Take a thousand addresses from the list you actually intend to mail, including the awkward domains. A vendor's own test corpus tells you about their test corpus, and a consumer Gmail sample is a far easier examination than your business data.
Compare the unresolved share first. If one service returns 2% unknown on a B2B list where another returns 20%, the difference is almost certainly guessing rather than superior infrastructure, and guessing shows up later as bounces.
Then check three pricing terms that rarely appear on the pricing page: whether unknowns consume a credit, whether duplicates are charged, and whether credits expire. Those decide what your money actually buys.
Ask whether the raw SMTP response is exposed in the result. A service that shows the server's own words lets you audit a verdict you doubt, and one that returns a bare label is asking you to take it on faith.
Finally, read the data terms. Retention period, whether your list is reused to improve a shared database, where processing happens. A verifier that keeps your contacts has made them part of a product, and you should know that before uploading.
A test sample with the answers hidden inside
A plain sample tells you how two vendors differ. It can't tell you who's right. Fix that by hiding known answers in the file. Take 800 addresses from the list you plan to mail. Add 100 that hard-bounced last month with a mailbox-not-found code. Then add 100 people who clicked a link in the past week. Shuffle the file well, and you now know the truth for 200 of the 1,000 rows while neither vendor knows which rows those are.
Score each vendor on a short card. How many of the 100 dead addresses did it catch? And of the 100 live ones, how many did it wrongly call invalid? What share of the other 800 came back unknown, and what share catch-all?
Add two more lines. Work out what your full list would cost under each vendor's rules on repeats and unknowns. And note whether the raw server reply appears on each row. One page like that beats any comparison table a vendor publishes.
Signals that don't predict quality, and one that does
Some things look like evidence and aren't. Star ratings on review sites mostly measure how hard a vendor asks for reviews. A speed claim means little without the list behind it. The number of integrations tells you about the sales team's priorities, and a single accuracy figure tells you about the marketing team's.
Here's a signal that works. Before you buy, email support with a precise question. Ask what verdict they return when a server answers 252, or how they handle a greylisting server that refuses the first attempt. You're testing two things at once: whether anyone there understands the protocol, and how long a real answer takes to arrive.
A good reply names the verdict and explains why. Weak replies paste a paragraph about industry-leading accuracy. You'll rely on that same inbox the day a result looks wrong before a big send. It's worth knowing in advance which kind of reply you'll get.
Related questions
What is the single best test of a verifier?
Run the same thousand addresses from your own list through two services and compare the unresolved share and the raw responses.
Should I trust published accuracy figures?
Only alongside coverage. An accuracy percentage without its denominator says nothing about how a list will perform.
What pricing terms should I check?
Whether unknowns are billed, whether duplicates are billed, and whether credits expire. All three are usually off the pricing page.
Does the cheapest option cost more later?
It can, when a low price is funded by guessing at hard addresses. The bounces arrive after the invoice.