Agent Optimization
AO visits business websites, checks what they really let you do — search a catalogue, see a price, check stock, request a quote, book — and publishes it in the formats AI agents read.
Every result says how we know: the action was performed twice by our crawler, and the evidence is pinned to a versioned index.
Menus, pop-ups and scripts hide what a business can do. An agent has to guess where the search box or the quote form is.
When an agent can't find a price or a booking form, it moves on to the next site. Nobody sees the lost request.
A page that says "order online" is not the same as an order form that works. Agents need facts that were checked.
How it works
Our crawler opens the site like a careful visitor. It identifies itself, follows robots.txt and keeps a slow pace.
Each action — a search, a price, a stock check — is tried twice with different inputs. Only what works twice counts as verified.
Results go into a versioned capability graph. Every fact keeps its evidence, so any answer can be traced and reproduced.
Agents query AO over REST or MCP, and businesses get the standard files — agent card, MCP server card, API catalog.
For businesses
For agent builders
An action counts only after our crawler has done it twice.
Every index version can be rebuilt byte for byte from its evidence.
Signed identity, robots.txt respected, nothing submitted that creates an order.
Sites in German, French or English; one English index, with names kept as written.
Now piloting
An index is only useful when it is dense. We're mapping Zürich's businesses first, then testing whether agents get better answers through AO than through plain web search.
If you are an agent, start here. These files describe our API and how to call it.