What Theia learned, week 2: 9 AI TikToks, 7 formats, one view ceiling

Every week we publish what Theia, our AI content studio, learned from her own films. She runs our TikTok account on auto-post: she picks the idea, makes the film, posts it, reads the analytics and changes her own code. Each film records the exact version of the code that made it, so every number below points back to one version.

This week covers the 9 films posted from 22 to 30 October. Last week's post is What Theia learned, week 1. For the whole story from the first slideshows, read How Theia learned to make TikToks. The numbers are real and unedited, including the bad ones.

The week in short

	This week
Films posted	9
Formats tried	6 (4 new, 2 climbed)
Views at 6 hours	224 to 250 on every film
Likes per 100 views	0.4 to 6.8
Best 2-second hold	58% (the drafts)
Shares	5, on 3 films
New followers from one film	2 (the lipstick)

The films

"Held at 2 s" is the share of viewers still watching two seconds in. "Code" is the version of Theia's code that made the film. Views and likes are the newest read, 1 to 8 days after posting.

Film	Posted	Format	Views	Likes	Held at 2 s	Avg watch	Code
The drafts	22 Oct	Phone screen, no face	273	18	58%	6.8 s	e305f17
Over him	23 Oct	Street vox pop	256	11	45%	6.3 s	0439ea5
The hoodie	24 Oct	Photo film	257	6	41%	3.7 s	5319e7b
The plant	25 Oct	Voice note over a walk	237	1	27%	3.4 s	5ca882d
The block	26 Oct	One take, a friend filming	254	6	30%	3.8 s	979046e
The lipstick	27 Oct	Selfie confession	252	13	41%	4.6 s	5ed27b2
The laugh	28 Oct	One take, a friend filming	235	16	41%	3.9 s	dff768d
The date	29 Oct	Selfie confession	251	15	not read yet	not read yet	af50207
The Netflix text	30 Oct	One take, a friend filming	250	4	not read yet	not read yet	6719667

The first four films finished the exploration week that started on 19 October. The last five climb the two formats that won it.

Seven formats, one ceiling

Over two weeks Theia tried seven very different formats, each with a written reason. Every one stopped at TikTok's first batch of viewers, about 230 to 250 views. None went past about 320 later.

For three of the films (over him, the hoodie and the plant), someone on our team liked and engaged with the post by hand in its first hour. Those three did no better on views than the four films before them, and worse on likes.

So on views, this week says nothing about format. What did change from film to film is who stayed and who liked.

ExplorationClimb

The drafts

250

Over him

241

The hoodie

238

The plant

230

The block

231

The lipstick

225

The laugh

224

The date

227

The Netflix text

240

Every film landed in the same band, whatever the format.

Engagement rate, film by film

We count TikTok engagement rate as likes, comments, shares and saves, divided by views, per film.

ExplorationClimb

The drafts

7.3%

Over him

5.1%

The hoodie

2.7%

The plant

0.4%

The block

2.8%

The lipstick

6.7%

The laugh

7.2%

The date

6.0%

The Netflix text

1.6%

Likes, comments, shares and saves per 100 views, newest read.

Likes are still most of it. But this week also brought the first shares on Theia's films since August: 1 on the drafts, 1 on the hoodie and 3 on the lipstick. The lipstick also brought 2 new followers.

What she changed this week

She wrote a verdict on the seven formats. After the seventh film had its 6-hour read, Theia ranked the formats on what moved: the hold, the finish and the likes. Her reading of it:

• A person speaking in real time holds this audience. The three exploration films most often watched to the end all had someone talking on camera.
• Stills and invented pictures do not. The photo film held 41% at 2 seconds and only 6% watched to the end. The voice note held 27%, the worst of the week.
• The likes go to a feeling the viewer has had herself: typing and deleting a text to an ex, keeping his toothbrush.
• She added one question to ask of every new format: does this video exist in real life, and who filmed it? Nobody films a silent selfie to go with a voice note, so that film's picture was invented.

She stopped exploring and started climbing two formats. The first is one take, a friend filming: the format with the best finish in week 1. The second is the selfie confession: the best like rate, and the only format that ever reached about 1,900 views on this account. Each climb film opens on a feeling the viewer has had, said in the first line.

She started posting every film to Instagram Reels too. The question was whether the ceiling is the films or the TikTok account. The same film on a second platform splits the two.

What the data supports

1. The ceiling is not only the TikTok account. Since 26 October each film also went to Instagram Reels. On all five, Instagram had fewer views than TikTok at the same point.

Film	TikTok at 6 h	Instagram, read at 4.5 to 5.7 h
The block	231	64
The lipstick	225	155
The laugh	224	130
The date	227	161
The Netflix text	240	62

On the four films with a read, Instagram viewers watched for longer, 5.9 to 7.0 seconds on average, but left almost no likes. Five films is not proof, and the two apps count views differently.

2. The climbs lifted likes, not reach. Three of the five climb films got 5 or more likes per 100 views: the laugh (6.8), the date (6.0) and the lipstick (5.2). The block and the Netflix text did not (2.4 and 1.6). Views stayed in the same band either way.

ExplorationClimb

The drafts

6.6%

Over him

4.3%

The hoodie

2.3%

The plant

0.4%

The block

2.4%

The lipstick

5.2%

The laugh

6.8%

The date

6.0%

The Netflix text

1.6%

3. The first two seconds are stuck. The lipstick and the laugh both held 41% at 2 seconds, and the block held 30%. The drafts, a phone screen with no face, still holds the record at 58%.

4. A long film loses the finish. The drafts held best at 2 seconds and had the longest average watch, 6.8 seconds. But it ran 23 seconds, and only 8% watched to the end. The 12-second climbs with a read finished at 12 to 16%.

What we cannot claim yet

• Why every film stops at about 240. Format, hand engagement in the first hour and a second platform did not change it. We do not know the cause yet, and we will not invent one.
• That the climbs work. Five films, two formats, mixed results. It is too early to call a winner.
• The last two films are incomplete. The date and the Netflix text do not have their retention read yet.
• Shares. Five shares on 9 films is better than none, but too few to learn from.

Next week

Theia keeps climbing the two formats, and works on the first two seconds, where the hold is stuck at about 41%. She also plans one test of a real swiped photo carousel, which TikTok counts and shows differently from a photo film. On 31 October two films went out ten seconds apart, by mistake, so neither gets a clean 6-hour score. Next week's post covers what went wrong.

If you want the same loop on your brand's account, start with a pilot month. For how this kind of loop works, and how it can go wrong, read Recursive self-improvement in practice: lessons from the STOP paper.

Next: Theia's first 100,000 views: four accounts, one week, and the format that broke out.

Week 3: What Theia learned, week 3: spoken scenes, three openers, and the first film past the ceiling.