Part 5: AI Vision Analysis
Part 5 of 11

AI Vision Analysis

Use AI to describe creative consistently and propose review questions—then validate performance hypotheses with observed data and human judgment.

⏱ ~10 min
🎓 Intermediate
🤖 Real Prompts Included
⚠ AI & Data Boundary

Vision models can miss context, invent details, and cannot infer why an ad performed from pixels alone. Upload only assets and performance data you are authorized to process; remove personal, confidential, and cross-client data unless the relevant agreement explicitly permits that use. Treat every score and recommendation as a reviewer aid, not a prediction.

Section 01

AI Vision Tools — What Can They See?

Multimodal capabilities, limits, interfaces and data terms change. Select a model through a dated evaluation on your own authorized task—not a permanent provider ranking.

💰 What This Means for Your Business

AI can create a first-pass description across many creatives, helping an analyst compare tags and flag items for review. A qualified human should verify the output, claims, brand context, accessibility, and rights before it influences a launch.

Some current models accept images or video-related inputs and can produce a first-pass description. Verify the exact input support in current official documentation, then test description accuracy, omission, hallucination, accessibility and rights handling on a representative, authorized set.

Upload Creative
AI Vision Processes
9-Element Breakdown
Human Review
ToolImage AnalysisVideo AnalysisBest For
Candidate AVerify current formats, limits, retention and regional availability in official documentation.Run the same representative cases and score against pre-defined criteria.
Candidate BInclude boundary, malformed, low-quality and adversarial examples.Record accuracy, abstention, latency, cost and reviewer effort.
Candidate CRe-test after material model, prompt or pipeline changes.Choose per task; different stages may need different tools.
💡 Choose by a Dated Task Evaluation

Check each provider's current official documentation for supported media, sampling behavior and limits. Use the same rights-cleared test set, rubric and reviewer; publish the date, version and observed tradeoffs. Do not generalize a single result into a permanent “best model” claim.

Section 02

Analyzing Images

Upload an authorized static ad or carousel frame for a fallible, structured description.

💰 What This Means for Your Business

Every authorized ad image can receive the same structured pre-launch review. This may catch visible issues, but a model score does not forecast performance and should not replace an experiment.

The Prompt Framework for Image Analysis

In an approved workspace, attach a rights-cleared asset to the candidate tool selected by your dated evaluation:

Describe this authorized ad creative as a fallible reviewer aid.
Break it down into these 9 elements:

1. HOOK: What grabs attention first? Type of hook used?
2. VISUAL: Lighting, color palette, composition, context
   (lifestyle vs product-only), human presence
3. COPY: What's the headline? Body text? What angle?
   (pain, benefit, social proof, scarcity, story, authority)
4. EMOTION: Primary emotional trigger? Secondary?
5. FORMAT: Static, carousel frame, or other? Aspect ratio?
6. CTA: What action is requested? Where is it placed?
7. CLAIMS: Which reviews, guarantees, awards, ratings, badges,
   origin or product claims appear? Mark every one unverified
   until linked to substantiation; never invent a claim.
8. AUDIENCE HYPOTHESIS: Who may understand or respond to
   this message, and what evidence would test that claim?
9. WEAKNESSES: What's missing or could be improved?

Flag uncertainty and any detail you cannot verify from the asset.
Do not predict performance or claim causal effects.
ai vision — analyzing a static ad
Input: [Premium leather bag ad image uploaded]
AI Analysis Output:
1. Hook: Observed: product in use in a café. Effectiveness unknown until tested.
2. Visual: Observed: natural light and warm tones. “Premium” interpretation is subjective.
3. Copy: Observed claim: “Handcrafted Italian Leather.” Verify materials, origin, substantiation and legal review before use.
4. Emotion: Hypothesis: some reviewers may read aspiration. Validate with representative users.
5. Format: Observed: static 1:1. Check current placement specifications before adapting.
6. CTA: Observed: “Shop Now.” Test task comprehension and destination consistency.
7. Evidence: No substantiation supplied. Never invent reviews, ratings, guarantees, awards, badges or customer language.
8. Audience: Not inferable from appearance. Use an eligibility-safe research hypothesis.
9. Unknown: Product truth, rights, offer, audience response and performance are not visible in the asset.
Reviewer status: description draft only; numeric score intentionally omitted.
Next step: verify every claim and right, then define one falsifiable change and a prospective test.
Section 03

Analyzing Videos

Frame-by-frame description, pacing observations, and testable retention hypotheses—not drop-off prediction.

💰 What This Means for Your Business

AI can describe the pacing of a short video and suggest moments to investigate. Only retention data from a comparable live test can show where viewers actually drop off, and only a controlled comparison can support a causal claim about the hook.

Video Analysis Prompt (for a currently evaluated video-capable tool)

Review this authorized ad video as a fallible structured
reviewer. Separate observations, hypotheses, and unknowns:

1. HOOK (0-3s): What happens? Is it compelling enough
   to stop the scroll? What type of hook is used?
   State what is visible and what evidence would test effectiveness.

2. PACING: How many scenes? Cut frequency? Does the
   energy build or stay flat? Where might viewers drop?

3. KEY FRAMES: Identify the 3 most important moments
   and explain what each communicates.

4. SOUND: What audio is used? Does it match the pacing?
   Does the video work WITHOUT sound (text overlays)?

5. STORY ARC: Beginning (hook) → Middle (value) →
   End (CTA). Is this structure clear?

6. CTA MOMENT: When does the call-to-action appear?
   Is it too late? Too subtle?

7. AUDIENCE HYPOTHESIS: Who may understand or respond
   to this message, and what evidence would test it?

8. RETENTION HYPOTHESIS: Which moments may create
   friction, how uncertain is that assessment, and what
   retention metric would test it?

9. TOP 3 TESTS: What specific, reversible changes would
   you test one at a time? Do not predict performance.

Fallback Frame-Sampling Method

If the currently evaluated tool does not accept the required video input, extract a documented sample of frames. This loses motion and audio context, so disclose the limitation:

Take screenshots of your video at these timestamps:
- 0 seconds (first frame — the hook)
- 3 seconds (end of hook window)
- 5 seconds (early middle)
- 10 seconds (mid-point)
- Last frame (CTA moment)

Upload all 5 frames with the prompt:
"These are frames from a 15-second ad video at 0s, 3s,
5s, 10s, and 15s. Analyze the progression..."
Section 04

Contrasting-Case Analysis

Compare adequately sampled cases using one pre-defined qualified outcome, and use the model only to surface visible differences for human verification.

💰 What This Means for Your Business

This comparison can help an analyst spot visible differences worth testing. It cannot determine which difference caused performance, especially when spend, audience, placement, dates, offer, and conversion lag differ.

Higher observed outcome

Example observation: lifestyle visual

Example observation: substantiated customer quote

Example observation: verified policy statement

VS

Lower observed outcome

Example observation: product-only visual

Example observation: feature-led copy

Example observation: no policy statement shown

The Comparison Prompt

I'm uploading 10 ad creatives. The first 5 are my
higher observed outcome cases. The last 5 are lower observed
outcome cases under the same pre-defined metric and lag.

For each group, analyze:
1. What visual patterns appear consistently?
2. What hook types are used?
3. What emotional triggers dominate?
4. What's the copy angle?
5. What trust signals are present/missing?

Then identify:
- The top visible differences and plausible confounders
- Which patterns are hypotheses rather than conclusions
- What metadata is missing (audience, dates, spend, placement,
  offer, sample size, conversion/return lag)
- A controlled next test with a pre-defined decision metric
best vs worst — pattern discovery
⚙ Reviewing two small synthetic comparison groups...
Three observations from this small synthetic sample:
1. Group A more often shows the fictional product in use; Group B more often uses a product-only composition.
2. The opening-copy categories differ; neither category is assumed superior.
3. A verified-policy statement appears unevenly. Presence is descriptive, not causal, and any real claim requires substantiation.
Candidate hypothesis—not a recipe:
The observed elements may correlate with the chosen outcome in this sample.
Confounders include audience, spend, placement, offer, time, product and measurement. Pre-register a prospective test; infer no performance probability.
Section 05

Batch Analysis at Scale

Describe and cluster an authorized creative set, then let a human review the tags and decide which hypotheses deserve testing.

💰 What This Means for Your Business

A batch pass can standardize first-review tags, provided the assets are authorized and a human verifies the output. Performance decisions still require comparable metrics, sufficient samples, and conversion lag.

Batch Analysis Prompt

I'm uploading an authorized set of creatives from one brand.
For each one, describe observable features and uncertainty:
- Opening device / hook category
- Visual style and composition
- Copy and CTA
- Verifiable trust signals

Then provide:
- Pattern clusters with the reason for each assignment
- Possible confounders and missing performance metadata
- 3 task-specific hypotheses for controlled testing
Do not rank likely performance from the assets alone.
Section 06

Competitor Creative Analysis via AI Vision

Use authorized public screenshots to summarize visible patterns and form differentiation hypotheses.

💰 What This Means for Your Business

Public ad-library screenshots can reveal visible executions and run dates, not spend, profitability, targeting, test design, or a competitor's strategy. AI may summarize observations and suggest differentiation questions.

Competitor Analysis Prompt

These are screenshots from my competitor's Meta Ad Library
showing their currently active ads.

Summarize only the visible execution:
1. What visual patterns repeat across their ads?
2. What hook types do they favor?
3. What emotional triggers dominate?
4. What formats are they using most?
5. What audience hypothesis does the message suggest,
   and why is that uncertain?
6. How long have their top ads been running?
   (longevity does not reveal spend or profitability)

Then identify:
- A tentative visible-pattern summary, not their internal strategy
- GAPS: What differentiated, brand-safe hypothesis could I test?
- DIFFERENTIATION: How to position against them visually
Section 07

Prompting Frameworks — Copy & Use

Ready-to-use prompts for every type of creative analysis.

💰 What This Means for Your Business

Use these prompts as reviewer aids. Benchmark task time, error rate, reviewer effort and decision quality in your own workflow; no fixed productivity saving is assumed.

Single Analysis Video Analysis Best vs Worst Batch Competitor Brief Gen

Quick Reference: 6 Essential Prompts

Prompt 1
Single Creative Description
A structured first-pass description with uncertainty and evidence fields. A qualified reviewer verifies it before launch.
Prompt 2
Video Analysis
Frame-by-frame description, pacing review, and retention hypotheses to validate.
Prompt 3
Best vs Worst Comparison
Compare adequately sampled candidates, surface confounders, and design the next test.
Prompt 4
Batch Review
Cluster an adequately sampled, authorized set and surface observations, uncertainty and confounders without a universal ranking.
Prompt 5
Public Creative Review
Authorized public screenshots summarized for visible patterns and materially original differentiation hypotheses—not inferred strategy.
Prompt 6
Original Brief Draft
Use supported observations to draft a materially original test brief. Review trademark, copyright, publicity, license and brand rights; do not recreate a competitor or apparent winner.
🌟 The AI Vision Advantage

Structured AI-assisted review can make descriptions more consistent, but it is not objective truth. Pair verified tags with adequately sampled performance data, lag-aware business outcomes, and controlled tests before saying a pattern outperforms.

AI Can See. Now Make It Count.

In Part 6, we'll join human-verified creative tags with lag-aware performance data to surface hypotheses for controlled testing.

Built by @itsmazinzaki — AVAMARTECH