If you match company names across two datasets, you will eventually produce a confident, wrong join. I'm Ruth Calder, Enforcement Editor at GLP Watchdog, and my job is matching online sellers to regulator records: FDA warning letters, outsourcing-facility registrations, state pharmacy filings. These are the three rules I use. They apply to any dataset where names are messy. (Editorial role, not a clinician; this post is about method, not medicine.)
1. A name alone is never a match
Our register holds pharmacies and FDA publishes a register of 503B outsourcing facilities. One pharmacy on our side is "Empower Pharmacy" and on FDA's it is "Empower Pharma." That pairing is only trustworthy because both are in Houston, Texas. We require a name and a city and a state before we call it a match, and when two candidates sit in one city we record no match at all.
def is_match(a, b): return (norm(a.name) == norm(b.name) and a.city.lower() == b.city.lower() and a.state == b.state) A resemblance between two names is a reason to keep looking, not a result.
2. Every "not found" needs a positive control
A search that returns nothing is as often a broken search as a clean record. Before I write "no FDA warning letter on record," I run the same query for a company I know has one. If the control comes back empty, the finding is about my query, not about the company.
assert search(known_letter_company) != [] # the control result = search(target_company) # only then trust an empty result 3. Log your own errors in public
When we counted warning letters per seller, four of our own reviews showed "4", "5" and "6" letters for companies with none. The count had been built from a field that also held general FDA notes. We fixed it, added an automated check that fails the build, and wrote it up. The write-up is part of What the FDA's GLP-1 Warning Letters Say About This Market, and the wider list of what a register can't verify is on What This Register Gets Wrong, and What It Cannot Check.
None of our errors was caught by review. Each surfaced when someone computed a number for an article. If you generate text from structured data, one wrong rule produces the same wrong sentence on many pages, in fluent prose, so test the rule, not the page.
Where this is used
The full pharmacy-check walkthrough, which relies on the same name-city-state rule, is How to Verify an Online GLP-1 Pharmacy. My author page: Ruth Calder.
This is general information, not medical advice.

