Feed it your worst table
You upload a financial statement or invoice whose largest table spans a page break, then check whether the rows after the break stayed in the right columns.
Why this one: Tables are the reason this category exists, and when a number lands in the wrong column the marked cell positions let you see where it came from rather than guessing. A parser that returns only text gives you no way to tell a misread character from a broken layout.
Tweak: Run the same file twice with one setting changed. Where the two runs disagree is where the document is genuinely ambiguous.
Photograph a page and parse the photo
You photograph one page of a printed form with your phone, upload it, and compare the result against the same page exported as a PDF.
Why this one: Reducto takes JPEG, PNG, HEIC and TIFF alongside PDFs, so one pipeline handles a phone snap and a born-digital file. That is the case that separates a parser from a text extractor: a clean PDF already contains its text and almost anything can read it.
Tweak: Then photograph the same page badly on purpose, at an angle and half in shadow, and find where it stops coping. Knowing the boundary is worth more than knowing the best case.
Ask for four fields instead of the whole document
You list the four values you would otherwise type into a spreadsheet from the document, then ask for exactly those four and nothing else.
Why this one: Parse and extract are separate products here with separate prices, because they are separate jobs: parse hands you the whole document, extract hands you the four values you were going to copy out of it anyway. Choosing the wrong one is the commonest way to overpay in this category.
Tweak: Ask for a field that genuinely is not in the document. How a tool behaves when the answer is absent tells you more than how it behaves when it is there.
Price your real workload before you decide you like it
You multiply your monthly page count by the published rate for the endpoint you need, then check that total against the free balance you start with.
Why this one: Reducto publishes a per-endpoint rate card — parse, extract, deep extract, split, deep split, classify and edit each priced separately — so you can cost a project off the pricing page without a sales call. Most vendors in this category quote by conversation, which makes the arithmetic somebody else's.
Tweak: Do it twice, once for the pages you have today and once for ten times that. The second number is what decides whether this becomes infrastructure or stays an experiment.
