Capturing and Characterizing the Decline of American Fast-Food

After ReviewBrah asked in a May 2026 video whether fast food is “nearing the end”, I recorded 754 of his reviews dating back to 2017. The mean rating has fallen only modestly over that span, while the year-over-year standard deviation has roughly doubled, which is the primary driver of the decline.

754 ratings analyzed | Last updated: May 2026

Ratings Over Time

Each review is a dot, colored by year. The black line is the 15-review rolling average and the dashed line marks the all-time mean (6.87). Hover any point for the item, chain, date, and score.

Distribution by Year

A box plot per year. The box spans the interquartile range, the center line is the median, the whiskers extend to 1.5 times the IQR, and dots are outliers. Hover a year for n, mean, median, and SD.

Spread by Year

Standard deviation of each year's ratings. Black whiskers are 95% confidence intervals on the standard-deviation estimate. Hover a year for the exact SD, confidence interval, mean, and n.

By Chain

One panel per chain with at least 15 reviews (13 chains, ordered by review count). Dots are individual reviews colored by year. Each black line is a least-squares trend with a shaded 95% confidence band. Hover any point for details.

Yearly Average by Chain

Each chain's mean rating per year, connected into a trajectory, with dots colored by year. Same 13 chains, ordered by the early-to-recent change in mean with the largest decrease at top. Hover a dot for the yearly mean and sample size.

Who Soured the Most

Each chain's mean rating for early reviews (2017 to 2021, blue) and recent reviews (2022 to 2026, red). Dot area is proportional to the number of reviews, and the faint colored whiskers are 95% confidence intervals on each mean. Rows are ordered by the size of the early-to-recent decrease. The dashed line marks the all-time mean (6.87). Hover a dot for mean, n, and CI.

Price Over Time

Each priced review with a parseable dollar amount at or below $30 (702 reviews), plotted by date and colored by year. The black line is a least-squares fit with a 95% confidence band. Hover any point for the item, chain, date, and price.

How I did it

The original approach used YouTube’s auto-generated captions via youtube-transcript, but YouTube started aggressively rate-limiting the transcript API endpoint. The updated approach bypasses this entirely by downloading audio and transcribing locally with OpenAI’s Whisper model.

Step 1: Listing All Videos in a Channel

Use yt-dlp to get video URLs and upload dates from the channel:

yt-dlp -s --flat-playlist 
  --print-to-file "%(upload_date)s %(id)s %(title)s" videos.txt 
  "https://www.youtube.com/@TheReportOfTheWeek/videos"

This gives you a file with one line per video containing the date, video ID, and title.

Step 2: Downloading Audio

YouTube’s transcript/subtitle endpoints are heavily rate-limited, but the audio stream endpoint is not. Download audio for each video:

mkdir -p audio
while IFS=' ' read -r date id rest; do
  yt-dlp --extract-audio --audio-format mp3 
    -o "audio/${id}.mp3" 
    "https://www.youtube.com/watch?v=${id}"
done < videos.txt

For ~57 videos this produces around 850MB of MP3 files.

Step 3: Transcribing with Whisper

Install mlx-whisper for fast transcription on Apple Silicon (about 20 seconds per video vs 30+ minutes with CPU-only Whisper):

pip install mlx-whisper

Then transcribe each audio file:

import mlx_whisper, os, glob

for mp3 in sorted(glob.glob("audio/*.mp3")):
    vid_id = os.path.basename(mp3).replace(".mp3", "")
    result = mlx_whisper.transcribe(
        mp3,
        path_or_hf_repo="mlx-community/whisper-large-v3-turbo",
        condition_on_previous_text=False
    )
    with open(f"transcripts/{vid_id}.txt", "w") as f:
        f.write(result["text"])

The condition_on_previous_text=False flag is important — without it, Whisper occasionally gets stuck in hallucination loops where it repeats the same phrase for minutes.

Step 4: Extracting Ratings with an LLM

Combine all transcripts into a single file with their dates prepended:

while IFS=' ' read -r date id rest; do
  echo "=== ${rest} (${date}) ==="
  cat "transcripts/${id}.txt"
  echo
done < videos.txt > combined.txt

Then feed the combined transcripts to an LLM. I used Claude’s CLI in batches of about 20 videos at a time to avoid output truncation:

claude --print -p "Extract all food ratings from these ReviewBrah transcripts.
Output a markdown table with columns: Item, Restaurant, Cost, Rating (out of 10),
Date (YYYY-MM-DD). Only output the table rows, no headers." < batch1.txt

Step 5: Deduplication and Cleanup

The LLM output across batches may have some inconsistencies. Merge the batch outputs and do a final cleanup pass:

cat batch*_output.txt | claude --print -p "Deduplicate these table rows,
normalize dates to YYYY-MM-DD, and output a clean markdown table."

Conclusion

By combining:

• yt-dlp to download audio (bypassing YouTube’s transcript rate limits)
• mlx-whisper for fast local transcription on Apple Silicon
• An LLM to extract structured ratings from free-form transcripts
• A final deduplication/cleanup pass

…you get a comprehensive, structured dataset from hundreds of YouTube videos without hitting any API rate limits.