A version of this story gets repeated in every publishing group chat I’m part of, and it goes something like this: AI answer boxes are eating the entire internet, every blog is doomed, pack it up.
It’s a tidy story. It’s also wrong in a specific, useful way, because it treats every blog as if it’s competing for the same kind of search, when the data shows something much more uneven happening underneath the panic.
The overall picture, before the split
Start with the baseline, because the shift is real even before you get to which categories are hit hardest. Researchers Athena Chapekis and Anna Lieb, publishing for the Pew Research Center, tracked nearly 69,000 real Google searches from a representative panel of US adults and found that “Google users who encountered an AI summary clicked on a traditional search result link in 8% of all visits.”
Users who never saw a summary clicked through roughly twice as often, in 15% of visits. That finding comes from a broad, methodologically careful look at ordinary search behavior across a real cross-section of searches, and it confirms the basic mechanism: when Google answers the question itself, fewer people keep going.
What that Pew data doesn’t do, deliberately, is break the result down by topic. It’s an average across every kind of search, which is exactly where the “every blog is doomed” story starts to fall apart, because an average hides enormous variation underneath it.
What actually decides who gets hit
A 2026 keyword study by marketing researchers Trevin Shirey and Macy Storm, who analyzed 2.37 million US search terms, gets at that variation directly. Some categories saw AI Overviews on roughly a third or more of their searches, among them pets and animal content, travel guides, mental health topics, and personal finance. Other categories saw them far less often.
The study’s own explanation for the gap is blunt: “If you rely heavily on informational queries to drive revenue, AI Overviews will hit you harder than an industry that relies more on transactional queries to drive revenue.”
That’s the one thing the split comes down to. The topic itself doesn’t decide it, and neither does how well-written the content is or how much domain authority a site carries in the old SEO sense. What decides it is whether the underlying question is one Google can fully answer on the results page, or one that still requires the reader to go somewhere and do something.
The same study found AI Overviews showing up on “39% of them producing AI Overviews” for informational searches, against 15.8% for transactional ones. A search for what a word means, how a condition is typically treated, or what year something happened is a question with a finished answer. A search meant to end in a booking, a purchase, or a specific account login generally isn’t, no matter how confidently an AI system tries to summarize it.
It’s worth sitting with why that split exists instead of just accepting the numbers. An AI system summarizing a page is essentially doing extraction. It pulls the checkable fact out of the writing and discards the rest, because the rest wasn’t the point of the query.
A reader asking what year a law took effect, or what the typical dosage range for a supplement looks like, wanted the fact, not the writer. A reader asking which specific hiking trail near a specific town is actually worth the drive, or whether a specific product held up after six months of real use, wanted a person who’d actually been there or used the thing. That second kind of question resists extraction because no single correct fact exists to pull out, only someone’s account of it.
The properties I watch don’t all move the same way
I manage content across several publishing sites at once, aimed at different audiences and built around different kinds of questions, and this pattern tracks with what I actually see in the dashboards, even without pretending I have a controlled study of my own.
The pages built around a pure factual question, the kind with one clean, extractable answer, behave differently from the pages built around a specific product, a specific place, or a specific person’s actual experience with something. The first kind is exactly what an AI summary is built to replace. The second kind is harder to compress into three bullet points and a citation, because the value was never really the fact itself. It was the specific voice delivering it.
A fair amount of my actual job is title generation and figuring out how to angle a piece so it earns attention instead of assuming it deserves it, and that work has gotten more interesting rather than less useful since this shift accelerated. A generic, fact-first headline is now competing directly with a machine that can answer the underlying question before a reader ever reaches the page. A headline built around a specific, opinionated angle, one an AI Overview has no equivalent version of because there’s no single fact to extract, is still doing the job a headline is supposed to do.
Where that leaves the rest of us
I don’t think this lets anyone off the hook, and I’m wary of the version of this argument that turns into “just make better content” as if that alone solves a structural shift in how search engines work. But panicking about a single average number, without asking whether your specific blog lives on the informational side of that split or the side that still requires a reader to click through and do something, wastes energy that could go toward an actual decision.
The honest question is whether the specific questions your blog answers are the kind a summary box can finish in two sentences, or the kind that still needs you, rather than whether AI answer boxes are a threat in some general sense.
I keep coming back to something I believe about most hard problems, which is that the mistake is usually treating a big, scary trend as one giant thing instead of breaking it into the specific pieces that actually apply to you. “AI is killing blogs” is the giant, scary version. “My mental-health explainer competes directly with an instant summary, and my review of a specific product doesn’t” is the version you can actually act on.
