AI is citing unsourced statistics, and I have the receipts

I track which domains get cited in AI answers across six B2B SaaS accounts I run. Somewhere in that data, sitting next to Gartner and Wikipedia, were sites like gitnux, wifitalents, and zipdo. If you have not run into them, they are “statistics” sites: pages titled “37 [Industry] Statistics for 2026,” stuffed with confident percentages, almost none of which trace back to a source you can actually check.

They were not a fluke in one account. The same species of site surfaced across all six verticals, from cybersecurity to legal to creator tools. An AI was answering real questions for real buyers, and part of its evidence was a page of numbers with no primary source behind them.

That is the uncomfortable finding. Now the useful part: why it happens, why it should bother you, and what it quietly teaches you about getting cited yourself.

What a statistics-roundup site actually is

These are pages built to rank for “[topic] statistics” by listing a pile of numbers, with no visible original research and no citations you can follow underneath them. The pattern is simple. Pick a topic anyone might search. Fill a page with fifty stat-shaped sentences, “62% of teams report,” “the market will reach $X billion by 2030.” Wrap them in headers and a publish date. Skip the part where you show where any of it came from. Then do it again for the next topic tomorrow.

The tell is always the same: you cannot trace the number. A well-sourced statistic points to a survey, a dataset, a study, someone who counted something. These point to another blog that points to another blog, or to nothing you can open. The percentage looks precise, which is exactly the effect. Precision reads as authority, and these pages read as authoritative while showing no work.

To be fair, I have not audited every claim on every one of these sites, and some individual numbers may well be accurate. That is not really the point. The problem is that you cannot tell, because the sourcing is not there to check, and neither can the AI citing them.

Why AI keeps citing them

Models have weak source discrimination on stat-shaped queries. When a system retrieves sources to answer “what percentage of X,” it is rewarding pages that look like clean, direct, confident answers, and a statistics roundup is built to look like exactly that. It is not weighing whether the number is verifiable. It is weighing whether the page is accessible, ranks, and hands over an extractable answer near the top.

Look at what actually drives AI citation and these pages tick the boxes by accident: crawlable, ranking for the long-tail stat query nobody else targets, answering the exact question, and structured into tidy, self-contained lines a model can lift. Those are real ranking factors, the same ones I break down in the factors that get you cited. None of them is “can you verify this.” Verifiability is not a retrieval signal. Format and rank are.

The page is not cited because anyone checked it. It is cited because it is shaped exactly like the answer, and nobody legitimate bothered to target that query.

That last clause matters. These sites win uncontested queries. No analyst writes a page called “43 niche-industry statistics,” so a thin roundup walks into an empty SERP and becomes the source by default.

The risk: you might be quoting a number that traces to nothing

Here is why this is your problem and not just a curiosity. If an AI cites one of these pages and you repeat what it says, you have just moved an unverifiable statistic into your own deck, your own blog, your own pitch. The number now has your name behind it. When a prospect or a journalist asks where it came from, the honest answer is “ChatGPT said it,” which is not an answer.

The “so what” is a habit, not a panic. Any statistic an AI hands you is a lead, not a fact. Trace it to the primary source before you use it. If the trail dead-ends at another listicle, or at a domain that publishes “statistics” on forty unrelated topics, leave it out. The specificity that made it feel citable is the same specificity that makes an unsourced number dangerous once it has your logo on it.

The uncomfortable lesson for your own content

Here is the part that pays you back. If a thin page gets cited for looking like a clean, specific, well-ranked answer, then real data structured the same way gets cited even more reliably, because it also survives scrutiny. These sites are accidentally showing you the mechanic. You just get to run it honestly.

So do exactly what they do, with one difference: show your work.

  • Own the specific stat query. These pages win because the SERP was empty. Publish the real number for the exact question in your niche and you take that slot with something defensible.
  • Structure it to be lifted. One clear stat per line, the finding stated plainly, near the top, in language a model can quote without editing. Extractable beats eloquent.
  • Make it a number only you have. A thin roundup serves an unsourced “58% of teams.” You publish “across six B2B SaaS accounts I run, a brand’s own site is only 3% of the sources AI cites about it,” the kind of first-party figure a model quotes and a competitor cannot fake. That is the entire first-party data moat, and it is why I published my citation numbers in full in the GEO statistics breakdown.
  • Cite your own source. Say where your number came from, how it was measured, over what window. The one thing these pages do not do is show their work. Do the thing they don’t.

This is the same theme as everything else in generative engine optimization: being cited is not the same as being named and trusted. Thin pages get cited on format alone. You want to be cited and verifiable, so that when a model leans on your number, it holds up. Beat the thin stuff by being the specific source that also shows its work. That is a bar most content never clears, which is exactly why clearing it works.

FAQ

Does AI make up statistics?

Sometimes directly, when a model fabricates a number outright, and often indirectly, by citing a source whose numbers you cannot verify. In citation data across six B2B SaaS accounts, statistics-roundup sites with no traceable sourcing surfaced as cited sources in AI answers. The number can be unsupported at the source and still arrive in the answer looking clean and confident.

How do I know if an AI statistic is real?

Trace it to a primary source. A well-sourced statistic leads to a survey, dataset, or study you can open; a thin one leads to another listicle or to a site that publishes “statistics” on dozens of unrelated topics. If you cannot find who actually counted something, treat the number as unverified and do not repeat it with your name on it.

Why does AI cite low-quality sites?

Because retrieval rewards accessibility, ranking, and a clean extractable answer, not verifiability. A page built to look like a direct answer can win an uncontested query and get cited even when its numbers cannot be checked. Whether a claim can be verified is not a ranking signal, which is why format-optimized thin pages sometimes sit next to real analysts in the sources.

Can I use this to get my own content cited?

Yes, honestly. Publish the real statistic for the specific query others ignore, structure it to be quoted, and source it clearly. You get the same extractability that gets thin pages cited, plus the sourcing that keeps you cited when anyone checks. First-party data you can actually defend is the version of this that compounds.