Google published a year of AI Mode behaviour data in May, written by Shivani Mohan, its vice president of data science and user research. The number everyone has picked up is that the average US AI Mode query is three times the length of a traditional search query. That is the least useful number in the whole report.
Here is the more useful one. The most common opening words in AI Mode queries are now what, how, I, is and can. Not nouns. Not product categories. Question words and pronouns. People are describing a situation and asking for help with it, the way they would ask a colleague who happens to be standing nearby.
Five verbs replaced the intent buckets
Google groups the new behaviour into five modes: explore, decide, learn, create and do. They are growing at noticeably different speeds. Explore queries, the open-ended brainstorming kind, are running around 30% ahead of AI Mode’s overall growth. Decide queries, built around comparison phrasing, about 40% ahead. Do queries, the planning kind covering everything from workout routines to household budgets, roughly 80% ahead over the past six months.
If you have ever built a content plan around the informational, commercial and transactional split, this is that framework being quietly replaced by something with more resolution. A comparison page and a getting-started guide can share a target keyword and still need completely different structures.
What most people will take from this, and why it is only half right
The advice going round is to put the answer first. That is correct, and it is not enough.
Answer-first at the page level is a 2015 idea in a new hat. Featured snippets taught everyone to front-load the intro paragraph, and plenty of sites did exactly that. What has changed is the unit of retrieval. AI Mode does not pull your page. It pulls a passage. Which means every H2 section needs to survive being lifted out on its own, with no surrounding context, and still make sense to somebody who never read your introduction.
Test this on your own best-performing article. Copy one middle section, paste it into a blank document, and read it cold. If it opens with something like “this is why the previous point matters”, or leans on a term you defined four hundred words earlier, it is not retrievable. It is prose that depends on being read in order, which is a fine thing to write and a bad thing to expect a machine to quote.
A before and after that takes two minutes
Take a small accountancy firm with a page called “Choosing a business bank account”. One section is headed “Fees and charges”, followed by three paragraphs building carefully towards a conclusion about monthly cost.
The retrievable version contains identical information with the order inverted. The heading stays plain. The first sentence states the actual finding: that most UK business accounts sit between zero and roughly fifteen pounds a month, and that the free tiers usually cap the number of free transactions. Then the nuance. Then the exception that catches people out.
Nothing was dumbed down and nothing was cut. The sentence order changed, and a section that previously needed the paragraphs above it now stands on its own feet.
The follow-up question is where the opportunity is
This is the part almost nobody is acting on. Follow-up questions in AI Mode are climbing quickly, and more than one in six searches now include something other than text, whether that is an image, voice, or a live back-and-forth exchange.
In practice that means your page should answer the second and third question, not only the one in the title. Somebody asking about business bank accounts asks next about switching, then about what happens to existing standing orders. If those three answers live on three separate thin pages built for three separate keywords, you are competing against yourself inside a system that would rather pull three passages from one source it already trusts.
One useful comparison doing the rounds is the telegraph, which pushed journalists into lead-first writing in the 1880s for reasons of cost and physical page constraints. The mechanism is different now. The pressure is the same.
Where this goes wrong
The failure mode is already visible: sites turning every article into a stack of disconnected answer blocks with no argument, no point of view, and nothing a person would choose to read. That is optimising for the retrieval system and losing the reader, which is a poor trade, because the reader is the one who buys something.
There is also a sourcing caveat worth stating plainly. This is Google’s own data, about Google’s own product, published by Google. The direction of travel is almost certainly right and matches what most people see in their own logs. The specific growth percentages are marketing, and should be treated that way.
And there is a cost question. Restructuring a two-hundred-page site for passage retrieval is a project, not an afternoon. For a small team I would take the ten pages that already earn something, fix those, and leave the rest alone until there is evidence it worked. Proving it worked is a separate problem, and counting mentions is not the same as counting outcomes.
Start here, not with a migration
Most sites do not have a content volume problem. They have a content order problem.
Pick your three highest-value pages. Read every H2 section in isolation and rewrite the opening sentence of any section that cannot stand alone. Nothing else. No migration, no new tooling, no rebuild, no consultant.
If it moves nothing in six weeks, you have lost an afternoon and learned something specific about your own site, which is more than most people get from reading another framework.
Editor’s note: This area changes quickly, so check the latest platform policy before making compliance decisions.