# Insights.

> HTML 版（正本）：https://voidvector.tech/insights/
> 語言：en · 來源：voidvector（虛空向量有限公司）

Notes on building AI systems that move the metric — and stay shipped.

- **A self-writing wiki only compounds knowledge if someone can catch the page it got wrong.** (2026-07-13) — “Knowledge compounding” is the pitch: every AI answer should make the knowledge base better, not vanish once the query ends. Almost nobody asks the more basic question — if that answer was wrong, what exactly compounds? (https://voidvector.tech/insights/llm-wiki-compounding-audit-gap/)
- **Metrics before models: three questions to answer before shipping an AI feature.** (2026-06-18) — Most AI features don't fail because the model is weak. They fail because nobody can say what success looks like. Answer three questions first, or don't start. (https://voidvector.tech/insights/metrics-before-models/)
- **When RAG retrieval disappoints, the model is rarely the culprit. Your knowledge base is.** (2026-05-27) — Teams tune embeddings and swap models for weeks when the real problem is a corpus full of duplicates, dead versions, and documents nobody owns. (https://voidvector.tech/insights/rag-fails-at-the-knowledge-base/)
- **Five users are enough for a usability test. Most teams don't even run zero properly.** (2026-04-30) — The classic finding still holds: five participants surface most usability problems. The real scandal is how many teams ship for years without watching a single user. (https://voidvector.tech/insights/five-users-are-enough/)
- **The UX of AI features: users don't want a chatbox. They want the thing done.** (2026-04-09) — Chat is the lazy default of AI interface design. It hands the hardest job — defining the problem — back to the user. Good AI UX hides the model inside verbs. (https://voidvector.tech/insights/users-dont-want-a-chatbox/)
- **How to read a retention curve — and why DAU is the most flattering liar in your dashboard.** (2026-03-12) — DAU rises whenever you buy traffic, which is exactly why it can't tell you if the product works. The retention curve — does it flatten, and how high — is the honest witness. (https://voidvector.tech/insights/how-to-read-a-retention-curve/)
- **Brand positioning isn't a slogan. It's what you dare not to do.** (2026-02-11) — If your positioning doesn't force you to turn down customers, features, or channels, it isn't positioning — it's decoration. A consultant's test for telling the two apart. (https://voidvector.tech/insights/positioning-is-what-you-refuse-to-do/)
- **Search traffic is moving house: between SEO and GEO, which do you do first?** (2026-01-14) — AI engines are answering questions your customers used to type into search. Panic-pivoting to GEO while your site's fundamentals are broken gets you neither. (https://voidvector.tech/insights/seo-to-geo-where-to-start/)
- **When content drives traffic but not revenue, the funnel is usually broken at the trust layer.** (2025-11-19) — Between 'they read your article' and 'they pay you' sits a layer most content strategies never build: evidence. Traffic isn't the top of the funnel — trust is. (https://voidvector.tech/insights/content-converts-when-trust-does/)
- **One research method, four practices: how we point UX research at marketing and operations.** (2025-09-24) — AI, UX, growth, and marketing look like four businesses. Underneath they're one method — observe behavior, distrust self-report, decide on evidence — aimed at four surfaces. (https://voidvector.tech/insights/one-research-method-four-practices/)