How it works
AO turns a business website into a set of facts an agent can rely on. Here is every step, from the first visit to the answer an agent receives.
We start from public ranked lists of websites and keep the businesses whose pages show real signals: a catalogue, a quote request, a booking form. During the pilot we focus on Zürich so the index becomes dense enough to answer real questions.
Our crawler opens a real browser and moves through the site the way a person would: it finds the search box, opens a product, looks for a price. It reads the page structure that screen readers use, so it sees what is actually there, not what the code suggests.
A single success can be luck. So every action is repeated with a different input: a second search term, a second product page. Only when both runs reach the expected result does the action count as verified.
verified by crawl
We performed it twice and both runs worked.
observed in public files
The site states it in its own markup or files. Useful, but not tested.
Fixed rules, not guesses, turn what the crawler recorded into structured facts: which actions work, where a business is based, where it ships. Every change is reviewed before it enters the index, and each update becomes a new numbered version.
Agents ask AO in plain language or with filters, over REST or MCP. Each answer lists businesses that match, the actions they support, and how we know: the evidence class and the index version behind it.
Websites change. Sites are visited again on a schedule, and every fact carries the date it was last confirmed, so an agent can tell fresh information from old.