YouTube refuses caption scrapers with a well-formed empty 200

The refusal happens at the caption fetch, not at the watch page. You can still read `ytInitialPlayerResponse` and pull a `captionTracks[].baseUrl` out of it ^[2]. The URL resolves and the request completes. The body has nothing in it ^[3]. The dev.to post that measured this names the missing ingredient as a PO Token, or Proof of Origin, generated by YouTube's BotGuard JavaScript, and describes the zero-byte 200 as a refusal ^{[4] [5]}.

"Two days of my own attempts were worth less than pip install," the post's author wrote ^[8]. The alternative is to run BotGuard yourself in a headless browser and mint tokens, which the post reports works and costs one browser process per request ^[6]. That bill grows with traffic. Take the library instead and you are depending on someone else's release cadence ^[7].

Either path leaves the address problem untouched. After about 30 transcript fetches from a home connection, youtube-transcript-api started raising `IpBlocked`, which the post identifies as ordinary rate limiting and not the token wall ^[16]. The same post's language test ran 39 long-form videos ^[10]. At one fetch per video, a single home IP crosses the 30-fetch mark with nine videos still to go ^[19].

Both headline results come from one operator on two consecutive days. The scrape test was seven variations on 2026-09-21 from a clean residential IP ^[3]. The translation test was four videos on 2026-09-22 through residential proxies, three retries apiece, each retry on a fresh exit IP ^[11]. Rotating the exit on retry rules out one poisoned address. Four videos is too small a sample to separate YouTube's auto-translate path from the luck of the draw. For the zero to transfer to your workload, the path is what has to be failing.

Two of the post's defensive checks are worth copying whatever route you take to the bytes. Naive SRT timecode math turns 59.9999 seconds into `00:00:60,000`, and some players reject the entire file over that one cue ^[13]. Rounding to milliseconds before splitting fixes it, and the listed boundary tests are 0.9995, 59.9999, 3599.9999, 86399.9999 and a negative ^[14]. The second is a length assert on batched self-translation: when the endpoint returns a different number of results than you sent, a silent `zip()` shifts every subtitle after that point by one cue, and the file, in the post's words, "still looks fine until someone watches it" ^[15].

Preferring a language instead of requiring one is the cheap fix. Most tools ask YouTube for English and give up, so a Korean video with a perfectly good Korean transcript comes back empty and you cannot tell whether it had captions at all ^[9]. Preferring took that 39-video mixed English/Korean sample from a pile of empty rows to 39 transcripts ^[10]. The egress side stays expensive: a residential exit per request, and a new exit on every retry ^[18]. Datacenter addresses are worse, with a fresh cloud container usually blocked on the first call ^[17].