GPT-5 Turbo’s ‘Citations’ Feature Is a Farce: Real Data from Live Publisher Logs

• OpenAI claims GPT-5 Turbo’s ‘Citations’ feature launched to “fight hallucinations” in June 2024.
• Analysis of 150+ outputs on ElephantNY: 109 references were broken, invented, or misattributed.
• Top offenders included fake URLs for The New York Times, phantom studies, and recycled Wikipedia scraps.

GPT-5 Turbo’s so-called ‘Citations’ are the kind of falsified backstops that lazy agencies and LinkedIn SEO grifters will sell you as “trust signals”—and it’s already polluting the entire AI search ecosystem. Within 72 hours of the feature hitting public endpoints, we found OpenAI hallucinating studies that don’t exist, citing ElephantNY stories that never ran, and even inventing non-existent Medium URLs with the confidence of a televangelist.

No, OpenAI, “referencing” a New York Times article with a URL that 404s isn’t “reliable.” That’s horseshit. We found 41% of citations going to dead pages, 24% pointing to content that had nothing to do with the summary, and 7% literally referencing sites that don’t exist. In one beauty editorial, GPT-5 Turbo cited a 2017 Pew Research Center study on TikTok—newsflash, TikTok wasn’t even in the U.S. then. The rot goes deeper than “oops, hallucinations”—this is institutionalized laziness.

Worse, the AI SEO cottage industry is already spinning this as a revolution. I’ve seen at least three LinkedIn “AI content strategists” this week sell GPT-5 Turbo’s citation-fodder to unsuspecting DTC brands as “fact-checked by AI.” No, Tom from Agency10X, your client’s traffic isn’t going to skyrocket because ChatGPT made up a Harvard Business Review article about their eyebrow gel. Anyone who’s actually managed a publisher stack knows: fake citations are worse than none at all. They gaslight editors, poison trust, and trash vertical search signals.

If you believe the Google Search Liaison fairy tale that “all AI citations will be scored for accuracy,” wake up. Their own SERPs are jamming SGE panels with the same flavor of hallucinated links, while plugin parasites like Yoast and Rank Math already push settings for “AI citation enrichment.” It’s a cargo cult. If you care at all about visibility, actual ranking, or not getting your brand sued, you need to audit every citation yourself—or, radical idea, stop stuffing pages with fake footnotes altogether.

The industry doesn’t want to hear it, but here’s the uncomfortable truth: every publisher, from the smallest indie blog to Condé Nast, should disable automated citations by default and force human review on every link. Yes, it’s more work. Yes, it’s slower. But you can’t scale trust with AI fairy dust. You either check every source, or accept that your “trusted content” is just another flavor of spam.

Frequently Asked Questions

How accurate are GPT-5 Turbo’s citation links in real-world publishing?

In ElephantNY’s test of 150+ outputs published this week, 72% of GPT-5 Turbo’s citations were hallucinated, misattributed, or linked to dead/nonexistent pages. Only 28% were both real and contextually relevant.

Who is promoting AI citation features as “trust signals”?

OpenAI’s own launch post, lazy agencies, and LinkedIn SEO “influencers” are hyping GPT-5 Turbo citations as fact-checking tools. In reality, these are widely inaccurate and mislead editors and readers.

What is the best way for publishers to handle AI-generated citations?

Disable automated citations by default. Require every AI-generated link to be manually checked by a trained human editor before publishing. There’s no shortcut—trust demands real verification, not AI fiction.