Semantic Search for TikTok Videos: Find Formats, Not Hashtags
The question hashtags cannot answer
You found a video that works. Now you want five more like it, because one data point is an anecdote and five is a format.
So you search the hashtags. And you get nothing useful, for two reasons that are both structural rather than fixable:
Format and hashtag are unrelated. The creator who built a great "big number on screen, no narration, fast cuts" video tagged it#crypto #trading #memecoins. The next creator using the identical structure tagged it #fyp #viral. Nothing connects them in the text.
The shared tags are the useless ones. When two videos do share a hashtag it is almost always #fyp or #viral, which match tens of millions of videos and tell you nothing.
Hashtag search finds topic. What you wanted was technique — and technique is the part that transfers to your niche.
What semantic search does instead
Search by meaning, as a distance calculation rather than a text match.
Each analysed video gets converted into a vector embedding — a numeric representation of its content, built from the transcript, the hook, the structure, the pacing. Finding similar videos becomes finding nearby vectors.
Two consequences that matter:
It crosses topics. A video whose structure is "claim a number, cut hard, never explain" sits near other videos built that way, whether they are about crypto, fitness or used cars. That is exactly the comparison you wanted and could not express as a query. It crosses languages. The embedding encodes meaning, not words. A Spanish video built on the same skeleton as an English one can surface as similar — something keyword search cannot do at all. In practice this matters a lot on TikTok, where the niches worth studying are frequently not in English.Using it
Over the API:GET /v1/videos/{id}/similar?limit=10Find videos similar to this one in my library.
The assistant calls find_similar_videos with the id and reads the results back in context — so you can immediately follow up with what do these have in common and get an answer grounded in the actual teardowns rather than in guesswork.
What to do with the results
The point is not the list. It is what the list lets you check.
Confirm a format is real. One breakout is luck. Five videos with the same structure, from five creators, is a pattern you can build on. Find the variable that moves. When five similar videos differ wildly in performance, the difference is usually one thing — hook length, whether there is a face, whether the payoff lands before the first cut. Compare the teardowns and it tends to be obvious. Find out whether your niche is saturated. If semantic search returns thirty near-identical videos, the format is known and you are late. If it returns four, there is room.Being honest about coverage
Two limitations, both real.
A video must be analysed first. Embeddings are generated from the analysis, not from metadata, so an un-analysed video is not searchable. The niche catalogue carries statistics for tens of thousands of videos; the analysis layer is smaller and grows as people request teardowns. About 63% of analysed videos currently carry an embedding. So an empty result genuinely can mean "not embedded yet". The API says which — it distinguishes "no embedding for this video" from "no similar videos found" rather than returning an empty list and letting you draw the wrong conclusion. That distinction matters more when an assistant is reading the response, because an assistant will otherwise confidently report that nothing similar exists.Why this is the right primitive for agents
Semantic search is a fast read, and that makes it the one discovery tool an assistant can use freely.
A fresh AI teardown takes about a minute because it runs a video model. Semantic search over already-analysed videos returns in well under a second. So the efficient agent workflow is: analyse one video properly, then use semantic search to explore outward from it — rather than analysing ten videos and hoping some were relevant.
In practice that turns a five-minute research session into about twenty seconds.
Where to go next
- Find viral TikTok videos in 4 steps — semantic search in the middle of a full workflow
- The MCP server — all 18 tools
- The REST API — endpoint reference
Frequently asked questions
What is semantic search for TikTok videos?
Search by meaning rather than keywords. Each analysed video is converted into a vector embedding of its content — transcript, hook, structure — and similarity is measured as distance between vectors. Videos built the same way score as similar even when they share no hashtags, captions or words.
Why not just search hashtags?
Because format and hashtag are unrelated. Creators using the same successful structure tag it however their niche tags things, and the tags that do get shared are broad ones like #fyp that match millions of unrelated videos. Hashtag search finds topic, not technique, and technique is the transferable part.
How does vector search work on video?
The video is analysed first — transcript extracted, hook classified, scenes timed — and that structured output is embedded as a vector. Similarity search then compares vectors rather than text, which is why it works across languages and across topics.
Does it work across languages?
Largely, yes, because the embedding encodes meaning rather than the literal words. A Spanish video built on the same structure as an English one can surface as similar, which keyword search would never find.
Is every video searchable this way?
No. A video has to be analysed first, and roughly 63% of analysed videos currently carry an embedding. An empty result can therefore mean 'not embedded yet' rather than 'nothing similar exists' — the API distinguishes the two rather than silently returning nothing.
Start free trial
See why any TikTok went viral with a frame-by-frame AI breakdown. 7-day free trial, then plans from $19/mo.
Prefer to explore first? Start free trial