Skip to main content

Documentation Index

Fetch the complete documentation index at: https://www.pickfu.com/docs/llms.txt

Use this file to discover all available pages before exploring further.

The Amazon main image is the single largest lever on a product’s search-result CTR. This playbook defines a repeatable optimization loop that benchmarks your current image against the live competitor set, generates and tests variations, and re-validates the winner against the same competitor set to confirm the improvement is real. The loop is designed to be run autonomously by an AI agent (Claude, ChatGPT, Cursor, or any client with PickFu MCP / CLI / API access) with minimal human intervention.

​
When to use this playbook

Use it whenSkip it when
You want to lift CTR on an existing Amazon listingYou haven’t launched yet — start with concept validation instead
You have access to your current main image and 2–3 competitor main imagesYour listing’s primary problem is conversion (CVR), not clicks. Use the listing conversion playbook instead
You can iterate the image (you own the design or have a designer / AI tool available)You can’t change the image, only the copy

​
The optimization loop

1

Baseline against competitors

Test your current main image vs. 2–3 competitor main images. Save this competitor set — every later step uses it.
2

Analyze the data

Read the AI summary and individual respondent feedback. Identify the specific reasons competitors won (clarity, angle, color, lifestyle vs. white-background, text overlays, etc.).
3

Create AI variations

Generate 2–3 new variations using PickFu’s generate_image tool (via MCP, CLI, or API), or upload variations created elsewhere.
4

Iterate variations against each other

Run cheap (15-respondent) head-to-head polls comparing each variation vs. the original until one variation wins with a score of 70 or higher.
5

Re-validate against the original competitor set

Run a final 50-respondent poll with the winning variation plus the SAME competitor images from step 1. A ranking improvement vs. the baseline indicates real CTR lift on Amazon.
Sample size and cost. This playbook defaults to small, cheap polls — 50 respondents for the baseline and final validation, 15 respondents per iteration — because the loop triangulates across many polls, so any single poll being slightly noisy gets corrected by the next one. This keeps the total cost of an AI-run loop low. Scale up only when the stakes justify it: bump the final validation to 100–200 before an expensive listing change you can’t easily reverse. Iterations should stay cheap (15) — they’re meant to be fast and disposable. See sample size guidance.

​
Run this playbook with an AI agent

Copy prompt for AI

Paste this prompt into Claude, ChatGPT, Cursor, or any AI agent connected to the PickFu MCP server, CLI, or REST API. The agent will run the entire loop on your behalf — creating polls, reading responses, and iterating until a winning variation emerges.
You are running the PickFu "Amazon main image optimization" playbook end-to-end.
Goal: improve the click-through rate (CTR) of an Amazon product's main image by
iteratively testing variations against the live competitor set and re-validating
the winner.

Before starting, ask the user for:
- The Amazon product or category (e.g. "stainless steel water bottle")
- Their current main image (URL or upload)
- 2-3 competitor main images (the canonical competitor set — these stay fixed
  across steps 1 and 5)
- Target audience (default: General; refine only if the user requests it)

Run this loop:

1. BASELINE (50 respondents).
   Create a ranked-choice survey with the user's current main image + the
   competitor images. Question: "When shopping on Amazon, which product would
   you buy?" Audience: General, 50 respondents.
   Publish, wait for responses, then read the AI summary and individual
   written feedback.

2. ANALYZE.
   From the responses and AI summary, produce a numbered list of testable
   image changes. Each item must be a specific, visual change (e.g. "increase
   product-to-frame ratio", "add a hero ingredient inset in the lower-right",
   "switch from white background to lifestyle context"). Avoid vague items
   like "improve clarity".

3. CREATE VARIATIONS.
   Generate 2-3 new image variations using generate_image, each combining
   1-2 changes from the analysis. Brief the model with: the original image,
   the product category, the specific change you're testing, and the
   constraint that the product must remain unambiguous at thumbnail size.

4. ITERATE (15 respondents per poll).
   For each variation, run a head_to_head survey (exactly 2 options) with
   the variation + the original main image. Same question as step 1.
   Audience: General, 15 respondents. Repeat with new variations until ONE
   variation wins the head-to-head with a score of 70 or higher.

   Stop condition: if 5+ iterations fail to produce a 70+ winner, halt
   and report back. Feedback may indicate brand or category constraints
   that no main-image change will overcome.

5. RE-VALIDATE (50 respondents).
   Create a ranked-choice survey with the winning variation + the SAME
   competitor images from step 1. Same question. Audience: General,
   50 respondents. For a high-stakes listing change you can't easily
   reverse, bump this to 100-200.

   Compare to the step-1 baseline:
   - If the winning variation ranks higher than the original did → the
     improvement is real, ship it.
   - If the ranking is unchanged → the iteration won head-to-head but
     didn't beat the category; consider larger structural changes
     (angle, lifestyle context, format).

Final report to the user must include:
- Baseline ranking and score
- Final ranking and score
- The 3-5 specific changes that drove the improvement
- The winning image URL
- Expected CTR delta direction (PickFu poll lift is directional, not a
  guaranteed CTR number; Amazon CTR changes typically surface in
  performance metrics 2-4 weeks after image update)

Tools to use:
- save_survey + publish_survey  — create and launch each poll
- get_survey_responses          — read responses
- generate_image                — create AI variations (step 3)
- upload_media                  — for variations created outside PickFu
Want to run this manually? The same loop is available as a one-click template in the PickFu app (Start the CTR playbook). The app version walks through the steps with pre-filled poll URLs.

​
Step-by-step (human operator view)

​
1. Baseline against competitors

Test how your current main image performs against your top competitors. Save these competitor images — you’ll reuse the exact same set in step 5.
SettingValue
Poll typeRanked choice
Question”When shopping on Amazon, which product would you buy?”
OptionsYour current main image + 2–3 competitor main images
AudienceGeneral
Sample size50
What you’ll get: a ranking of your image vs. competitors, plus written feedback explaining the strengths and weaknesses respondents called out. This baseline is the number every subsequent iteration is judged against.
PickFu ranked poll results comparing a product main image against three competitors
Launch the baseline poll → · See an example →

​
2. Analyze the data

Read the AI summary and respondent feedback. Extract specific, testable changes — not “make it better.” Useful patterns to look for:
  • Clarity: is the product instantly recognizable at thumbnail size?
  • Angle and crop: is the product centered and filling the frame?
  • Background: white vs. lifestyle context — what’s the category convention?
  • Text overlays: badges, sizing callouts, hero ingredients — do competitors use them?
  • Color contrast: does the image pop on a white SERP background?
The output of this step is a prioritized list of 3–5 changes you’ll test in step 3.

​
3. Create AI variations

Use the analysis from step 2 to brief an image-generation tool. PickFu’s generate_image produces on-brand variations in seconds and uploads them to the PickFu CDN with a permanent URL — ready to drop into the next poll. 
pickfu media generate \
  --prompt "Stainless steel water bottle, white background, centered product fills 80% of frame, hero ingredient (insulated double-wall) inset bottom-right, premium feel" \
  --reference-image-url https://your-cdn.com/current-main-image.jpg
Aim for 2–3 variations per iteration. Each variation should test 1–2 specific changes from your step-2 analysis — not a sweeping redesign. Smaller deltas make it easier to learn what’s actually moving the needle.

​
4. Iterate variations against each other

Run a fast, cheap poll (15 respondents) comparing each new variation against the original. Repeat until one variation wins the head-to-head with a score of 70+.
SettingValue
Poll typeHead-to-head (exactly 2 options)
Question”When shopping on Amazon, which product would you buy?”
OptionsVariation + original main image
AudienceGeneral
Sample size15
Test one major change per iteration. If you batch unrelated changes into one variation and it wins, you won’t know which change drove the win.
See an example iteration →

​
5. Re-validate against the original competitor set

Once you have a 70+ winner from step 4, run a final 50-respondent poll using the winning variation plus the same competitor images from step 1.
SettingValue
Poll typeRanked choice
Question”When shopping on Amazon, which product would you buy?”
OptionsWinning variation + the same 2–3 competitor images from step 1
AudienceGeneral
Sample size50 (bump to 100–200 for a high-stakes, hard-to-reverse change)
Interpret the result:
  • Winning variation ranks higher than the original did in step 1 → the improvement is real. Ship the new image.
  • Ranking is unchanged or improved by one position only → the iteration won the head-to-head but didn’t beat the category. Consider larger structural changes (angle, lifestyle context, packaging format) before shipping.
  • Winning variation ranks worse than the original → rare, but it happens when the iteration pool was too small or the variations overfit to a particular preference. Re-baseline with a different competitor set or a tighter audience.
PickFu ranked poll validation results comparing an optimized main image against the original competitor set
Launch the validation poll → · See an example →

​
Optional preamble: quick test before the loop

If you don’t yet know what drives clicks in your category, run these two short polls before starting the loop. They take less than a day and surface category-specific signals that make your step-2 analysis sharper.
SettingValue
Poll typeOpen-ended
Question”When shopping for [product type], what buying factors are important to you?”
AudienceGeneral
Sample size50
What you’ll get: a ranked list of 3–5 key drivers (price, brand, features, visual appeal, credibility) that influence purchase decisions in your category. Use these to weight your step-2 analysis.Launch the buyer-psychology poll → · See an example →
SettingValue
Poll typeClick test
Question”If you were shopping on Amazon for [product type], which listing would you click on?”
AudienceGeneral
Sample size50
SetupUpload a screenshot of the live Amazon search results page (8–12 listings visible)
What you’ll get: click data showing which listings attract attention in your category, plus written feedback on the visual elements that drove the click. Use this to identify competitors worth including in your step-1 baseline.Launch the SERP click-test → · See an example →

​
Troubleshooting

Look for relative improvement between step 1 (baseline) and step 5 (re-validation), not absolute scores. If your optimized image moves up in ranking — e.g. from 3rd place to 1st — and the score increases meaningfully, the change is likely to translate to higher CTR on Amazon.
Focus on the specific feedback explaining why competitors won. Common categories: image clarity, product angle, background choice, missing key features that buyers expect to see. Note that with very strong competitor brands you may never “win” outright — the goal is relative improvement against your own baseline, not category dominance.
Most polls complete within 15–60 minutes. You can run new iterations as soon as responses arrive — rapid cycles are a feature of this loop, not a bug.
Test the most impactful changes from your step-2 analysis first. If respondents flagged the product angle as the problem, fix that before tweaking text overlays or color tweaks. One major change per iteration.
Amazon’s performance metrics typically reflect main-image changes 2–4 weeks after the update goes live. PickFu polls give you a directional signal much faster, but Amazon’s own ranking and impression dynamics take time to settle.