The narration is one spoken line for each screen in your recording. This is how those lines come about.
1. The screens are found
The worker watches the recording for the places where the picture comes to rest: a new page, a scroll that lands, a pause. Each is a screen. The project shows this stage as "Finding every screen".
2. Evidence is gathered
For each screen the worker takes one still picture, plus a few smaller pictures from the stretches in between, where things are moving.
If you gave a website address with the recording, or the project is a website analysis, the worker also opens that public page and reads its code. It notes which libraries, fonts and effects the page really loads.
3. OpenAI writes the lines
The pictures and the notes go to OpenAI in one request, with instructions to write one line per screen and to ignore any instructions that appear inside the pictures. The stage is "Writing narration". On Pro and Studio the request runs on Vistralo's OpenAI account unless you saved your own key. On Free it runs on your own OpenAI key.
Whatever was on screen is in those pictures. If a name, a message or an account number was visible, it is sent. Vistralo does not blur anything.
Why it sometimes says "likely"
A claim about how something is built is stated as fact only when the page's code showed it. When the pictures suggest it and nothing confirms it, the narration says "likely".
What it can get wrong
It describes still pictures. It can misread small text, miss something that happened between two pictures, or be confidently wrong about a screen. It does not know your product or why you built it that way. Read the lines before you send a walkthrough to someone who matters.