One of the many counterintuitive things I’ve learned studying the sources that shape AI’s recommendations is that unlike most primary research, larger sample sizes don’t mean better data. If you’re conducting a survey everyone knows you need a representative sample, but when you’re analyzing citations - I often find more is not more. Take the recent mega studies of tens of millions of AI answers and hundreds of millions of citations - they found Wikipedia and Reddit were among the most cited. If you read that you might think your GEO questions have been answered and there’s nothing further to glean - but I’d posit otherwise. Sample size doesn’t matter if the sample isn’t your customer.
In three recent client audits of 5 platforms and 500 answers each spanning consumer retail and B2B tech - I found neither Reddit nor Wikipedia served as material sources (1.8% and 0.5% of citations, respectively). The most important third-party sources were an enthusiast review community the client hadn’t meaningfully engaged (cited in nearly three of four answers), and niche trade publications that weren’t high on either company’s target media list. If they’d stopped at the meta data of those reports - they might’ve missed the opp to engage the outlets actually influencing AI (and arguably consumer) perception.
Averages across millions of answers tell you what the internet looks like to AI. They don’t tell you what AI says when a buyer asks about your category, or which pages it leaned on to say it.
I built a free tool to show you what I mean. So try one prompt. Type your brand and the question your buyer might ask, and watch six agents answer, every link they cite get resolved, checked, and read, and the sources sorted by type. Free, no login, about two minutes.
Try it here: https://signal.innatec3.com/one-prompt
