Teardown No. 126 · Social

171 to 4,813 Twitter followers: How Yan Wang handed his blog's distribution and growth to AI

Yan Wang says AI ran the growth of his blog yage.ai from late February to late May 2026, and that in that time he never manually posted a single tweet. The AI rewrites each article into a five-to-six tweet thread in Chinese and queues it through the Typefully API. It also writes a daily AI briefing that goes out through the Kit API, and it reviews per-post results using a Reach, Traffic and Newsletter framework that it proposed itself. His own part was paying for services, reverse-engineering one X analytics endpoint, picking topics for under two minutes a day and checking the direction every few weeks. He reports Twitter followers growing from 171 to 4,813 and weekly active users growing from 2,500 to 7,000.

Try this step

Daily AI briefing newsletter. Every day the AI pulls significant AI developments from arXiv, Twitter, Hacker News and WeChat public accounts. It selects three to five items and writes a 200-400 character Chinese summary with a brief judgment. Then it sends the briefing to all subscribers through the Kit API at a scheduled time. It remembers the previous issue to avoid repeating topics. Wang told it which judgment lens to use, and the AI wrapped the Kit API calls and configured the cron job.

The problem

In early March 2026, 77% of yage.ai’s traffic was Direct, which he describes as “untraceable dark social sharing via WeChat, Telegram, Slack.” Organic Search was 15% and social media was nearly zero. In his words, “the site depended entirely on other people manually sharing links.” The site already had 370+ articles but only 2,500 weekly active users. He puts it this way: “The problem wasn’t a lack of content. It was that content had no way to reach readers.”

The blueprint (as he describes it)

  1. Daily AI briefing newsletter. Every day the AI pulls significant AI developments from arXiv, Twitter, Hacker News and WeChat public accounts. It selects three to five items and writes a 200-400 character Chinese summary with a brief judgment. Then it sends the briefing to all subscribers through the Kit API at a scheduled time. It remembers the previous issue to avoid repeating topics. Wang told it which judgment lens to use, and the AI wrapped the Kit API calls and configured the cron job.
  2. Article-to-thread conversion. The AI reads an article from yage.ai and rewrites its core arguments into a five-to-six tweet thread. The first tweet is a hook. The middle three to four tweets develop independent points, and the final tweet carries the article link. All threads are in Chinese with a fixed style: “rational and direct, no hashtags, no clickbait language.” In early March it generated 30 threads from the back catalog in one batch. Every new article since then gets a thread automatically.
  3. Scheduling through Typefully. The AI queued the 30 threads through the Typefully API at one every two days, at UTC 20:00. It picked that slot as the dual-peak overlap window it identified in GA4 data. Each new thread is appended to the queue automatically.
  4. Per-post review with R/T/N. At first only Typefully’s account-level data was available. Wang reverse-engineered X’s backend analytics endpoint and gave the AI the browser cookie and session token. The AI then built its own calling logic on top of that. After per-post data came online in late April, the AI proposed classifying posts into three types:
    • R (Reach), measured by impressions and follow growth.
    • T (Traffic), measured by URL clicks and GA4 campaign sessions.
    • N (Newsletter), measured by subscription conversions.
  5. Attribution with UTM tags. All Twitter thread links carry utm_source=twitter&utm_medium=thread&utm_campaign=<slug> and newsletter links carry utm_source=kit&utm_medium=email. For WeChat, the AI noticed that bounce rates were above 70% but session duration rose from 30 seconds to 80-100 seconds. That judgment led Wang to keep sharing links in WeChat groups rather than opening a public account.
  6. Weekly reports. Every Sunday or Monday the AI pulls data from GA4, GSC, Kit and Typefully. It writes a Markdown report covering channel trends, content performance, search changes, Twitter review and next-step recommendations. The following week it checks whether those recommendations were implemented and whether the data changed as expected.
  7. SEO. The AI added Allow rules for AI crawlers in robots.txt. In April it noticed referral traffic starting from claude.ai and chatgpt.com. In response it restructured pages so every section had a standalone conclusion paragraph. In May it found two pages with high impressions and low CTR, and it rewrote their titles, meta descriptions and opening summaries itself.
  8. What Wang does. He pays for the services. He picks two or three topics a day from a dozen or two candidates the AI lists, in under two minutes. Every few weeks he spends about ten minutes reading a weekly report to check that the AI’s judgment “hasn’t drifted.”

Stated results

  • Twitter followers grew from 171 at the end of February to 4,813 by the end of May 2026.
  • Weekly active users grew from about 2,500 to around 7,000, which he calls “nearly a 3x increase.” Daily users went from roughly 255 to about 1,000.
  • Newsletter subscribers grew from 420 to 1,183.
  • Posting volume was 15 posts in March and 180 posts in April (six per day), which was the period of fastest follower growth. May had 74 posts, with 458,000 impressions, 22,190 engagements and a 5.05% engagement rate.
  • By late May, Organic Social was 30-50% of traffic, depending on posting volume, and Direct had dropped to about 35%.
  • UTM-tagged Twitter traffic grew from zero to 10,000+ sessions per month. One weekly report cites “Twitter thread campaigns drove 10,445 sessions.”
  • Search reached 369,000 monthly impressions and 2,300 clicks by late May, with an average position of 7.6.
  • In May the newsletter had daily open rates of 40-55%, click rates of 3-6% and unsubscribe rates below 0.1%.
  • Delaying the Kit subscribe form until a reader had scrolled through 50% of the page improved subscription conversion by roughly 40%.
  • He spent “less than ten hours total” on the system over three months.

Caveats

  • All figures are self-reported for one personal blog. The post does not isolate how much of the growth came from AI rather than other factors.
  • The post does not name the AI tool or model it used.
  • Most of the distributed content is in Chinese. That includes the threads and the daily briefing summaries.
  • The post footer says it is “100% AI-generated.”
  • The footer links to Superlinear Academy, a community Wang co-founded that sells AI courses.
  • The 1,300 additional monthly clicks from the two rewritten pages are the AI’s estimate. The source does not state the actual result of those rewrites.
  • The baseline conversion rate behind the roughly 40% form improvement is not stated.
  • Per-post X data comes from a backend analytics endpoint he reverse-engineered, not from a public API.
  • Some channels did not work. Reposts to Zhihu and CSDN brought almost no effective traffic, and Chinese articles got 15 to 115 times the pageviews of their English counterparts.
  • He says the AI “won’t actively examine whether the framework itself is sound.” Framework-level judgment still needs a human.
  • Handles: X @grapeot (verified: linked from the post itself). GitHub grapeot (verified: linked from the about page). No other platforms checked.

FAQ

Did Yan Wang post anything himself?

He says that over the three months he never manually posted a single tweet, wrote a single newsletter issue or did any SEO work. He picks two or three topics a day for the AI to write.

What does the setup cost?

He pays for Typefully Creator at $12.50/month, Twitter Premium and the Kit paid plan, which he says comes to under $50 per month total. The source does not give separate prices for Twitter Premium or Kit.

How much time does he spend on it?

He says he spent less than ten hours total over three months. Daily topic picking takes under two minutes, and a direction check every few weeks takes about ten minutes.

Which AI model runs it?

The post does not say. It describes the system as Python scripts the AI wrote, a few cron jobs, a collection of prompts and data flowing through APIs.

Sources


Want agents like these working your distribution? Get It Built. Get new teardowns weekly: subscribe (double opt-in).

Get one teardown a week.

Same format, a new operator every issue. Double opt-in.