Read this first
This is the short form of a research paper. The full paper has the sources, tables and math. Each part here links to the part that backs it up. The main market is the US.
The facts were read up to 11 November 2026. Facts that may change are checked again on 11 December 2026. Any change will be added as a dated fix.
What an ad check does
An ad check asks a set of questions. Was the ad in view? Did the traffic look valid? Did the ad run next to content the brand wished to avoid?
Each check has a scope. It may cover a site, an app, a page or a video. A check on a site does not prove what was next to each ad. An ad in view does not prove a person paid heed to it. A brand must state which content it wants to block. Chapter 1 sets out the terms.
A seal has limits
The MRC is a body that checks how ad counts are made. It signs off on a named service and its scope. That scope can change by screen, ad type, place and language. Its letters do not give a rate for how often the tool is wrong.
The paper's grid has 108 units signed off out of 298 sold. That is 36.2%. A unit means one product, one kind of check and one place it runs. This is a share of those units. It is not a share of all ads or all spend.
Some sales claims can be read more broadly than the MRC's scope. That is the author's reading of the words. It is not a ruling by the MRC. Chapter 2 has the scope and the caveats.
Who checks if the checks are right?
A fair test needs a set of ads with known labels. It needs to state how many were tested. Then it can show what the tool missed or got wrong.
In the sources read for this paper, no outside test met that rule for a scored product's sold task. Some reports show bad cases. Some show how often tools agree. Those are useful facts. They do not tell us how often a tool is right.
A low rate of bad ads in a screened group is not the same as a low miss rate. The same tool may have screened the ads and judged the result. Chapter 3 weighs the tests and replies.
Does a check change the bill?
At times it can. The Trade Desk has terms that give some outside reports a final say on what it owes a seller. Microsoft Invest had a refund route based on a partner's report. Microsoft said that route would end.
On the paper's fixed list, 2 of 16 terms that could be read gave a report that force. That is 12.5%. The author uses that rule for the main finding. A wider set of contracts gives a less clear result. So the finding must keep its scope.
No payment settled on such a report was found in the sources screened. That does not rule out private deals. Nor does a tool that blocks an ad prove the buyer got cash back. Chapter 4 follows the money.
The platform can shape the view
Meta, YouTube and Amazon pass data to firms that check ads. What the firms can see depends on that route. A partner badge does not prove raw access.
TikTok has a dated route in which the checker's own tag runs through a kit in the app. TikTok still controls that kit. The old route was not fully restated in a new source. Other platforms name firms but leave parts of the route unclear. Chapter 5 shows what each route supports.
TV needs its own check
A check passed for web ads may not cover TV ads. Some TV view counts cover only named types of apps. An ad can play while the screen is off. The MRC does not require tools to find all such plays, though known cases must be removed.
A kit's release is not proof that all apps use it. A count of firms with the kit is not a share of all ads. Chapter 6 has the TV and app detail.
Who pays the checker?
Both buyers and sellers pay ad-check firms. That calls for clear rules on how the teams are kept apart. It does not prove that a count is biased.
The author accepts DoubleVerify's public pledge to keep its quality team apart from sales as a safeguard. The pledge leaves its scope and enforcement unclear. It does not cover the pending new owner's interests. Chapter 7 sets out the payers and owners.
Which content rules apply?
GARM closed, but its content rules still have uses. ISBA has a later guide with a floor and risk tiers. It pledges to review the guide and names who oversees it. The author counts that as a change process. A stricter test would leave the finding open.
A firm that says it is GARM-aligned may not state which dated rules it uses. Ask for those rules. Chapter 8 explains the choice.
A claim is not a court finding
A letter can ask for a probe. A suit can make a claim. Neither proves a tool failed. A court can let a suit go on without deciding who is right. The paper keeps those steps apart. They do not add or take points from the scores. Chapter 9 has the record.
Can AI buy ads by these checks?
The tools can name a check. Some have links that work and tests that run. Firms also claim live use. Reports add some buyer and campaign facts.
But the sources read do not show which actual check made the AI buyer change a live buy. They also show no named pact that makes the check decide the bill. This question stays open. Chapter 10 sets out the proof still needed.
What the scores mean
The paper scores 17 products on 12 things. It groups like services with like. A high score means more of the service is backed by public records under the set rules. It does not mean the tool is more often right.
Some groups are too small for bands. Some bands move when weights change or gaps are filled in a test. A gap is not a zero. Read the cells and flags with the score. Chapter 11 shows the work.
What to ask before you buy
Ask for the exact scope letter. Ask for a test with known labels and a count. Ask whose data the tool sees. Ask who can change a result. Read which report sets the bill, who pays, and when you must claim. These are checks to make before a buy. Chapter 12 turns them into requests.
How this was made
Claude and OpenAI models did the work under the author's lead. Some checks used a separate AI. For the added score rows, the same model scored and then tried to refute its work. The author checked the work. No human gave a second set of scores.
The author works at Samba TV, now called Samba. It competes with Nielsen in TV counts. DoubleVerify's sale to Nielsen's parent was still pending when read. He helps lead a group that works on AdCP. His firm has agents in that group's registry. He once worked at Verve Group, Hearts & Science and Lotame. No vendor paid for, read or saw the paper first. Chapter 13 has the full method and interests.