AI Social Media Growth
0/15 complete

Module 04 · Platform-Specific Growth Tactics

TikTok and Instagram Reels: The Algorithm in Plain English

TikTok and Instagram do not have one secret switch called “the algorithm.” Their changing recommendation systems predict which eligible video a viewer may value. Your job is to make a clear promise, deliver it, and learn from native analytics—not reverse-engineer a permanent recipe.

By the end, you will have a completed three-post experiment sheet. It will separate what you changed from what the platform decided, record comparable per-post data, and turn that evidence into one modest next hypothesis—not a reach guarantee.

// concept

What Distribution Signals Really Mean

TikTok groups recommendation inputs into user interactions, content information, and user information. Its help page says watching in full, skipping, liking, sharing, and commenting can influence the For You feed. It says watch time is generally weighted more heavily for most users, while weights can change.

Instagram exposes views, watch time, average watch time, reach, likes, comments, saves, shares, and follows in Reel Insights. Those measurements show what happened to your Reel. They do not reveal a fixed ranking formula or prove that one metric caused more distribution.

Creators often say a video is “tested with a small audience” and then distributed in waves. Use that only as a mental model. Neither platform promises a test-group size, sequence, or second wave. An eligible post receives viewing opportunities; behavior supplies signals; personalized predictions continue.

Creator-controlled leverSignal you can inspectWhat you cannot control
First line and frameEarly retention or average watchInitial viewers
Pace, captions, and lengthWatch time, completion, replaysSignal weights
Payoff that fulfils the hookRetention shape, saves, sharesCompeting posts and viewer mood
Topic, on-screen text, caption, soundSearch/traffic source where shownFeature and policy changes
Relevant call to actionFollows, saves, commentsGuaranteed reach or sales

Early interaction is evidence, but “first-hour comments unlock reach” is not a safe rule. Do not buy engagement, use comment bait, or run follow/unfollow schemes.

// concept

Read Retention Before Reach

Check the current official help page because labels and availability vary by account, app version, and region.

  • TikTok: tap More insights on a post, or use Profile → Menu → TikTok Studio → Analytics → View all. Record duration, views, average watch or retention data if shown, watched-full percentage if shown, shares, and traffic source.
  • Instagram: Insights require a creator or business account. Use Profile → Reels → select Reel → View insights. Record views, reach, average watch, follows, saves, and shares. Meta says some metrics are estimated or developing.

Read a retention graph from left to right:

  1. A sharp opening drop suggests the first frame or line missed expectations; it does not prove the topic is bad.
  2. A drop at one timestamp points to slow setup, confusion, or a late payoff. Rewatch it.
  3. A flatter section suggests those viewers kept watching. Identify what was demonstrated there.
  4. A late spike may mean replaying, scrubbing, or confusion. Inspect the video before interpreting it.

For a rough Instagram retention proxy, divide average watch time by duration: 9 seconds ÷ 18 seconds = 0.50. It is not an official completion rate; views may include replays. Compare the same account, platform, topic, duration band, and observation window.

// concept

Design a Three-Post Test

A useful experiment changes one creative variable while holding the rest reasonably stable. Three posts cannot prove an algorithm law; they can tell you whether a change deserves another test on your account.

Use these sheet columns:

PostPlatform/dateConstant coreVariableHypothesisDurationAvg watchCompletion/proxySavesSharesReach/viewsContextNext decision

Choose one variable such as the hook. Keep the topic, footage, payoff, length band, account, and conditions similar. Record all three after the same observation window. Posting time affects availability, but there is no universal “magic hour.” Log changes in Context.

This prompt can help interpret the sheet without pretending to know the ranking code:

// prompt — copy me18 lines
You are reviewing a three-post short-video experiment.
Use only the sample table between <data> tags.

Task:
1. Calculate average-watch-time / duration for each post.
2. Describe differences, not causes.
3. Identify confounders in the Context column.
4. Recommend exactly one next test that changes one creator-controlled lever.

Rules:
- Do not claim a platform algorithm rule.
- Do not promise reach or treat three posts as statistically conclusive.
- Call missing fields "unknown"; do not invent them.
- Separate observation, interpretation, and next hypothesis.

<data>
[paste your three rows, without private customer data]
</data>

// worked_example

Worked Example

This is a labelled hypothetical sample, not a real result. A Lahore home baker makes three 20-second Reels about custom cake quotes. Footage, body, payoff, caption purpose, and observation window stay fixed; only the opening changes.

Hypothesis: A specific customer question will retain more viewers than a general announcement because it immediately identifies the problem being answered.

PostSample opening testedDurationAvg watchRetention proxySavesSharesAccounts reachedContext
A“Custom cake pricing explained”20s6s30%31240Sample data; normal weekday
B“Why does the same-size cake have two prices?”20s10s50%53315Sample data; normal weekday
C“The decoration, not just size, changes this quote”20s9s45%72290Sample data; uploaded after a power cut

After receiving the table in the prompt above, the AI produced this sample output excerpt:

Observation: B has the highest average-watch proxy; C has the most saves. Interpretation: the question opening may deserve another retention test, while C's saves may reflect useful wording. Confounder: C's upload conditions differed. Next hypothesis: repeat A-versus-B style openings on three new pricing topics, keeping the 20-second format.

Draft one wrongly said, “Instagram rewards question hooks.” That made a platform-wide causal claim. The fix: state the observed metric, sample, and confounder, then repeat across new topics. The decision is keep testing B's question structure, not B cracked the algorithm.

// failure_cases

Failure Cases to Diagnose

6 cases to diagnose

  • Everything changes at once

    changing hook, topic, duration, audio, and time hides the useful variable. Recut the same core and change one lever.

  • Views are treated as retention

    a view definition is not the same as meaningful watching. Pair reach or views with average watch time, completion where available, and the actual retention curve.

  • Raw watch time is compared across unequal lengths

    10 seconds on a 12-second clip and 10 seconds on a 40-second clip mean different things. Record duration and use a labelled ratio alongside raw seconds.

  • A drop is blamed on a shadowban

    check eligibility notices, account status, policy warnings, removed audio, and privacy. Appeal an actual restriction; folklore is not a diagnosis.

  • One high post becomes a law

    a single outlier may reflect topic demand, returning viewers, news timing, or chance. Turn it into a new hypothesis and repeat it.

  • Posting time is called magic

    timing changes audience availability, not a guaranteed distribution switch. Test time slots against your own follower activity and log load-shedding, holidays, and live events.

// pakistan_angle

Pakistan Angle

English, Urdu, and Roman Urdu can produce different retention patterns. Test one language variable at a time. Keep Urdu readable on low-cost Android screens and review auto-captions: Pakistani names, code-switching, “qarz,” and “bachat” may be transcribed incorrectly. Never upload customer WhatsApp numbers, addresses, order screenshots, or CNIC details into AI.

Posting conditions vary with load-shedding, mobile data, Ramzan, and Maghrib/Isha. Log them; do not declare a Pakistan-wide best time. Export a compressed vertical copy on Wi-Fi, keep the original offline, and check whether native scheduling is currently available. Karachi and Peshawar audiences can behave differently even in Roman Urdu.

// hands_on

Hands-On Exercise

6 steps

Build and complete one three-post experiment sheet.

  1. Choose TikTok or Instagram Reels and one existing topic that can support three honest variations.

  2. Write one falsifiable hypothesis: Changing [one lever] may improve [one native metric] for [this account], measured after [same window].

  3. Produce three posts with the same core and payoff. Change only the hook, pacing segment, or CTA you named.

  4. Publish under normal conditions. Record any differences—outage, collaboration, paid boost, trend, or holiday—in Context.

  5. After the same observation window, copy native metrics into the sheet. Use unknown for unavailable fields.

  6. Run the analysis prompt, verify its arithmetic, correct causal language, and write one next-test decision.

// completion_rubric

Completion Rubric

5 checks — tick as you verify

0/5

// sources

Sources

// check_yourself

Check yourself

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

0/4
  1. 1 / 4 · diagnose

    Your work shows this failure mode: “Raw watch time is compared across unequal lengths.” The lesson describes it like this: “10 seconds on a 12-second clip and 10 seconds on a 40-second clip mean different things.” What does the lesson tell you to do about it?