PostHog

What PostHog is

PostHog started in 2020 and has grown into an unusually wide platform: product analytics with autocapture or manual instrumentation, a GA-style web analytics dashboard, session replay for web and mobile, feature flags, experiments, error tracking, logs, surveys, a data warehouse that syncs from tools like Stripe, data pipelines out to a couple of dozen destinations, and observability for LLM-powered apps that captures traces, latency and cost. The README lists all of it behind one install, and you can drive the workspace from Slack, the web app, a desktop client, or an editor through its MCP server rather than only from a dashboard. The repo is a Python-led monorepo and is genuinely active — the most recent push at capture was the same day.

The licence and hosting details matter more than they usually would, because this is the sort of tool you either adopt for years or rip out. The repository is available under the MIT expat licence except for its ee directory, which carries its own licence, and a separate posthog-foss repository exists with the proprietary parts stripped out for anyone who needs the whole thing to be free software. Every tool has its own monthly free allowance on their cloud — the README lists the first 1 million events, 5k recordings, 1M flag requests, 100k exceptions and 1500 survey responses each month — after which billing is usage-based. Self-hosting is one line of Docker with 4GB of memory recommended, but read the project's own framing before you plan on it: they scope open-source deployments to roughly 100k events per month, recommend migrating to their cloud beyond that, and state plainly that open-source installs come with no customer support and no assurances.

What to instrument first

The temptation with a platform this wide is to switch everything on in an afternoon, which produces a workspace full of dashboards nobody opens. PostHog's own getting-started guide walks activation, then retention, then revenue, and that order is worth following literally:

1. One named activation event Decide the single moment a new user first gets the thing they came for, and capture that with a name you chose. Everything downstream is interpreted against it, and activation rate is the one number that stays meaningful when you only have a handful of users.
2. A retention view on that event Not on signups. A signup chart flatters you because it only ever goes up; a return chart tells you whether the product is worth opening twice.
3. Revenue attached to the same identity Until money sits on the same person as the events, you cannot separate an expensive channel from a good one, which is the whole question behind whether your marketing is working.
4. Session replay, on the one step that leaks Replay is the fastest way to learn why a step fails and the easiest way to lose an afternoon. Point it at the funnel step your retention view already flagged, watch five sessions, then close it.

Autocapture is the reasonable exception to instrument-later, because it records interactions without you having to name them in advance — which is genuinely useful while you still do not know what matters, and genuinely noisy once you do. Feature flags and experiments are worth wiring the day you have enough traffic for a result to mean anything, and not before: an experiment on a hundred visitors mostly measures your patience.

What we learned running PostHog on our own funnel

This part is from operating rather than from reading. AgentCeres — the AI Growth Officer at agentceres.com — uses PostHog for pageviews and signup conversion, and the most expensive lesson in it was structural: a browser-side event is a request the browser is free to decline. Roughly half of our signups never produced the client event we were counting, and the shortfall was worst in precisely the cohort we were paying to acquire, Brazilian paid search, where ad-blocking is common. The number was not noisy, it was absent, and an absent event looks identical to a user who did not do the thing. For one field a product decision genuinely rested on, we stopped asking the browser at all and wrote it server-side into our own database at the moment it happened.

The general rule we took from it: any number a decision rests on should be recorded where you control the write, and analytics is for the shape of behaviour rather than the ledger. It is also worth being clear about what a tool like this does and does not do for growth. PostHog will tell you, with real precision, which step of your funnel leaks and who left — it will not write the post, send the email, or fill the funnel back up. That gap is the work; a lighter alternative if all you need is traffic and page-level numbers is Umami, and the execution side — drafting the content and outreach, with a human approving anything that goes out — is what AgentCeres exists to cover.