CLEAR SIGNALS. HONEST LIMITS.
Understand what
you’re looking at.
Takedown helps you examine suspected copies using material you supply. Every score is a comparison signal for human review. It is not a finding of ownership, unauthorized use, counterfeiting, or infringement.
The baseline: explainable comparisons
The heuristic overlap score combines title wording (25%), description wording (45%), and shared three-word phrases (30%). Available signals are reweighted when inputs are missing. Meaningful word frequencies are compared with cosine similarity, after case and punctuation normalization. Common English words are excluded from word comparisons.
Image comparison uses a 64-bit difference hash. An uploaded image is reduced to a small grayscale representation, then neighboring pixels are compared. The image signal reports matching fingerprint bits. Only agreement above 65% contributes to the overall score, with an additional weight of 0.4. This method can flag resized images, but different images can share a fingerprint, and crops or changed angles can reduce similarity.
High overlap starts at 70/100; medium starts at 35. These are unvalidated review thresholds, not accuracy claims. Short or repetitive text may be marked as insufficient. Brand-name mentions are reported separately and do not change the overall score.
The AI: meaning, on your device
When you choose “Run AI analysis,” Takedown downloads the pretrained all-MiniLM-L6-v2 model and the Transformers.js runtime. The model computes sentence embeddings in a browser worker. No listing text is sent to a hosted model API.
We compare the first 960 whitespace-separated words on each side in passages of up to 80 words. Passage embeddings are averaged and normalized, then compared by cosine similarity. The output is clamped to a 0–100 indicator and displayed separately from the heuristic score. A tokenizer may further truncate unusually long tokens. Related product categories can produce high similarity without copying.
The model is best suited to English text. Results have not been validated on a brand-protection benchmark. Initial download, browser compatibility, and network access can affect availability. If AI fails, the baseline comparison remains usable.
What the product does not do
- Search the web, monitor stores, crawl marketplaces, or connect to seller accounts.
- Fetch or independently verify source URLs, capture screenshots, or establish publication dates.
- Recognize logos, authenticate goods, decide design rights, or confirm seller authorization.
- Provide cloud case storage, team accounts, managed legal review, or automatic removal.
- Submit a takedown request. Drafts need a rights holder’s review and platform-specific completion.
Build a record that holds up to review
Save original files and dated screenshots yourself. Add source URLs, rights documentation, permission context, and your observations. Evidence exports contain user-supplied text, optional image fingerprints, methods, scores, and notes; they are not certified captures. Image bytes are not saved in cases or included in exports.