How much of an AI citation is just your Google ranking?

In June we published GEO vs SEO and wrote that much of the underlying work overlaps, "though nobody has measured how much". That sentence is now out of date. The measurements exist, several of them, and they point the same way.

They also make a decent part of the GEO industry look unnecessary.

1. The number that changed under us

Correction, 6 November 2026. The first version of this post said 76.1 percent of AI-cited URLs already rank in Google's top ten, and gave citation rates by position (33.07 percent at position one, 13.04 percent at position ten). The 76.1 percent figure is from Ahrefs' July 2025 study. Ahrefs published an update on 2 March 2026, so newer data already existed when we wrote this on 21 September. We should have used it. We also could not trace the 33.07 and 13.04 percent figures to Ahrefs, so they are removed. We headlined "Most of it" and that no longer follows.

Here is what Ahrefs' update (read 6 November 2026) says. Across 863K keyword SERPs and 4M AI Overview URLs, 37.9 percent of cited URLs appeared in the first ten blocks of the results page, and 37.1 percent counting blue links only. Beyond position 100, it found 31.0 percent on the all-blocks count and 36.7 percent on blue links.

Ahrefs says the comparison with 2025 is not like for like, and suggests fan-out queries and a model change as possible causes, without proving either.

So the leaderboard still matters, and it is no longer most of the story. Roughly three in eight cited URLs sit in the top ten, and a similar share sit past the top hundred.

We are the agency with the most incentive to say GEO is separate work, and the corrected number leans that way. We would rather tell you the number moved against our earlier claim than leave it up.

2. The complication: a third of citations come from far down the page

Here is where it stops being simple, and it is the part we got wrong first.

The same Ahrefs update puts the share beyond the top 100 at 31.0 percent (all blocks) and 36.7 percent (blue links only). We cited a version of that ourselves in why every AI visibility tool gives you a different number.

Read quickly, that looks odd. About 38 percent of cited pages rank top ten, and about a third sit beyond the top hundred.

Both hold, because the citations are measured against the query you typed, and the engine answers a wider set. The engine takes your question and fans it out into sub-questions you never see. A page can be invisible for "best CRM for agencies" and rank first for "CRM that does retainer billing", and get cited through the second door.

So ranking still counts, but for a query you did not think to check as often as for the one you did.

3. What that does to keyword research

It kills the flat keyword list, which is most of what gets sold as GEO strategy.

If citations arrive through generated sub-questions, the useful unit is not a head term with volume next to it. It is the set of narrow, specific, slightly awkward questions a buyer would ask once they already know what they want.

Those questions usually have no search volume attached, because no volume tool measures a query the model invented. Absence of volume stops being a reason to skip a page.

We build those panels for clients by writing out the buying conversation and pulling every sub-decision from it. It is unglamorous and it is the part that works.

4. The strongest predictor is the one you cannot buy

The single biggest correlation in the data is not on-page at all.

Branded web mentions correlate 0.664 with AI Overview citation, per Ahrefs Evolve data from October 2025. That is other people writing your name on sites you do not own.

No schema change touches that number. No llms.txt file touches it, and we published the evidence that llms.txt does close to nothing already.

Mentions come from doing something worth writing about, being quotable, and being easy to find when a journalist or a forum thread needs a name. That is a twelve-month job. It does not fit in a three-month GEO retainer, which is exactly why it is missing from most of them.

5. Freshness is the cheapest lever left

Seer Interactive reported in October 2025 that 65 percent of AI bot hits went to content published within the past year, and 89 percent to content updated within three years. AI-cited content ran 25.7 percent fresher than traditionally cited content.

That is the most actionable finding in the set, and it costs almost nothing.

Updating a page that already ranks beats publishing a new one that does not. Most sites have twenty pages earning citations on two-year-old facts, and a content plan pointed entirely at new URLs.

We reorder that on day one for most NovaTechRay clients. Fix what already ranks, then write.

6. What we no longer say in a pitch

We used to describe the GEO layer as its own discipline sitting beside SEO. The data does not support that framing and we have stopped using it.

What survives is narrow, and we listed it in most GEO is a scam: crawler access, server-rendered HTML, a direct answer near the H1, mentions on sites you do not own. Three of those four are cheap. The fourth is slow.

Everything else in the category is either ordinary SEO with a new invoice line, or a tactic with no causal evidence behind it.

We did not know this when we started selling GEO. We were reasoning from mechanism rather than measurement, and the mechanism story was more flattering than the numbers turned out to be.

7. How to test any GEO proposal against this

Ask the agency what share of their proposed work would still matter if AI search disappeared tomorrow.

If the honest answer is most of it, that is a good sign and a badly named service. If the answer is none of it, ask which study supports each tactic, by name and date.

Then ask what they plan to do about branded mentions. The silence there tells you whether they have read the same data.

The short version

• Treat your Google ranking as one input, not the whole answer, because about 38 percent of cited URLs rank top ten (Ahrefs, 2 March 2026).
• Stop building flat keyword lists and start writing out the sub-questions a buyer asks late in the decision.
• Treat branded mentions as the real programme, and budget a year for it.
• Update pages that already rank before you publish new ones.
• Do the cheap GEO plumbing once: crawler access, server-rendered HTML, an answer near the H1.
• Ask any GEO vendor which of their tactics has a dated study behind it.
• Ignore any score that cannot show you its spread.

This is the work we do at novatechray.com, and most of it has an older name.