Matt Newsome

build web data extractors, mostly for e-commerce marketplaces.

My focus is the unglamorous part: whether the data is actually correct, and whether you can tell when it isn't. Most scrapers fail quietly, they hardcode CSS class names that break on the next redesign, return half a dataset without saying so, and bill you for duplicate rows. The tools I publish are built the other way round:

• Structured data first - Extraction reads schema.org JSON-LD and documented JSON endpoints before it ever falls back to CSS selectors, because sites do not break the markup that feeds Google Shopping.
• Honest field coverage - If a source does not expose a field, it is documented as null rather than filled with a plausible guess. A dead image URL is worse than an empty one.
• No duplicate billing - Results are deduplicated before anything is emitted or charged. Ask for 1,000 products from a keyword that only has 600 and you get 600 real rows, not 1,000 with 400 repeats.
• Visible degradation - Every run reports its yield. A partially blocked run looks different from a clean one in the log, instead of looking identical.

Before building anything I measure cost to serve, block rate, bytes per request, items per request, and whether prices appear at listing level or only on detail pages. That last one is worth roughly 50x. Skipping this step once cost me a complete build against a marketplace whose compute cost exceeded the market price per result.