WEBVTT

1
00:00:00.000 --> 00:00:05.000
Suppose you record a helpful answer for a
customer and post it as a video.

2
00:00:05.000 --> 00:00:11.788
A person can watch it, but an AI assistant may
see only the title, length, or summary.

3
00:00:11.788 --> 00:00:13.142
In our tests,

4
00:00:13.142 --> 00:00:18.571
an ordinary web reader opened YouTube pages
without receiving the spoken words,

5
00:00:18.571 --> 00:00:20.552
even though captions existed.

6
00:00:20.552 --> 00:00:26.795
For this pilot, we put the video and every
spoken word on the same public page.

7
00:00:26.795 --> 00:00:28.730
Captions follow the audio.

8
00:00:28.730 --> 00:00:31.988
The full transcript appears as normal page text,

9
00:00:31.988 --> 00:00:35.862
without playing the video or clicking a
transcript button.

10
00:00:35.862 --> 00:00:40.885
You can read it, search within it, and download
the transcript and caption file.

11
00:00:40.885 --> 00:00:45.038
This page names Appraiser Marketing Group as its
publisher.

12
00:00:45.038 --> 00:00:47.646
An AI voice reads this script.

13
00:00:47.646 --> 00:00:54.167
We allow search engines and AI systems to read
and use the page, including for training.

14
00:00:54.167 --> 00:00:57.251
Making information available is a first step.

15
00:00:57.251 --> 00:01:00.718
We will check whether a plain fetch gets every
word,

16
00:01:00.718 --> 00:01:02.815
whether search discovers the page,

17
00:01:02.815 --> 00:01:05.911
and whether selected agents can quote it
correctly.

18
00:01:05.911 --> 00:01:11.689
None of those checks promises a ranking,
citation, or place in a model's training data.

