We Extracted Hundreds of TikTok Recipes — Here's What We Learned
After running AI extraction on hundreds of TikTok recipe videos, here's what works, what fails, and what surprised me about extracting TikTok recipes.
I built ReelToMeal because I kept losing TikTok recipes I wanted to cook. I'd save a video, forget about it, and three weeks later have no idea why I bookmarked a clip of someone tossing cubed feta into a pan. So I made a tool that watches the video and writes the recipe down for me.
After running AI recipe extraction on hundreds of TikTok cooking videos — both for testing and for actual weeknight dinners with my kids — I've collected enough data and bruises to share what's really going on. Some of this surprised me. Some of it confirmed what I already suspected. And some of it is just genuinely embarrassing for the AI.
The numbers I keep seeing
Let me start with patterns from the extractions themselves.
TikTok recipes are short — really short
The average TikTok recipe video I processed runs 38 to 75 seconds. Compare that to YouTube cooking videos, which are typically 8 to 15 minutes. That brevity is both a blessing and a curse for AI extraction. Less audio to transcribe, but every second is dense — creators cram entire recipes into 60 seconds with rapid cuts and on-screen text flying by.
Ingredient counts cluster tight
Most TikTok recipes I extracted had 6 to 11 ingredients. Anything under 5 tends to be a "hack" video (one-pot pasta, three-ingredient cookies). Anything over 12 is almost always a chef account, not a home cook.
Cuisine breakdown
From my sample, the most common cuisines were:
- Italian-American (pasta, pasta, more pasta)
- Asian fusion (Korean, Chinese, Thai blends)
- Middle Eastern and Mediterranean (this is where I cook from most as an Israeli, so I notice every shawarma and shakshuka video)
- Mexican and Tex-Mex
- American comfort food
What's underrepresented? Indian, French, and most African cuisines. Possibly an algorithm bias on my end, possibly a real gap in viral TikTok food content.
Where the AI nails it
Recipe extraction works beautifully on certain types of videos. After hundreds of attempts, here's what I've learned produces clean results:
- Clear voiceover narration where the creator says ingredient names and quantities out loud
- On-screen ingredient text that stays visible for more than half a second
- Single-cuisine recipes with familiar techniques
- Videos with a recipe card at the end — even a 2-second flash of a list dramatically improves accuracy
- English or Hebrew narration (those are the two I personally test most, and both are well-supported)
When all of these line up, I get a clean recipe with proper measurements, ordered steps, and accurate cook times. I made a roasted tomato pasta from a 47-second video last Tuesday and the extraction was so clean I didn't have to edit a single word.
Where it falls apart
I want to be honest about this because anyone telling you AI extraction is 100% accurate is selling something.
Fast cuts kill context
When a video has 25 cuts in 40 seconds, the AI sometimes loses track of which ingredient went where. I've seen extractions list "2 tablespoons of olive oil" twice because two different shots showed two different oil pours.
Heavy accents and slang
Strong regional accents trip up the audio transcription. So does cooking slang — "a glug," "a knob," "a splash." The AI tries to convert these to measurements and sometimes guesses wrong. A "knob of butter" became "1 tablespoon" in one extraction and "50 grams" in another, from the same video.
Off-screen narration with no visual confirmation
If the creator says "add a teaspoon of cumin" while showing a closeup of a different bowl, the AI sometimes can't link the audio to the action. Step ordering gets scrambled.
Text overlays in stylized fonts
Cute hand-drawn fonts and rapidly animated text are surprisingly hard for OCR. Block sans-serif text on a solid background? Perfect every time. Cursive script that fades in and out over a sizzling pan? Coin flip.
The "vibe" recipes
Some TikTok creators don't actually give you a recipe. They show vibes. "Just add a bit of this, a bit of that, you know." The AI does its best, but garbage in, garbage out. I extracted a viral chili oil video where the creator never specified a single quantity. The output was a recipe of pure guesses, and I had to throw it out.
What this means if you're using the tool
A few practical things I've learned:
- Pick videos with clear narration or visible recipe cards. The extraction quality is directly tied to the source.
- Always read the output before cooking. Treat it as a strong first draft, not a final recipe. I usually skim it once and adjust quantities based on common sense.
- Trust your cooking instincts on weird outputs. If the AI says 3 tablespoons of salt for a single chicken breast, that's wrong. You know it's wrong.
- Save the original video link. Sometimes you want to rewatch a technique even after you have the written recipe.
The honest caveat
Recipe accuracy on TikTok extractions sits around 85 to 92% in my experience — meaning most ingredients and steps are right, but you should expect to fix something on roughly 1 in 8 recipes. That's much better than my old method (no method, just hoping I remembered), but it's not perfect and I don't pretend it is.
The bigger win for me isn't perfection. It's that I now actually cook the things I save. Before, my saved folder was a graveyard. Now it's a meal plan.
If you've got a folder of TikTok cooking videos collecting dust, try pulling a recipe out of one with ReelToMeal and see what comes back. Worst case, you spend two minutes and learn something about how the AI thinks.