ppl.studio's own project · AI companion app
Bae on TikTok: What 128 Posts Made Entirely with ppl.studio Did
Bae is an AI companion app we build. Since 28 September 2026 every post on its TikTok has been made by ppl.studio presets. Here is what happened, including the part that didn't work.
3 min read
The setup
Bae is an AI partner that remembers you, aimed at people who feel lonely. Its TikTok account is @ppl.studio. A scheduled job calls ppl.studio's public API several times a day: it picks a topic in the relationships niche, starts a run of one preset, waits for the render, asks for a caption, and hands the finished post to a phone that publishes it.
The mix, from 30 September, was six posts a day:
- two text-message videos (an on-screen chat with Bae, five chat shapes taking turns),
- two TikTok carousels,
- one app list carousel (“N apps that…” with Bae's real screenshots among others),
- and one paid video, rotating through an animated story, a talking AI creator, an app demo and a lip-sync clip.
Posting ran between 5 and 12 a day as the schedule was tuned. Every caption's first line says the companion is an AI, and posts with a generated person carry TikTok's AI label.
Views: every post got the first batch, and none went further
| Period | Posts a day | Median views per post, by day | Best post |
|---|---|---|---|
| 28 Sep – 7 Oct | 5–12 | 777–818 | 1,114 |
| 8 Oct – 13 Oct | 5–12 | 262–286 | 1,810 |
This is the honest headline: views stalled. Every post got TikTok's first test audience, about 800 views until 7 October and about 270 from 8 October, and not one broke past it. For You traffic stayed at 98–100% and no post was flagged as restricted, so the account was not being suppressed; the posts simply never earned a second push.
The drop on 8 October came three days after posting went from 5–8 to 10–12 a day. That is a guess, not a finding, so we cut the schedule back to 8 a day on 14 October to test it.
A side effect: while every post gets the same first batch, views tell you which week a post went out, not whether its format is any good. A format posted the week before the drop looked three times better than one posted after it.
Retention: where formats really differ
| Format | Average watched | Watched to the end | Likes per post |
|---|---|---|---|
| Text chat: “what do I say” (25 posts) | 46% | 18% | 11.5 |
| Text chat: sweet / recommendation | 43% | 14–19% | 11–23 |
| Text chat: parent-text prank | 30% | 10% | 8.7 |
| Talking: AI-message proof rant | 13% | 1% | 0.8 |
| Talking: mortified crush confession | 8% | 0% | 0.5 |
| Lip-sync: escalating answer | 6% | 1% | 0.7 |
| Explainer: history | 4% | – | 1.5 |
The text-chat videos held people: 43–46% of each video watched on average. The talking and lip-sync videos, our copies of formats that do well for real creators, lost almost everyone in the first seconds (6–13% watched). Looking frame by frame, the copies barely followed the original format: wrong length, a random setting, a monologue where the original had four short lines. So we rebuilt the copying step to follow a source video shot by shot, timing, framing and on-screen text included. An early faithful copy of the lip-sync format cost $0.19 to render; it has not been posted long enough to report on.
Carousels get no watch-time figure from TikTok, so they can only be judged on views and likes, and saves and shares were close to zero on every format.
Other platforms
The same posts went to YouTube Shorts, Instagram and Facebook. On YouTube, half the Shorts got 0–1 views; chats did best there (median 357 at 24 hours) and carousels posted as slideshows got a median of 3, so carousels now skip YouTube. Instagram's median was 135 views. Facebook gave about 5 views a post whatever the format.
What it cost
From our cost ledger (25 September to 6 October, at the model providers' list prices): a text-message video costs close to nothing, since it is text and a server render; a six-slide carousel with AI pictures about $0.37; a 26-second talking video about $1.31; an animated story about $2.21. Bae's routine and the advent.center carousel routine together came to about $195 per 30 days of rendering, at six posts a day.
What we take from it
- Volume alone did not grow the account. More than a hundred posts in two and a half weeks got the same first batch every time.
- Judge formats on retention while views are flat. Watch time separated the formats when views could not.
- A bad copy is not a bad format. The talking formats failed as we built them; whether they fail when copied faithfully is the next test.
The tools used here are the same presets every ppl.studio account gets at /workflows, driven through the same public API. For how another of our projects uses ppl.studio, see advent.center's carousels.
Founder of ppl.studio. Building AI tools for product marketing teams who need visual content at scale without the production overhead.