What decides which sources Google’s AI Overview quotes

Google now answers a large share of B2B queries at the top of the page, and the buyer never scrolls.

There is no click to win. The only position that exists is being one of the handful of sources the answer cites, and four things decide whether that is you. Only one of them is content.

What changed, and why ranking stopped being the goal

Ranking third used to mean a share of the clicks. Now ranking third can mean nothing at all, because the question was answered above you by a summary that quoted two other companies.

The pattern is worth naming precisely. Your page can be indexed, ranked, technically healthy and completely invisible, because the buyer got what they needed without leaving the results page. We have written about the broader zero-click shift before. This post is about the narrower question of what gets you cited when it happens.

ai overview optimization is one of the fastest-moving terms in this space. In the twelve months to July 2026 it grew 900% year on year, and 900% again over the most recent three months. That is not a mature market with settled players. It is a category forming now.

How Google decides which sources to cite

AI Overviews are retrieval-based, which is genuinely good news.

The model is not answering from what it memorised during training. It runs a search, pulls current pages, and synthesises an answer from them with links. That means the inputs are live pages, and live pages you can change.

Two consequences follow.

Changes show up in weeks, not model generations. Unlike getting into a model’s training data, this is a system you can actually iterate against.

Being indexed is necessary and nowhere near sufficient. Google is choosing a handful of sources from everything it could have used. Ranking well helps. It does not decide it.

I would be sceptical of anyone claiming a repeatable method that guarantees citation. Overviews do not fire for every query, they vary between users, and Google changes the triggering logic regularly. What I can tell you is which inputs correlate with being cited.

Lever one: answer the question in the first two sentences

This is the single biggest difference between content that gets cited and content that gets ignored, and it is almost entirely a formatting decision.

An Overview is assembled from passages. If your answer to the question arrives in paragraph six, after context and background and a story about why this matters, there is no clean passage to lift.

Put the answer immediately under the heading. Then explain, expand, add nuance. The reader who wants depth will keep going. The system extracting a passage gets what it needs from the first fifty words.

Most B2B SaaS content does the opposite, because most of it is written to demonstrate expertise rather than to answer a question.

Lever two: use the question as the heading, word for word

Not the keyword. The question.

“How long does SaaS onboarding take” is a heading. “SaaS onboarding timeline considerations” is a keyword wearing a heading’s clothes, and it matches nothing anyone typed.

The mechanical reason: Overviews fire on questions. A page structured as a series of real questions with immediate answers gives the system a direct match at passage level. A page structured around topics does not.

Look at what your own buyers type into support tickets and sales calls. That phrasing is more valuable than anything a keyword tool returns.

Lever three: be specific enough to quote

There is a mechanical reason hedged writing loses.

When a system builds an answer, it either quotes a source with attribution or blends it into a general summary. Specific, checkable statements get quoted and the source gets named. Vague ones get absorbed into the average and the source disappears.

So:

• Use real numbers. “Significantly improves retention” is unquotable. “Companies above 100% net revenue retention grow 43.6% a year against 13.1% below 60%” is quotable, and ChartMogul gets the citation.
• Name your frameworks. A named thing can be referenced. A list of tips cannot.
• Take a position. “It depends on your business” is true and useless. Systems summarise hedges and quote claims.

Every one of those is in this post, which is deliberate.

Lever four: a page Google can actually parse

Short and mostly boring, which is why it gets skipped.

• Check your crawler access. Google-Extended controls whether your content can be used for AI training and grounding. I keep finding companies blocking it by accident, usually because someone blocked “AI scrapers” a year ago without distinguishing between training, grounding and retrieval.
• Make content exist without JavaScript. If your key passages render client-side, they may not be there when the page is read.
• Add FAQPage structured data where you have a genuine question-and-answer section. It removes ambiguity about which passage answers which question.
• Keep the answer above the fold in the HTML, not just visually. Passage extraction does not care about your layout.

How to check where you stand

Ten minutes, no tooling.

1. Take the twenty questions your buyers actually ask. From sales calls and support tickets, not from a keyword tool.
2. Search each one, logged out, in an incognito window. Note three things: did an Overview appear, were you cited, and who was cited instead.
3. Read the pattern rather than the count. The most common result I see is being cited on questions about your own company and absent from every question about the category. That is a specific diagnosis. Google knows you exist and does not consider you a source on the subject you sell into.

Writing more posts will not fix that. It is a corroboration problem, and it usually means the people being cited have third-party coverage that you do not.

What to be honest about

Overviews do not fire on every query and Google changes when they do.

Results vary between users and between sessions. Nobody can guarantee citation, and I have not seen a credible method that does. Anyone selling one either does not understand the mechanism or is hoping you do not.

What you can do is become a more obvious source than you are now, and measure whether it moved.

How long it takes

Weeks rather than months, because this is retrieval rather than training.

Restructure a page so the answer sits at the top, fix crawler access, add a genuine FAQ section, and you can see movement inside a quarter. That is much faster than anything involving a model’s training data, and it is why I would start here rather than with the broader AI visibility work.

Common questions

---

Close

Measuring this is the easy half, and there is a category of software that will do it and email you a weekly score.

Working out why Google cites four other companies on the question you sell into, and changing it, is the work. That is what we do in SEO and AI visibility.