TLDR Marketing

TLDR MARKETING 2026-05-11

5 sections13 entrieslinks blockcomplete parse

Parse notes

  • Subscription/account footer truncated before parsefooter_truncated

02STRATEGIES & TACTICS

WHAT TO DO IF YOUR CUSTOMER COUNT IS FLAT BUT YOUR BASE IS FALLING APART

Flat customer growth can mask a weakening base and declining profitability. In one case, retained customers fell 18% while net-new customers rose 31%, yet margin-per-customer dropped 14%. To surface this early, segment customers into retained, reactivated, and net-new groups, then track margin-per-customer over time. Monitor CLV decile migration to see whether high-value customers remain loyal and mid-tier customers move up. Margin-per-customer is the key leading indicator, often signaling base erosion 6-9 months before revenue declines.

linkedin.com1 min read

03RESOURCES & TOOLS

HOW PERPLEXITY WINS THE 10X KNOWLEDGE WORKER

Perplexity Computer is not a chatbot or search tool, but a system for building and running workflows end to end. Users describe what they want, and it assembles the stack and executes it with no code. It turns multi-step tasks into automated systems that run in the background, like pulling Reddit trends, ranking posts, generating hooks, and syncing outputs into Notion on a schedule. Its advantage is orchestration, combining multiple models and tools into one workflow that handles real work.

links.tldrnewsletter.com4 min read

COURSE CORRECTION: GOOGLE TO LINK MORE SOURCES IN AI OVERVIEWS

Google is adding several new link formats to AI Overviews and AI Mode in an effort to drive more traffic back to the websites its AI summarizes. Changes include a "Further Exploration" section at the bottom of AI answers linking to relevant articles, an "Expert Advice" block surfacing snippets from news outlets and forums, and preview pop-ups when users hover over source links. Google is also testing a subscription integration that would surface a user's subscribed sites more prominently in AI results, with early tests showing higher click-through rates.

links.tldrnewsletter.com3 min read

04MISCELLANEOUS

EXPERIMENT ESTIMATION STUDY: HOW WELL CAN 1,391 GUESSES PREDICT A/B TEST EFFECTS?

A/B test impact is often overestimated, making gut-based prioritization unreliable. Across 1,391 predictions, the median error was 2.7 percentage points, and 56% of guesses were too high. Experience had a limited effect, with teams of 10+ years performing similarly to newer practitioners. Familiarity with similar tests improved accuracy, raising directional correctness to 65.1% and reducing median error from 3.2 to 2.5 points. Small groups performed better than individuals, with crowds of five cutting error to 2.3 points, though a consistent positive bias remained.

goodui.org7 min read

THE WORK OF KNOWLEDGE IN THE AGE OF AI REPRODUCTION

Walter Benjamin's idea of mechanical reproduction is extended to knowledge, where AI collapses the scarcity of knowledge work in the same way mass media did for art. Consulting decks, medical analysis, and legal drafts can now be produced instantly, reducing the value of the artifact while leaving judgment and accountability as the main sources of value. This creates a barbell effect where low-cost AI output floods the market and value concentrates in institutions and credentials that can be held responsible for decisions. As skills become easier to replicate, signaling and provenance rise in importance, similar to the sheepskin effect, where degrees matter more than marginal skill gains.

digitalnative.tech6 min read