Nine observations on the Threads algorithm

Experiments Published 2026-11-20 · · 11 min read

Five months of automated posting, more than 1,200 posts, 2.94 million views. We cross-checked Threads' official insights, GA4, and our own database: the top 10% of posts took 86% of views, posting more diluted reach, forcing comments didn't work, and only 1 to 2 people per thousand views clicked through to a website.

Hundreds of thin copper-colored lines fan out from the left, most breaking off midway, with only a few continuing to the right, one of which grows thicker and turns green

TL;DR

  • In mid-June, I connected my own Threads account to an automated schedule. By 11/19 it had reached about 2.94 million viewsand more than 1,200 posts. In the three years before automation, most months came in under 5,000 views
  • But 88% of views came from a single line(the quiz bait posts that funnel traffic to the mobile quiz game AQ ). The other nine scheduled lines combined made up less than 10%
  • Posting volume went from 20–30 posts a week to 70–85,and total views actually fell by 40%
  • Views, followers, and site visits are three different things, each driven by different content, with almost no crossover. "The Post-KMT Era" drew 81,000 views and 806 shares, yet brought only 33 people to the website
  • Out of 2.94 million views, about 3,800 actually clicked through to our various websites.For every thousand views, 1 to 2 people
  • The not-so-flattering conclusion: for me right now, Threads is "a free intake for one product" plus "a megaphone for personal opinions". It is not yet a business channel for the company

In August I wrote a piece called "After writing four insider pieces about ad agencies, I put them on Threads and watched the data." Back then the sample was 39 posts.

Three months on, the sample has grown to more than 1,200 posts, and the scheduled lines went from two to ten, with four shut down along the way. This time I re-checked the numbers across three sources: Threads' official insights API, GA4, and our own product database.

To be clear about what kind of account this is: my personal @saominhuang, with 330 followers. Not a big account. Every conclusion below only represents what "one small account running automated posting for five months" saw.


The scorecard

Account-level numbers from Threads' official insights:

Month Views Likes Reposts + quotes
June (automated from 6/14) 167,000 816 2
July 500,000 2,171 12
August 606,000 1,862 33
September 345,000 712 8
October 330,000 1,093 11
November (through 11/19) 996,000 6,810 430

Broken down by line:

Line Posts Total views Likes Shares
AQ quiz bait 446 2 million 8,009 496
Posted manually 207 136,000 2,999 837
Blog carousel 171 61,000 504 156
Just Eat This (就吃這)(discontinued) 77 30,000 33 2
Satsuma service ads (discontinued) 86 20,000 101 23
Five-Star Dungeon 81 16,000 44 4
Relative Tarot 88 14,000 322 1
Sim Taipei election bulletin (discontinued) 35 6,000 3 0
Taipei 456: I Am the Mayor of Taipei(discontinued) 64 3,000 3 0

First glance: one line of bait posts carries the whole show. Second glance: the 446 bait posts add up to 496 shares. My manually posted "The Post-KMT Era" got 806 on its own.


Nine observations

I've labeled each one with its strength of evidence. The ones marked "backed by data" I'm willing to stand behind. The ones marked "correlated" are patterns I saw without ruling out other causes.

Backed by data

1. Winner takes all. The top 1% of posts took 34% of views, and the top 10% took 86%. 38% of posts didn't even reach 100 views. On Threads, averages are meaningless. Results are decided by a handful of posts.

2. A post finishes growing within 48 hours. After that it barely moves. So there's no need to wait a week to judge performance. It's settled in three days.

3. Comments grow alongside reach, but questions "designed for comments" don't get pushed because of it. My bait posts come in two types: quick-answer ("What is Taiwan's second-highest mountain?") and pick-a-side (getting people to choose a side and fight it out). Through the end of October, pick-a-side posts had 6 times the comment rate of quick-answer posts, yet only a third of the median views: 271 versus 875. Of the eighty-odd questions along the lines of "savory or sweet tangyuan," the vast majority stayed under 1,000 views. Conversely, the most-viewed post (sun cake filling, 88,000 views) had only 13 comments from start to finish. Hardly anyone commented, and it got pushed anyway.

But November brought one set of exceptions: "Which district is the best to live in?"

Question Views Comments from others Shares
Where's the best place to live in Hsinchu 68,000 89 196
Which district of Tainan is the best to live in 40,000 143 41
Which district of New Taipei is the best to live in 27,000 98 10
Which district of Kaohsiung is the best to live in 11,000 82 16
Which district of Taoyuan is the best to live in 1,506 5 0
Which district of Taichung is the best to live in 985 5 0

Same sentence pattern: the ones with comments shot up, and the ones without stayed put. I save a growth curve every 15 minutes, and this is how the Hsinchu post grew: 4 hours after posting, 4,114 views, 7 comments, 4 shares; at 8 hours, 19,000 views, 46 comments, 63 shares; at 24 hours, 72,000 views, 142 comments, 206 shares (comments here include the replies underneath, so the figure is higher than in the table). Comments, shares, and views all took off in the same window, and you can't tell which led which.

How I read it now: comments are the result of "people care about this question," not a switch you can force. Tangyuan doesn't start a real argument, but where to live does. The difference is that the comments on the latter are useful to other people, and someone will forward the whole thread to a friend who's house-hunting. What sets the Hsinchu post apart from the other questions most is its share count, not its comment count.

4. Shares and reposts are what's tied to follower growth. Opinion pieces get more than 20 times as many shares per thousand views as bait posts. Every noticeable jump in followers happened the day after a highly shared post. The biggest was +32 in a single day, against a normal rate of less than 1 a day.

5. Threads keeps people on the platform. For every thousand views, only 1 to 2 people click a link to leave. People finish reading long posts right on Threads. "The Post-KMT Era" got 81,000 views on Threads, while the blog version of the same piece got only 33 visits.

6. Questions that call out a specific group are more likely to get pushed. The same kind of opener keeps showing up in posts that pass 10,000 views: "Tainan people, come out," "Hsinchu Science Park people," "Only true fans can answer this." Name an identity, and that group shows up to claim it.

Correlated, but other causes not ruled out

7. Posting more dilutes reach. As daily posts went from 3 to 11, the median views of bait posts fell from 877 to 356, and the account's total views dropped by 40%. Weaker questions may have been a factor at the same time, and I can't separate the two. But at the very least, "post more, get more exposure" did not happen.

8. One viral post lifts the whole account. In the two weeks after "The Post-KMT Era," the weekly median for bait posts rose from 300–500 to 1,323, then to 3,336, and in a single week 9 of 21 posts passed 10,000 views. But I also changed question types at the same time (from knowledge quizzes to pick-a-side questions like "Which district is the best to live in"), so this one only counts as an observation.

9. Time of day matters less than you'd think. The three evening slots were virtually identical. 8:30 a.m. did slightly better than 11:00, with the gap coming mainly from the top few posts. I ran a time-slot experiment for two months, and the conclusion is that it's not worth changing the schedule for.


Three different things

Compress five months of numbers into one table and you get this:

Desired outcome What drives it Evidence
Views AQ quiz bait 2 million views
Followers, reposts Opinion pieces written by hand "The Post-KMT Era": 806 shares, +32 followers the next day
Site visits Content where "there's something there the moment you click" See below

The three barely cross over.

  • The bait posts that bring views bring almost no followers.
  • The long posts that bring followers bring almost no site visits.
  • The most efficient driver of site visits is one of the lowest-view lines: the dungeon logs from Five-Star Dungeon. Its 81 posts totaled just 16,000 views, yet brought 265 people to the site to try the game, staying an average of 55 seconds. That's about 16 people per thousand views, ten times the overall average. The reason isn't hard to guess: what you click into is something you can play right away, not another article to read.

In the August piece I wrote, "Views are traffic, followers are an asset." Now I need to add one more line:Site visits are a third kind of thing, and they have little to do with the other two.


Traffic that actually reached the sites

Visits with "Threads as the source" in GA4 (6/1–11/19):

Site Visits Share of site traffic
Satsuma website 1,886 About 20%
AQ(grabquiz.com) 1,700 About 50%
Relative Tarot 249 About 10%

Of the 1,886 visits to the company website, 543 came from the August methodology piece alone. Take that out, and it holds steady at around 300 a month. Of those, visits that reached "Services"-related pages total about 250 over five months.

That's the not-so-flattering conclusion. 2.94 million views sounds like a lot, but when it comes down to "someone came to see what we sell," it's 250.


The lines I shut down, and why I started them

I shut down four in five months. Looking at the data afterward, all of them deserved it:

  • Just Eat This (就吃這): 77 posts, 30,000 views, site visits: 2。
  • Taipei 456: median of 18 to 59 views.
  • Sim Taipei election bulletin: 35 posts, 3 likes.
  • Satsuma service ads: 86 posts, about 150 views each, 101 likes in total.

What they had in common: they were all "what I wanted to say," not "what someone wanted to see." Scheduling makes posting free, so it's easy to keep opening new lines. But an account's exposure is limited, and every extra line nobody reads is competing for space with the one people do read (see observation 7).

Automation saves the effort of posting. It doesn't save the judgment of whether a post is worth publishing.


Limitations of this data

  • This is an account with 330 followers, and it is the only sample.
  • Observations 7 and 8 are correlation, not causation. Verifying them would take a reduced-volume test, which I haven't done yet.
  • GA4 was only installed on the Satsuma website and AQ in late July and August, so referral traffic for June and July isn't visible.
  • "From Threads" only counts visits where GA4 can identify the source. Threads' in-app browser sometimes strips the source, so the real numbers may be somewhat higher.

FAQ

Does automated posting on Threads get your reach reduced? My data shows no sign that "automated posting" itself is penalized: for the same scheduled bait posts, November's median was the highest in five months. What does make a difference is posting volume. As the number of posts per day goes up, the median per post goes down.

How many posts a day is best on Threads? There's no universal answer. On my account, the median per post was highest at 3 posts a day. After going up to 11, it fell by more than half, and total views didn't increase either. For small accounts, I'd suggest starting low and adding gradually, watching whether the median drops.

Do links in Threads posts reduce reach? I always put links in the comments and never ran a comparison with links in the main post, so I can't answer this one. What I can say for sure is that the click-through rate is very low: 1 to 2 clicks per thousand views.

Can Threads bring clients to a company? Based on my five months of numbers: not yet. What it brings is views and a small number of followers, and very few people reach the services pages. To turn it into a business channel, the content has to shift from "ads" to "opinions with a stance" plus "things people can try right away," and it has to be measured in months.

Does chasing comment counts work? Forcing it doesn't work. I posted eighty-odd pick-a-side questions designed specifically to start arguments. Their comment rate was 6 times that of regular questions, but their views were only a third. The most-viewed post, at 88,000 views, had only 13 comments. But comments aren't meaningless: questions that drew 4 or more comments within two hours of posting ended up with a median view count roughly six to nine times that of other questions (the sample is only 36 posts). The difference is whether the comments grew on their own or were forced by the question. What gets pushed are questions where "the comments are useful to other people," such as "Which district is the best to live in," which also bring in a large number of shares.