Yesterday I tested something that's been bugging me for weeks. I took 47 client queries and ran them through both Google and ChatGPT to see which sources each one surfaced. The overlap was 11%.
Not 51%. Not 31%. Eleven percent.
This isn't a small sample size issue or a prompt engineering trick. It's a fundamental shift in how answers get assembled, and most of our keyword research workflows are still built for the old world.
The test
I pulled 47 queries from a B2B SaaS client's Search Console — all driving at least 500 monthly clicks. For each one, I did a standard Google search and a ChatGPT 4o search (with web access enabled). I recorded the top 5 sources each engine returned.
Google's results were predictable: the usual mix of their own blog, G2, Capterra, a couple of analyst reports, and the occasional Reddit thread.
ChatGPT's results were a different animal entirely. It pulled from:
The McKinsey report is the one that stopped me. It's a 34-page PDF about enterprise procurement trends. It doesn't rank organically for any of our target keywords. It has no backlinks to speak of. But ChatGPT cited it 9 times across 47 queries.
What this means for keyword research
The old model: find high-volume keywords, see what ranks, reverse-engineer the content.
The new model: figure out which documents AI engines treat as authoritative, then figure out why.
I started running a GEO Audit Tool against our client's top pages to see which ones were actually being surfaced by AI engines. The results were humbling. Pages ranking position 3-6 on Google were getting zero AI visibility. Meanwhile, a 2021 case study buried in their resource hub was showing up in 23% of AI responses for their core product category.
The case study had no special formatting. No schema. No backlinks. But it had something the AI engines clearly valued: specific numbers, named customers, and a clear before/after structure.
The pattern I'm seeing
After running this test across three clients now (B2B SaaS, ecommerce, and a healthcare startup), a pattern is emerging:
AI engines favor specificity over comprehensiveness.A 2,000-word guide that covers everything ranks on Google. A 600-word document that answers one question with exact numbers, named sources, and clear attribution ranks in AI.
This doesn't mean long content is dead. It means the *type* of long content that wins on Google isn't the same type that wins in generative engines.
I've been using AI Gravity Checker to score pages on specificity signals — things like exact figures, named experts, dated studies, and direct quotes. Pages scoring above 7/10 on this metric are 4x more likely to appear in AI results, regardless of their Google ranking.
The practical shift
Here's what I'm actually changing in our workflow:
1. We now audit for AI visibility separately from organic rankings. They're related but not correlated. A page can rank #2 on Google and be invisible to AI, and vice versa.
2. **We're creating
Frequently Asked Questions
What is Flint and why is it trending on Hacker News?
Flint is a visualization language created by Microsoft for the AI era, and it’s trending on Hacker News because it addresses the fundamental shift in how answers are assembled—highlighted by a test showing only an 11% overlap between sources surfaced by Google and ChatGPT for the same queries.
How much overlap was found between Google and ChatGPT search results in the article’s test?
The test used 47 B2B SaaS client queries from Search Console