Guides / Writing
We tested Cuely against ChatGPT on 10 YouTube videos
It is easy to say your tool writes better scripts than ChatGPT. So we measured it, and we are sharing how, including what the test cannot tell you.
How we tested
- 10 topics a small channel might film: posting every day for 30 days, beginner guitar mistakes, a $600 gaming PC, meal prep on £25, a weekend without a phone, and five more.
- The same request for both: a full script for a 5 to 7 minute video, for a small channel where the creator films alone.
- ChatGPT got that request as a plain prompt. Cuely ran its new Writing Agent: draft, retention check, rewrite the weak parts, remove what sounds like AI.
- A judge model read both scripts blind, in random order, and scored them 1 to 10 on the hook, retention, specificity, voice, filler, style and length.
- We also counted AI signs with code: AI-sounding words, long dashes, "-ing" tails, "not X, it's Y" and lists of three, per 1000 words.
What we found
The judge picked Cuely's script in all 10 videos. Average scores, Cuely against ChatGPT:
- Hook: 9.0 against 6.6
- Retention: 9.0 against 7.2
- Specificity: 9.1 against 7.4
- Sounds like a person: 9.1 against 7.0
- No filler: 8.8 against 6.5
- AI signs per 1000 words: 1.7 against 16.3
The biggest gap is the last line. ChatGPT's scripts had about 16 AI signs per 1000 words, roughly one every 60 words. Cuely's had fewer than 2.
The judge's reasons were mostly the same: a hook with a real promise, more concrete detail, and fewer phrases that sound like nobody in particular.
What this test does not prove
We would rather say this ourselves than have you find it out:
- The judge is also an AI model, from the same family Cuely runs on. A model judging its own family may be kinder to it. A human panel is next.
- Our checklist is close to our own rules. Cuely is built to avoid long dashes and lists of three, and the test counts them. That is fair to what viewers notice, but it is our checklist.
- ChatGPT got one plain prompt. A person who writes a careful prompt and edits the result will get a better ChatGPT script than this.
- Sentence rhythm was close. ChatGPT's sentence lengths varied slightly more than Cuely's, so there is still work to do there.
- Claude is not in this test yet.
What is next
We will run the same 10 videos past real creators, blind, and ask one question: which one did a person write? When people can no longer tell, we will call it done. We will publish that result too, whichever way it goes.
Questions
Which model does Cuely use?
The Writing Agent runs on a GPT-4.1 class model. The difference in this test comes from how it works: checking its own draft and fixing it in steps, not from a different model.
Can I see the full results?
Yes. We keep the scores for every topic and will update them each time we rerun the test.