How to appear in Google AI Overviews

Two years ago an AI Overview was a curiosity at the top of a few queries. Today it sits above a large share of informational searches, AI Mode has rolled out well beyond the United States, and for a growing set of questions the person never scrolls to the links at all. The links are still there. They are just lonely. If you sell anything that people research, the question is no longer whether to care. It is how to be the page the answer is built from.

We have been working on this for our clients since the first rollout, and the good news is that most of it is craft, not magic. Here is what we have learned.

What Google actually pulls from

Every study we have run or read points the same way: AI Overviews cite pages that already rank well for the query or its close variants. Not exclusively, but overwhelmingly. Google is not inventing a new index; it is reading its existing one more carefully and picking passages rather than pages.

Three consequences follow.

• Being on page one is still the entry ticket. There is no GEO without SEO underneath it.
• The unit of competition has shrunk from the page to the passage. A great paragraph on a good page can beat a great page with no quotable paragraph.
• Third-party pages count. Reddit threads, review sites, industry publications, forums. If the model trusts them more than your product page, it will cite them instead of you, and often about you.

Write the answer, then the argument

Most pages are written like an essay: context, build-up, conclusion. AI systems read like an impatient expert: they want the answer in the first two sentences under a heading that matches the question, then the evidence.

So for every question your buyers ask, the page needs a heading phrased the way they ask it, a direct answer of forty to sixty words immediately under it, and only then the nuance. Do this for the ten questions that matter and you have done more than most of your competitors will do all year.

Be specific enough to be quoted

Vague pages do not get cited because there is nothing to cite. "Our software improves productivity" is not a fact. "Teams using the workflow feature cut approval time from four days to one, on average, across 120 accounts in 2025" is a fact, and it is the kind of sentence a model lifts into an answer with your name attached.

Numbers, dates, named methods, defined terms, clear comparisons. If it would survive a fact-check, it is the right kind of sentence.

Make the structure legible to a machine

This is the unglamorous part and it matters.

• One idea per section, one question per heading.
• Lists where the content is actually a list. Tables where it is actually a comparison.
• Schema markup that describes what the page is: FAQ, HowTo, Product, Organization. Not because it guarantees a citation, but because it removes ambiguity.
• Author and organisation pages that exist and say something. Models weigh who is speaking.
• Fast, crawlable, no content locked behind interaction.

Earn mentions where the models look

If your category is discussed on Reddit, in trade press, on comparison sites, you need to be present there in a way that is accurate and useful. Not seeded reviews; the models are getting better at spotting those and Google has been penalising them for years. Real answers to real questions, from people who work at your company, under their names.

Three habits to drop

Chasing keywords you cannot answer. If you have nothing distinctive to say about a query, a thousand words of filler will not get cited. Pick the questions where you have a genuine edge.

Hiding the answer to keep people on the page. The "read on to find out" structure was a bad habit in 2019 and is fatal now. Give the answer. The people who want the depth will keep reading.

Measuring only clicks. Being cited without a click still moves the brand, and for many commercial queries the citation leads to a branded search later. Track citations and branded demand alongside sessions, or you will cut the thing that was working.

How we measure it

We track a fixed set of queries per client across AI Overviews, AI Mode and the main assistants, weekly, and log whether the client is cited, where, and next to whom. Over the past year that number has risen by 244% on average across the accounts we do this for. The method is not secret; it is the discipline of doing it every week, which turns out to be rarer than the method. If you would like to know where you stand today, that is a thirty-minute conversation and we will show you the sheet.