Publish
Every hour, the counts per video are recomputed and the site refreshed, with the corpus coverage, the known limits and the corrections log.
- 1Every hour, the counts per video are recomputed from the base.
- 2The site’s pages are rebuilt, in French and in English.
- 3The same figures are served as JSON and CSV; the texts go through the explorer and the API, with a key.
- 4Coverage, the known limits and the corrections log are published alongside.
- Cadence
- hourly
- Measures
- comparisons between orientations as means per video
- Writes
- this site, the JSON and CSV files, the explorer and the API
Every hour, the site recomputes itself
The pages of this site are not written by hand. They are rebuilt from the base, in French and in English, as the previous steps progress.
The counts per video
Every hour, a table of facts is recomputed for each video. It holds its sentences, political sentences, themes, comments, speakers, audience and possible removal. Wherever one orientation is compared with another, the measure is a mean per video, each video counting once.
The pages and the open data
The figures of the videos of the week, the themes, the public figures, the trends, the channels and the removed videos are also served as JSON and CSV, without a key. Transcripts, classified sentences and comments go through the explorer and the API, with a key and an access tier.
Meaning-based search
On the central server, each sentence is converted into a numerical vector. The explorer uses these vectors to find sentences with a similar meaning, beyond the exact words.
The state of transcription
Of the 72,107 videos tracked, 35,481 have been extracted and 29,502 transcribed. In the last seven days, 1,592 videos were transcribed.
A transcribed video is analysed once its sentences have been split and classified. The gap between the last two columns corresponds to the videos still waiting for this step.
Last detection: 5 h ago. Last transcription: 13 min ago.
| Year | Videos tracked | Share transcribed | Share analysed |
|---|---|---|---|
| 2016 | 1,855 | 43% | 43% |
| 2017 | 1,749 | 40% | 39% |
| 2018 | 1,496 | 41% | 40% |
| 2019 | 1,380 | 50% | 49% |
| 2020 | 2,237 | 52% | 51% |
| 2021 | 3,902 | 65% | 64% |
| 2022 | 5,610 | 69% | 68% |
| 2023 | 7,432 | 74% | 74% |
| 2024 | 10,569 | 65% | 65% |
| 2025 | 10,915 | 28% | 29% |
| 2026 | 11,318 | 25% | 24% |
The limits of the data
Transcription
Automatic transcription contains errors, more frequent on noisy recordings, overlapping speech and proper names.
Classifiers
Their results are estimates.
Read on
The theme classifiers detect a subject, whatever the position expressed. A percentage computed on too few sentences is not published.
Coverage
The share of videos transcribed varies by year and orientation.
Read on
In early October 2026, few French far-right videos published in 2025 and 2026 have been transcribed. A single Quebec channel supplies most of the far-right political sentences for those two years.
Orientations
These are research categories assigned to whole channels.
Read on
A single video may depart from them.
Deletions
A deleted video has become unavailable.
Read on
It was removed by the platform or the author, made private, or its channel closed. The reason is known only if the platform states it.
Automatic texts
Summaries, arguments by theme, records of deleted videos and section texts are written by our local language model.
Read on
They are not reviewed and may contain errors of detail.
October 2026 audit
Collection errors were found.
Read on
Like counts under-reported since March 2026, missing dates on part of the comments, comment deletions wrongly counted. The bulk corrections are logged below.
Corrections log
Main bulk corrections applied to the extracted data, with the rule used.
| Date | Data | Rule applied | Rows |
|---|---|---|---|
| 3 October 2026 | Comment sentences, YouTube comments | Dates of comments captured with the French interface (“il y a 3 mois”): recomputed from the time of capture, as for the English interface. | 6,238,021 |
| 3 October 2026 | Audience measurements | Empty audience readings, recorded for seven minutes by a faulty version of the collector: removed, they are not observations. | 15 |
| 3 October 2026 | TikTok comments | TikTok comments: the source channel, replaced by a technical identifier during distributed collection, is restored from the video. | 4,702 |
| 3 October 2026 | Like and view counts, Comment sentences | Video likes shown with comments: the value is read from the corrected video record instead of the copy kept with each comment. | — |
In addition, the like count is not shown when it is below 0.05% of the views of a video viewed at least 10,000 times. Such values come from a reading error during collection.