In Google's AI Mode, the Most-Cited Source Is Google

On May 19 at I/O, Google made AI Mode the default. Not a tab you could choose, not an experiment behind a toggle: the thing that happens when you type into the box. Google put it past a billion monthly users at the same announcement.

Almost every guide written since then is the advice people were already giving for ChatGPT with the brand names swapped. Most of it isn't wrong. It just skips the one finding that should decide what you do first, and that finding is uncomfortable if you sell websites for a living.

Google's biggest source in AI Mode is Google

Between April 15 and June 30, across more than 32 million citation instances, the AI-search analytics firm Profound found that google.com had become the second most-cited domain in AI Mode, with citations to it growing 8.4 times over those eleven weeks.

The interesting part is which Google pages. Almost none of that growth was ordinary search result pages. It came from two Google-hosted surfaces: Business Profiles and Product Knowledge Panels.

SE Ranking arrived at the same shape from a different direction. Google's share of AI Mode citations rose from 5.7% in June 2025 to 17.42% in February 2026, and it was the single most-cited domain in 19 of the 20 industry categories they measured.

Read that as a local business and it says something blunt. When someone asks AI Mode who to call for a physiotherapist in Mississauga tonight, the page it quotes for your hours, your rating and your address is a page Google hosts about you. Your website is where it goes for everything that card can't hold.

That isn't an argument for neglecting your site. It's an argument about order.

Why it works that way: one question becomes sixteen

AI Mode doesn't run your question as a search. It decomposes it, then runs the pieces in parallel. Google calls this query fan-out, and the counts people have measured run up to sixteen sub-queries, with most prompts landing somewhere between five and eleven.

Each of those sub-queries is a check the model has to pass before it's willing to name anyone. Does this business actually do the thing that was asked for. Is it near the person asking, and is it open right now. Is it any good. Will they take this insurance, and what does it cost.

Split that way, the strategy falls out on its own. The checks Google can settle from its own data, it will, because that data is structured, current, and already in the building. The checks it can't settle that way get answered from a passage on somebody's website. Being that somebody is the part you compete for.

So the hours question is a profile question. The "do they treat rotator cuff injuries, and what does a first session cost" question is a website question. Two different jobs, and most businesses are only doing the second one.

Ranking first is not the same as being cited

This is where the older SEO instinct misleads you.

Retrieval in AI Mode happens at the passage level, out of a pool that query fan-out has already widened well past the ten results a person would ever see. Position on page one is an input, not a qualification. A Ziptie study found that ranking in the top ten gave a page roughly a 25% chance of turning up in an AI Overview, which means three times out of four the page that ranks isn't the page that gets quoted.

It gets stranger. Across 730,000 responses, AI Mode and AI Overviews reached the same conclusion 86% of the time while their citations overlapped only 13.7%. Two surfaces, one company, the same answer, almost entirely different sources. There is no single list to get onto.

What does carry across every study is the shape of the text that gets lifted. Kevin Indig's analysis of 1.2 million ChatGPT responses found 44.2% of citations came from the first 30% of a page, and that quoted passages skew heavily toward definitive, entity-rich sentences. Not "we pride ourselves on personalised care." Closer to "a first physiotherapy assessment runs 45 minutes, costs $120, and we direct-bill to most major insurers."

One of those can be lifted into an answer without modification. The other can't be lifted at all.

What to do, in order

Fix the profile first, because it's the surface being quoted. Categories, hours, service area, services, photos, reviews. Recency carries real weight here, and a profile whose newest review is from 2023 reads as a business that might not exist any more. The six mistakes that cost you local customers covers the fields that actually move something, and the popular tactic that gets listings suspended.

Then answer the sub-questions on your own pages. Go back through your last twenty enquiries and write down what people asked before they booked. Those are the checks. Give each one a heading and a first sentence that answers it outright, with the number, the name and the place in it, near the top of the page rather than at the bottom of a scroll.

Then make the two agree. Name, address, phone, hours and services have to be identical across your site, your profile, your schema markup and your directory listings. A model that finds two versions of a fact tends to drop the claim rather than pick a side, and dropping the claim means dropping you out of an answer you were about to be in. That's the argument behind the schema markup that earns its keep and llms.txt, and it's why we generate both from one copy of the facts instead of maintaining three.

Before any of it, check that your pages can be read at all. If your content paints in after load, nothing above applies to you yet. There's a one-minute test for that, and we failed it on our own site once.

The part nobody can tell you

SparkToro's analysis of Similarweb clickstream data found that between January and April 2026, only 0.34% of US Google searches moved into AI Mode.

That was measured before the May default switch, on a tab most people never clicked. Which is the point: the surface changed shape a few weeks after the last clean measurement of it, and nobody quoting you a percentage today has a real read on how many of your customers now get an answer instead of a page of links. Anyone who says otherwise is guessing with more confidence than the evidence supports.

You can't watch it in your own analytics either. GA4 files AI Overviews and AI Mode clicks under Organic Search along with everything else Google sends, so no report isolates them. The measurement article covers what each tool actually counts, and what to report instead.

What you can do is run the questions yourself. Pick ten things a customer would ask before choosing you, ask them in AI Mode, and write down who gets named. Repeat monthly. It isn't a rank tracker and never will be, because there's no ranked list to track. It's the closest honest equivalent, and within about half an hour it tells you whether you have a visibility problem or a competitor problem.

Three things not to do

Don't rebuild the site for this. Nothing in query fan-out rewards a framework, and a redesign is the most reliable way to lose the rankings you already have when the redirects are an afterthought.

Don't buy AI Mode placement. It isn't sold. The emails offering it are selling something that doesn't exist, which is a different thing from the advertising Google and OpenAI genuinely do sell on these surfaces.

Don't publish forty FAQs nobody asked. Fan-out doesn't reward volume, it rewards a clean answer to a question a real person had. Ten good ones beat forty padded ones, and the padding makes the good ones harder to find.

The reason we run local search and AI visibility as one programme rather than two is on this page. For a local business the profile and the site answer different halves of the same question, and doing either one alone leaves the other half unanswered. That's the AI SEO service. If you'd rather know where you stand before talking to anyone, the ten-question test above costs an afternoon and no money.