Build an AI-Powered YouTube Script & Audience Retention Analysis System

# Build an AI-Powered YouTube Script & Audience Retention Analysis System ## Project Overview I run an established YouTube channel focused on language learning, and I am looking for a developer to build a private analytics system that connects my historical YouTube performance data with the scripts/transcripts of my videos. The goal is to determine, using actual channel data, which elements of my video scripts correlate with stronger or weaker audience retention and then use those findings to improve future scripts. I am NOT looking for a generic AI tool that simply asks ChatGPT whether a script is "good." I want a data-driven system that combines: * YouTube Analytics API data * YouTube Data API data * YouTube Reporting API data where appropriate * Timestamped video transcripts/captions * Audience retention curves * Python-based statistical analysis * OpenAI API analysis/classification The ultimate objective is to create an evidence-based "YouTube Writing Playbook" specific to my channel and audience. ## Core Workflow The system should work approximately like this: YouTube channel → Import all historical videos → Import video metadata and performance metrics → Import audience retention curves → Import/download timestamped captions/transcripts → Align transcript sections with retention data → Use AI to classify structural/script elements → Compare those elements with actual performance across videos → Identify recurring patterns → Generate individual video postmortems → Generate/update channel-wide evidence-based script-writing rules ## Phase 1: Import Historical YouTube Data The system should automatically retrieve all eligible videos from my YouTube channel. For each video, store information such as: * Video ID * Title * Publish date * Duration * Views * Watch time * Average view duration * Average percentage viewed * Subscribers gained/lost * Likes * Comments * Shares * Traffic-source information where available * Impression/CTR information where available I would also like performance snapshots to be stored when practical so that new videos can eventually be compared at standardized ages, such as 7 days and 30 days after publication. ## Phase 2: Audience Retention Data This is one of the most important parts of the project. Use the YouTube Analytics API to retrieve audience-retention data for individual videos, including metrics such as: * audienceWatchRatio * relativeRetentionPerformance * startedWatching * stoppedWatching * totalSegmentImpressions The system should preserve the retention curve and map elapsed-video ratios to actual timestamps based on each video's duration. The objective is to know what was being said in the video when unusually strong or weak retention behavior occurred. ## Phase 3: Timestamped Transcripts Retrieve timestamped captions/transcripts for videos where available. The system should store transcript segments with: * Video ID * Start timestamp * End timestamp * Transcript text If automated YouTube caption retrieval is not possible for certain videos, the system should make it easy for me to import my original script/transcript manually. ## Phase 4: AI Script Classification Use the OpenAI API to analyze scripts/transcripts using structured output. Important: The initial script-classification stage should NOT see the retention/performance results. I want to reduce hindsight bias. The AI should objectively identify things such as: * Hook * Core promise * Time to core promise * First actionable value * Personal stories * Concrete examples * Abstract/theoretical explanations * Credibility statements * Questions directed at viewer * Open loops * Payoffs * Topic transitions * Repetition * Calls to action * CTA timing and duration * Major script sections Each identified element should ideally have: * Classification/type * Start timestamp * End timestamp * Relevant transcript text * Confidence score where useful I am open to recommendations from the developer for additional useful script features. ## Phase 5: Retention + Script Analysis The system should align the AI-generated script classifications with the actual retention curve. For example, if audience retention falls unusually quickly during a certain part of a video, I want to know: * What the speaker was saying * What type of script section it was * Whether similar sections perform the same way across other videos * Whether the result appears unusual relative to the channel baseline I do NOT want the system to assume that correlation automatically means causation. Ideally, the system should distinguish between: * Observation * Hypothesis * Supporting evidence * Confidence level * Recommended future experiment ## Phase 6: Cross-Video Analysis This is the most valuable long-term component. I want to be able to investigate questions such as: * Do videos that state their main promise earlier have better 30-second or 60-second retention? * Does a longer hook hurt early retention? * What happens to retention during abstract explanations? * Do personal examples tend to stabilize or increase retention? * How quickly should an abstract explanation be followed by a concrete example? * How does CTA length affect retention? * How does CTA timing affect retention? * Which types of transitions cause viewers to leave? * Which script structures consistently outperform others? Where appropriate, Python/statistical analysis should perform the quantitative comparison rather than relying entirely on an LLM's interpretation. The AI can then interpret and summarize those results. ## Phase 7: Video Postmortems The system should be able to generate a report for an individual video containing information such as: ### Strongest Sections Retention regions that significantly outperform the channel baseline. ### Weakest Sections Retention regions that significantly underperform. ### Possible Script Factors What was happening in the script during those periods. ### Comparisons How similar script elements performed in other comparable videos. ### Hypotheses Possible explanations worth testing. ### Future Experiments Specific changes to try in future scripts. The report should distinguish actual measured observations from AI-generated hypotheses. ## Phase 8: Channel Writing Evidence Base The system should maintain a living set of channel-specific writing principles. For example: **Hypothesis:** Concrete personal examples perform well immediately after theoretical explanations. **Evidence:** 12 qualifying instances across 9 videos. **Average retention behavior without example:** X **Average retention behavior with example:** Y **Confidence:** Medium **Recommendation:** When explaining an abstract language-learning concept, introduce a concrete example within approximately X seconds. These rules should be updated as additional videos are published and analyzed. The objective is for this evidence base to eventually become an input into my AI YouTube script-writing workflow. ## Technology I am flexible on implementation, but my current preference is to keep Version 1 relatively simple. Possible stack: * Python * SQLite initially, or PostgreSQL if justified * YouTube Data API v3 * YouTube Analytics API * YouTube Reporting API * OpenAI API * Pandas / appropriate Python statistical libraries I do NOT initially need: * A complicated SaaS application * Multi-user authentication * A sophisticated public-facing dashboard * Mobile applications * Model fine-tuning * Large-scale cloud infrastructure This is initially an internal analysis tool for one YouTube channel. A simple local or private web dashboard could be useful, but data collection and analysis are much more important than visual polish. ## Desired Deliverables At minimum, I would like: 1. Automated YouTube video importer 2. Historical metrics importer 3. Audience-retention importer 4. Transcript/caption importer 5. Database containing the resulting information 6. Script/transcript-to-retention timestamp alignment 7. OpenAI structured script-classification pipeline 8. Cross-video quantitative analysis 9. Individual video postmortem generator 10. Channel-wide writing-rule/evidence generator 11. Clear documentation explaining how to run and maintain the system 12. Source code owned by me and provided at project completion The system should be designed so new videos can easily be added and analyzed in the future. ## Ideal Developer You would ideally have experience with several of the following: * Python * Google APIs / OAuth 2.0 * YouTube Data API * YouTube Analytics API * YouTube Reporting API * Data engineering * Pandas/data analysis * Statistical analysis * OpenAI API * Structured LLM outputs * Transcript/timestamp processing * YouTube analytics or creator analytics systems Direct previous experience with the YouTube Analytics API, particularly audience-retention reports, is a major plus. ## Important I am particularly interested in finding someone who understands the difference between: **"Have AI look at my videos and tell me how to improve them"** and **"Build a structured dataset that allows us to quantitatively investigate which script characteristics are associated with actual audience behavior."** The second is what I want. I would rather have a technically solid, relatively simple Version 1 than an impressive-looking dashboard built on weak analysis. ## When Applying Please answer these questions: 1. Have you previously worked with the YouTube Analytics API? If so, what did you build? 2. Have you specifically retrieved audience-retention data through the API? 3. How would you align timestamped transcript segments with YouTube's retention data? 4. How would you prevent the AI from simply inventing explanations for random changes in a retention curve? 5. What database/architecture would you recommend for Version 1, and why? 6. Which parts of this project do you anticipate being restricted by YouTube API availability? 7. Please describe a similar analytics, data-engineering, or AI project you have built. 8. Please include your proposed approach and estimated project cost. Please begin your application with the words **"RETENTION DATA"** so I know you read the full description.