Skip to main content
One tool. It answers a question you cannot answer from your own analytics: which YouTube videos does Google’s AI Overview cite when someone searches for the things you care about?

​
get_cited_youtube_videos

Give it a list of search intents. For each one it reads Google’s AI Overview, pulls out the YouTube videos the overview cited, and returns them with their transcripts.

​
Division of labour

The agent writes the intents; the tool only reports what Google actually showed. That split is deliberate — an agent that knows your business can generate forty relevant intents far better than a fixed keyword rule could, and keeping generation out of the tool means the tool never invents a query you did not ask for.

​
Parameters

Ten intents per call is a context limit, not a rate limit. Forty full transcripts in one response would not fit in the conversation, so the cap forces the agent to triage: sweep wide on "summary", then re-call with "full" for the handful of videos that matter.

​
What comes back

A table for you to read, plus structured data for the agent: Each video carries url, videoId, title, channel, publishedAt, citedIn, transcript, transcriptStatus, transcriptChars and transcriptTruncated.

​
Which market gets searched

With country: "auto" the market is worked out in order: your workspace’s own country, then the country implied by your website’s domain (.in, .co.uk), and only then the US. The result always says which market was used and where that choice came from — workspace, tld, explicit or default. When it lands on default, the result says so and names the parameter to change, because a US search nobody asked for should never pass as one they chose.

​
Where the search runs

Google has no worldwide search — every query is geo-targeted to one country. So "worldwide" runs one search per country and costs a credit for each: ten intents across five countries is fifty credits, not ten. The countries sampled are the large English-language markets on purpose, because an intent is one string in one language; running English text against Japan measures how Google handles a foreign-language query there, not whether the topic is covered. The same intent genuinely differs by market. “how to fix a leaking tap” cited 9 YouTube videos in the US, 8 in India and 6 in the UK — different videos each time. Every result carries the country it was observed in, where that choice came from (explicit, workspace or default), and when.

​
Four outcomes, not a yes/no

An intent that returns no videos is not one finding but four, and the tool names which:
AI Overviews are volatile. The same query can cite ten sources on one check and none fifteen minutes later. Treat a single observation as a sample, not a verdict — especially before concluding that a topic is uncontested.

​
Transcript language

Transcripts come back in the language of the country searched, so country: "DE" returns German without asking. Pass transcript_lang to override. Each video reports transcriptLang — the language it actually came back in — and transcriptLangsAvailable, every language that video has captions for. If a video has no track in the language you asked for, the tool says which language it returned instead and what else was on offer, rather than handing you a transcript you cannot read and calling it success.

​
When captions cannot be read

Some videos have no caption track at all. Rather than an empty result, each video then carries:
  • quote — the passage Google itself quoted from the video, verbatim. Short (about 133 characters) but genuinely spoken.
  • coverage — a summary and topics synthesised from that quote plus the uploader’s own description, with derivedFrom naming the sources used.
coverage is derived, not spoken. It is never a substitute for the transcript and is never presented as the video’s words. Where the source material is too thin, the summary says so instead of guessing — because a fabricated account of what a cited video says would corrupt the very decision this tool exists to inform.
transcriptStatus is explicit about failure rather than returning an empty string: Searches that cited nothing are reported by name with which of the four outcomes applied, so “no videos” is never confused with “no overview” or with a failed lookup.

​
Cost

One keyword-lookup credit per intent, per country searched. A ten-intent call in one country spends ten; the same call with country: "worldwide" spends fifty. If your remaining allowance is smaller than the batch, the tool searches what you can afford and names the intents it skipped rather than failing the whole call or quietly overspending. Results are cached, so re-running the same intent inside the caching window is free.

​
Cost at a glance

See Plans, credits and limits for the allowances behind each credit.