What AI can and can't see in a video
Start with the limits. A language model works with text: the spoken transcript, the title, the description and the comments. Editing, visuals, thumbnails and publishing time stay invisible unless you add screenshots. Numbers like views and retention exist only in the channel owner's analytics, and a model shouldn't be allowed to guess them.
So the right question is not "why did this video take off?" but "how is this video built?" You draw conclusions about causes yourself, by comparing several examples.
Choosing a tool and getting the transcript
Some services accept a YouTube link. Gemini can answer questions about a video's content, and NotebookLM takes a video as a source and answers from its transcript. Availability depends on your account and region and changes over time, so don't build a routine around links alone. The dependable method works almost anywhere: open the transcript under the video, copy the text and paste it into ChatGPT, Claude or another chat.
Videos in Uzbek need extra care. Automatic captions may be missing or full of mistakes. In that case, generate the transcript with a separate speech-to-text tool and read it through yourself, otherwise the model analyzes a distorted text.
If you only pasted a link and the answer feels generic, the model probably never opened the video. The test is simple: a real analysis quotes specific moments and phrases from it.
Breaking down a single video
Paste the transcript and ask for the skeleton of the script: how the video opens and what it promises in the first seconds, which blocks it consists of, where the call to action sits and how it ends. Ask for one short quote per block so you can check the answer against the text. Then follow up: which retention techniques appear, such as a question, a story, a contrast or a promised result, and where the pace drops.
Say you run a renovation studio in Tashkent and you are studying a competitor's apartment walkthrough. The skeleton might show that the host states the client's problem right away and leaves the budget talk for the end. That is an observation, not a recipe, but it tells you what to compare next.
Comparing several competitor videos
One video is an accident; several are a pattern. Upload three to five videos from one channel or from different channels in your niche and ask:
- which topics repeat and how they are worded in titles;
- what the openings of the most visible videos have in common;
- which audience questions the authors answer and which they avoid;
- who is on screen, an expert, a host or a voiceover, and how the product is shown.
Using comments as demand research
Copy public comments without names or personal details and ask the model to group them by meaning: questions, objections, thanks, requests for new topics, complaints. For each group you want a few typical quotes, and separately a list of questions the author never answered. If comments mix Uzbek and Russian, tell the model to group by meaning rather than by language.
That list becomes website FAQ entries, post topics, replies for managers in Telegram and objections for sales scripts. If viewers keep asking about installments or paying through Click and Payme, that is a reason to answer it inside your next video.
Moving the findings to Instagram and Telegram
A strong script structure isn't tied to one platform. An opening with a promise, one clear block and a plain call to action work in a long YouTube video, in a Reel and in a clip for a Telegram channel. Based on the gaps you found, ask the model for topics, several opening lines in your own tone, and a video outline built on your examples.
After that comes ordinary work: filming, delivery, publishing and comparing results in your analytics. The model is an assistant, not a replacement for the expert on camera. Don't copy a competitor's script with swapped words. Viewers notice, and your experience is what sets you apart.
Common mistakes and where to go next
Most problems come from one of the points below. Run through them before you build a content plan on the analysis. Then repeat the review once a month across several channels in your niche, and topics for your plan will accumulate on their own.
If you want to turn the findings into regular Reels and posts, markdev can help with Instagram management for businesses in Uzbekistan. Tell us about your task and we will work out where to start.
- drawing conclusions from a single video;
- treating the model's "it works because..." as fact;
- trusting quotes without checking them against the transcript;
- uploading clients' private materials or internal recordings to public services;
- forgetting that visuals and thumbnails are absent from a transcript.
FAQ
Can AI actually watch a YouTube video?
Some services accept a link and answer from the content, but this depends on your account and region. A safer way is to paste the transcript as text and check that the answer quotes real phrases from the video.
What if a video in Uzbek has no captions?
Use a separate speech-to-text tool and correct the result by hand. Errors in the text directly damage the analysis.
Will this tell me why a video got views?
No, only hypotheses. Views and retention live in channel analytics, and editing and thumbnails are not in a transcript. Confirm any pattern across several videos.
Is it safe to upload internal company videos?
Avoid sending private materials or client data to public services. Use public videos and comments stripped of personal details.
Let’s look at your task for free
Leave your contact — in a 15-minute call we’ll show where leads get lost and what to fix first.