Loading tutorials…
Loading tutorials…
Cohorts and Personas are where Amplitude beats GA4 by a mile. But most teams build sloppy cohorts and end up with overlapping segments that conflict. Here's the discipline that keeps them clean.
Who this is forProduct, marketing, and growth teams who want to segment users by behavior — not just demographics. Especially relevant for SaaS teams running activation and retention experiments where "power user" needs a real definition, not vibes.
What you'll need
Step 1
Cohort = behavioral group ("did X in last 30 days"). Persona = ML-generated group from user properties. Audience = a synced cohort/persona exported to ad platforms.
A Cohort is a user list defined by behavior: "Users who completed `Signup Submitted` AND `First Feature Used` in the last 30 days but did NOT do `Subscription Started`."
A Persona (Plus plan and above) is an ML-generated segmentation of your user base based on behavioral + property patterns. Amplitude clusters users automatically — useful for "what natural groups exist?"
An Audience is a Cohort or Persona synced via Amplitude's Data Destinations to Google Ads, Meta Ads, LinkedIn Ads, Klaviyo, or any other downstream tool.
Most teams build Cohorts first, then turn them into Audiences for paid retargeting. Personas are best used to discover unknown segments, not for direct targeting.
Cohort definitions are mutable and recompute daily. Audience syncs to ad platforms run on Amplitude's schedule (typically every 1-6 hours depending on destination).
Step 2
Navigate to Amplitude → Users → Cohorts → "New Cohort." Use the visual builder to define behavior + time window + property filters.
Open Amplitude → Users → Cohorts → click "Create Cohort" (top-right).
Use the visual builder. Step 1: "Users who performed [event] [count] times in [time window]." Example: "Users who performed `Feature Used` at least 5 times in the last 30 days."
Add filters: properties on the user OR on the event. Example: filter event property `feature_area = "dashboard"` to get only dashboard power users.
Add exclusion clauses: "AND did NOT perform `Subscription Cancelled` in the last 90 days." Critical for cohorts you intend to retarget — don't advertise to churned users.
Save with a descriptive name: "Dashboard power users — active L30, not churned L90." Vague names ("VIP users") create confusion two months later when you forgot the definition.
Step 3
Set up these 6 cohorts as the foundation: New Users, Activated Users, Power Users, Churning Users, Paid Converters, Free-Trial At-Risk.
New Users L7: signed up in the last 7 days, fewer than 3 sessions. Target: activation onboarding.
Activated Users: completed your North Star action (varies by product) at least once. Target: feature expansion.
Power Users: top 20% by event count in last 30 days. Target: case studies, advocacy programs.
Churning Users: were active in 60-90 days ago but not in last 30. Target: win-back campaigns. (Distinct from churned — they're still salvageable.)
Paid Converters: completed `Subscription Started` in last 30 days. Target: upsell, retention emails.
Free-Trial At-Risk: in trial period, <2 sessions in trial L7. Target: in-app nudges, sales outreach. Typical save rate: 5-15%, worth $20-50 per saved trial.
Document each cohort's definition in a shared doc. Amplitude's description field is fine but rarely checked.
Step 4
Personas (Plus plan and above) uses unsupervised ML to cluster your users into 3-10 natural groups based on behavior + properties. Useful for finding segments you didn't know existed.
Open Amplitude → Users → Personas → "Create Persona."
Select the property to cluster on (e.g., "behavior in last 30 days"). Amplitude's ML clusters users into 3-10 groups.
Review each persona: Amplitude shows the defining behaviors (e.g., "Cluster 3: high `Feature A` usage, low `Feature B`, paid plan").
Useful pattern: build a Persona, then convert the most interesting cluster into a hand-defined Cohort with explicit rules. This bridges discovery to actionable targeting.
Personas are NOT for direct ad targeting — the cluster definitions can shift as user behavior changes. Use them for product discovery, hand-build cohorts for activation.
Step 5
In Amplitude → Data → Destinations, connect Google Ads, Meta Ads, LinkedIn, Klaviyo. Then mark cohorts for sync. Audiences refresh every 1-6 hours.
Open Amplitude → Data → Destinations. Add the platform (Google Ads, Meta, LinkedIn, Klaviyo, Customer.io, Iterable, etc.). OAuth into your ad account.
In the cohort detail page → "Sync to Audiences" → select the destination. Amplitude hashes user emails/IDs before syncing (privacy-safe).
Sync frequency: Meta + Google = ~6 hours. Klaviyo + Customer.io = ~1 hour. LinkedIn = ~24 hours.
In your ad platform, the synced audience appears as a Custom Audience (Meta) or Customer List (Google Ads). Use it as a Targeting or Exclusion list.
Common pattern: sync "Power Users" cohort to Meta as a Lookalike Audience source — Meta builds a 1-3% lookalike from your top users. ROAS on these is typically 2-3x retargeting baseline.
Step 6
Cohort definitions drift as events get renamed, new features ship, and your business model shifts. Audit every 60-90 days.
Calendar reminder: quarterly cohort audit. Open each saved cohort and verify the definition still maps to current product reality.
Watch for events that were renamed, deprecated, or replaced. Amplitude's Data → Catalog flags deprecated events; cohorts using them silently shrink.
For ad-platform-synced cohorts, check Amplitude → Data → Destinations → sync logs for failures. A failed sync means your retargeting is silently broken — typical loss is $200-1,500/mo in wasted impressions.
Archive cohorts you no longer use. Amplitude Plus plan caps active cohorts (typically 200 per project). Bloat slows the UI and confuses analysts.
Document cohort changes in a changelog so analysts know when a definition shifted (and historical comparisons may be apples-to-oranges).
Common mistakes
Vague cohort names like "VIP users"
What goes wrong: Two months later, no one remembers the definition. Analysts re-define "VIP" on the fly and create three overlapping cohorts. Marketing campaigns target conflicting lists; ad spend gets duplicated. Typical waste: $500-2,000/mo on cross-targeting.
How to avoid: Cohort names must include the behavior + time window: "Dashboard power users — L30 5+ sessions, not churned L90." Verbose is fine; ambiguous is expensive.
Forgetting to exclude churned users from retargeting cohorts
What goes wrong: You retarget "Active Users" cohort to upsell premium. The cohort includes users who churned 45 days ago. They get ads. They reply angrily on Twitter. Brand damage + wasted spend (~$50-200 per case).
How to avoid: Every retargeting cohort needs an explicit "AND did NOT perform Subscription Cancelled in the last 90 days" exclusion.
Using event count without a time window
What goes wrong: Cohort "Users who used Feature X 10+ times" includes users from 3 years ago. List balloons to include long-inactive users. Email open rates drop, ad ROAS drops, deliverability suffers. Cost: $300-1,500/mo in lost campaign efficiency.
How to avoid: Every behavioral cohort needs a time window: "in the last 30 days," "in the last 90 days." Always bound by recency.
Treating Personas as targetable audiences
What goes wrong: You sync a Persona to Meta. Two weeks later the ML re-clusters and your audience shifts. Your retargeting campaign now targets a different group than you launched against. ROAS reporting becomes meaningless.
How to avoid: Use Personas for discovery only. Once you find an interesting cluster, hand-build a deterministic Cohort with explicit rules and sync that.
Audience sync set up once, never monitored
What goes wrong: Amplitude's OAuth to Meta expires. The sync silently fails. Your "Power Users" Lookalike on Meta is stuck at the December audience while your real user base has 3x'd. Lost growth: 20-40% of potential reach for paid social.
How to avoid: Monthly: open Amplitude → Data → Destinations → check sync logs for each destination. Set Slack alerts on sync failures.
Building 50+ cohorts and never deleting any
What goes wrong: Active-cohort cap on Plus plan hit (typically 200). New experiments can't create cohorts. Team uses spreadsheets instead, and the cohort approach silently collapses.
How to avoid: Quarterly cleanup: archive cohorts unused in 60 days. Use Amplitude's Last-viewed timestamp on each cohort to find candidates.
Recap
Done — what's next
How to set up Amplitude event tracking the right way
Read the next tutorial
Hand it off
Defining cohorts is judgment work — not just product knowledge but business-model thinking. A specialist who has built cohort libraries for 30+ SaaS products will catch definitional gaps that cost months of bad campaign decisions. Initial cohort build typically $200-500; ongoing maintenance $200-600/mo at $14-16/hr.
See cohort-setup rates
Google Ads Customer Match: 1,000+ matched users minimum. Meta Custom Audiences: 100+ minimum. For Lookalike Audiences, you want a 1,000+ seed for stable lookalikes. If your cohort is too small, widen the time window or relax the event count threshold.
Yes, if you sync CRM data to Amplitude as user properties (via Segment, Census, or Hightouch). Once `arr_band` or `industry` are on the user profile, you can filter cohorts by them. Without the sync, Amplitude has no way to know.
Behavioral cohorts (event-based) recompute daily by default. You can change to hourly on Growth plan. Cohorts synced to ad destinations re-export on the destination's cadence (Meta: ~6h, Google: ~6h, Klaviyo: ~1h, LinkedIn: ~24h).
Personas only segment users already in Amplitude — they're an existing-user analysis tool, not a prospecting tool. For prospecting, use Lookalike Audiences in Meta/Google seeded from your "Power Users" cohort.
Amplitude cohorts for tactical (campaign-ready, sync-to-destination) segments. Data warehouse (Snowflake/BigQuery + reverse ETL) for complex business logic that needs SQL. Most teams use both — Amplitude for the 80% case, warehouse for the long tail.
Amplitude
Bad event tracking is the most common reason Amplitude projects fail. Here is the naming convention, the SDK code, and the Data Guard rules that keep your taxonomy clean for years — not weeks.
Amplitude
Funnels are the most-used chart type in Amplitude and the most-misused. The defaults assume an attribution window that doesn't match most SaaS sales cycles. Here's how to build funnels that match reality.
Amplitude
Retention is the single most important SaaS metric. But Amplitude has three retention models and they answer different questions. Pick the wrong one and you'll think you're growing when you're actually leaking users.
Amplitude
DIY Amplitude is a great idea — until your taxonomy gets out of control or your charts disagree with reality. This is the honest framework for when the math flips toward hiring.