TikTok Semantic Search & Audit

Find TikTok videos that match a topic, understand why they match, and export the evidence. SignalProof combines live candidate collection with AI-powered relevance analysis and a visual report.

Use it for creator research, content discovery, campaign checks and finding posts about a specific consumer problem.

Start a search

Enter your topic and click Start. The default collects up to 30 real candidates and returns up to 20 matches. AI creates up to three short search phrases in the language of your topic; supply searchQueries to control these yourself.

{
  "query": "skincare routine for oily acne-prone skin",
  "maxCandidates": 30,
  "maxResults": 20,
  "minRelevance": 0.65,
  "upstreamMaxCostUsd": 1
}

What you receive

• Videos ranked by relevance, with direct TikTok links.
• An explanation in the language of the topic and topic labels.
• An evidence excerpt checked against the original text.
• Creator, caption, publication date and available engagement metrics.
• JSON data, CSV and an HTML report in Portuguese.
• Source dataset and collection run identifiers.

Analysis uses captions, hashtags and transcript text supplied by the source, up to 12,000 characters per video. It does not inspect video frames, transcribe audio or analyze comments. Missing metrics stay null.

Sources and costs

Source	Input	Collection cost
live (default)	Topic or campaign URLs	Separate Clockworks TikTok Scraper run on your account
dataset	Select an existing datasetId	No new collection
inline	Supply a videos array	No new collection

Live collection is billed separately from SignalProof. upstreamMaxCostUsd caps the child scraper's event charges. SignalProof's run charge limit applies to this Actor. Source usage/storage may add costs according to its pricing. Discarded candidates can still incur collection charges.

AI analysis is ready to use with the hosted Actor's default configuration; no API key is required. Advanced users can supply their own key through the optional encrypted API key field. External AI usage on a supplied key is paid by its holder. Keys are never included in results. Select rankingEngine: "semantic" for local multilingual AI without an external key.

Import a dataset

Set source: "dataset", select datasetId in the form, and describe the topic. The picker grants read access under limited permissions. Standard nested Clockworks records and flattened exports are supported. Keep TikTok IDs as strings to preserve precision.

Audit campaign URLs

Set mode: "audit", add canonical URLs to videoUrls, and optionally describe the campaign topic. Omit query for a coverage-only audit. Short links and profile URLs are rejected.

Audit preserves every requested URL, including posts that fail filters. unavailable_in_source means the loaded source returned no usable record for the ID; it does not establish deletion. Set maxCandidates high enough to cover the requested URLs, especially when reading a dataset.

Relevance and filters

AI relevance scores range from 0–1: around 0.3 for a passing keyword mention, 0.5 for a partial match, and 0.7 or above for direct relevance. Default threshold: 0.65. Scores are AI assessments, not probabilities. Read the explanation and evidence when making a decision.

semantic uses a pinned multilingual MiniLM model cached in the Docker image. It compares the first 400 query characters with four excerpts totaling up to 1,600 text characters, subject to tokenizer limits. lexical measures exact normalized word overlap. Errors never silently switch methods.

Filters support required/excluded terms, source-supplied language, minimum views and inclusive UTC publication dates. Unknown fields fail the corresponding active filter. Duplicate IDs keep the first usable record and combine source query labels.

Failed or interrupted collection is labeled partial. Source runs, records and completed AI batches are checkpointed. Restarts reuse them and skip delivered IDs. This reduces repeat work; it is not an exactly-once billing guarantee across abrupt platform failures.

Output

The default dataset contains selected search results or requested audit rows. Fields include relevanceScore, relevanceReason, topics, evidence.excerpt, evidence.evidenceVerified, metrics, filter outcomes and source provenance.

Open Visual report, Download CSV or Coverage report (JSON) in the run output. The CSV includes explanations and protects spreadsheet formula prefixes. The JSON report records analyzed/selected/delivered counts, source status and AI token usage.

Analysis pricing

US$0.01 per unique collected video analyzed and delivered (US$10 per 1,000 videos). Missing audit URLs, invalid records, duplicates and rejected search candidates have no analysis charge. Collected audit rows are billable even when they fail filters.

There is no startup fee. Platform usage for SignalProof is included in the analysis price. Live collection charges remain separate. Set a maximum cost per run to limit the number of paid results delivered, and use the collection cost field to control the separate scraper run. Check the Pricing tab for the current price.