A Threads post nobody hit "send" on drew 99 comments: which district in New Taipei is the best place to live?

At 8:30 p.m. on October 27, our Threads account published this post:

Which district in New Taipei is the best place to live?

Banqiao / Zhonghe-Yonghe / Xinzhuang / Xindian / Sanchong

Locals, speak up

Nobody hit "send." A scheduler picked it from our question bank and posted it automatically.

Two days later, it had 24,974 views, 99 comments, and 95 unique participants. We pulled every comment back, read each one, and turned them into the table below.

The results first

Of the 95 people who commented, 83 clearly backed a district. One vote per person, counting the first district they named:

The five votes outside the listed options: Xizhi 2, North Xinzhuang 2, Sanxia 1. Another 12 people didn't pick a side—mostly "none of them are great," "depends on what you need," or just a "+1."

A few things that stood out:

• Banqiao is both the most recommended and the most criticized district.The case for it came down to one word: convenience. A three-line rail hub, department stores and cinemas, "more convenient than some districts in Taipei." The complaints were just as specific: traffic jams, MRT trains too packed to board, roadwork everywhere, and the annual Christmasland festival—three comments brought it up, and one suggested splitting Banqiao into two options: "with Christmasland" and "without Christmasland."
• Xindian runs on a different set of values.Xindian fans talked about quiet, mountains and rivers, and being just a bridge away from Taipei City. Everyone agreed on its one drawback: humidity. As one put it, "Banqiao is sunny and Xindian is still raining—two completely different worlds."
• Within Zhonghe-Yonghe, Yonghe beats Zhonghe.Yonghe was singled out in 9 comments, Zhonghe in 4—and one person made a point of saying "definitely not Zhonghe." Nearly everyone who backed Yonghe added, "but it's way too crowded."
• Half of Sanchong's votes were reverse recommendations from locals.Our favorite: "Sanchong resident here. I recommend everyone pick another district. Why? I'd like fewer people around." We counted it as a vote for Sanchong—that's love.
• People's priorities are clear.Tallying every reason mentioned across the comments: transit convenience 19 times, crowding 8, weather 6, redevelopment zones 6, safety and neighbors 4, and food, housing prices, and quiet 3 each. Transit alone outnumbered the next two combined.

I didn't publish this post

This account posts three quiz questions every day to drive traffic to a quiz game we built. The questions live in a single question-bank file—389 so far—in two types:

• Quick-answer questions: trivia with a correct answer, like "What was Yangmingshan called before it was renamed in 1950?"
• Debate questions: questions with no right answer that make people want to pick a side, like this one.

Every evening at 6:30, 8:30, and 10:00, a scheduler (a program set to run automatically at fixed times) picks the next question from the bank and posts it.

The posting program runs in three steps: pick a question and write the content into a staging file, post from that file, then immediately record "this question has been used."

(Why split it up? If picking and posting happen in the same run, any failure midway means the next retry picks a new question—so the version you previewed and the version that actually goes out might not be the same post. Putting a file in between locks the content in the moment it's written.)

The system publishes 90 posts a month. The time we spend on it is mostly topping up the question bank every few weeks.

The real value of automation is in the data coming back

Automating posts is cheap; there are tutorials everywhere. What few people build is the other half:whether the data comes back automatically after you post.

We have a separate program that runs every morning at 6:10, pulling views, likes, and comment counts for every post on the account into a database. Once a post is three days old, its numbers are frozen. That gives us 375 posts from the same series to compare against.

Against that baseline, this post's position is clear:

• For posts in this series,the median view count is 521(line up every post by views; the median is the one right in the middle). This post got 48 times that.
• Debate questions actually underperform quick-answer questions on average: their median is just 258 views, andmore than half of debate questions got zero comments.。
• This post's 83 genuine comments rank second out of the account's 1,000-plus posts.

("Genuine comments" is a metric we defined ourselves: only top-level comments from other people, excluding our own. The platform's comment count includes the links we post ourselves, which makes low-view posts look busier than they are.)

In other words, this post didn't take off because "debate questions perform better." Debate questions are weaker overall; this one is the outlier. Without those 375 posts as a baseline, we could easily have drawn the wrong conclusion from this single post and switched the whole question bank to debate questions.

Automated posting gives you volume. Automated data collection gives you judgment. Do only the first half, and you're just shooting in the dark more efficiently.

Turning 99 comments into a table

The official Threads API (an interface that lets programs exchange data directly with the platform) has an endpoint that returns an entire conversation in one call, including second-level replies. All 99 comments, retrieved at once—no browser, no copying them out one by one.

The hard part is the next step:interpretation.

The intuitive approach is keyword matching—if a comment contains "Banqiao," count it as a vote for Banqiao. That goes badly wrong:

• "If you love food, skip Banqiao": mentions Banqiao, but it's not a recommendation.
• "Sanchong resident here, I recommend everyone pick another district": literally not a recommendation, actually a recommendation.
• "Neihu Township, Taipei County 😎": a joke—Neihu is in Taipei City.
• "North Xinzhuang": could mean the northern part of Xinzhuang, or Beixinzhuang in Sanzhi—people use it both ways.
• "Xindian > Zhonghe-Yonghe > Xinzhuang > Banqiao > Sanchong": one person ranked all five—how many votes is that?

So we set the rules first, then had AI interpret each comment:

1. One vote per person: if someone left two comments, they count once.
2. First pick only: for rankings, take the top choice; separately, we tallied total mentions as a recommendation for reference.
3. Criticism counted separately: comments that clearly called out a district's downsides were logged as "criticized," without subtracting from its votes.
4. Zhonghe, Yonghe, and Zhonghe-Yonghe merged, since that's how the original question grouped them, with the breakdown kept separately.
5. Jokes, off-topic replies, and non-answers don't count as votes, but the original text is kept so a human can check back.

Once interpreted, every comment traces back to its original text. The summary table is just the last step—the comment-by-comment record before it is what can actually be audited.

A few things to be honest about

• This isn't a poll.Commenters chose to comment; they weren't randomly sampled. People who scroll past a quiz question at 8:30 p.m. and bother to type a reply don't represent New Taipei residents.
• The options steer the answers.The question listed five districts, and only 5 of 83 votes went outside them. Had it listed Xizhi, Linkou, and Tamsui, the results would look completely different.
• AI interpretation has gray areas.We didn't force the two "North Xinzhuang" votes into Xinzhuang, and we classified "just passing through the Xingfu redevelopment zone" as no answer. Someone else might land a vote or two differently, but the ranking wouldn't change.

What this confirmed for us

People in social media always ask: how do you write a post that goes viral?

Our answer is increasingly this:You can't write one. But you can make it more likely to happen—and know when it does.

Whether a post takes off is hard to predict in advance. Half of the debate questions in our bank didn't get a single comment. What you can do is make posting cheap enough to do every day, have the data come back on its own, make sure the few posts that do take off get noticed, and then treat the comments under them as data, not just buzz.

Those 99 comments are 95 people taking the time to tell you what they care about: transit, crowding, weather, Christmasland. That's hard to get out of a survey.