My scrapers broke on a Tuesday.
Not dramatically. No 403s, no rate limits, no captcha walls. The HTML just came back shaped differently, and the selectors I’d been maintaining for years pointed at nothing. That was my first real signal that Google had changed into something else. May 2026 is when Search stopped being a list of links and became an assistant that talks back.
On 19 May 2026 at I/O, Google announced what it called the biggest changes to Search in 25 years. AI Mode became the global default. The search box was rebuilt for conversational prompts. And persistent background agents entered the picture, meaning Search can now go do things while you’re not watching.
If you build bots for a living, that last part should make you sit up.
The weirdness is structural, not cosmetic
Every few years Google ships a redesign and people complain about spacing. This isn’t that. When AI Mode is the default rather than a tab you opt into, the primary output of the world’s largest information system changes from a set of pointers to a generated answer. Those are different data types. My tooling assumed the first one.
Google’s own framing is that Search now “puts the world’s information to work for you” and can do more than ever before. Read that as a product person and it sounds like ambition. Read it as someone who writes parsers and it sounds like a warning: the contract between your code and the page just got renegotiated without your input.
Then there’s the update cadence. The May 2026 broad core update landed fast and hit hard, a sharp contrast to March, which analysts described as a strange update without much power behind it. June 2026 brought a spam update, with the page indexing report delayed two weeks and autocomplete experiments running in the background. Three months, three different personalities. That’s not a stable platform, that’s a system being actively rewired while it’s running.
What this actually breaks in a bot stack
I’ve spent the past several weeks going through my own projects. The damage falls into a few buckets:
- Result extraction. Anything that assumed ten blue links and a predictable DOM tree needs rethinking. A generated answer has no stable position or structure to anchor to.
- Attribution logic. If your bot cites sources it found via Search, the source set is now mediated by a model. You’re citing a summary’s references, not your own retrieval.
- Freshness assumptions. With indexing reports delayed and core updates hitting quickly, “how recently was this indexed” became a much shakier input to any ranking or filtering logic you’d built.
- Agent collision. Persistent background agents mean Google’s automation and yours may be working the same surfaces. I don’t have data on how that plays out at scale yet. Nobody does.
The uncomfortable part for builders
Here’s what I keep circling back to. A lot of us built bots that treated Search as infrastructure. Cheap, stable, boring infrastructure you could point at and get structured-enough data from. That was never a guarantee Google made. We just got used to it.
The move to an AI-driven assistant makes Google a competitor to a whole category of tools people built on top of it. If Search can run a background agent that watches a topic and reports back, that overlaps with a lot of side projects I’ve seen shipped in the last two years. Not fatally, but noticeably.
SEO and marketing teams got the loudest wake-up call, since the whole discipline of optimizing for a ranked list assumes a ranked list exists. But the bot-building side has a quieter version of the same problem: our inputs changed and our abstractions didn’t.
How I’m adjusting
Nothing clever, just discipline:
- Stop treating any single source as infrastructure. I’m putting an adapter layer between my bots and whatever search backend they use, so swapping providers is a config change rather than a rewrite.
- Test against change, not correctness. My integration tests now fail loudly when page structure shifts, instead of silently returning empty arrays.
- Prefer APIs with contracts over pages with layouts. Slower, sometimes costlier, dramatically more predictable.
- Log everything about retrieval. When behavior changes again, I want the before-and-after in my own data rather than in a forum thread.
So when did Google get weird? Somewhere around mid-May 2026, at least officially. But the honest answer is that it got weird the moment it stopped being a directory and started having opinions. If your bots depend on it, build like that shift will happen again. Because based on the past few months, it will.
🕒 Published: