I'm Matheus Matos, and in 2016 I built a chatbot called Findera — find + era — that answered questions about local businesses in natural language. You'd type "where's a good place for bacalhau near me?" and it would come back with a real recommendation.

It wasn’t a random bet. Between 2012 and 2015, my father, brothers and I ran a printed magazine for local businesses. I spent years watching small businesses struggle to be found. The chatbot was the next attempt at solving the same problem.

In April 2016, at F8, Zuckerberg opened the Messenger Platform to developers. Amazon had Alexa. Microsoft had Cortana. Google was building Assistant. Every major tech company was converging on the same idea: that the next computing interface would be conversational — the same kind of shift we'd seen with the personal computer and with mobile. Natural language instead of clicks. I believed it then, and I still do.

The product worked, technically. But NLP wasn't mature enough for the experience to feel right, and the Messenger bot wave crested and broke faster than anyone expected. I shelved it.

The idea didn't go away. It came back in 2020 as a different shape: local feeds. Each place in the world would have its own activity feed — Copacabana in Rio de Janeiro; Manhattan; Mission District; Tatuapé, my neighborhood in São Paulo, one of the largest cities in the world. Businesses would post to their local feed. Locals would read it. The monetization layer would be ads targeted by geopath instead of by device ID.

That version died on the classic cold start problem: you need users to attract businesses, and businesses to attract users. The unit economics of building a social network are brutal, and I wasn't willing to raise the kind of money it would take to win that game.

The shift that made it click

By 2023, LLMs changed the question. The problem was no longer "how do we build a consumer product around local data?" The problem became "why don't AI agents know what's happening near me?"

ChatGPT, Gemini, Perplexity — they all struggle with the same blind spot. Ask them what's happening in your neighborhood today and you get either silence or something pulled from a review platform updated six months ago. The data they need doesn't exist in a format they can consume.

That's the gap I built the Open GeoFeed Protocol to fill. Not a consumer app, not a social network — a protocol. A standard for how local actors publish structured activity feeds, and how AI agents discover and consume them. The same idea from 2016, but finally in the right form.

What OGP is (and isn't)

OGP is infrastructure, not a product. At its core is the GeoPath — a hierarchical identifier made of GeoSubdivisions, from country down to neighborhood: bra/sp/sao-paulo/tatuape. On top of that: a root registry that maps GeoPaths to content servers, and a feed format based on ActivityStreams 2.0 anchored to a place. No dashboard, no login, no pricing page.

The registry at geofeed.ai is the root — the DNS of the protocol. It doesn't serve content. It tells you who does. The protocol is Apache 2.0 — permanently.

Timeline