Screenshots happen in bursts, and bursts are the strongest pre-decision intent signal available to any AI.

The 5 Numbers That Matter

The Activation Metric

Pool's "7 friends in 10 days": 3 bursts in 7 days

At the extreme: users with 16+ first-week bursts average 2,282 lifetime screenshots (3.4x the 0-burst group).

56.5% of users burst on day 0 or day 1. Bursting is pre-existing — people already do it. Pool just needs to capture and deliver value from it immediately.

Validated against product metrics (Mixpanel, Mar 8–24)

Bursters don't just screenshot more. They search 2.6x more, tap intents 5.5x more, and open links 5.5x more. They are the core product audience.

Churn signal

Churned users had less than half the early burst activity. Early burst frequency is a leading indicator of retention.

What Bursts Are

The data

A burst = 5+ screenshots with <10 min gaps between each. A session of focused saving.

5,618 total burst sessions across 209 users (81% of active users).

Speed within bursts

Marathon users screenshot every 8 seconds. The AI can't react per-screenshot. It must detect the burst, wait for it to end, then process the whole session.

When bursts happen

• Peak: 2pm–9pm (afternoon/evening browsing)

• Every day of the week (slightly higher midweek)

• Top power user bursts on 64% of active days

• Average power user: ~25% of active days

Growth

9x sessions, 4x users in 2 years. TikTok culture: see it, save it, move on.

What's Inside Bursts

Burst types (5,129 classified sessions)

95% of bursts are classifiable

First 3 screenshots predict the burst

With regex only. LLM classifier would push to 90%+.

6 burst archetypes

1. Shopping/browsing — products, prices, style inspo. AI resolves to buy links, creates mood board.

2. Restaurant hunting — saving places rapidly. AI ranks by taste, adds booking links.

3. App evaluation — testing 5-10 apps in one session. AI compares and recommends.

4. Course/content capture — screenshotting slides and tips. AI organizes into summary.

5. Research rabbit hole — deep focus on one topic. AI auto-saves as collection.

6. Trip/event documentation — saving during travel. AI creates trip diary.

The Social-Shopping-Food Triangle

Burst-to-burst transitions

Social, shopping, and food feed into each other. Travel is the trigger that activates all three.

Shopping bursts escalate before resolving

The 2nd and 3rd shopping bursts are peak intent. That's when to intervene.

Repeat rates

Multi-burst days (611 days with 2+ bursts)

Top same-day combos: social+social (79), shopping+social (35), food+social (24), food+shopping (18). The triangle holds at the daily level.

After a burst, users come back in ~2 hours

That's the delivery window. Detect burst → process → have results ready in 2 hours.

Additional Signals

City detection from content (no GPS needed)

720 city-identifiable sessions across 18 cities. The AI knows where you are from what you're screenshotting.

Time-of-day patterns

Seasonal patterns

• Shopping peaks in January (sales season), not November/Black Friday

• Travel peaks in October and December (autumn + holiday)

• Food peaks in summer (July)

• Fitness is flat year-round (no "new year resolution" spike)

Cross-session repeats

185 items were screenshotted across 4+ separate sessions by 40 users. Examples: TEMU cart (558 days, still not purchased), domain name searches (93 days of startup naming), dating profiles (613 days). These are the highest-confidence intent signals — the "nudge" goldmine.

Intent feedback: 65% acceptance

When we turn a screenshot into an action link, 65% of users tap it. The resolution model is validated.

Couple & Household Detection

We can detect couples purely from screenshot patterns — no social graph needed.

Discovered pair: 373 shared titles, 1,117 screenshots within 10 seconds of each other, syncing on 87% of active days.

What they screenshot simultaneously:

• Shopping together (Saint Laurent sunglasses, Ffern perfume)

• Shared hobbies (jewelry making, photography)

• Home life (recipes, coffee shops, furniture)

• Same product from different angles — same item, same second

Product opportunities: Gift signals, shared shopping, recipe sync, joint interest recommendations, relationship health tracking via sync rate.

The "Other" Bucket (24% of bursts)

Unclassified bursts contain real patterns our regex missed:

With an LLM classifier, the "other" bucket drops from 24% to under 5%.

The Product Architecture

The burst is the atomic unit of intent

Screenshot #1  →  save it
Screenshot #2  →  save it (8 sec later)
Screenshot #3  →  BURST DETECTED — classify (77% accuracy)
...
[10 min gap]

Three product modes

1. Live Mode — Burst detected, AI activates: shopping → buy links + mood board. Restaurant hunt → ranked list + booking links. Course capture → organized summary.

2. Morning Brief — "Yesterday you saved 16 products in Amsterdam. Here are the 3 worth buying." Generated from real recent bursts.

3. Deep Memory — "What restaurants did I save in Amsterdam?" Full history, instant recall.

Context architecture

General context (built once, benefits all users): brand table for shopping, venue table for restaurants, route table for travel, app directory for app evaluation. The burst classifier routes to the right table.

Personal context (per-user): taste profile, history, preferences. Gets richer over time.

Level 1: Screenshot + general context → "Here's a buy link" Level 2: + personal context → "That bag is 20% off, you've been watching it since January" Level 3: + live data (email, calendar) → "...and you have dinner Friday where you'd wear it"

The moat

Screenshots are a pre-decision signal. Email is a post-decision confirmation. The window between screenshot and email receipt is the window Pool owns.

No other AI has:

• Access to the pre-decision intent signal

• Burst-level classification of what users are actively doing

• Context tables that resolve intent to action in real-time

• The behavior growing at 9x in 2 years

Case Study: Amsterdam, March 12, 2026

One user, 32 screenshots in one session: 16 products (Samsoe Samsoe, Jacquemus, Acne Studios, Zara), 6 Amsterdam restaurants, 3 style TikToks. They were in Amsterdam, restaurant-hunting and shopping simultaneously.

What the AI should have done:

1. Detect burst at screenshot #3

2. Classify: shopping + food, location = Amsterdam

3. When burst ends: rank restaurants by taste + add booking links. Cluster products by brand + find buy links.

4. Morning brief: "Amsterdam haul — 16 products, 6 restaurants. Here are the 3 worth buying and where to eat tonight.