Keyword research has worked the same way my whole career.
Pull the list. Sort by volume, highest to lowest. Cross out anything you have no business
competing for. Build the content calendar from what's left.
Somebody has been doing a version of that since the early 2000s, and for a long time it was the right call, because volume was a decent proxy for opportunity.
However, Google just published data that makes me think that proxy is breaking down. But it’s not because volume stopped mattering.
It’s because the moments people are bringing to AI aren't the same moments they're typing into a search box, and the gap between the two is bigger than I expected.
What Google’s ATLAS Data Measures
The report is called AI & Economy ATLAS, from researchers at Google and Google DeepMind.
It covers 14.65 million interactions across the Gemini app, AI Mode in Search, and the Gemini API.
The clever part is what they compared it against. They took US non-work conversations and lined them up against the American Time Use Survey, which tracks how Americans actually spend a day.
So you get a direct read on which subjects come up in AI conversations more or less often than they occupy people's time.
That comparison is where it gets useful, because it tells you what people specifically choose to bring to AI rather than just what's popular.
Google demonstrates how AI Mode handles open-ended questions and follow-up searches, one of the AI experiences included in its ATLAS analysis of user interactions:
Which High-Friction Topics People Bring to AI
Some categories come up wildly more than you'd predict from how much time people spend on them.
Government services and civic obligations, meaning licenses, taxes, fines, and voting, run at roughly twenty to one.
Professional and personal care services, which covers doctors, lawyers, banks, and salons, sits above seven to one.
Education runs close to six to one. Buying things runs about three to one.
Now flip it. Eating and drinking takes up an enormous share of people's day and comes up at about one to eighteen. TV and movies, washing and dressing, cleaning the house, cooking, all at the bottom.
So people aren't handing AI their daily routine. They're handing AI the stuff they're stuck on.
Google's own framing for the high side of that gap is "high-friction," and there's one more detail I keep thinking about.
About half of those medical, legal, money, and government questions came in outside working hours, at night, early morning, and on weekends, per Search Engine Journal.
That's a person at eleven at night trying to figure something out that they couldn't get answered during the day.
Google demonstrates how Gemini can be used for deeper research when a question requires more than a quick answer:
Why AI Search Intent Changes Keyword Research
Here's the practical read.
If people reach for AI when a decision is hard, has real consequences, and requires weighing options, then the content that shows up in those moments isn't your highest-volume page.
High volume usually means a simple, well-defined question with a short answer, which is exactly the kind of thing AI handles without needing to cite anybody in particular.
The moments that matter are messier, such as “which option fits my situation?”
What happens if I pick wrong? How do I compare these two things that aren't really comparable? What does this cost once you include the parts nobody mentions upfront?
In my experience at ForeFront Web, most brands have thin content there, because that content has never scored well on a keyword tool.
It doesn't have clean volume, the intent is hard to categorize, and it's a pain to write since it requires actual expertise rather than a summary of what already ranks.
That's also why it's worth doing. The stuff that's tedious to produce is the stuff your competitors skipped too.
Google SVP Nick Fox discusses how search behavior is moving toward longer, more detailed questions as AI changes the way people look for information:
What Content Brands Should Build for AI Search
I'd start by looking at your existing content and asking a blunt question about each piece. Does this answer a question, or does it help somebody make a decision?
A page that answers a question is complete when the question is answered. But a page that
supports a decision has to do more.
It has to lay out the options honestly, name the tradeoffs, say who each option is wrong for, and be specific about cost and timeline and what goes sideways.
That last part is where most brand content falls apart. Nobody wants to publish the sentence that says “this isn't right for you if X.”
But that sentence is exactly what makes a page useful in a decision, and it's what makes it worth citing when a system is helping somebody weigh choices.
If your category shows up on the high side of Google's gap, and healthcare, financial services, legal, education, and anything involving a licensing or paperwork process all do, this is probably where your next twenty pages should go.
Google Search Central explains how website content can become eligible to appear in AI-powered search experiences:
What Google’s AI Data Can’t Tell Marketers
I don't want to oversell this dataset. It covers two weeks inside Google's own products, and it has no click data at all.
So, while it can tell you what people ask Google's AI about, it can't tell you whether any of that sent a single visitor to a website.
That's a real limitation and the same one that showed up in Google's Merchant Center query pilot, where retailers got told what people were asking without being told whether it drove anything.
So treat it as directional, not as a plan. What it's good for is a gut check on how you're prioritizing.
If your content calendar is sorted by AI search volume, you're optimizing for the moments people find easy. The evidence suggests they're bringing AI the ones they find hard.
