PostHog
PostHog is a platform for product and web analytics with session replay, feature flags, experiments, error tracking, surveys and observability for AI applications. Model calls are captured with prompt, response, tokens, cost and duration, and linked to usage data.
PostHog combines product analytics, session replay, feature flags and experiments with observability for language models — model calls can be traced back to the affected user. An EU cloud region is available.
Explanation
Its distinctive angle is joining two worlds that usually sit apart: how people behave in the product and what the model does meanwhile. Anyone wanting to know whether an AI assistant not only works technically but actually leads to more completed tasks needs exactly that link. Feature flags and experiments add the practical part: a new prompt variant ships to ten percent first and is judged on completion rate rather than gut feel. For European requirements it matters that an EU cloud exists alongside the US one. Self-hosting is possible under an open licence but is explicitly unsupported by the vendor — so not a recommended route for production loads.
Marketing Relevance
For marketing and product, PostHog is where AI features are measured against business numbers rather than model metrics. For accountable owners the EU region choice and the ability to attribute token spend to its source also count.
Example
A new answer logic in the website chat ships to a subset of visitors via a feature flag. PostHog shows the variant completes more inquiries but consumes more tokens per conversation — the decision is documented as a trade-off.
Common Pitfalls
Session replays and prompt logs are personal data: they need consent, masking of input fields and retention periods. And self-hosting looks cheap but, without vendor support, is an operational commitment at your own risk.
Origin & History
PostHog launched as an open-source alternative to classic product analytics and grew into a suite of tools for product teams. As AI features spread in products it added observability for language models, including evaluations on live traffic and access via the Model Context Protocol.
Marketing Use Cases
Measure the impact of AI features on completion rates.
Ship prompt and model variants as experiments.
Attribute token cost per feature and user group.
Trace faulty answers back to the affected session.
Frequently Asked Questions
Can PostHog be used in a privacy-compliant way?
With three prerequisites: EU region, consent before loading analytics and replay, and masking of sensitive input fields. Without consent no recording may start.
Does PostHog replace a specialised LLM platform?
For an overview of cost, usage and impact, often yes. For deep evaluations, test datasets and prompt versioning, specialised tools are stronger.
Is self-hosting sensible?
Only for testing. The vendor explicitly does not support self-hosting; for production loads the hosted EU option is the more dependable route.
Related Services
Go deeper: Measurement & attribution · Model comparison 2026