Read this first
Start with one short question. What did the seller match? Was it a row in a file, a phone, a web page or a home? Then ask what came back. Can you see the link? Can your ad system use it? Can you check if it is right? These are three tests.
OpenAI models did the research, wrote the paper and ran the source checks. No independent human peer review is documented. Read how the work was done.
We kept the tests apart. We looked at what firms sell in the US. We read the source for each claim. We kept a note when we had to change what the paper said. Read the full paper.
Count what goes in
A shop sends a list of sales. A firm says it found a match for most of them. Ask which sales went in the base. Did the firm leave out rows with no name? Did it count the same sale twice? Ask what it had to find for that sale to count.
One shop case gave us those parts. The firm said how it tried to link a sale to a buyer. Its rate told us how much of the file got through. It did not tell us how many links went to the right buyer. Both facts matter if you plan to spend money on the list. Read the match test.
Check the links
Think of two people who share an email account. You may want to send one note to that inbox. Or you may want to know which one bought a coat. The same email match could help with the first job and give the wrong answer to the next.
Ask for a test with people or homes whose links you know from a source of your own. Ask who picked the test rows and what they left out. A high match rate will not fill that gap. Our search left part of this test open. Read how the test works.
Follow the choice
A person opts out. What stops? One site might stop asking for a new ID. That does not tell you what a firm can do with an old one. Ask the seller to show the whole path. Who hears the choice? Who acts on it? What happens to a file that has left the site?
We found clear paths for some IDs. Too few showed the full path to meet the test in the main group we could check. The answer could change if more firms belong in that group. Follow a choice.
Know the home
A TV can be linked to a home. A home can have more than one viewer. To test who saw an ad, first say which of those you need to know. Then ask when the link was made and when it stops.
Only one US home-graph service, Blockgraph, met the buying test. Its guides gave no time limit for a live link. We could not tell what most firms in the US disclose.
A file can be new while a link in that file is old. A firm can look up a past IP address without saying how long it keeps the link to a home. Ask for the rule on the link you plan to buy. Read the TV cases.
Check what the tool can do
A plan for a message tells you what it can hold. Code can show that a tool can read the message. A sale needs its own proof. In one project we found code that can use an ID in a check. The case did not show that a live sale used that same check. Ask for the proof that fits the claim. Read the agent test.
Read the score
The score tells you how much proof a buyer can see for a set of tasks. It groups firms that sell the same kind of help. Some products had too few checked facts to enter the score table. Some groups had too few products to show a band.
A gap in public proof is a good question to take to a seller. It is not proof that the firm cannot do the work. Look up the row, read the source and ask for the test that matters to you. Read about the score. Use the buyer’s guide.