AI Product MatchingAI Product Matching › LLM matching
Can ChatGPT or other LLMs match products?
Yes, for small batches and hard cases. A large language model can read two listings, compare brand, model, size and variant, and explain its decision. For thousands of products it is slow, costly and not perfectly consistent, so the usual pattern is embeddings and rules for the bulk, and an LLM to review the uncertain pairs.
Updated 10 November 2026 by the ScrapeWise team.
Quick picks
For a quick experiment
Ask an LLM to compare pairs
Request a reason and a confidence.
For production
Embeddings plus LLM review
LLM only on the grey zone.
For no pipeline at all
ScrapeWise (our product)
Scraping and matching as a service.
Compared
| Approach | Speed | Cost | Best for |
|---|---|---|---|
| LLM only | Slow | Highest | A few hundred pairs, one-off checks |
| Embeddings only | Fast | Low | Large catalogues, first pass |
| Embeddings plus LLM review | Fast for most, slow for few | Medium | Production matching with a quality check |
| LLM attribute extraction, then rules | Medium | Medium | Messy titles that hide size and model |
| ScrapeWise matchingOur product | Fast | €0.15 per 1,000 products (text) | Matching competitor data without building a pipeline |
ScrapeWise matching
Matching runs are currently free. Results can be read by AI agents through the ScrapeWise MCP server.
Prompting an LLM to match products
- Give both listings in full
- Title, brand, price, size and attributes for each.
- Ask for a fixed answer
- “Same product: yes or no; reason; confidence 0–100.”
- State the rules
- “Different sizes, pack counts or model years are different products.”
- Check a sample
- Compare its answers with hand-labelled pairs before trusting it.
Where LLMs fail
- Near twins
- They may call size 42 and 43 the same product unless told otherwise.
- Missing facts
- If neither listing states the size, the model may guess.
- Inconsistency
- The same pair can get different answers on different runs; fix the settings and log results.
Frequently asked questions
Can ChatGPT match products?
Yes, for small batches. It can compare two listings and explain its answer, but it is slow and costly for large catalogues and can be confidently wrong.
Is LLM product matching accurate?
It's accurate on clear cases and helpful on messy titles, but it makes mistakes on sizes, pack counts and model years unless you state those rules.
How much does LLM product matching cost?
It depends on the model and the length of the listings. It is usually far more per pair than embeddings or rules, which is why LLMs are used only for uncertain pairs.
What's the best way to use LLMs for matching?
Use embeddings and rules for the bulk, and an LLM to review the uncertain middle, asking for a structured yes/no, reason and confidence.