Between late February and late May 2026, yage.ai’s weekly active users grew from 2,500 to 7,000, newsletter subscribers from 420 to 1,183, and Twitter followers from 171 to 4,813. These numbers make it look like the blogger was busy. The reality is different. Over these three months, I never manually posted a single tweet, wrote a single newsletter issue, or did any SEO work. AI handled all of it.
What I did was minimal: paid a few bills, helped it reverse-engineer one API, spent two minutes a day picking topics, and checked the direction every few weeks. Everything else was AI.
This is a retrospective, laying out what AI did, how it did it, what I contributed, and what worked and what didn’t.
GA4 weekly active users tell the clearest story.
The early March baseline was about 2,500 weekly active users. By late May, it had stabilized at around 7,000, nearly a 3x increase. The path was not smooth: every new article triggered a pulse of traffic, and occasionally a piece that spread widely pushed the numbers higher for a week. But these pulses came and went quickly. What matters is the platform left behind after each pulse subsided. It never fell back to the starting point.
The daily baseline tells the same story. Early March was roughly 255 daily users. Late May was about 1,000 daily users, a 4x increase. This platform was not driven by any single article. It came from several supporting systems working together: a Twitter account built from scratch, newsletter subscribers growing from 420 to 1,183, Google Search delivering 2,300 monthly clicks, and the /share/ directory becoming the site’s primary content surface. Continuous content output created continuous distribution opportunities, and these systems converted the attention left behind after each pulse into a repeatable asset.
The channel composition shifted fundamentally. In early March, 77% of traffic was Direct (untraceable dark social sharing via WeChat, Telegram, Slack), Organic Search was 15%, and social media was nearly zero. By late May, Organic Social accounted for 30-50% (varying with posting volume), Direct had dropped to about 35%, Organic Search held steady at 15-20%, and Email and Referral each contributed 2-3%. In March, the site depended entirely on other people manually sharing links. By May, it had a multi-channel distribution system.
Search data reinforces the structural argument. In early March there was no GSC data (it hadn’t been connected yet). By late May, yage.ai was receiving 369,000 monthly impressions and 2,300 clicks, with an average position of 7.6. The two pages with the most impressions were ollama-cloud-vs-api-vs-subscriptions-en (74,194 impressions) and mlx-apple-silicon-en (58,108 impressions), both English-language technical comparison pieces. These weren’t pages I deliberately optimized for SEO keywords. They were byproducts of AI running its own research and writing workflows, indexed and ranked by Google without any per-page SEO work from me.
The Twitter trajectory is equally clear. The account had 171 followers at the end of February and 4,813 by the end of May. The posting cadence went through three phases: March was a cold start, one thread every two days, 15 posts for the month; April entered a high-frequency phase, 180 posts (six per day), which was also the period of fastest follower growth; May settled into a steady rhythm, 74 posts (two to three per day), generating 458,000 impressions and 22,190 engagements at a 5.05% engagement rate. For a tech account with zero manual interaction, that engagement rate reflects content relevance, not operational effort.
Below is a breakdown by system. Each one represents a specific task AI executed autonomously.
This was the first system to go live. Every day, AI automatically pulls the day’s significant AI developments from multiple sources (arXiv, Twitter, Hacker News, WeChat public accounts), selects three to five items with the most signal, and writes a 200-400 character Chinese summary with a brief judgment, then sends it to all subscribers via the Kit API at a scheduled time.
The key here isn’t that AI can write summaries. It’s that it does three things: first, it runs automatically every day without human triggering; second, it summarizes with a specific judgment lens, focusing on engineering practice rather than paper announcements, on concrete mechanisms rather than grand narratives; third, it remembers what it wrote in the previous issue to avoid repeating topics.
I told it what judgment lens to use. It wrapped the Kit API calls, configured the cron job, and ran. After that, I never touched it. I glance at open rates and unsubscribe numbers once a month. In May, daily open rates ranged from 40-55%, click rates from 3-6%, and unsubscribe rates were below 0.1%. These are above-average numbers for newsletters, but more importantly, the whole thing runs on autopilot.
There was one optimization AI discovered on its own. Kit’s subscription form popped up immediately when a page loaded. AI noticed elevated bounce rates in the GA4 data and proposed delaying the form until the user had scrolled through 50% of the page. This change improved subscription conversion by roughly 40%. I didn’t think of it. AI spotted it in the data and proposed the fix. Delayed form triggers are a known industry practice, but the point is AI didn’t need me to know that. It saw the problem, formed a hypothesis, and tested it.
This system has three components: content generation, publishing scheduling, and performance review.
For content generation, AI reads an article from yage.ai, extracts the core arguments, and rewrites them into a five-to-six tweet thread. The first tweet is a hook (a counterintuitive claim or sharp question), the middle three to four develop independent points, and the final tweet carries the article link. All content is in Chinese, with a style directive: rational and direct, no hashtags, no clickbait language. In early March, AI generated 30 threads from the back catalog in one batch. Every new article published since then automatically gets a corresponding thread.
For scheduling, AI queued all 30 threads via the Typefully API, one every two days, at UTC 20:00 (4 AM Beijing time, noon Pacific, the dual-peak overlap window identified from GA4 data). Each new article’s thread is automatically appended to the scheduling queue.
Performance review was the last piece to come online. Initially, only Typefully’s account-level data (total impressions, total engagements) was available, with no per-post visibility. After connecting X’s per-post analytics API in late April, AI could finally compare data at the post level. This is when it proposed a conceptual framework: classify posts into three types. R (Reach), responsible for expanding the audience, measured by impressions and follow growth. T (Traffic), responsible for driving site visits, measured by URL clicks and GA4 campaign sessions. N (Newsletter), responsible for capturing high-intent readers, measured by subscription conversions. This framework shifted the review question from “how did this post perform?” to “did this post accomplish what it was supposed to?”
My contribution here was two things. First, I paid for the services AI recommended: Typefully Creator Plan ($12.50/month) for automated thread scheduling, Twitter Premium to unlock full analytics access. Second, I helped it solve one technical obstacle. X has no public per-post analytics API, but AI needed post-level data for content review. I reverse-engineered X’s backend analytics endpoint and fed it the browser cookie and session token. It then built its own calling logic on top of that. Everything else, AI did: reading articles and writing threads, designing the format specification, setting UTM tagging rules, managing the publishing calendar, analyzing per-post data, proposing the R/T/N framework, and writing the Twitter section of every weekly report.
The biggest data blind spot in early March was that 77% of Direct traffic was unexplainable. GA4 labels all links shared in WeChat, Telegram, and Slack as Direct because these shares carry no UTM parameters. I had no idea where traffic was actually coming from, and AI had no basis for deciding where to invest effort.
Building the attribution system happened in stages. Stage one was
adding UTM tags to all outgoing content. Links in Twitter threads all
carry
utm_source=twitter&utm_medium=thread&utm_campaign=<slug>.
Newsletter links carry utm_source=kit&utm_medium=email.
This separated Twitter traffic from Direct. Verified over several
months: UTM-tagged Twitter traffic (twitter / thread) grew from zero to
10,000+ sessions per month.
Stage two was WeChat attribution. WeChat-shared traffic appears in GA4 as weixin110.qq.com referral, but with bounce rates above 70%, because WeChat’s in-app browser renders external links poorly. AI noticed that while bounce was high, session duration was steadily improving, rising from 30 seconds initially to 80-100 seconds, indicating that the readers who stayed were reading more. This judgment led me to keep sharing links in WeChat groups rather than opening a public account, saving the cost of content moderation and dual-platform maintenance.
AI also took over weekly growth analysis. Every Sunday or Monday, it automatically pulls data from GA4, GSC, Kit, and Typefully and generates a Markdown report covering overall assessment, channel trends, content performance, search changes, Twitter review, and next-step recommendations.
The first report, from March 3, had only GA4 data, ran over 400 lines, and its core findings were ~300 daily users, 77% Direct, negligible search, zero social. By the May 25 report, it covered five data sources, GSC page-level analysis, Twitter per-post R/T/N review, AI referral tracking, and newsletter activation diagnostics. The report evolved from “how much traffic did we get” into a decision system.
AI developed its own judgment language in these reports. It doesn’t say “traffic dropped 50%, needs improvement.” It says “May is a controlled retreat from the April peak, with social cooldown and base stability happening simultaneously.” The difference isn’t rhetorical. It’s cognitive depth. The first statement reports a number change. The second provides a judgment framework. That framework allows the next week’s AI to build on the previous week’s conclusions rather than starting from raw data every time.
AI also handled SEO entirely. I was just the person who nodded at the end.
In early March, AI noticed that robots.txt didn’t explicitly allow AI crawlers, so it added Allow rules for GPTBot, ClaudeBot, and PerplexityBot. The immediate effect was negligible, but looking back, every page that later got cited by AI models originated from this change.
In April, AI spotted a new pattern: claude.ai and chatgpt.com were starting to bring referral traffic. The volume was small, only 100-200 sessions per month, but session durations reached 200-500 seconds with bounce rates below 50%. This was higher-quality traffic than most search referrals. AI judged that AI citation traffic would continue to grow and restructured pages on its own, ensuring every section had a standalone conclusion paragraph that AI tools could easily extract and cite. It applied this change across all subsequent English-language articles.
In May, AI identified two SEO problem pages. ollama-cloud-vs-api-vs-subscriptions-en had 74,000 monthly impressions at position 6.5 but a CTR of only 0.24%. mlx-apple-silicon-en had 58,000 impressions at position 7.1 with a CTR of 0.09%. AI’s diagnosis: Google already considered these pages worth showing to a large audience, but the titles and descriptions weren’t adequately capturing search intent. It rewrote the titles, meta descriptions, and opening summaries for both pages itself. By its estimate, raising the CTR from 0.2% to 1% on these two pages alone would bring in roughly 1,300 additional clicks per month, a higher-leverage move than writing new articles.
Three months of experimentation. Some things were validated. Some were falsified.
What worked.
First, automated distribution matters more than content volume. In March, yage.ai already had 370+ articles but only 2,500 weekly active users. After Twitter auto-distribution launched in April, the same content pool stabilized around 7,000 weekly active users. The problem wasn’t a lack of content. It was that content had no way to reach readers. AI solved the distribution problem.
Second, what AI does in data analysis and pattern recognition has far more leverage than what it does in content creation. Writing a newsletter summary, AI versus human doesn’t make much difference. But finding the signal “Direct traffic has settled to a stable baseline while WeChat share is growing against the trend” buried in hundreds of lines of GA4 data, doing that manually requires monitoring every channel for half an hour each. AI does it by running a script and applying an analytical framework in the prompt. The time and precision aren’t even in the same category.
Third, feedback loops matter more than one-off optimizations. The form delay optimization itself isn’t remarkable. What’s remarkable is that AI noticed the bounce rate anomaly, formed a hypothesis, tested it, and verified the result, all on its own. It didn’t need me to tell it “go check the form conversion.” It saw the problem in the data. Once that loop is established, it can keep optimizing, not just fix things once.
Fourth, attribution gives decisions a foundation. Before UTM tagging, I had no idea which channels brought traffic and had to adjust strategy by feel. With UTM tags, AI could write in the weekly report: “Twitter thread campaigns drove 10,445 sessions, WeChat share grew counter-trend to 3,097.” With those numbers, its recommended actions, adding separate WeChat tracking, categorizing Twitter posts by R/T/N, all had data support.
What didn’t work.
First, English-language content had very low direct traffic ROI. Chinese articles got 15 to 115 times the pageviews of their English counterparts, and English pages had higher bounce rates. But English content has an indirect value: it gets indexed and cited by Google and AI search engines. Sessions arriving from claude.ai had some of the highest engagement durations on the site (500+ seconds). So English content shouldn’t be abandoned, but its positioning needs to shift from “translation for English readers” to “corpus for AI indexing.” This judgment itself came from AI analyzing the data.
Second, using total impressions to evaluate Twitter content quality. Before the R/T/N framework, AI would sort posts by impressions, but this ranking would put high-impression Reach posts without outbound links at the top, while Traffic posts that actually drove site visits ranked lower. Reach posts have broader hooks that trigger algorithmic recommendation more easily, but that doesn’t mean they achieved the growth objective. After switching to R/T/N review in May, content strategy finally aligned with the objective function.
Third, Chinese social platforms had unclear ROI. Reposting content on Zhihu and CSDN brought almost no effective traffic. WeChat public accounts were explicitly evaluated and rejected because maintenance costs and content moderation time exceeded the benefit. Sharing links in WeChat groups, while high-bounce, costs nothing and occasionally triggers organic spread.
This distinction matters because the common narrative is “I used AI to help me write articles.” That’s not what happened here.
What I did falls into four categories.
Category one: I paid. AI recommended I subscribe to Typefully Creator ($12.50/month) for automated thread scheduling, upgrade to Twitter Premium for full analytics access, and maintain the Kit paid plan for newsletter distribution. These are the system’s operating costs, under $50 per month total.
Category two: I helped it break through one technical obstacle. X has no public per-post analytics API, but AI needed per-post data for content-level review. I reverse-engineered X’s backend analytics endpoint and fed it the browser cookie and session token. It built its own calling logic on top and then had data. This was my only technical intervention in the entire system. Everything else, it did on its own.
Category three: topic selection. Every day, AI lists a batch of candidate topics in the daily briefing, usually a dozen or two. I scan the list, pick two or three for it to write, occasionally adding a one-line directional hint like “write from this angle.” This is my entire daily editorial work, under two minutes.
Category four: occasional framework calibration. AI generates judgments and recommendations based on data, but every few weeks I read through the weekly report to confirm its judgment direction hasn’t drifted. This takes about ten minutes each time.
Beyond these four things, AI did everything. It wrote the calling scripts for GA4, GSC, Kit, and Typefully. It set the UTM tagging rules. It designed the thread format specification. It proposed the R/T/N classification framework after analyzing the data itself. It produced every weekly growth report from data pull to final draft.
But the most important thing isn’t the volume of work. It’s that the work forms a closed loop. AI doesn’t execute a task and then stop and wait for the next instruction. It pulls data every week, writes a report, makes next-step recommendations in the report, and then the following week checks whether last week’s recommendations were implemented, whether the data changed after implementation, and whether the change matched expectations. If it did, continue. If not, adjust. The Kit form delay optimization came from exactly this process: it saw elevated bounce, proposed a hypothesis, changed the trigger timing, and then compared subscription conversion rates before and after in the next report. The entire cycle of “spot problem, propose solution, implement change, verify effect” runs without me.
The diagram below captures the full loop.
Doing all this manually would take at least five to ten hours per week, with inconsistent analytical depth and judgment standards. AI’s output is higher in both volume and consistency.
yage.ai is a personal technical blog with a clear reader profile (AI and tech practitioners in the US and China), a stable content production cadence (one long-form article every 12 days, one daily newsletter), and well-defined distribution channels (Twitter, WeChat, Newsletter, Search). These conditions make the operational difficulty relatively low for AI: it’s operating in a system with high determinism and few edge cases.
What happens on a different project with a different category, a more dispersed audience, or more varied content formats? My assessment: as long as distribution can be decomposed into high-determinism subtasks (generating threads, scheduling posts, pulling data, comparing metrics, writing reports), AI will outperform a human on each subtask. AI’s ceiling isn’t determined by how smart it is. It’s determined by how large the toolbox you give it is.
But there’s one thing AI can’t do: it won’t question the analytical framework you give it. If I say “the weekly report header should include a channel comparison table,” AI will execute but won’t ask “is that table really the most important information here.” Framework-level judgment still requires human calibration. This isn’t a model capability limitation. It’s that current agent systems generally lack built-in metacognition. AI can run efficiently within a given framework, but it won’t actively examine whether the framework itself is sound.
Over the past three months, I spent less than ten hours total on this growth system. Mostly two minutes a day picking topics and a check every few weeks to confirm the direction hadn’t drifted. Everything else — writing scripts, pulling data, publishing content, doing attribution, writing reports — AI executed automatically. In terms of input-to-output ratio, this is the most efficient of all my AI automation projects.
It’s not glamorous. No frontend interface, no product logo. It’s a bunch of Python scripts that AI wrote, a few cron jobs, a collection of prompts, and a lot of data flowing through APIs. But over three months, it grew a blog from 2,500 to 7,000 weekly active users, built a Twitter account and newsletter from zero, and went from having no idea about channel attribution to producing a multi-source cross-analysis report every week. All I did were four things: paid, helped it reverse-engineer one API, picked topics daily, and checked direction occasionally. AI did the rest.