Analytics cannot tell you why a viewer behaved a certain way. It can show where attention changed, how people reached a video, and what some viewers said. Your job is to connect those signals to the video’s intended job without turning correlation into a confident story.

By the end, you will have a one-page review of five comparable videos. Each row will contain an observation, a cautious interpretation, and one testable change for the next script—not a promise that the change will improve performance.

// concept

Begin With the Video’s Intended Job

Write the intended behaviour above the metrics: “Help a beginner understand one CapCut caption fix and save the video for editing day.” A narrow tutorial should not be judged as if its only purpose were broad reach.

Compare videos from the same platform, format, and publishing age where possible. Record duration, topic, traffic source, device mix, language, and promotion. A 45-second Roman-Urdu Short receiving WhatsApp traffic is not comparable to a six-minute English tutorial discovered in Search.

In YouTube Studio, open Content → select a video → Analytics. Use Engagement for retention and Reach or channel-level Content for impressions and CTR. Confirm the current path in YouTube’s official guide if the interface moves.

// concept

Map Each Signal to a Craft Decision

Use the metric closest to the craft choice you can change.

Signal	What you can observe	Craft choice it can inform	What it cannot prove
First-seconds retention	How many starters remain after the opening	Hook wording, speed to payoff, title-to-opening match	That one phrase caused exits
Retention-curve shape	Cliffs, gradual decline, dips, flat sections, or spikes	Section order, repetition, explanation length, pattern breaks	Why each individual left or replayed
Impressions CTR	Share of counted thumbnail impressions that became views	Title/thumbnail promise and audience fit	Total click rate from every source, or video quality
Impressions, reach, traffic sources	Where and how often the platform surfaced the video	Topic choice and distribution context	That the platform “likes” or “hates” a topic
Comments and questions	Exact language used by a self-selected subset of viewers	Missing explanation, confusing term, desired follow-up	The opinion of the silent majority
Device, geography, language	Viewing context reported for an eligible audience	Caption size, pacing, examples, language version	A person’s motivation or identity

For a manual CTR check, keep numerator and denominator in the same reporting scope:

// prompt — copy me2 lines

sample CTR = views from counted impressions ÷ counted impressions × 100
sample calculation = 72 ÷ 1,200 × 100 = 6%

Do not divide total views by impressions: some views do not come from counted impressions. Prefer the platform-reported CTR.

// concept

Read Retention as a Shape, Not a Verdict

A cliff near three seconds shows an opening loss, not its cause. Replay it: Did the first line restate the title? Did a logo delay the result? Did captions begin late? Test one fix, such as opening on the finished result.

A mid-video sag points to a section, not automatically a bad topic. Mark its start and end, then inspect the definition, B-roll, unfamiliar term, or slow setup. A script fix might be: “Move the completed example before the definition and cut two repeated sentences.”

Spikes may reflect rewatching or sharing, but also confusion. Watch the segment and read nearby comments before repeating its style. Five videos can surface a testable pattern; they cannot prove a universal rule or cause.

// concept

Turn One Insight Into One Script Change

Use four labels in every review row:

Intent: What was this video supposed to help the viewer do?
Observation: What does the report literally show?
Interpretation: What are two plausible explanations, written as “may” or “could”?
Next-script change: What single line, section, or cue will change?

Choose one change per video, then one priority for the next upload. Rewriting the hook, middle, topic, and thumbnail together makes the result uninterpretable. Keep packaging and script tests separate where practical.

AI may organize exported data and supplied comments, but must not invent a diagnosis. Remove viewer names and identifiers.

// worked_example

Worked Example

The following is sample data for a hypothetical Pakistani faceless channel, not real performance. These similar-length explainers share a seven-day review window.

Video	Intended job	CTR	Retained at 3s	Curve/context note	Sample comment theme
A	Show one caption fix	6.0%	42%	Immediate 3s cliff; WhatsApp traffic present	“Show the result first”
B	Explain voice pacing	5.8%	68%	Sag at 24s during definition	“Can you give an example?”
C	Compare two hook styles	6.1%	66%	Flatter curve; mostly Search	“Which one suits tutorials?”
D	Fix Roman-Urdu TTS	5.9%	64%	Spike during pronunciation list	“How do I spell this word?”
E	Export for mobile	6.0%	65%	Dip during repeated menu steps	“Does this work on Android?”

Here is the actual prompt used to structure the review:

// prompt — copy me14 lines

You are an analysis assistant. Use only the labelled sample data below.
For each video, return: intent, literal observation, two possible explanations,
one check I can perform in the video, and exactly one next-script change.

Rules:
- Never say a metric proves a cause.
- Do not compare CTR without noting traffic-source context.
- Quote comments exactly; do not treat them as representative.
- If evidence is missing, write “unknown.”

Priority question: Which ONE script change should be tested in the next upload?
<sample_data>
[paste the five rows above plus anonymized comment text]
</sample_data>

An incorrect conclusion was: “Video A’s weak hook caused its low CTR, so change the thumbnail and opening.” CTR describes registered impressions becoming views; the hook comes after the start. The claim also ignores traffic mix and changes two variables.

The corrected review says: “Video A shows the largest first-three-second loss. Its opening may delay the result, or WhatsApp starters may expect something different. Replay the opening against the shared caption. In the next script, show the finished caption effect in frame one and remove the logo intro.” CTR remains outside this script test. B moves its example before the definition; C names the tutorial use case in the hook; D spells tested Roman-Urdu words on screen; E replaces repeated narration with one step card. The next upload prioritizes A’s frame-one test.

// failure_cases

Failure Cases to Diagnose

6 cases to diagnose

Calling a dip a cause
“People left because the explanation was boring” exceeds the evidence. Name the timestamp, inspect the content, and list at least two possible explanations.
Using total views to calculate CTR
External or other views may not come from counted impressions. Use the platform’s reported CTR and inspect traffic sources.
Comparing unlike videos
Different durations, formats, audiences, promotions, or publishing ages distort the comparison. Build a five-video set with the closest available context.
Treating one comment as audience consensus
Preserve the exact comment as qualitative evidence, then look for repeated questions without inventing a percentage.
Reading a spike only as approval
Rewatching can indicate value or confusion. Review the segment and nearby comments before reusing the pattern.
Changing everything at once
Multiple script and packaging changes destroy the learning value of the next cycle. Choose one scripted change and log the rest as later hypotheses.

// pakistan_angle

Pakistan Angle

For a Pakistani audience, inspect device, language, geography, and traffic source. A large mobile share favours readable captions and immediate demonstrations over tiny interface text. WhatsApp arrivals may behave differently from YouTube Search or Home; record that context instead of calling viewers “low quality.”

Roman-Urdu comments can use several spellings for one word. Cluster meaning manually before giving anonymized comments to AI; never upload usernames, phone numbers, or WhatsApp screenshots. During load-shedding or costly mobile-data periods, export the table once on a stable connection, save low-resolution curve screenshots, and finish the review offline.

// hands_on

Hands-On Exercise

6 steps

Build a one-page analytics review for five of your own comparable videos. If you have not published five, use your available videos and mark the missing rows “not yet available”; do not fabricate data.

Write one intended behaviour for each video.
Record the same window, format, duration, traffic source, CTR where available, first-seconds retention, and one curve timestamp.
Add up to three anonymized comment quotes per video; “no comments” is valid evidence.
For every row, write a literal observation and two possible explanations.
Replay the relevant timestamp and write one next-script change for that video.
Circle one priority for the next upload; hold other variables steady where practical. Done means all five rows separate observation from interpretation, each has one script change, and the next upload has one priority test.

// completion_rubric

Completion Rubric

6 checks — tick as you verify

0/6

// sources

Sources

4 official sources — check every claim yourself

// check_yourself

Check yourself

4 questions · answers and options are taken word-for-word from this course

0/4

1 / 4 · diagnose
Your work shows this failure mode: “Calling a dip a cause.” The lesson describes it like this: ““People left because the explanation was boring” exceeds the evidence.” What does the lesson tell you to do about it?