Skip to content
Analyst work

Data study · October 2026

TikTok STEM feed

Question

Why does TikTok’s US STEM feed trail Douyin in China, and what would make it work?

Headline finding

Douyin users each watched about 959 science videos in a year. TikTok’s STEM tab drew 0.84 views a year for each user it reached, less than one video each. The two counts cover different scopes. My reading is that where the feed sits and how teen limits are enforced drive the gap.

My recommendation

Make a capped teen mode the home screen for verified teens, and pay for more STEM videos to fill it. Speed up review and publish the minutes moved into STEM.

Tools
  • CSV tables with a source URL on every row
  • JavaScript on Node for the derived rates
  • D3.js for the first dashboard
  • React and SVG for the charts on this page

The question

The client

The team that runs TikTok’s STEM feed in the US. TikTok announced the feed in March 2023, and in January 2026 the TikTok USDS Joint Venture was formed to run and retrain the US algorithm.

The decision they face

What should change so the feed reaches US teens? The candidates were where the feed sits, how teen settings are enforced, how much STEM content exists and how success gets measured.

The data, 3 tables

  • metrics.csv37 rows

    Every figure on TikTok’s STEM feed, Douyin’s science content, teen use and time limits. Each row has its market, period, scope and a flag for whether the platform reported it.

    Where it came from
    Press coverage of TikTok and Douyin announcements, TikTok Newsroom, Pew Research Center, a CMMR survey reported by Gigazine and press reports of the Kentucky attorney general’s filing.
    Why I chose it
    Each figure gets an ID and a scope, so a count for one tab can’t pass for a count across the whole app.
    What it cannot tell us
    A like-for-like rate. TikTok publishes no yearly For You STEM views. 22 of the 37 figures come from the platforms themselves, and 8 rows reach the table through a second outlet.
  • events.csv14 rows

    Dated events from 2021 to 2026 on China’s minor-protection rules, Douyin’s youth mode, TikTok’s STEM feed and what US lawmakers and courts did about teen use.

    Where it came from
    Press coverage, plus TikTok’s own newsroom.
    Why I chose it
    The launch and report dates set the window behind finding 1, and the Texas law near the end is what option A rests on.
    What it cannot tell us
    Cause. TikTok announced the feed 9 days before its CEO’s House hearing, and the dates alone can’t say why. 3 of the rows are dated to the month.
  • variables.csv9 rows

    Policy and product variables compared across Douyin and TikTok US, each scored 0 to 2 for how far it pushes users toward science content by default.

    Where it came from
    My own coding, with the row IDs or source names each score rests on.
    Why I chose it
    Each option targets one of the gaps in this table.
    What it cannot tell us
    Stand as measurement. The scores are judgment. 7 of the 25 citations in the evidence column name a source with no row in the other two files, so a reader can’t open them from the data.

The process

  1. One row per figure

    Each figure and each event went into a row of its own with a source name and a URL. Metric rows also carry a market, a period, a scope and a flag for whether the platform reported the number. That came to 60 rows across 3 files.

  2. Scope

    TikTok’s count covers the STEM tab alone. Douyin’s covers science views anywhere in the app. I kept both and put the scope on every row, so the mismatch stays visible wherever the two numbers meet.

  3. Join

    TikTok gave its 200 million tab views as a running total. I joined the launch and report dates in events.csv (rows E05 and E11) to get a window, stored it as 19 months in row T14 and annualized the views over it.

  4. Assumptions

    Views spread evenly across the window. The window starts on March 14, 2023, the day TikTok announced the feed. Douyin’s monthly users from Feb 2024 stand in for its whole year of views. The age band that ends at 19 includes two years of adults.

  5. Coding

    I scored 9 variables from 0 to 2 for each platform, where 2 pushes users toward science content by default. The gap is Douyin’s score minus TikTok’s. This table is judgment.

  6. Check

    I added three checks after the dashboard was done. Douyin’s daily figure times 365 comes out 0.015% under its yearly figure. The launch and report dates sit 589 days apart, which is 19.4 months against the 19 I stored, and the exact window moves the tab rate from 0.84 to 0.83. The third check splits the coding table’s evidence column and is shown below.

Three evidence cells before and after the split.

Before, three rows of variables.csv (four of six columns shown)
variable,douyin_score,tiktok_score,evidence_ids
Hard time cap,2,0,"D16; T11; T12"
Age signal,2,1,"Sigma; Wiley Rein"
Can the user leave,2,0,"Pandaily; TikTok Support"
After, one entry per variable
{
  "variable": "Hard time cap",
  "rows": ["D16","T11","T12"],
  "named": []
}
{
  "variable": "Age signal",
  "rows": [],
  "named": [{"name":"Sigma","has_row":false},{"name":"Wiley Rein","has_row":true}]
}
{
  "variable": "Can the user leave",
  "rows": [],
  "named": [{"name":"Pandaily","has_row":false},{"name":"TikTok Support","has_row":false}]
}

One cell can mix row IDs with source names. After the split, a citation that matches a row ID lands in “rows”. Any other citation is a name, flagged by whether some row cites that source. 16 of the 25 citations are row IDs.

The code for the key steps

The rate behind finding 1, for the STEM tab and for Douyin. derive.mjs
export function reach(metrics) {
  const byId = pick(metrics);
  const tt = byId("T05") / byId("T10") / (byId("T14") / 12);
  const dy = byId("D01") / byId("D12");
  return [
    { lens: "TikTok STEM tab (feed only)", views_per_user_per_year: tt, ratio_to_tiktok: 1 },
    { lens: "Douyin natural science (all feeds)", views_per_user_per_year: dy, ratio_to_tiktok: dy / tt },
  ];
}
The split of the coding table’s evidence column into row IDs and source names. derive.mjs
export function evidenceCheck(variables, metrics, events) {
  const ids = new Set([...metrics.map((r) => r.metric_id), ...events.map((r) => r.event_id)]);
  const names = [...metrics, ...events].map((r) => r.source_name);
  return variables.map((v) => {
    const cited = v.evidence_ids.split(";").map((s) => s.trim()).filter(Boolean);
    return {
      variable: v.variable,
      rows: cited.filter((c) => ids.has(c)),
      named: cited.filter((c) => !ids.has(c)).map((name) => ({ name, has_row: names.some((n) => n.includes(name)) })),
    };
  });
}

Dead ends

  • No yearly For You figure

    TikTok does not publish yearly For You STEM views, so I could not set an all-feed TikTok year beside Douyin’s yearly count. The closest figure is 110 billion hashtag views counted up to the tab’s launch with no start date. Finding 1 uses the tab alone.

  • A mixed denominator

    The 200 million views cover the US, UK and Ireland up to the report date. The 150 million users reached come from a later statement about Europe and the Americas. The tables hold no user count for the same markets and dates, so the tab rate divides one scope by another.

  • Sources with no row

    9 citations in the coding table are source names where the rest are row IDs, and 7 of those names match no row in the other two files. I have not added rows for them, so those scores can’t be traced from the data on this page.

The 3 findings that change the decision

  1. Finding 1

    0.84

    STEM tab views a year for each user it reached, against about 959 per Douyin user

    TikTok’s STEM tab drew 0.84 views a year for each user it reached, and Douyin users each watched about 959 science videos across the whole app.

    Dataset metrics.csv
    Query derive.mjs, reach()

    The STEM tab drew 0.84 views a year for each user it reached

    The scopes differ. Douyin’s bar counts science views anywhere in the app, divided by monthly users. TikTok’s counts the tab alone, divided by users the tab reached and annualized over 19 months. TikTok publishes no yearly For You STEM views.

    Science views per user per year

    Source: metrics.csv, derive.mjs, reach().

    Show the numbers as a table
    Platform and scopeScience views per user per yearViewsDivided bySources
    TikTok STEM tab, feed only0.84200 million in 19 months150 million users reachedTechCrunch and Eastleigh Voice
    Douyin natural science, all feeds958.81723.9 billion in a year755 million monthly usersXinhua (Douyin 2024 Natural Science Report) and Statista via RSIS review
  2. Finding 2

    4.5%

    Douyin users aged 4 to 19, CMMR survey, 2021

    Users aged 19 or under were 4.5% of Douyin’s audience in 2021. Youth mode serves only a small slice of the people behind its 723.9 billion science views.

    Dataset metrics.csv
    Query derive.mjs, audience()

    Users aged 19 or under are 4.5% of Douyin’s audience

    CMMR survey figures for 2021, reported by Gigazine. Youth mode locks in verified users under 14, who sit inside the first two bars. The third bar is what remains after those two.

    Share of Douyin users (%)

    Source: metrics.csv, derive.mjs, audience().

    Show the numbers as a table
    Age bandShare of Douyin users (%)Where it comes from
    Aged 4 to 120.34%Row D14, CMMR survey
    Aged 13 to 194.18%Row D15, CMMR survey
    Everyone else95.48%100 minus rows D14 and D15
  3. Finding 3

    1.4%

    Drop in teen minutes a day, 108.5 to 107

    TikTok’s 60-minute prompt cut average teen use from 108.5 to 107 minutes a day, a 1.4% drop, while Douyin locks verified under-14s out at 40 minutes.

    Dataset metrics.csv
    Query derive.mjs, limitsCalc()

    The 60-minute prompt cut teen use by 1.4%

    The two averages come from a TikTok internal test described in the Kentucky attorney general’s filing. The two limits are settings, shown for scale. The filing says TikTok planned to revisit the design if use fell by more than 10%.

    Minutes per day

    Source: metrics.csv, derive.mjs, limitsCalc(), derive.mjs, limitsRows().

    Show the numbers as a table
    MeasureMinutes per dayWhat it isSource
    Teen average before the prompt108.5Measured in TikTok’s internal testKentucky AG filing via NPR/AP
    Teen average with the prompt107.0Measured in TikTok’s internal testKentucky AG filing via NPR/AP
    TikTok’s default limit, under 1860.0A setting that a passcode extendsEngadget
    Douyin’s youth-mode cap, under 1440.0A setting the user cannot liftEngadget / Bloomberg

My recommendation

Give verified teens a capped STEM home screen, and publish how many minutes it moves into STEM.

  1. A

    Verified teen mode as the home screen

    Use the age category that app stores pass to developers under Texas’s App Store Accountability Act. Verified teens land on a curated mode with a hard daily cap that needs a parent to lift.

    Evidence
    Finding 3. A prompt teens can extend with a passcode cut teen use 1.4%, and Douyin holds verified under-14s to 40 minutes. On my 0-to-2 scale for a hard cap, Douyin scores 2 and TikTok scores 0.
    Main risk
    The Texas law took effect in January 2026 while under appeal. If the appeal succeeds, the verified age signal goes with it.
  2. B

    Pay for more STEM videos

    Add a STEM multiplier to creator payouts and give verified accounts to US museums and universities.

    Evidence
    Douyin reports 856 verified museum accounts and 101 academicians presenting through science museums. TikTok reports 7,000 STEM creators worldwide and no count of institutions.
    Main risk
    Supply may not be what holds the tab back. TikTok counts almost 15 million STEM videos published in the 3 years to Feb 2025, and the tab still drew 0.84 views a year for each user it reached.
  3. C

    Speed up review

    Pre-approve credentialed creators so their uploads skip the two-partner review queue, with spot checks after posting.

    Evidence
    Every STEM feed video has to pass review by Common Sense Networks and Poynter, and Douyin’s reviewers curate a youth pool of 543,600 videos.
    Main risk
    Review is the one factor I scored the same for both platforms, 1 and 1 on my 0-to-2 scale. By my own scoring, faster review closes no gap.
  4. D

    Change the scorecard

    Measure minutes moved into STEM content and publish the figure.

    Evidence
    Finding 3. The Kentucky filing describes TikTok judging its teen time tools by press coverage and public trust.
    Main risk
    TikTok would report the figure itself. 10 of the 21 TikTok figures in this study already come from the company.
  • The teen mode puts verified teens on a home screen with a daily cap only a parent can lift.
  • Paying for STEM videos fills that screen. Faster review lets credentialed creators post sooner, though by my own scoring it closes no gap on its own.
  • The scorecard tells TikTok whether any of it works. Today it judges its teen time tools by press coverage.

Limits and next steps

What the data cannot support

  • The two rates in finding 1 cover different scopes. Douyin’s counts science views anywhere in the app and TikTok’s counts one tab.
  • The TikTok rate divides views in the US, UK and Ireland by users reached across Europe and the Americas, from two statements made at different times.
  • 22 of the 37 metrics come from the platforms themselves. 8 rows cite their source through a second outlet, among them the Kentucky filing and the CMMR survey.
  • Douyin’s age split dates from 2021 and its youth-mode pool size from 2020.
  • “Trails” means the tab fell short of Douyin’s reach. TikTok’s own figures for the tab read well, with 25% of visitors back the next day and 96% viewer satisfaction (rows T03, T04).
  • Nothing here follows one teen from the tab to a science video. Placement and enforcement are my reading of platform totals and a table I coded by hand.

What I would test with more access

  • With TikTok’s Research API I would pull view counts by hashtag and year, which would put a For You figure beside Douyin’s.
  • I would pull Google Trends search interest in “TikTok STEM” as a signal of demand.
  • With TikTok’s own data I would measure minutes moved into STEM content per teen, the number option D asks TikTok to publish.
  • I would add a row and a URL for every source the coding table names, and read the Kentucky filing itself.

TL;DR

  1. Studied

    How TikTok’s US STEM feed compares with Douyin’s science content, and what would close the gap.

  2. Found

    TikTok’s STEM tab drew 0.84 views a year for each user it reached, against about 959 science views per Douyin user.

  3. By the numbers

    STEM tab views a year per user reached
    0.84
    Douyin users aged 4 to 19, 2021
    4.5%
    Drop in teen minutes from the 60-minute prompt
    1.4%
  4. The catch

    The two rates cover different scopes. Douyin’s counts the whole app, and TikTok’s counts one tab.

  5. Next

    With TikTok’s Research API, pull view counts by hashtag and year to set a For You figure beside Douyin’s.