the limitations part is the one section that actually matters here, bot walls and skipping social platforms are real constraints not just disclaimers. schema design is what decides accuracy once you're in context, a plain price field will happily return the sale price, a monthly price or whatever number sits next to it and call it done. descriptions and a price_raw field do more work than people think and validation only checks shape not that you grabbed the right value. distillation misses lazy loaded rows too, that's usually just a render or scroll timing issue nobody tested for. for sites I hit repeatedly I cache selectors and only fall back to the LLM when something drifts, way cheaper, but for true one off long tail jobs the schema approach is the whole pitch. also the langchain demo defines required_schema_description and never actually uses it, and the pricing table people keep citing is out of date, current pack runs a dollar for 160 prepaid credits