Every internet era creates a layer that organizes the mess left by the one before it. For the public web, that layer was search. Whoever owned the index owned the web.
Google’s index was central because the data was public. Everyone saw roughly the same web, so one index could serve everyone.
AI is creating a new organizing layer. Many people assume it will look the same: one assistant, in the cloud, owned by the first company to get there. But the data is different. The next layer does not organize the public web. It organizes you.
1. The Web Keeps Re-Aggregating
At first there were a few pages, reached by address. Then pages became cheap to make, and the pile grew too fast to browse. The first organizer was human: the Yahoo directory, a hand-built table of contents. It failed because people cannot index a pile that grows faster than they can read.
Search automated the job, and for a while the mess looked solved.
Then came platforms. They grew around the things search could not reach: social graphs, commerce, video, anything live or logged in. And they walled themselves off. A platform is an organizer for one slice of the web, kept out of the open index so users stay inside. By the end of the web era, the web was fragmented again, this time on purpose.
So the web does not move steadily toward order. A mess becomes an index, the index helps create platforms, the platforms create a new mess, and the next organizer has to deal with that.
2. The Layer Follows the Data
An organizing layer goes where the data is, and takes the shape of that data.
Public substrate
Personal substrate
AI organizes a different kind of data. It does not just organize documents you read. It organizes actions you want to take across apps and websites. To do that, it needs your files, history, accounts, credentials, and unfinished plans.
There is no shared copy of that data. My context and yours cannot be merged into one useful index. So the next organizing layer has no reason to gather in one place. It belongs where the context already is: on the device.
The platform walls reinforce this. To organize the platform web, an agent has to log in as you. A central service logging in as millions of people looks like a bot farm. It gets CAPTCHA’d, rate-limited, or blocked.
So the walls that kept search out also keep a new central aggregator out. The only aggregator those walls accept is the one already inside them: software running on your authenticated device.
Put the data and the walls together, and the claim is stronger than “local is nicer.” The AI organizing layer is hard to monopolize by structure.
What remains is a shape: local aggregation, one instance per person. The winner, if there is one, is the standard those instances share - a protocol, a commons, a surface owned by everyone who uses it - not a gate with an owner.
That is why a local agent like Mango is not almost the next Google. It is one local instance of a layer that cannot have one owner.
3. The Last Index That Hasn’t Come Home
There is one gap. The organizing layer is moving local, but intelligence mostly has not. The model still runs in someone else’s cloud. The center has not disappeared; it has moved down one level, from the index to the model.
That is why running the model on your own hardware is not a side quest. It is the same move. The index has already followed the data home. The model is the last index that has not - and when it does, there is nowhere left for the toll to sit.