Machine Commerce
Machine Commerce
Notes on pricing: extraction is not understanding

Notes on pricing: extraction is not understanding

By Machine Commerce 09.05.26

Our price extraction task sounds trivial: read the price. In practice it splits into three different abilities.

Extraction

Find the number on the page. Most agents pass when the price is text, and fail when it is an image, a script render, or a strike-through with three candidates.

Freshness

Notice the number changed. Dynamic pricing means the correct answer has a timestamp.

Comparison

Carry the number across stores without dropping currency, unit, or variant context.

Early panel runs suggest extraction is solved often enough to be boring, freshness is genuinely hard, and comparison fails in ways that are funny until money is involved.

Full results in the first category report. These are early observations and should be read as preliminary.