TikTok Search API: Four Things to Build With Keyword Search

The simplest useful job: every morning, what got posted about your category yesterday that anyone actually saw? A date bound plus a view floor turns a firehose into a short list you can read over coffee.

Because results come back newest first, you do not need ranking logic on your side. Run it on a schedule, diff against what you stored yesterday, and alert on what is new.

Influencer discovery usually starts from a follower-sorted list, which is the wrong sort. Start instead from posts that performed on your topic, then look at who made them — you get creators proven on your subject rather than creators who are merely large.

Widen coverage for a single keyword with parallel_search_slices, group the results by author, and rank by average views. Anyone appearing several times with strong numbers is a candidate; from there, run each handle through the profile scraper to check they are consistent rather than lucky.

This is the step conventional social listening cannot take. On TikTok the substance is spoken, not written — captions and hashtags are packaging. If you are only reading text fields, you are analysing the label on the tin.

Every search result carries its video URL, so transcription is the next loop. Now “what are people saying about us?” has a literal answer, in 99+ languages, including the complaints nobody bothered to type in the caption.

Captions are optional, not required. Ask for the transcript and you get the creator's captions when they exist; when they do not — which on TikTok is most of the time — speech-to-text reads the audio instead. You never get a blank row because someone skipped the subtitles. Both paths are documented: transcript and transcribe.

Transcripts answer questions one video at a time. To answer them across a whole keyword you need columns, so define the fields you care about and get them back typed for every result.

Hook, brands mentioned, sentiment, main claim. That is a table, and a table answers questions prose cannot: which claim shows up in the negative posts, which hook pattern dominates the top decile by views, which competitor is being named most this month.

A single search returns roughly 90 to 120 unique videos for about 0.05 credits. That is the right setting for monitoring, where you want today's posts rather than an exhaustive archive.

When you need broader coverage of one keyword, parallel_search_slices runs several searches concurrently and deduplicates them, reaching up to roughly 480 unique videos at its maximum of 4. Cost scales close to linearly, so reach for it on a launch-week sweep, not on a daily job.

Transcription and extraction are billed per video when you chain them on. Filter hard first and you only pay to analyse the posts that matter — see the pricing page for the full picture. Failed searches are reversed automatically.