, then search that page for the number in question. Those thirty seconds are the difference between “it has a source, so it’s probably fine” and “I checked, and the evidence is there.” As information gets cheaper, the habit of looking past the form to the substance is what holds its value.

Sources

• META TOUR own experiment. All 753 unique citation URLs (427 domains) from our published reference lists were requested; a 12-link sample was opened and compared against our cited figures. Run 2026-07-10. (Method and criteria disclosed in the article and disclosures.)
• Pew Research Center. When Online Content Disappears (38% of 2013 webpages unreachable by 2023; 25% of pages collected 2013–2023 gone). 2024-05-17. Accessed 2026-07-10. View source
• Stanford Digital Economy Lab. Canaries in the Coal Mine? (roughly 16% relative employment decline for ages 22–25 in the most AI-exposed jobs — verified as a full match in our sample). Accessed 2026-07-10. View source
• Indeed Hiring Lab. Experience requirements have tightened (share of tech postings requiring 5+ years rose 37% → 42% — verified full match). 2025-07-30. Accessed 2026-07-10. View source