Seventy-six percent of marketing leaders now spend at least three hours a week editing, fact-checking or correcting AI-generated output — and when asked what creates the most extra work, 48% named fact-checking and hallucination review, ahead of every broken-tooling complaint on the list (Optimizely, 2,003 marketing leaders across seven markets, fielded May–June 2026). "AI was supposed to give marketers room to think," said Optimizely's SVP of Marketing Tara Corey. "What most teams got instead was more to manage."
Three hours a week is the tax. This post is about paying less of it without lowering the bar — a repeatable 20-minute pass over a draft that catches the errors most likely to embarrass you, in the order that costs you the least time.
The organizing idea is simple: most of the damage in an AI draft comes from its citations, not its opinions. A wrong-ish opinion gets edited in a sentence. A citation that points at nothing gets screenshotted, sanctioned or retracted. And citations have a property almost nothing else in a draft has — they are mechanically checkable. As one 2026 audit of hallucinated references put it, a reference "either resolves to a real scholarly work with compatible authorship, or it does not" (Russinovich et al.). That auditability is why verification should start with the links.
There isn't one failure mode. There are three, and they need different checks.
1. The URL never existed. Across 10 commercial models and deep-research agents, 3–13% of citation URLs were hallucinated and 5–18% failed to resolve at all, measured over 53,090 URLs in one dataset and 168,021 in another (Rao, Wong & Callison-Burch, University of Pennsylvania). The paper is precise about the relationship between the two numbers: "Hallucinated URLs form a strict subset of non-resolving URLs." Every fabricated link is broken; not every broken link was fabricated.
And the configurations you'd expect to be safest weren't: the single highest hallucinated-URL rate in the audit, 13.3%, came from a deep-research configuration that reads the web before answering. Retrieval helps; it does not solve. The authors' own summary is worth pinning above a desk: "Generating more citations does not compensate for, and may exacerbate, per-citation quality problems."
2. The URL existed and died. A quarter of all webpages that existed at some point between 2013 and 2023 were no longer accessible as of October 2023 — 38% of pages from 2013, and 8% of pages from 2023 already gone within that same year (Pew Research Center). Pew also found that 23% of news pages contain at least one broken link, and 54% of English-language Wikipedia pages have at least one dead link in their references — which matters because Wikipedia is one of the most heavily cited sources an AI model draws on.
The practical consequence: "verified once" is not "verified forever." A verification report has a timestamp for a reason.
3. The URL resolves perfectly and doesn't say that. This is the big one by volume, and the one a link checker will never catch. In a study of seven models across 800 medical questions and 58,000 statement–source pairs, "between 50% and 90% of LLM responses are not fully supported, and sometimes contradicted, by the sources they cite." Even the best performer — GPT-4o with web search — left roughly 30% of individual statements unsupported and nearly half of its responses not fully supported (Wu et al., Nature Communications, 2025). A separate 2025 audit of deep-research systems put citation accuracy in a 40–80% band, with "large fractions of statements unsupported by their own listed sources" (DeepTRACE, Salesforce AI Research).
One case contains all three failures at once. In a Sixth Circuit matter, counsel filed a brief with more than two dozen fabricated citations, plus citations that lacked the language quoted from them, plus real citations that didn't support the proposition they were attached to; the attorneys were ordered to pay $15,000 each (Sixth Circuit Appellate Blog). Invented cases, invented quotes, real cases that say something else. That's the checklist in miniature.
And the failure isn't limited to models being asked to find sources. A December 2024 letter in Intensive Care Medicine carried 15 references; only five could be matched to real articles. Reference 11 cited a nonexistent paper supposedly published in the very journal printing the letter. The editor-in-chief's explanation: "The authors have stated that these non-existent references resulted from the use of generative AI to convert the PubMed IDs of cited articles into a structured reference list." The letter was retracted on 29 November 2025 (Retraction Watch). The model wasn't researching. It was reformatting — arguably the most constrained citation task there is — and ten of fifteen references came out wrong.
Start here because it's mechanical, fast, and it removes the errors with the worst consequences. Do not read the draft yet. Extract every URL and check what each one returns.
The cleanest mental model available comes from the Penn audit, which sorts every citation into four states:
| State | How it's determined | What it means |
|---|---|---|
| LIVE | HTTP 200 | The page exists. That is all it means. |
| DEAD | HTTP 404 and an archived snapshot exists | Genuine link rot. Swap in the archive link or re-source. |
| LIKELY HALLUCINATED | HTTP 404 and no archived snapshot | The URL probably never existed. Cut the claim or re-source it. |
| UNKNOWN | 403, 429, timeouts, connection failures | Needs a human or a real browser. Do not auto-delete. |
The 404-plus-archive test is the part most checklists miss, and it's the difference between "this link rotted" and "this link was invented." The Internet Archive exposes it as a single unauthenticated GET: http://archive.org/wayback/available?url=example.com, optionally with a ×tamp=YYYYMMDDhhmmss. A hit returns an archived_snapshots.closest object with available, url, timestamp and status; a miss returns exactly {"archived_snapshots":{}} (Wayback Availability API docs). The reference implementation from the Penn paper, urlhealth, is 83 lines of Python that issue an HTTP HEAD request plus that Wayback lookup.
Two traps sit on either side of this step, and both produce confident wrong answers.
Treating 403 as fabrication. HTTP 403 is the single biggest source of false positives in automated link checking. When the Penn team re-checked 403 responses with a headless browser, 99.7% of those pages were live — the blocks came overwhelmingly from academic publishers refusing bot-like traffic, not from dead pages. Cloudflare's own documentation describes the same class of false positive: bot-management rules routinely trigger on legitimate monitoring and checking services because of their user agent or cloud IP range (Cloudflare WAF troubleshooting). If your checklist deletes every 403, you will delete good citations. (We hit exactly this while verifying the Sixth Circuit link above — 403 to an automated fetch, fine in a browser.)
Treating 200 as verification. A server can return HTTP 200 on an error page, a paywall interstitial or an empty template — the classic soft 404 (background on 403s and soft 404s). A 200 tells you a page exists at that address. It tells you nothing about whether your claim is on it.
This whole step is a machine's job, and machines are now very good at it. When the Penn team gave models a URL-checking tool plus a self-correction loop, non-resolving citations dropped below 1% — GPT-5.1 from 16.0% to 0.6%, Gemini 2.5 Pro from 6.1% to 0.1%, Claude Sonnet 4.5 from 4.9% to 0.8%, reductions of 6× to 79×. The catch they flag honestly: the loop only works if the model actually acts on the tool's output.
That structural gate is how we built Magic Share: a draft isn't saved until every cited URL resolves, and the verification report ships attached to the draft so the reviewer starts from a list of what was checked rather than an empty page. Roughly 20 sources go into a post, which is about 20 URLs to resolve — a tedious five minutes by hand, a few seconds automated, and the residual 403s flagged for a human instead of quietly deleted.
Now open the draft. This is the judgement half, where the 50–90% unsupported-statement problem lives, and it cannot be fully automated. It can, however, be made fast — and the evidence on how to make it fast is unusually clear.
When researchers compared professional fact-checkers with PhD historians and Stanford undergraduates on live web evaluation tasks, the fact-checkers "arrived at more warranted conclusions in a fraction of the time." The advantage wasn't knowledge; it was technique. Fact-checkers read laterally — they leave a page after a quick scan and open new tabs to establish who's behind it — while everyone else read vertically, staying on the page and reading harder (Wineburg & McGrew, Teachers College Record). Applied to an AI draft: for each cited domain, spend seconds working out what it is, then go straight to the sentence that carries your claim. Do not read the article.
Browsers give you a free verification tool here. A text fragment appended to a URL scrolls to and highlights an exact string:
https://example.com/report#:~:text=between%2050%25%20and%2090%25
The syntax is #:~:text=[prefix-,]textStart[,textEnd][,-suffix]. It's case-insensitive, needs percent-encoding for spaces and punctuation, works only on user-initiated navigation, and — the useful part — silently falls back to the top of the page when the text isn't found (MDN reference). That makes it a self-testing citation. If the highlight lands, the sentence is there. If you get dumped at the top of the page, the sentence isn't. (Sites can opt out with a Document-Policy: force-load-at-top header, so a non-landing highlight is a strong signal rather than proof.)
You will not check every sentence. Check in this order — the ordering is our reasoning, but each rung is justified by something measured.
The most common way a correctly sourced number becomes a false published claim is that it quietly grows. Two worked examples from this very post:
When you check a number, check its scope: US or global, one industry or all, one year or a decade, sampled or complete. Same number, wider claim, is a factual error.
The last five minutes catch what the first two steps structurally cannot.
Is the source actually a source? NewsGuard had identified 3,749 AI-generated news and information sites across 16 languages as of 23 June 2026 — sites with substantial AI-produced content, no meaningful human oversight, no disclosure, and a presentation designed to look like a legitimate outlet (NewsGuard AI Tracking Center). A live URL on a professional-looking site is not evidence of anything. If you don't recognize the domain, spend fifteen seconds finding out who runs it.
Does the draft cite the source or a copy of it? Aggregators, press summaries and roundups drift from the original. If a statistic matters enough to headline a section, follow it back to the study, the filing or the press release and cite that instead.
Does the draft claim more than the source claims? Read your own sentence and the source's sentence side by side and ask whether a hostile reader could say "that's not what it says." Hedges the source used — "associated with," "in this sample," "self-reported" — should survive into your version.
A post that oversells its own checklist has the same problem as a draft that oversells its sources. So, plainly:
A complete link audit plus a triaged claim check is not a full verification. Fabricated URLs run 3–13%; unsupported-but-real citations run 30–50% and up even in the best-behaved configurations. Twenty minutes buys you the whole link problem and the highest-value slice of the claim problem. It does not buy you an academic review.
Some AI output can't be checked this way at all. Jonathan Bailey made the strongest version of this argument in June 2026: testing six platforms on a general research prompt, he found only two supplied citations at all, and argued that without sources you must "examine every single claim independently" — which makes the tool less useful than advertised (Plagiarism Today). He's right about sourceless generation. The answer isn't a better checking habit; it's insisting on a different artefact. A draft that ships with a citation per claim converts an unfalsifiable output into a checkable one — which is why the checking step belongs inside the writing step rather than bolted on after.
And the SEO case is weaker than most posts on this keyword pretend. Broken outbound links are not a documented direct ranking penalty; Google has said outbound links aren't a ranking signal on their own, and low-quality ones are generally ignored (Search Engine Roundtable). The real exposure is elsewhere. Google's helpful-content guidance asks self-assessment questions that fabricated citations fail directly — "Does the content have any easily-verified factual errors?" and whether the page shows "clear sourcing, evidence of the expertise involved" (Search Central). Google's spam policies separately target scaled content abuse — publishing many pages without adding value — which we unpacked in our close reading of the scaled content abuse policy. Verify links for credibility and legal exposure. Any ranking benefit is a side effect.
The instinctive response to fabricated citations is "review more carefully." The evidence says that isn't enough on its own.
Papers at NeurIPS and USENIX Security go through three to five expert reviewers. In 2025, roughly one in twenty accepted papers at those venues still contained at least two likely hallucinated academic references — even though per-reference hallucination rates stayed below 1%, proceedings are large enough that paper-level failures show up (Russinovich, Siva Kumar & Salem). Their automated checker costs about $0.04 per paper to run. Expert human review, without a resolution step in front of it, is not a filter.
The wider trend runs the same direction. A Columbia University team analysing more than two million papers and 97 million citations found fabricated-reference rates climbing from roughly 1 in 2,828 papers in 2023 to 1 in 458 in 2025, and 1 in 277 in the first seven weeks of 2026 (STAT News on the Lancet study). "Citation practices are changing with generative AI use," Northwestern's Mohammad Hosseini told STAT. "Engagement with the literature is becoming increasingly more superficial."
It also isn't a bug waiting on the next model release. OpenAI researchers argue that hallucination persists because evaluations reward guessing: "Like students facing hard exam questions, large language models sometimes guess when uncertain, producing plausible yet incorrect statements instead of admitting uncertainty" (Kalai et al.). Their proposed fix is changing how benchmarks score abstention — a socio-technical change, not a patch.
Meanwhile, assuming a newer or paid model handles it is its own trap. The Tow Center tested eight AI search tools with 1,600 queries and found they "provided incorrect answers to more than 60 percent of queries," with rates ranging from 37% to 94% by tool. One tool returned error pages for 154 of the 200 citations it produced. And premium tiers "paradoxically also demonstrated higher error rates" than their free equivalents, because they answered confidently instead of declining (Columbia Journalism Review).
So the working order is: machine resolution first, human judgement second. Poynter's newsroom guidance lands in the same place — all AI-generated information requires human verification, inside a human-in-the-loop process with clearly assigned oversight (Poynter). That's the shape of a review that works: let software prove the links resolve, then spend your scarce attention on whether the pages say what the draft claims.
It's also the design behind Magic Share's approval gate. Drafts grow in a greenhouse, not in production: the agent researches, writes, resolves every cited URL and then stops, holding the post until a person approves it. Nothing publishes on its own. The goal isn't to remove you from the loop — it's to make sure that when you arrive, the mechanical work is already done and your 20 minutes go entirely to judgement.
Copy this into your review template.
Minutes 0–5 — resolve (machine):
Minutes 5–15 — match (human):
Minutes 15–20 — sniff:
How do I tell a hallucinated citation from a dead link?
Check the archive. A 404 with a Wayback Machine snapshot is genuine link rot — the page existed and has gone, so swap in the snapshot. A 404 with no snapshot anywhere means the URL most likely never existed, which makes the underlying claim suspect too. The Wayback availability API answers this in one request and returns {"archived_snapshots":{}} when nothing was ever archived.
Is a broken link in my post an SEO penalty? Not a documented direct one — Google has said outbound links aren't a ranking factor on their own. The exposure is credibility and, in regulated or legal contexts, liability. Google's helpful-content guidance does ask whether content contains easily-verified factual errors and whether sourcing is clear, so fabricated citations undermine you on the quality assessment even without a link-specific penalty.
Can I just trust a newer or paid model to get citations right? The measurements say no. Premium tiers in the Tow Center study showed higher error rates than free ones, and in the Penn audit the highest hallucinated-URL rate belonged to a deep-research configuration that reads the web before answering. Search augmentation and self-correction loops genuinely help — non-resolving citations dropped below 1% when models were given a URL-checking tool — but the reduction comes from a verification step, not from model size.
Twenty minutes for a 3,000-word post — is that realistic? Yes, if the link resolution is already done when you sit down. That step is the bulk of the mechanical work and the whole of the automatable part. If you're resolving 20 URLs by hand as well, budget closer to 45 minutes.
Most AI citation errors are not arguments you disagree with. They are addresses that lead nowhere, or addresses that lead somewhere real and irrelevant. That's good news: both are checkable, one by a script and one by a reader who knows what to look at first. Resolve every URL, sort the dead from the invented, then spend your remaining minutes on quotes, numbers and named works — in that order.
If you'd rather not do the first half by hand, that's the part we automated. Magic Share researches your niche, drafts a 2,000–4,000-word post sized to the topic, verifies that every cited URL resolves before the draft is even saved, and then waits for you — reviewing a draft takes a coffee, not an afternoon of fact-checking. Start free — your first three posts are on us.
Magic Share researches, writes and fact-checks posts like this for any site — point it at your URL and review your first draft today.
Plant your first post