Ranked #1 on Google, Invisible in AI Answers

An export business owner in outdoor gear sent me two screenshots the other day. The first showed his website's ranking on Google: a folding table-and-chair set he was pushing hard, holding steady on page one for an English search term. The second showed the same question put to ChatGPT and Gemini — the AI listed five suppliers, start to finish, none of them him.

He couldn't work it out. The ranking was there — so why wouldn't AI mention him?

I said: those two things used to be the same thing. They are now splitting apart — and faster than most people think.

A measurable shift: in AI search optimization, ranking and citation are decoupling

This isn't a hunch. Research firms have been measuring it continuously.

Ahrefs ran two comparative studies. The July 2025 round found that 76% of the pages cited by AI Overviews — Google's AI summaries — also ranked in the traditional top 10. Eight months later, in March 2026, they widened the sample to 863,000 keywords and 4 million AI Overview citation links — and the same share had fallen to 38%. Going from "strongly correlated" to "roughly a coin flip" took eight months.

Now look at where the citations actually come from. The 2026 study broke down like this.

That last line is the key one: more than a third of AI citations come from pages that don't rank on Google at all.

Other firms point the same way. BrightEdge tracked nine industries over twelve months and found only about 17% of AI Overview sources ranked in the top 10 — a figure that barely moved across the whole observation window. A separate Ahrefs cross-platform study tested 15,000 long-tail questions across ChatGPT, Gemini, Copilot and Perplexity: on average, only 12% of cited links appeared in Google's top 10.

Here I should mark an honest boundary: these studies use different samples and methods, so the numbers conflict (6.82%, 12%, 17%, 38%). And Search Engine Journal points out that part of the 76%-to-38% drop comes from Ahrefs' own citation detection getting better — the 2025 round may have undercounted citations, so the two waves are not strictly comparable.

What can be confirmed is the direction, not the precise magnitude: ranking well is increasingly not the same as being cited by AI. That is the first premise anyone doing AI search optimization has to accept.

Why they decouple: one question gets split into a dozen

The answer lies in a mechanism called query fan-out.

Google has confirmed its AI systems use it. Put simply: you ask one sentence, and it doesn't search just that sentence — it breaks it into several related sub-questions, searches each one separately, then stitches the results into a single answer.

Google Vice President of Product Robby Stein explained the mechanism in July 2025. His example: ask "what is there to do in Nashville with a team" and the system generates its own follow-ups — best restaurants, best bars, where to go with kids — then searches each. By his account, a single question typically spawns 8 to 12 parallel sub-queries, more when the question is complex.

After Gemini 3 became the default model behind AI Overviews on January 27, 2026, the mechanism's effect grew noticeably.

That is the mechanism behind the decoupling. Your page doesn't need to crack the top 10 for the original term — if it is "the best answer" to some sub-question, it has a shot at being picked.

The reverse holds too: rank first for the head term, but say nothing about a sub-question, and AI won't use you. That was exactly the outdoor-gear owner's problem — his page said "bulk customization supported, reliable quality, on-time delivery", yet the questions buyers actually spin off (what's the minimum order, can the material be changed, which certifications, how long does sea freight take) had not a single concrete number on the page.

So the difference between AI search optimization and SEO isn't technical — it's in what you're optimizing.

One more point: this mechanism isn't unique to Google. China's Doubao, DeepSeek and ERNIE Bot follow the same "retrieval plus generation" route — they don't care where you rank on Baidu either; what they weigh is whether a passage can serve as an answer. So the gap you hit in AI search optimization is the same problem at home and abroad.

An easily overlooked channel: video

One finding in the same dataset deserves a mention of its own.

Among pages cited by AI Overviews that don't rank in Google's top 100, 18.2% are YouTube links. YouTube accounts for 5.6% of all AI Overview citations. Ahrefs' ongoing tracking also shows YouTube is now the most-cited domain in AI Overviews, up 34% in six months.

For B2B factories this matters: it's a completely separate exposure channel — and one that AI search optimization usually skips entirely. A video's spoken script, title, subtitles and description are all in scope for retrieval. If you don't publish it, someone else will, and the version AI learns is theirs.

Google's first official measurement entry point for AI search optimization

The most awkward part used to be that you couldn't measure it. You had no idea how often you showed up in AI answers — you could only ask by hand.

On August 31, 2026, Google rolled out a "Generative AI performance report" in Search Console, worldwide. What can it show?

The main metric is impressions: how many times your links were shown inside AI features (AI Overviews and AI Mode). It can be grouped by page, country, device and date, and exported.

What it can't show needs stating too: there's no query-level breakdown (you can't see which search term triggered it), AI Overview and AI Mode data are merged, and an impression is not a recommendation — it only means your link was displayed, not how AI described you. The report may also stay hidden when impressions are too low, which is normal.

Still, the report's biggest value is that AI search optimization can, for the first time, look at "ranking" and "AI visibility" separately. The two used to be conflated; now they can be pulled apart. For the first time, people doing AI search optimization have official comparative data.

Clearing up one commonly confused switch

Google's official line on "how to appear in AI features" hasn't changed, and it's simpler than many assume: a page just needs to be indexed and eligible to show a snippet. There is no extra technical requirement, and no special markup or schema is needed.

To control whether your content appears in AI features, you use two other things: the page-level nosnippet tag, or the "Search generative AI" setting in Search Console (default is include; you can set include, exclude or inherit).

One widespread misunderstanding is worth correcting: Google-Extended in robots.txt does not control AI Overviews or AI Mode. It governs Gemini training and grounding, which is not the same as AI features in Search. So some people think they've blocked it when they haven't at all — and the reverse is true too: some think they've allowed it, when that door was never its to open.

So what should you actually do?

Back to the outdoor-gear owner. My advice: don't scrap your existing SEO — that's the foundation — but add one thing: turn your pages from "introducing yourself" into "answering questions".

Three moves, which are also the most basic set in GEO.

First, list the sub-questions. Under your target term, write down what buyers actually spin off: how much does it cost, what's the minimum order, how long is delivery, which certifications, how does it compare with another material, how do I choose. The source is easy to find — asking AI is itself the best method: the answer structure it returns is the structure it believes the question should have.

Second, give each sub-question a passage that stands on its own — with concrete numbers, dates and nouns, not adjectives. There's academic backing: the widely cited KDD 2024 paper tested adding source citations to content, and pages originally ranked 5th gained 115.1% in visibility, while pages originally ranked 1st dropped 30.3%. In AI answers, a prior positional advantage gets diluted; whether something "works as evidence" is the variable.

Third, spread the same answers beyond your own site. AI's sources aren't only your website — they include third-party content platforms, industry communities and video platforms. This is where GEO and SEO differ most in practice: SEO mostly happens on your own site, while half of GEO happens off it.

One closing boundary, so it isn't misread: this doesn't mean ranking is useless. In fields where trust matters most — healthcare, insurance, education — BrightEdge measured overlap between AI citations and traditional rankings still at 68% to 75%. Ranking still helps; it's just no longer the only ticket in.

Three questions about SEO and GEO, answered

Q: Was the SEO work wasted?

A: No. The two optimize different things; they don't replace each other. Crawling and indexing are shared foundations, and being indexed is still a prerequisite. What changed is the test: it used to be ranking; now it also includes whether you're used as an answer.

Q: How long does AI search optimization take to work?

A: No timeline, no promised results. Google's official position has always been that it doesn't guarantee content will be crawled, indexed or shown. If someone gives you an exact number of days, that's a sales pitch.

Q: Where's the most cost-effective place to start?

A: Open the Generative AI performance report in Search Console and get a clear picture — which pages already get impressions in AI features, which get none. With that baseline, decide which page to fix first. It beats rewriting everything from the outset.

In closing

To wrap up: ranking well and being cited by AI used to be almost the same thing; they are becoming two. In eight months the overlap fell from 76% to 38%. That is the premise AI search optimization has to accept first.

The reason: AI splits one question into a dozen, and what it looks for isn't "the page ranked first for that term" but "the best answer to one small question". So AI search optimization means turning content from self-introduction into answers that can each be lifted out on their own — the biggest divide between GEO and SEO.

Yihang Network, based in Shenzhen, previously worked in Alibaba.com-related services and expanded into GEO from 2024, serving many exporters across the Pearl River Delta. The method is written up at https://www.chyihang.com/en/geo/.

Have you ever had good rankings but no AI mentions? Tell me your industry and product in the comments and I'll take a look at which layer the problem sits on.

1. Ahrefs, "38% of AI Overview Citations Pull From The Top 10", 2026-03-02, https://ahrefs.com/blog/ai-overview-citations-top-10/
2. Search Engine Journal's coverage of the above study (noting the detection-method improvement and that the two waves are not strictly comparable)
3. Google Search Central, official guidance on AI features (pages must be indexed and snippet-eligible; no extra technical requirement), checked 2026
4. Google Search Console, Generative AI performance report, launched worldwide 2026-08-31 (via multiple outlets and media; official announcement page not directly accessed)
5. Aggarwal et al., KDD 2024, Generative Engine Optimization, DOI 10.1145/3637528.3671900
6. BrightEdge industry tracking (2026-02, nine industries over twelve months)

Data and scope note: The studies cited here differ in sample scope, statistical method and time window, so their percentages cannot be compared directly; only directional conclusions are drawn. The comparability of Ahrefs' two waves is qualified (citation detection improved) and this is stated in the body. Search Console's Generative AI report is still evolving; the latest official guidance prevails. Experimental figures reflect relative changes under lab conditions and do not represent inevitable real-world outcomes. No method guarantees being cited or recommended by AI. This article contains no pricing or package information.