Review TikTok and Instagram Content with AI: From Data to Your Next Shoot

Standardise TikTok and Instagram metrics and use calculation examples and prompts to identify practical changes for your next videos.

A video that does not take off does not necessarily have the wrong topic, and high views do not guarantee relevant enquiries. I separate content performance from business outcomes and finish with a few changes the next shoot can actually implement.

This tutorial uses teaching data for a drawer organiser account. You will build a consistently defined content table, calculate interpretable metrics and ask AI for the next filming plan. All numbers are calculation examples, not results from a real account.

01 · Define what this round of content should achieve

If the aim is explaining size selection, relevant questions, saves and later specification enquiries may matter most. If it is improving attention, examine early drop-off and viewing behaviour. Do not judge every video by the same metric.

Write one objective: “Use measuring demonstrations to attract people with drawer-fit questions and direct them to specifications.” Separate metrics into four stages: seen, watched further, action taken and business conversation.

Platforms define views, reach, watch time and profile visits differently. Start by comparing the same platform, similar content types and equal observation windows rather than mixing metrics.

02 · Find your own account data and leave unavailable fields blank

For TikTok, use TikTok Studio or your business account's analytics. On Instagram, use available Insights or the professional dashboard. Do not claim access to another account's private watch-time or conversion data.

Choose a fixed date range, examine account trends, then individual posts. Preserve exports when available; otherwise record manually with relevant screenshots. Leave unavailable fields blank rather than entering zero.

Instagram insights depend on conditions such as professional-account eligibility. TikTok menus may vary by account or region. Check help and eligibility before concluding that a missing metric is lost data.

YouTube also distinguishes metrics such as views and engaged views. Do not combine them with TikTok views in one ranking. Analyse each platform separately before comparing business outcomes.

03 · Record what the video contains alongside its numbers

Without content structure, AI can only suggest generic improvements. Record the topic, first-three-second visual, duration, main shots, closing action, paid promotion status and observation date for every video.

Content ID | Platform | Original URL | Publication time | Observation time | Observation window
Topic | Opening type | Duration in seconds | Closing action | Paid promotion
Views | Reach | Average watch seconds | Completion rate if available
Likes | Comments | Saves | Shares | Profile visits, labelled account-level or post-level
Landing-page visits | Attributable enquiries | Sales-qualified enquiries | Notes

Do not copy account-level profile visits into every post when post-level data is unavailable. Mark enquiries with unknown sources as unattributed rather than assigning them to the most-viewed video.

Use a fixed window, such as 72 hours after publication. Comparing a three-hour-old post with a seven-day-old one confuses elapsed time with content performance. This is an internal comparison window, not a platform distribution rule.

04 · Clean the data before calculating denominators

Standardise time zones, dates and number formats. Mark rounded figures such as 1.2K and prefer exact backend data. Use one percentage convention rather than mixing 0.12 with 12.

Deduplicate content IDs and check observation windows. Separate paid and organic posts and flag giveaways, trends or special collaborations. Keep missing values blank rather than averaging them as zero.

Define common calculations as follows:

Average watch ratio = average watch seconds ÷ video duration
Saves per 1,000 views = saves ÷ views × 1,000
Attributable enquiry rate = attributable enquiries ÷ landing-page visits with the same scope
Qualified enquiry rate = sales-qualified enquiries ÷ all enquiries

Average watch ratio is not completion rate and may exceed 100% with loops. Do not divide one video's enquiries by all-channel landing-page visits and call it that video's conversion rate. Leave the result blank if the denominator is zero or missing.

05 · Use three teaching records to decide what to change

Assume all three videos were posted organically on TikTok and measured after 72 hours:

A: self-introduction opening; 30 seconds; 1,000 views; 6 seconds average watch; 5 saves.
B: drawer measurement opening; 25 seconds; 800 views; 12 seconds average watch; 16 saves.
C: before-and-after organisation; 20 seconds; 3,000 views; 8 seconds average watch; 12 saves.

A has a 20% average watch ratio and 5 saves per 1,000 views; B has 48% and 20; C has 40% and 4. This small sample suggests observing B further, not that measurement openings are definitely better: topics and durations also differ.

Next, compare introduction-first and action-first openings with similar topics, duration and filming conditions. Keep records and use equal observation windows. A single result does not establish causation.

If C attracts many irrelevant comments while B produces a few genuine fit questions, B may deserve more attention for lead generation. Preserve the actual questions, not just comment totals.

06 · Ask AI for evidence, not slogans

Analyse this content table. The goal is to help people with drawer-fit questions view specifications and enquire.
First check fields, units, observation windows and missing data. Flag inconsistent definitions.
Compare only records from the same platform and type with comparable windows. Separate promoted content.
Show numerators and denominators. Do not treat missing values as zero or call average watch ratio completion rate.
Return:
1. Three data-supported observations with content IDs.
2. Three questions that remain hypotheses, not causal findings.
3. Three filmable next videos: user question, opening, main shots, one change and measurement.
4. Additional data needed.
Do not use universal viral thresholds, invent conversions or count enquiries as orders.
Data: [paste anonymised table]

If AI says B converts best, check whether conversion data exists. With saves alone, discuss saves only. Replace vague advice such as increase engagement with a specific question, such as whether viewers need width or height measurement guidance.

07 · Turn the review into a filming list

The next round could retain B's action-first opening and clarify height measurement, compare two real sizes, then answer an actual fit question. Assign shots and factual sources to each.

Record one principal change per video. Real publishing is rarely a strict experiment, but changing topic, language, shots, duration and audience together makes it impossible to attribute results to a title.

Content ID: next-round-01
Keep: tape-measure opening, same product, same language
Change: add a height-measurement close-up and remove repeated speech
User question: why will the drawer not close even when width and depth fit?
Evidence: real measurements and opening/closing footage of our drawer
Observe: relevant questions, saves and retention curve if available
Window: 72 hours after publication, with a separate one-week observation later

08 · Diagnose different performance patterns differently

Low views: check visibility, review status and recommendation notices, then opening clarity. Do not declare suppression without a platform notice. Keep observing small samples rather than deleting every low-view post.

Views without action: check audience relevance, clarity of the next step and consistency between profile and content. Merely repeating “follow me” more loudly is not a solution.

Clicks without enquiries: test the landing page and form before reviewing specifications and field burden. A post-click technical failure is not evidence that the video is poor.

Enquiries without progress: classify fit problems, shipping conditions, purchasing stage and unclear replies. Specific bottlenecks tell the next video which question to answer.

09 · Keep five conclusions from each weekly review

Record the week's objective, most relevant content, supported findings, remaining unknowns and three changes for next week. Keep raw data and versioned conclusions, then check whether the changes actually happened.

A polished report is not the end. A useful review tells the filming team what to shoot, sales what to answer and everyone where to find comparable data next time.

10 · Build a checkable table rather than a viral ranking

Continue the illustrative A/B/C records, all assumed organic posts on one platform observed after 72 hours. A: duration 30 seconds, 1,000 views, average watch 6 seconds, 5 saves. B: 25 seconds, 800 views, average watch 12 seconds, 16 saves. C: 20 seconds, 3,000 views, average watch 8 seconds, 12 saves. Retain raw fields and add calculated columns.

A has a 20% average-watch proportion and 5 saves per thousand views; B has 48% and 20; C has 40% and 4. B is higher on these measures in this dataset, but average-watch proportion is not completion rate, and different topics prevent a causal conclusion about openings.

For aggregate saves per thousand views, sum first: 33 / 4,800 × 1,000, approximately 6.88. Averaging 5, 20 and 4 gives approximately 9.67, an equal-weight average across posts. That answers a different question and is not the aggregate rate. Label the method.

Do not replace a zero denominator with one or missing saves with zero. Leave the calculated cell empty and explain why. Without attributable inquiry records, do not claim which video converts best.

Analyze this teaching CSV without accessing or changing an account or claiming real performance.
content_id,platform,window_hours,duration_seconds,views,average_watch_seconds,saves
A,TikTok,72,30,1000,6,5
B,TikTok,72,25,800,12,16
C,TikTok,72,20,3000,8,12
Assume organic content but different topics: A introduction, B measurement, C before/after arrangement. No click, inquiry or order data is available.
Retain raw columns and add average-watch proportion and saves per thousand views with formulas and units. Compute the aggregate from total saves and total views and distinguish it from an equal-weight mean of post rates. Give three observations and three unsupported conclusions; do not rename average-watch proportion completion rate or discuss conversion. Propose one measurement video, specifying what to retain, the single change, missing footage and an actually available metric. Explain missing values and zero denominators without defaulting to zero or replacing the denominator.
Output analysis.csv and next-content.md, or complete copyable contents if file creation is unavailable. Label all numbers illustrative and never claim to have read a real account.

11 · Turn an observation into the next filming decision

“Make more videos like B” is too vague. Write a production card: explain usable drawer height; retain product, language and measurement context; make the height action clearer; capture a close-up from the placement surface to any potential obstruction and actual drawer movement. Do not stage a fake failure where no obstruction exists.

Editorial narration: “Check the usable height, not just the width. Measure from the surface where the organiser will sit to anything it could catch on. Compare this with the product dimensions, then check the drawer movement.” Match the real drawer structure rather than presenting universal fit assurance.

The handoff specifies three actions: clear tape-measure starting point, clear possible contact point and continuous drawer movement after placement. Use short captions such as “Usable height” and “Check the movement”. Time actual narration before assigning shot durations.

Observe the next post using the same window. If language, product, duration or publishing method changes too, record those differences instead of attributing the entire result to the height shot. Keep conclusions within the evidence even when production tests cannot fully control variables.

Add an owner and evidence-for-review field to each recommendation. Assign the height close-up to the filming owner and look for continued measurement confusion in actual comments and inquiries, not views alone. Use real assignments and dates, leaving unagreed timing unassigned.

If a recommendation was not implemented, later results cannot prove it failed. If the observation window is incomplete, mark it in progress. Correct processing mistakes in the working table with an explanation while preserving the original export.

Keep three deliverables: traceable raw data, an analysis table with formulas and limits, and a filming-ready production card. That connects AI-assisted review to the next production cycle.

Official resources

TikTok: TikTok Studio and analytics

Instagram: account insights

YouTube: content performance reports