This website uses cookies

Read our Privacy policy and Terms of use for more information.

Sponsor: Troutman Amin LLP, Deserve to Win.

Rich Kahn read yesterday's Field Kit and wrote back before most of the industry had finished its coffee. He runs Anura. He sells invalid traffic detection, which you should hold in your head for the whole of this piece. His note ran four paragraphs and ended like this.

"The Programmatic industry has been built on fraud. It's a shame."

His claim was that HUMAN runs a general invalid traffic scan only at pre-bid, that general invalid traffic is one to three percent of the real problem, and that nobody has said whether a sophisticated invalid traffic scan happens at all. He wrote that clients on The Trade Desk are not protected, and he capitalized the not.

He is wrong about the scan. He is right about the protection, and the documents that prove him wrong on the first point prove him right on the second.

We spent the morning in the paperwork. Here is what is actually written down, by the vendor, by its accreditor, and by a United States senator.

The part Rich got wrong

HUMAN's pre-bid product is accredited for sophisticated invalid traffic. The Media Rating Council's accreditation letter for HUMAN, dated April 4 2025, covers the Ad Fraud Defense pre-bid platform for both general and sophisticated invalid traffic requests across desktop, mobile web, mobile in-app and connected television.

All four environments. No carve-out for CTV, which is where BADBOX 3.0 lived.

HUMAN's own dashboard documentation says the same thing from the other direction. The MediaGuard screen reports flagged sophisticated invalid traffic requests and flagged general invalid traffic requests as separate lines, in counts and percentages.

So the sentence "HUMAN is doing a GIVT scan only in pre-bid" does not survive the accreditation letter. Print it that way and the company answers with the letter, and your week goes sideways.

Now read the verb in that dashboard line, because it is the whole story. Flagged.

What the accreditation actually accredits

The accredited statistics for HUMAN's pre-bid platform are a list of counts. Total requests. Invalid traffic requests. General invalid traffic requests. Sophisticated invalid traffic requests. Numbers and percentages, correctly labeled.

The accreditation certifies that HUMAN can tell you how many fraudulent requests it saw. It does not certify that one of them was stopped.

This is not a loophole somebody found. It is what measurement accreditation is. The letter speaks the language of requests and percentages because counting is the thing being audited.

So when a platform tells you it uses an MRC-accredited pre-bid partner, the honest translation is that the scores it receives are counted accurately. What happens next is not in the letter, and was never meant to be.

HUMAN wrote down who decides, and it isn't HUMAN

This is the part that should have ended the argument years ago, and it has been sitting on a public developer documentation site the entire time.

HUMAN's MediaGuard overview describes the product as making real-time predictions about the validity of the user behind each bid request, and then providing the platform with "a recommendation to bid or not bid on the associated ad inventory."

A recommendation. Their word, on their own documentation, for the thing being sold to buyers as protection.

The same page gives an example of what a customer might do with it. You might configure your bids to automatically drop any request that comes back with an IVT value of true. Might. The page explaining how the pre-bid and post-bid products fit together is even plainer: based on that response, you can then choose to bid or not bid.

Might configure. Can choose. That is the vendor describing its own product, and it is not describing a shield. It is describing a consulting opinion delivered in under twelve milliseconds.

Integral Ad Science has published the same admission in nearly the same words. Its fraud methodology page states that "DSP partners decide whether or not to bid" on a given request, and describes IAS receiving selected fields, running its detection, and telling the platform whether it would consider the impression invalid.

Two of the three largest verification companies in this business have written down, in their own marketing material, that they cannot stop a bid.

There is a code for when the scan doesn't answer

This is the detail that stopped us cold, and it is the quietest damning document of the lot.

HUMAN's post-bid product collects a field recording what happened at pre-bid. It is an integer, and the documented values include: the bid request was sent and the scan answered successfully. The request was sent and the scan timed out without responding. The request was sent and the scan returned an error.

Somebody wrote code for the scan not answering, which means the scan sometimes does not answer, which means somebody had to decide what happens when it doesn't.

Nothing public says what that decision is. The documentation records the failure. It does not require a behavior.

So here is a question no buyer has ever put to a DSP, and it may be the best one in this entire story. When the fraud scan times out, do you bid anyway?

There are only two answers. Bid anyway, and your protection has an outage window nobody has ever disclosed to you. Don't bid, and you are discarding inventory every time a vendor has a bad afternoon, which is a cost you also were never told about.

Ask. Then ask how many times it happened last month.

The body that accredits it says not to do it

The MRC published a statement on pre-bid invalid traffic in April 2025. Its position: it does not require pre-bid filtration or blocking. Back-end filtration is the requirement, applied impression by impression to everything measured, whether or not a pre-bid tool was used at all.

And it does not merely decline to require pre-bid blocking. Its standards discourage it, for two reasons worth understanding.

The first is that blocking in real time teaches the fraudster what you caught. The MRC calls upfront detection prone to telegraphing detection techniques. Block a bid and the operator on the other end learns, instantly and for free, which fingerprint is burned. He changes it tonight.

HUMAN agrees, in writing. Its best-practices page warns platforms that fraudsters are eager to reverse-engineer the tools and instructs them to ensure "none of MediaGuard's predictions are exposed to the outside world", specifically that an invalid verdict must never reach the device the request came from.

Read those two documents together and the industry's position becomes very strange. The vendor and the accreditor both say that acting visibly on a pre-bid verdict degrades the detection. The sales material sells acting visibly on a pre-bid verdict.

The second MRC objection is the one that costs you money. Pre-bid blocking, it warns, may mask the occurrence of invalid traffic in back-end reporting. Filter it before the bid and it never enters the numbers you use to grade a seller. Your fraud rate improves because the evidence left before the count began.

Then the sentence, with the company named. The decision to bid or serve rests solely with the platforms, and "Neither DV nor HUMAN have control over these decisions."

A senator asked Rich's question in March 2025

Rich thinks nobody hears him. Somebody with subpoena-adjacent power asked his exact question nineteen months ago, and the industry's answer is still outstanding.

In March 2025, Senator Mark Warner wrote to the chair of the Federal Trade Commission and to the Attorney General, citing Adalytics research, and named Integral Ad Science, DoubleVerify and HUMAN Security. He named Google's DV360 and The Trade Desk as the demand-side platforms.

Among the questions he asked the FTC to investigate was whether the vendors' pre-bid technology can actually stop ads from serving to declared bots on the industry's own published bot list, or whether it merely prevents ads from serving on "website domains with historically high levels of bot traffic."

That is Rich Kahn's email, on Senate letterhead, in March 2025. Not whether the vendors are competent. Whether the product is the thing its name implies, or a reputation filter on domains wearing the costume of a real-time scan.

He asked the FTC two more things worth your attention. Whether the vendors can even access the user agent field in a real-time auction. And whether claims made without that access amount to deceptive practice.

To the Justice Department he asked something sharper. Whether those firms misrepresented their capabilities to federal clients, and whether they violated the False Claims Act by charging the government or its contractors for services they did not deliver.

No public finding has resolved any of it. The request is documented. The answer is not. Nineteen months, and the largest connected TV fraud operation ever disclosed ran through most of them.

What is actually on the buyer's control panel

If you want Warner's distinction made concrete, open the help page for the pre-bid fraud setting in the biggest buying platform in the world.

Google's documentation for IAS media quality in DV360 describes the ad fraud prevention control as a setting to "exclude impressions if there's a risk (moderate or high)" that a website, or the ads appearing on it, are associated with deceptive activity such as impression fraud. The same help documentation elsewhere explains that IAS provides data about each site, and the buyer uses that data to decide whether to purchase that site's inventory on a pre-bid basis.

Read it as a buyer rather than as a marketer. The control is a site risk level. Moderate or high. About the website.

That is a reputation filter. It may be an excellent reputation filter. It is not a real-time per-impression bot catcher, and the distance between those two products is the entire subject of a federal referral.

And on connected television, where there is no page and often no JavaScript, a site-reputation control is being asked to do a job it was not built for, against an operation that was impersonating legitimate apps.

The threshold is the product

One line contains the whole category, and nobody will answer it.

At what point does your DSP decline to bid?

That is the policy. Everything upstream of it is instrumentation. The scanner can be accredited, the signals excellent, the partnership announced with a chief executive quote, and none of it touches a bid until somebody at the platform decides what to do with a recommendation.

Nobody publishes that decision. It is not in the accreditation letter, because the MRC does not audit it. It is not in the partnership announcement, which is a marketing document. It is not in your contract, which talks about invalid traffic the platform identifies, a different thing again.

And if the rule changed on Wednesday afternoon, after HUMAN published, nobody owes you a notice, because there is no disclosure obligation attached to a setting that officially does not exist.

Three questions, and they fit in one email

To your DSP. When HUMAN or your verification partner returns an invalid verdict on a connected TV bid request, do you drop the bid automatically, every time, on every supply path? If not, what are the exceptions?

To your DSP again. When the pre-bid call times out or errors, do you bid anyway? How many times did that happen on my campaigns last month?

To your verification vendor. Your documentation describes your pre-bid output as a recommendation and says the platform decides. When your sales material says my campaign was protected, which of those two companies are you describing?

The last one has no comfortable answer. Stand by the documentation and the protection claim becomes somebody else's to defend. Stand by the sales material and you are contradicting your own developer docs, your accreditor, and a letter sitting at the Federal Trade Commission.

The mic drop

Rich Kahn apologized for the size of his audience. He wrote that his soap box is small, so nobody hears him shouting from the rooftops, and that he has been at it for decades.

He got one label wrong and the thing underneath it right, which beats the week the press releases are having.

Because here is where the documents leave us. The scanner is accredited to count sophisticated invalid traffic and describes its own output as a recommendation. The accrediting body says the scanner cannot control the decision, does not require the blocking, and warns that blocking can hide the problem from the report you use to check the work. The platform owns the only decision that costs a fraudster anything and will not say what it set that decision to. A United States senator asked whether the product is what its name implies and got no public answer. And forty billion fake requests a day moved through all of it for two and a half years.

Every company in that paragraph can truthfully say it did its job.

That is not a failure of the system. Read the documentation again. That is the system, working exactly as written.

How we reported this. Built from HUMAN Security's MRC accreditation letter dated April 4 2025, the MRC's April 2025 statement on pre-bid invalid traffic requirements and its invalid traffic definitions, HUMAN's public MediaGuard and FraudSensor developer documentation and its Prebid module documentation, Integral Ad Science's published fraud methodology, Google's DV360 help documentation for IAS media quality, Senator Mark Warner's March 2025 letters to the Federal Trade Commission and the Department of Justice and his office's summary of them, HUMAN's sales material quoting The Trade Desk, and correspondence sent to ADOTAT by Rich Kahn, chief executive and co-founder of Anura Solutions, which sells invalid traffic detection and competes with companies named here. Kahn's figures are his own and are not independently verified; we have asked him for the sample base and window. His claim that HUMAN scans only for general invalid traffic at pre-bid is contradicted by the accreditation letter and appears here as a corrected claim, not a finding. Warner's letters are requests for investigation; no agency finding has been published. Nothing here is a finding that The Trade Desk, HUMAN Security, Integral Ad Science, DoubleVerify or any named company carried, bid on or billed fraudulent traffic. Right of reply is going to HUMAN, The Trade Desk, Integral Ad Science and the MRC today; this piece will be updated with what they say or with a note that they declined. ADOTAT is reader-supported. No paid placement in editorial.