A 5.35 Million-Citation Study Just Measured the SEO-to-AI Citation Gap

Kingsworth Research published a study on September 8, 2026, built on 5.35 million citations pulled from ChatGPT, Claude, Gemini, Perplexity, and Grok. The headline number: non-paid, third-party sources account for 94-95% of all links AI engines cite, and journalism alone drives roughly 27% of citations across every wave of the study. Kingsworth's managing director summarized the finding bluntly: "traditional SEO rankings are no longer a reliable proxy for AI visibility."

That's a big, citable claim, and it's mostly right. It's also being packaged, in some of the coverage, in a way that overstates what the data actually supports. Worth separating the two.

What the data actually shows

The core finding is a decoupling, not a replacement. Ranking position in classic search results and citation likelihood in an AI answer are measuring different things, and a page that ranks well doesn't automatically get cited - because the two systems select sources on different criteria. Google's ranking is substantially about relevance and authority signals accumulated over time at the URL and domain level. An AI answer engine, synthesizing a response in real time, is selecting from whatever it can retrieve and trust for that specific claim - which favors third-party corroboration (reviews, comparisons, press coverage, forum discussion) over on-page optimization the way classic SEO measures it.

That lines up with what we found citing independent research when Adobe launched Brand Visibility in June: AI answers draw an estimated 57-92% of citations from third-party editorial content, versus 41-45% in traditional Google results. Two independently run analyses, months apart, pointing at the same structural gap is a stronger signal than either one alone.

Where "SEO rankings don't predict AI visibility" isn't the same as "SEO doesn't matter"

The Kingsworth framing that's spreading in trade coverage - SEO rank "no longer a reliable proxy" - is accurate as a measurement claim and easy to overread as a strategy claim. The two aren't the same thing:

• As a proxy metric, ranking position is unreliable. You cannot infer your AI citation rate from your Google position anymore, if you ever reliably could. That's the actual, defensible finding.
• As underlying practice, the work SEO and AEO share still matters. Technical crawlability, clear entity signals, and content structured so a system - human or model - can extract a direct answer are still foundational. We've made this case before: SEO, AEO, and GEO aren't separate funnels - ranking well in classic search is still the foundation AI agents build on, even though it stops being sufficient on its own.

The mistake would be reading "rank doesn't predict citation" as "stop investing in the fundamentals that make you crawlable and legible in the first place." The correct reading is narrower: stop measuring AI visibility by proxy, because the proxy has broken down.

The measurement implication

If ranking position can't tell you whether you're getting cited, the only way to know is to measure citation directly - which is exactly the gap we've written about before: traditional analytics stacks, including rank trackers, weren't built to see what happens inside an AI-generated answer. A 5.35-million-citation study is a research artifact, not a live dashboard for your brand specifically - it tells you the gap exists industry-wide, not whether your brand is on the winning or losing side of it this month.

That's the operational takeaway. GetBotRank's Visibility Tracker runs real buyer-intent questions against ChatGPT, Claude, and Gemini on a recurring schedule and reports your actual mention rate as a trend line - independent of, and no longer inferable from, your search rankings. If you've been treating a good SERP position as reassurance that you're also showing up in AI answers, this is the study that says to stop assuming that and start checking.

Run a free AI visibility check to see where the structural gaps are before you try to close the measurement one.