Ranking in Google's top 10 does not get you into its AI answer

Most SEO advice about Google's AI answers boils down to one sentence: get into the top 10 and the AI will quote you. It sounds sensible. Google's AI summary does pick from pages it has already found, so being high up should help.

A study presented at the 40th annual conference of the Japanese Society for Artificial Intelligence tested exactly that, and the answer is more interesting. Being high up is not what separates a quoted page from an unquoted one, because they compared pages that were at the same height. What separated them was how precisely the page answered the question.

What they did, in plain words

The researchers, from Yokohama City University and CyberAgent, wanted to measure something slippery: how well a page's text matches the question somebody typed. You cannot measure that with a ruler, so they built a stand-in.

They took 100 questions and, for each one, the first 100 results Google gave. Then they trained a language model on that: show it two pages and it learns to tell you which one Google put higher. In effect, they taught a machine to imitate Google's judgement.

To check the imitation was any good, they tested it on questions it had never seen. It reproduced Google's ordering with an accuracy of 0.91 out of 1. Not perfect, but close enough to use as a stand-in for "how well does this page answer this question".

The clever part: they removed position from the equation

Here is where the study earns its keep. If you just compare "pages the AI quoted" with "all Google results", you learn nothing, because the quoted pages tend to be the ones ranking well anyway. You would only be rediscovering that good pages rank.

So they matched them by position. Each page quoted by the AI got a number depending on where it appeared in the summary: first quoted, second quoted, and so on. Then, for each one, they pulled a normal Google result from the same position at random. First against first, third against third. That gave them 714 pairs where the position is identical and the only thing that can differ is the page itself.

The result

Scored by the model, the pages the AI had quoted matched the question better than the normal results in the same position.

The gap is 0.058 points. That looks small on a scale of 0 to 1, and it is fair to ask whether it is just noise. The researchers ran two different statistical checks and both said no: with 714 pairs, a gap that size coming up by chance is roughly a one in two hundred thousand event.

So the headline is not that 0.915 is a target to hit. It is that when you hold position steady, the pages the AI picks are still the ones that answer the question more exactly. Ranking gets you into the room. Answering the question gets you quoted.

Four things this study does not say

We would rather tell you the limits than have you find them later. The authors are upfront about most of these.

One more detail worth knowing: they stripped out all the HTML before scoring, so titles, headings and structured data were invisible to the model. The paper notes this as something to fix in future work. In other words, the 0.058 gap comes from the words alone, with every technical signal removed. If anything, that makes the finding cleaner.

What to actually change on your pages

The practical reading is simple. Stop writing pages about a topic and start writing pages that answer a question.

Five changes follow from that, in the order we would do them:

• One question per page or per section. A page trying to cover six intentions will lose to six pages covering one each.
• Use the real question as the heading. Not a keyword stuck in a heading, the sentence a person would actually say out loud.
• Answer straight away, underneath. The specific figure, the specific condition, in the first two or three lines. Not after four paragraphs of context.
• Cut the generic filler. Any sentence that could sit on a competitor's page without changing a word is dragging your average down.
• Look at who is being quoted instead of you. For each of your key questions, check which pages the AI summary is actually citing. That tells you the standard you have to beat, question by question.

That last one is the step people skip, and it is the only one that turns the study into something you can act on. The rest is guesswork until you know which pages are winning the answers you want.

Why this fits what we see elsewhere

This study is about Google's AI summary, but it lines up with what other research says about AI answers in general. Work on ChatGPT's citations found the same preference for pages that give a precise answer: comparison articles, reviews and reports get quoted the most, and product and price pages the least. Different engine, same underlying idea. The page that answers the exact question wins the quote.

It also explains something that confuses people when they first measure their AI visibility: they see themselves ranking well on Google and absent from the AI answer, and assume something is broken. Nothing is broken. Those are two different competitions, and the second one is judged on how directly you answer.

Questions people ask

Does ranking first on Google get you into AI Overviews?

Not on its own. The comparison in this study held position equal on both sides, and the quoted pages still scored higher. Ranking helps you be found. It does not decide whether you get used.

Is a gap of 0.058 worth caring about?

On its own it looks tiny. What makes it meaningful is that it held across 714 pairs and survived two separate statistical checks. Read it as a direction, not as a target number to hit.

Does this apply outside Japan?

Strictly speaking it describes Japanese-language search. The mechanism, preferring pages that answer the exact question, has nothing Japanese about it, but nobody has repeated the measurement in English or Spanish yet.

How do I know which pages the AI is quoting for my questions?

You have to look, question by question, and keep looking, because the sources change between one day and the next. That is the job a tracker does: it asks your questions every day and records which pages fed each answer.

Source: T. Ogawa, S. Kimura and T. Koshinaka, An Analysis on Citation Strategies of AI Summarization in Search Engines, 40th Annual Conference of the Japanese Society for Artificial Intelligence, 2026. Related reading: tracking your brand in AI Overviews and the same for ChatGPT.