I’ve been following DuckDuckGo’s numbers this week with more interest than the story seems to deserve at first glance. Last week the company launched extensions for Chrome and Firefox that set noai.duckduckgo.com —their AI-free search results page— as your default search engine with a single click. In itself, it’s a minor release. What caught my attention is why it landed, and what’s happening with traffic.
Since Google announced in May its revamp of AI-first search —the biggest change to that product in over 25 years, where AI Overviews and an AI chat mode now sit above the old “10 blue links”— DuckDuckGo’s AI-free page has been climbing. Visits tripled in a single day at the end of May, a new record since Google’s announcement, and the company carefully clarifies that it’s not an isolated spike: traffic is holding steady around 84% above baseline. Installs in the US iOS app hit a peak of nearly 70% week-over-week growth. That’s not a protest. It’s a change in behavior.
I want to be precise about what this is and isn’t, because the headline framing—“anti-AI sentiment”—is the least interesting read. DuckDuckGo isn’t an anti-AI company; it runs its own chatbot and a paid plan with access to frontier models. The people switching aren’t rejecting AI. They’re rejecting AI as a non-negotiable default. Those are very different things, and that difference is the whole story for us.
Why developers feel it first
Here’s my thesis: when you install a summarization layer in front of the source, developers are the canary in the mine. We’re the ones who notice when the synthesized answer is subtly wrong—not obviously-hallucinated-wrong, but plausibly-wrong in a way that costs you forty minutes. We’re the ones who needed the real method signature in the real version we’re actually running, and got a confident paraphrase of a deprecated API. We’ve all been through the experience of an AI Overview cheerfully summarizing a Stack Overflow thread while burying the only comment, three answers down, that had the real fix.
For most searches, a good summary is a genuine improvement. “What’s the capital of such-and-such country”, “convert this unit for me”, “what year did X happen”—there the abstraction layer is pure gain. The problem is our queries aren’t those. Our queries are version-specific, edge-case-specific, and often about exactly the thing the documentation got wrong. For that kind of work, the distance between “the answer” and “the source” is where bugs live.
The same instinct, already in our workflow
What’s interesting is we’ve been solving this within our tooling without calling it the same way. Context7, which we have on our radar for a future piece, exists precisely because feeding an agent stale or generic documentation produces stale or generic code—so it brings live, version-specific docs straight into the model’s context. RAG pipelines over internal docs are the same move. The instinct to keep the primary source close, rather than trust a summary of it, is something good engineering teams already practice. The DuckDuckGo shift is that same instinct showing up in consumption behavior.
So the question I’d ask my team isn’t “is AI-powered search good or bad?”. It’s a more useful one: who controls when the abstraction layer gets between you and the source? When you can toggle it—summarize when I want speed, get out of the way when I need ground truth—it’s a tool. When it comes soldered in as a default with no exit, it’s a liability for exactly the kind of precision work we do all day.
The DuckDuckGo extension is a one-click toggle. The more important version of that toggle is the judgment to know which mode your task actually needs.
The takeaway for teams
I’m not telling anyone to switch search engines. I’m saying: notice the reflex. The teams that will navigate the next two years well are the ones that build the discipline to ask themselves, on each task, whether they want the summary or the source—and the tooling habits to get to ground truth fast when the summary isn’t enough. That’s a workflow decision, not a search engine preference. The browser extension is the trivial part. The judgment is what’s worth developing.
Have you already changed how you search for documentation since AI got mixed into the results? Do you feel it as help or as another layer you have to dodge?
Sources: