northstar — a free product management library of 124 case studies, 30 book reviews, 72 answers and 18 playlists

AARRR is a five-stage funnel framework — Acquisition, Activation, Retention, Referral, Revenue — coined by Dave McClure. It gives a startup one metric per stage so problems can be located rather than guessed at. The order matters: fixing acquisition while retention leaks is the classic misuse.

Activation is a user reaching first value — the moment your product proves it works for them. Retention is them coming back after that. Activation is a one-time event you can design, retention is an ongoing verdict on whether the value holds. Fixing retention with an activation problem underneath never works.

MRR is monthly recurring revenue — the predictable subscription revenue you bill in a month. ARR is annual recurring revenue, usually just MRR multiplied by twelve. The distinction that matters is not the timeframe but the word recurring: one-off fees, services and usage overages belong in neither, and including them is the most common way these numbers get inflated.

Bottom-up starts with individual users adopting a product and spreading it inside the company until a purchase becomes formality. Top-down starts with an executive buying for the organisation and rolling it out. Bottom-up needs a product one person can love; top-down needs a business case one executive can defend.

Build what differentiates you, buy what doesn't. The test isn't cost — it's whether customers would ever choose you because of this component. Teams get burned by underestimating the permanent maintenance cost of building, and by buying the one thing that was actually their edge.

Bundling combines several capabilities into one offering; unbundling splits a capability out of an incumbent bundle and does it much better alone. Startups typically unbundle to get a wedge, then re-bundle as they grow — the cycle is real, but the failure mode is unbundling your own product for internal reasons and calling it strategy.

DAU/MAU divides daily active users by monthly active users to estimate how many days a month the average user shows up. 20% is a common baseline, 50%+ is exceptional. It only means something for products people should use daily — applying it to a tax product or a job board produces nonsense.

Use RAG when the model needs knowledge it does not have — documents, current data, anything that changes. Fine-tune when it needs a behaviour it will not follow reliably from a prompt — a format, a style, a narrow classification task. Knowledge problems are almost always RAG problems, and most teams reach for fine-tuning to solve one and are disappointed.

A free trial gives full access for a limited time; freemium gives limited access forever. Trials suit products with fast, obvious value and a clear buying moment. Freemium suits products with network effects or slow-building value, and only works if the free tier costs you little to serve.

Horizontal SaaS solves one function across every industry — CRM, payroll, ticketing. Vertical SaaS solves many functions for one industry, like software for dental practices or construction firms. Horizontal has a larger market and fiercer competition; vertical has a smaller ceiling and far better retention.

Vendors promised 80% deflection and no more agents. The teams who believed it are now drowning in escalations and churned customers. Here's what's true.

Traditional SaaS companies enjoy 80%+ gross margins. AI companies are grappling with the massive variable cost of inference. Here is why the business model must change.

The days of launching a buggy, ugly 'Minimum Viable Product' and charging early adopters for it are over. AI has fundamentally changed consumer expectations for baseline software quality.

The 'prompt engineer' title peaked and crashed in 18 months. Clever phrasing stopped mattering. Specifying problems clearly never will.

ChatGPT proved that horizontal, general-purpose AI is incredible for brainstorming. But to capture true enterprise value, AI must be verticalized into specific industry workflows.

Selling AI to enterprises in 2026 isn't a sales motion — it's a gauntlet of security review, data governance, and a buyer terrified of looking foolish.

'RAG is dead' is the hot take of 2026, fueled by long context and agents. The truth: naive chunk-and-retrieve is dying. Retrieval itself is fine.

The doom take says AI ends entry-level engineering. The reality is messier: the tasks juniors learned on vanished, and nobody rebuilt the ladder.

Everything is an 'agent' now, which means nothing is. Here's the test that separates products that actually decide and act from chatbots with a thesaurus.

The frontier model race grabs headlines. The real shift in 2026 is small, fast, on-device models quietly handling the 80% of tasks that never needed a giant brain.

You can't ship a serious AI product on vibes. Evals are becoming the core engineering discipline of applied AI — and the moat nobody copies.

Ranking #1 means nothing when the answer renders above the links and nobody clicks. The funnel didn't shrink — it moved inside the model.

Per-seat pricing assumed labor was the cost. When the model does the work, your COGS becomes a variable you don't control. Here's the math.

Apple Intelligence was supposed to be the iPhone moment for on-device AI. In May 2026, it's an embarrassment. The flagship Siri rebuild is delayed again, adoption is rounding to zero, and Apple's structural problem is now visible: they refuse to use third-party models, and their own are 18 months behind.

AI engineer went from niche role to most-hired role in tech in 18 months. They don't train models. They don't write ML papers. They don't do statistics. Here's what they actually do, why the role exists, and what it means for hiring.

'AI-native' used to mean 'we use ChatGPT for marketing copy.' In 2026 it means something specific. Here are the six operating decisions AI-native startups make differently — and why each one compounds over time.

Operator launched in early 2025 to mixed reviews. 18 months later, it's the most-used agentic browser by 4x. Here's what's actually working, what's still broken, and what OpenAI should ship next to keep the lead.

Three production-grade video generation models, each better than the entire 2024 market. For marketers shipping ads, they're not interchangeable. Here's what each one is genuinely good at, where they fail, and how to pick for your use case.

OpenAI has the brand. Anthropic has the trust. In enterprise AI deals in 2026, the gap is closer than the headlines suggest — and Anthropic is winning the segments that matter most for long-term durability. Here's the honest read.

Google's Antigravity v2 dropped quietly in 2026 — the most ambitious AI coding platform from Big Tech since GitHub Copilot. Built on Gemini agents, vertically integrated into Google Cloud, designed to leapfrog Cursor on enterprise. Here's what's actually inside and who should care.

Anthropic's Model Context Protocol went from 'interesting experiment' to 'standard everyone implements' in 18 months. Cursor, Claude Desktop, OpenAI Agents SDK, Cline, Continue, Zed — all support MCP. Here's what it is, why it matters, and what your product should do.

Cursor crossed $200M ARR in 2026, built by a small team that started with a VS Code fork. GitHub Copilot had Microsoft, Azure, OpenAI access, and a billion-user funnel. How did the upstart win? Three structural decisions, made early.

Two months in with both Claude 4.7 and GPT-5 in production. The 'which is better' answer isn't what benchmarks suggest. Here's the honest read on what each model is genuinely good at, where they fail, and how to pick for your use case.

Six browsers launched agent capabilities in the last 12 months. Operator, Comet, Dia, Arc Browse, Brave Leo, and Chrome extensions. Most will lose. Here's what's actually different about each, who they're for, and what to watch for the category to settle.

AI agents went from demo to actual tool in 2026 — OpenAI Operator, Claude agents, and the Universal Commerce Protocol changed the game. Here's what's real, what's still hype, and the concrete moves PMs, marketers, and founders should make.

Google's May 15, 2026 documentation is the closest thing to an official rulebook for the AI search era. Here's what it actually says — and the concrete moves PMs, marketers, and founders should make this week.

Google's 2026 keynote reshuffled the AI roadmap for every product team. Here's the 5 announcements that actually matter — and the specific moves PMs, marketers, and founders should make this quarter.

Linear AI, Cursor, NotebookLM, Notion AI, Granola, Loom AI — the PM toolkit reshuffled in 2026. Here's what's actually worth paying for, what's hype, and the stack we'd build today.

Claude Code, Cursor, Replit Agent, Bolt — the AI-assisted build stack matured in 2026, and a wave of non-engineers (and engineers without frontend experience) are shipping real products solo. Here's what the workflow actually looks like, what works, and what breaks.