Performance learning

Performance Learning — Pinterest Compound Intelligence

Purpose

Turn pin results into reusable intelligence about what image-copy-board-timing combinations work for this specific brand. Pinterest's long content lifespan (3-4 months average) means data accumulates slowly but compounds powerfully. By pin 50, the system knows which specific image types paired with which title patterns on which boards drive the highest save rates — a level of cross-variable analysis that no human tracks manually and no scheduling tool provides.

Why This Is the Product

The composed pin is the deliverable. The learning loop is the product. Any tool can generate pin copy. Any designer can make a vertical image. The value is in knowing — with increasing confidence over time — that "vertical lifestyle images paired with problem-solution titles on the Content Operations board = 4.2% save rate, 3.8x the rate of product shots with the same title pattern on the same board." That kind of compound intelligence is the defensible advantage. It makes every subsequent batch meaningfully better than the last.

Why Pinterest Specifically

Pinterest is the platform where the learning loop compounds fastest because content lives longest:

• Instagram post: peaks in 48 hours. A learning from 3 months ago is ancient.
• Tweet: dies in 43 minutes. Learnings decay immediately.
• Pinterest pin: drives traffic for 3-4 months. Top pins last years. A learning from 3 months ago is still actively relevant because pins from 3 months ago are still actively performing.

This means: by pin 50, the system has a genuine map of what works. And because each pin is a long-lived asset, the intelligence helps retroactively understand why old pins are still performing — or why they stopped.

The Learning Loop

/create (Batch 1) → composed pins → post → 7 days → /track (early signal)
                                              → 30 days → /track (full picture)
                                                              ↓
                                                       learnings extracted
                                                       (image-copy combos,
                                                        board patterns,
                                                        title effects)
                                                              ↓
/create (Batch 2) → reads learnings → better image matching → /track
                                                                  ↓
                                                           more learnings
                                                                  ↓
/create (Batch 5+) → reliable predictions → optimized combos → /track

Recommended tracking cadence:

• 7-day check: early signal. Catches obvious winners and losers.
• 30-day review: full picture. This is where real learnings live.
• Quarterly review: long-tail. Which pins are still driving traffic months later?

Confidence Levels

Pins Tracked	Confidence	What the System Knows
0	Baseline	Pinterest vertical benchmarks only. No brand-specific intelligence.
1-10	Low	Directional. "Lifestyle images seem to save better." Single-variable observations only.
11-30	Medium	Board-level patterns. Image type preferences. Title pattern effects. Enough data to spot 2-3 reliable patterns.
31-75	High	Multi-variable patterns. Can predict which image-copy combos will outperform. Posting time optimization. Search term effectiveness.
76+	Very High	Full predictive capability. Recommends specific image + title pattern + board + time combinations with tracked accuracy.

Always state confidence level. A Low confidence prediction is a guess with data — present it as such.

Learning Categories

Image-Copy Pairing Effectiveness (Primary Intelligence)

This is the new layer that makes the learning loop compound faster. Track which combinations drive results:

Image Type	Title Pattern	Board	Save Rate	CTR	Close-ups	Sample
Lifestyle + vertical	Problem-solution
Lifestyle + vertical	How-to
Lifestyle + vertical	Listicle
Product shot	Product spotlight
Product shot	Problem-solution
Diagram / checklist	How-to
Diagram / checklist	Listicle
Quote card	Aspirational
Behind-scenes	Behind-scenes
Seasonal	Seasonal

Minimum sample: 5 pins per combination before treating a pattern as reliable. Below 5, label "emerging pattern."

What to look for:

• Does the same image type perform differently with different title patterns? (Usually yes — this is the insight scheduling tools miss.)
• Does the same title pattern perform differently with different image types? (Also usually yes.)
• Are there board-specific preferences? (A board's audience may prefer lifestyle images even for how-to content.)

Image Type Effectiveness

Track overall performance by image category:

Image Type	Save Rate	CTR	Close-ups	Engagement	Sample
Lifestyle (vertical)
Lifestyle (square)
Lifestyle (horizontal)
Product shot
Diagram / infographic
Quote card / text overlay
Team / workspace
Seasonal

Key hypotheses to test:

• Vertical outperforms horizontal by ~60% (Pinterest baseline claim)
• Lifestyle outperforms product shots for saves (common but not universal)
• Diagrams drive close-ups (users zoom in) but fewer saves
• Bright, high-contrast images outperform dark/muted ones

Title Pattern Effectiveness

Pattern	Avg Impressions	Save Rate	CTR	Close-ups	Sample
Number + specific
How-to + outcome
Question
Year + topic
Direct value
Problem-solution
Aspirational

Board Effectiveness

Board	Total Pins	Avg Impressions	Avg Saves	Avg CTR	Best Image Type	Best Title Pattern	Trend
[Board A]
[Board B]

Board health assessment:

• Growing: impressions per pin increasing
• Stable: consistent
• Declining: impressions dropping — needs content refresh or description update

Timing Effectiveness

Day × Time	First-24hr Impressions	7-day Saves	Sample
Monday AM
Monday PM
Wednesday AM
Wednesday PM
Friday AM
Friday PM
Weekend AM
Weekend PM

Pinterest timing caveat: Timing matters less on Pinterest than social media. The algorithm distributes over days. Title keywords and image quality matter more. Don't over-optimize timing until sample size exceeds 30.

Search Term Effectiveness

Track which target terms actually drove discovery:

• Which terms matched real search queries?
• Which terms drove saves vs. clicks?
• Unexpected search terms surfacing? (Opportunity for new concepts)

Pin Lifespan Tracking

Pin	7-day Save Rate	30-day Save Rate	90-day Save Rate	Still Active?	Classification

Classifications:

• Evergreen — still driving traffic at 90 days. Highest value. Make more like this.
• Flash — strong at 7 days, dead at 30. Tied to a trend. Good but not the foundation.
• Slow builder — weak at 7 days, strong at 30+. Algorithm needed time. Don't kill concepts based on 7-day data.

Prediction Engine

Step 1 — Find Similar Pins

Search pin history for matches on:

• Same image type (most predictive)
• Same title pattern
• Same board
• Same or similar search terms
• Similar seasonal timing

Step 2 — Calculate Base Prediction

• 5+ similar pins: use their average save rate and CTR
• 2-4 similar: use average, widen confidence interval
• 0-1 similar: use board average + image type average. State it's extrapolated.

Step 3 — Apply Adjustments

• Image reuse: if image was used before, discount by 10-15%
• Board saturation: if board got 5+ pins in last 7 days, note dilution
• Seasonal alignment: matching seasonal trend = positive adjustment
• Proven combo: if this exact image type + title pattern + board combo has tracked data, use that specific rate
• Orientation penalty: horizontal image = -30% from vertical baseline

Step 4 — State Honestly

"Predicted save rate: 2.8% (Medium confidence, based on 22 similar pins). This vertical lifestyle image with a problem-solution title on the Content Operations board matches your strongest-performing combination."

Or: "Predicted save rate: 1.2% (Low confidence — first diagram image on this board. Could range from 0.5% to 2.5%). Track at 7 and 30 days."

Handling Underperformers

When a pin underperforms (>25% below prediction or board average):

1. Check the image first. Horizontal? Dark? Low resolution? Wrong category for the concept?
2. Check the image-copy alignment. Did the image and title tell different stories?
3. Check the title keywords. Front-loaded? Searchable? Specific?
4. Check the board. Is the board itself healthy?
5. Check timing context. Competing cultural moment that week?

Record specifically: "Product shot (horizontal) with how-to title on Content Operations board: 0.4% save rate (board avg 2.1%). Likely causes: horizontal orientation + product-shot-as-how-to mismatch."

Memory Files

Written to ~~docs. Plain markdown. Human-readable, human-editable, no database.

pin-log.md — chronological record of all batches with per-pin metrics, images used, and outcomes.

pin-learnings.md — cumulative learnings by category: image-copy pairings, image types, title patterns, boards, timing, search terms, lifespan arcs.

pin-baselines.md — current state: overall averages, per-board, per-image-type, per-combination baselines, confidence level, image library health, last tracked date.

Why Plain Markdown

Transparent, human-readable, human-editable, no database required. Works in plugin mode (reads from Drive) and teammate mode (reads from workspace). The client can open pin-learnings.md and see exactly what the system knows. The system can read it and apply what it knows. Same files, same format, same intelligence.

The Compounding Advantage

No scheduling tool — not Tailwind, not Buffer, not Later — builds a cumulative intelligence layer about what specific combinations of image + title pattern + board + keyword + timing work for a specific brand. They optimize scheduling. They don't learn.

By pin 50, this system has a genuine map. By pin 100, it has institutional knowledge about this brand's Pinterest performance that the team itself may not have noticed — because it cross-references every variable systematically without forgetting. That map gets more detailed with every tracked batch. And because Pinterest pins are long-lived assets, the intelligence doesn't just help future pins — it helps understand why old pins are still performing.

That's the product. The composed pin is the deliverable. The learning loop is the value.

Teammate Protocol Compatibility

Same skill, different trigger:

Plugin mode: User runs /track manually after 7+ days.

Teammate mode: Standing orders schedule automatic performance reviews:

RECURRING TASKS

Weekly
- Task: Track pin performance (7-day check)
  Day: Monday
  Skill: performance-learning

Monthly
- Task: Track pin performance (30-day review)
  Day: First Monday
  Skill: performance-learning

Quarterly
- Task: Pin lifespan review and evergreen identification
  Skill: performance-learning

The teammate reads analytics (via ~~pinterest API if connected, or from pasted results in standing orders), runs the analysis, writes learnings, and reports in the morning email. Fully autonomous.