How Baiki stories are made

1 · A language model writes the draft

We give it a city and a request — “Soho, hidden courtyards” — and it proposes the stops and writes a first draft of each story. That draft is machine-written. We do not pretend otherwise.

2 · Every place is checked against Google Maps before a word is written about it

Ask a language model for a place and it will invent one; ask a search API about an invented place and it will confidently hand back a real one somewhere else. We measured an invented “Palazzo dei Sospiri Perduti, Roma” resolving to the Bridge of Sighs — in Venice, 500 km away.

So every proposed stop is looked up and sorted into one of three kinds:

• On the map — Google Maps lists this exact place.
• Real, but not the kind of thing a map lists — a blue plaque, a memorial tablet, a carving, a bricked-up doorway, the site of a demolished building.
• Not there at all.

The last kind is deleted. The middle kind is most of what makes a walk worth hearing, and we keep it.

3 · Checks, and a person reading a sample

Before a story is offered to a search engine it has to clear a fixed checklist — no judgement involved, so it is the same for every page. It needs a finished story of real length, its own audio, a Maps verdict from step 2 that is not “not there at all”, and a walk it actually belongs to. Any leftover scaffolding from the model — a note it wrote to itself before starting — blocks the page outright.

On top of that, the Baiki Guide team reads a sample from each city before that city is published: a spread across different walks and every language it is written in. We read for invented facts, for text that has drifted from the place it describes, and for anything the model left behind. What we find gets fixed, or that page is refused.

We are not claiming every one of these pages has been read line by line, and we would rather say so than imply otherwise. For London, published on 17 August 2026, we read 70 of the 364 stories that cleared the checklist — 19%, across 31 walks and all three languages — and refused three of them. The rest were published on the strength of that sample and the checklist above.

Two things have happened since, and they are different, so we list them separately. Every story written while the product was being built — everything on this site before 13 September 2026 — had been read or listened to by the team as it was made, on top of the sample above. Stories added from 13 September 2026 onwards are published as soon as they clear the checklist; the team samples them afterwards rather than reading each one before it goes out. We would rather say which is which than let this page imply that every one of them has had a reader.

A city nobody has sampled yet is not published here. Those pages stay available to anyone who has the link — they are simply not offered to search engines.