LLMs are picking winners. Here's how to become one.
PostHog's LLM traffic grew 41x in two years and converts better than almost anything else, but even they struggle with uneven AI visibility across products—here's their brutally honest playbook for Answer Engine Optimization when the entire discipline runs on vibes.
Read Original Summary used for search
TLDR
• Added a conditional onboarding question asking users which AI prompt led them to PostHog—collected 9,214 real prompts in 3 months, creating a goldmine of first-party data that costs $0 to implement
• Traditional SEO still works, but the unit of consumption shifted from pages to "chunks"—content needs to be citable in pieces with scannable structure, natural language, and concrete specifics over vague claims
• Most AEO is correlation dressed up as causation—be twice as suspicious when your dashboard shows flattering results, and scrapped 6 months of their own data when tracked prompts had zero overlap with real user behavior
• Clean your house first before chasing Reddit threads and Wikipedia pages—making owned content genuinely great is step one, everything else comes after you have capacity
• Use MCPs to pull from multiple sources (Gauge for prompt visibility, GSC for rankings, PostHog for conversion) since no single data source tells the full story yet
In Detail
PostHog's traffic from LLMs grew 41x in two years with conversion rates better than almost any other channel, but they face a visibility problem: while most models know them for product analytics and feature flags, fewer acknowledge their 14 other products like data warehouse and AI observability. The author, responsible for AEO at PostHog, shares a practitioner's guide to navigating this new discipline built mostly on assumptions and vibes.
The breakthrough came from adding a conditional question to their onboarding flow: whenever someone says they heard about PostHog through AI, they ask which prompt was used. In three months, they collected 9,214 real prompts from actual users with credit cards—first-party data that feeds directly back into strategy. Traditional SEO fundamentals still apply (good structure, clear writing, fast load times), but the unit of consumption shifted from pages to "chunks." Content needs to be citable in pieces with scannable subheadings, natural language, direct answers in first sentences, and concrete specifics over vague gesturing. The rule: clean your house first—make owned content genuinely great before worrying about Reddit presence or Wikipedia pages.
The hardest lesson was scrapping six months of historical data when they realized their tracked prompts had little overlap with what real people actually used. In a discipline this new, adaptability isn't optional—models change constantly, and attachment to your v1 can be fatal. Most AEO case studies are correlation dressed up as causation, so be twice as suspicious when your own dashboard shows flattering results. Since no single data source tells the full story (referrer traffic, self-reported attribution, bot analytics, prompt visibility tools all have gaps), they use MCPs to pull from Gauge, Google Search Console, and PostHog to see the bigger picture. The takeaway: expect constant change, question everything that makes you look good, and be the most honest person in every room.