We fact-checked an AI-researched site against primary sources. Roughly a third of the leads didn't hold up.

aibubblequestion.com asks whether the 2023 to 2026 AI investment boom is a bubble. It measures the boom against 200 years of documented bubbles: railways, 1929, Japan in the 1980s, dot-com, telecom and housing. Every statistic links to its original source. Contested figures are flagged. The case for and the case against are both laid out.

It’s also a demonstration of our research method, and the research was done with AI’s help. Which makes it a fair test of a question a lot of business owners are now asking: if AI helps write your website, how much of it is right?

So we checked.

The challenge

A site like this is made of numbers. Numbers go stale. They get repeated from one article to the next until nobody remembers where they came from. And an AI research tool will hand you a confident figure with a link beside it, which looks like a source but isn’t proof.

We wanted to know how much of the site would survive a line-by-line check against the original documents, and how to ship the fixes without breaking anything.

The choice: three stages, kept apart

We split the work into three jobs and didn’t let them blur.

1. Orientation. Perplexity found the leads: claims, figures, and where they supposedly came from.
2. Evidence. Each lead was then checked against the primary document it pointed at. The filing, the report, the regulator’s own record.
3. Verdict. An editor decided what changed.

The headline finding came from step two. Roughly a third of Perplexity’s leads were refuted by the very documents they pointed at. Not “couldn’t confirm”. Refuted.

What we found

The corrections ledger ended with 58 findings. By type:

• Stale: 17
• No primary source: 14
• Mischaracterised: 9
• Inconsistent with another page on the site: 7
• Wrong when written: 4
• Scope mismatch: 3
• Misattributed: 2
• Figure since changed: 1
• Prediction since resolved: 1

By severity: 11 high, 33 medium, 14 low.

The pattern surprised us a little. Only four figures were wrong when they were written. The bigger problems were old numbers (17) and numbers nobody could trace (14). The AI wasn’t mostly inventing things. It was mostly repeating things that had stopped being true, or had never been pinned down.

Deleting is a correction

Eight figures came out entirely, because no primary source existed for them. We counted each removal as a correction, not as a loss of content. A number you can’t source is a liability on the page, however good it sounds.

Publishing the conflict instead of tidying it

Two items didn’t resolve, and we published them as open questions.

• The Georgia Public Service Commission’s own materials give two different numbers for the same vote: 9,885 MW and 9,985 MW.
• The comparison of “114 GW announced versus 15 GW under construction” has no locatable method on either side.

Picking one number would have made the page look cleaner. It would also have made it less accurate.

Where the approval gate had to go

This is the part that’s easy to get wrong.

The site keeps two copies of its facts. One is a knowledge bundle written for machines: the files AI tools read. The other is the database the human-readable pages pull their text from, live, every time someone loads a page.

The knowledge bundle had already been corrected. The database had not. If we had shipped only the code, AI tools would have read the corrected version while people read the old one.

And because the pages read the database live, any write to it publishes instantly. So the approval step went in front of the database write, not in front of the code deploy. 58 corrections went into the knowledge bundle. 57 corrections plus 5 new additions were staged, approved and published through the CMS to the pages people read. Each page now carries a dated corrections block, so a reader can see what changed and when.

What we don’t know yet

We don’t know whether a corrected, openly sourced site gets treated any differently by search engines or AI assistants. That’s the bet, and it’s too early to call. The “roughly a third” figure is also our working estimate from the pass, not a precise count.

What it means for your site

If AI drafted any part of your website, assume some of it has gone stale or was never sourced. Statistics are the riskiest part: the “most homeowners” and “studies show” lines. Check each one against the original. If you can’t find the original, delete the line. Nobody will miss it.

And if the same facts live in more than one place, correct every copy before you call it done. Fixing one and leaving the other is how you end up telling different readers different things.

More on how we build and check sites on how it works.