This website uses cookies

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

Sponsor: Troutman Amin LLP, Deserve to Win.

Yesterday we published a piece arguing that the number which actually governs programmatic fraud, the point at which a platform declines to bid, is never published by anyone.

That was wrong, and the correction is worse than the claim.

The platform's own internal bid threshold is still unpublished. But the setting the buyer controls is published, in full, with constant names, in Google's developer documentation for the largest demand-side platform in the world. It is a ladder. You can climb it.

The top rung reads "Ad Impression Fraud 75% or higher."

That is an avoidance option. Selecting it means you have asked your campaign to steer clear of sites and apps where at least three quarters of the traffic is classified as fraud. It also means you have not asked it to steer clear of anything below that.

The ladder, in full

Google's structured data file reference for third-party verification in DV360 lists DoubleVerify's ad impression fraud tiers as constants. Here they are, in the order the documentation gives them.

Unknown. 2 percent or higher. 4 percent. 6 percent. 8 percent. 10 percent. 25 percent. 50 percent. 75 percent or higher.

The lowest available rung is 2 percent. There is no option below it. A buyer cannot say "avoid anything with measurable fraud," because the vocabulary does not contain that sentence.

This is not a Google quirk. IQM, a smaller demand-side platform, publishes the same DoubleVerify ladder in its own help center, describing segments that exclude sites or apps by the share of fraud detected, "ranging from insufficient data to 100% IVT."

Read that last phrase slowly. There is a tier for a property whose traffic is entirely invalid. It exists because somebody needed it to exist.

What a setting like that actually does

Set your avoidance at 50 percent and here is the arrangement you have entered into.

A site where forty-nine percent of the traffic is classified as fraud is, as far as this control is concerned, inventory. Your money goes there. Your impressions count. Your dashboard reports a campaign that ran.

The control did its job. It was asked to avoid places above a line, the place sat below the line, and the ad served. Nobody malfunctioned. Nobody lied to you. You picked a number and the machine honored it.

Which raises the only question that matters about the ladder: who picked yours?

Most buyers have never seen this list. The tier is set inside a platform, often by an agency trading desk, sometimes years ago during onboarding, sometimes by whatever the default was. It is a number with enormous consequences and almost no governance.

And it is a site-level number. It describes a property's historical fraud rate, not the request in front of you. A clean request on a dirty site is blocked. A fraudulent request on a clean site goes through. That is what a reputation filter does, and it is what Senator Mark Warner asked the Federal Trade Commission about in March 2025, in the sentence we quoted this morning.

The second number, and where DoubleVerify said it

In May 2026 a federal judge in Maryland described what DoubleVerify had conceded in its own complaint against Adalytics.

The court's account: the complaint acknowledged that DoubleVerify's pre-bid services do not prevent all ads from reaching bots, because it stated that those services filter out "99% of unwanted GIVT impressions."

Hold the number. Look at the letters.

GIVT. General invalid traffic. The MRC defines it as what routine filtration catches by applying lists and standardized parameter checks. Declared bots. Known data centers. The traffic that announces itself.

Sophisticated invalid traffic is the other category, the one the standards define as requiring advanced analytics and human intervention, built specifically so that a list will not catch it.

When DoubleVerify went to court to defend the effectiveness of its pre-bid product, the unit it reached for was GIVT.

Its public response to Adalytics uses the same frame. With pre-bid avoidance enabled, it says, known self-declared bot traffic shows up on under three hundredths of one percent of programmatic impressions, which it calls a reduction of 75 to 98 percent "depending on client settings."

Depending on client settings. That is the ladder again, named by the vendor, in the same breath as the performance claim.

This is not an accusation. It is a scope.

Nothing here says DoubleVerify's pre-bid product cannot see sophisticated invalid traffic. DoubleVerify is MRC accredited for sophisticated invalid traffic detection. So is HUMAN, pre-bid, across all four environments, as we laid out this morning. The capability is real and the accreditation is real.

What the documents show is narrower and more useful. When these companies quantify what pre-bid accomplishes, in a courtroom or in a rebuttal, the number they produce is a general invalid traffic number. The easy category. The one with a list.

Rich Kahn wrote to us yesterday that pre-bid scanning is a general invalid traffic exercise and that general invalid traffic is one to three percent of the real problem. We told him this morning that he was wrong, and we were right to. HUMAN's accreditation letter says otherwise about capability.

But he was describing the output, not the capability, and on the output the paperwork is closer to him than to the marketing.

He runs Anura, which sells the thing he says is missing. That has not changed and you should keep weighing it. It is simply less relevant when the corroboration arrives from the defendant's own pleading.

The case that is now going forward

DoubleVerify sued Adalytics in May 2025 over the March 2025 pre-bid bot report, alleging defamation and false advertising. Adalytics moved to dismiss, calling the suit an effort to shut down protected criticism of an industry.

Judge Theodore Chuang let the case proceed.

The reasoning on the false advertising claim is worth understanding, because it is going to shape what gets published about this industry for years. The claim survived because DoubleVerify alleged that Adalytics had an economic motive for producing the report. Research published by a company that also sells services can be framed as commercial speech, and commercial speech gets less protection.

On defamation the court noted that most of the challenged statements were not alleged to be literally false, which is why the theory in play is defamation by implication, a harder road under Maryland law.

Adalytics is also being sued by Colossus SSP.

Set aside who should win. Look at the structure. The people with the capital to run this kind of forensic work at scale are, almost by definition, people with something to sell. If selling something converts your findings into commercial speech, the only safe researcher is one with no business model, and there are not many of those with petabytes of bid stream data.

That is worth saying out loud by a publication that takes sponsorship money and investigates the same industry. We have an interest too. It is at the bottom of this page, where it belongs.

Nineteen months, no answer

In March 2025 Senator Mark Warner asked the Federal Trade Commission whether pre-bid technology can actually stop ads reaching declared bots or merely avoid domains with high historical bot traffic, and asked the Justice Department whether vendors billed the federal government for services not delivered.

He wrote that the failures amount to "far more than simple contradictions of their marketing puffery."

We can find no public FTC or DOJ finding resolving any of it. That is a search result, not proof that nothing is happening behind a closed door.

What has happened in those nineteen months is that one vendor sued its critic and survived a motion to dismiss, a second vendor's critic was sued by an exchange, and the largest connected TV fraud operation ever disclosed ran for most of the period.

And it is not a new question. Adalytics' report cites earlier mystery shopping work by Shailin Dhar in which a well-known verification vendor, in testing, "was only able to catch 14% of bot traffic."

That was 2016.

Go look at your own setting

This one takes ten minutes and you do not need anyone's permission.

One. Open your DSP and find the pre-bid fraud and invalid traffic avoidance setting on a live campaign. On DV360 it is the third-party verification section. Write down the tier that is selected.

Two. Ask who selected it, and when. If the answer is the agency, ask for the date and the person. If the answer is nobody, you are on a default.

Three. Ask what the tier costs. Google's own help documentation notes that using multiple brand safety controls means being billed for the combined total of the types used. Find out what you pay for the setting, then compare it to what the setting excludes.

Four. Ask the same three questions about connected TV specifically, where there is no page, often no JavaScript, and a site-reputation control is doing work it was not designed for.

Five. Ask your verification vendor to restate your last six months of campaign fraud rates as though your tier had been set at the lowest available rung. If the answer is that they cannot, you have learned what the dashboard is.

The mic drop

This morning's piece ended on the idea that the system works exactly as written. Here is how well it is written.

The scanner describes its own output as a recommendation. The accrediting body says the scanner cannot control the decision and discourages the blocking. The platform owns the decision and does not publish its threshold. The buyer owns a different threshold, which is published, and whose most permissive setting tolerates a property that is three quarters fraud. And when the vendor quantifies the whole apparatus under legal pressure, the unit is the easy category.

Every one of those is a defensible engineering choice. Nobody has to have done anything wrong for all of it to be true at once.

That is the part that should bother you. There is no villain required. There is only a stack of reasonable settings, each one owned by somebody who can point at the next one.

Go find out what yours is set to.

How we reported this. Part two of a piece published this morning, and a correction to it. Built from Google's published DV360 structured data file reference for third-party verification, which lists the DoubleVerify ad impression fraud avoidance tiers and their constant names; IQM's published help documentation for DoubleVerify pre-bid segments; the May 2026 opinion of US District Judge Theodore Chuang in DoubleVerify's suit against Adalytics, as reported by MediaPost and analyzed by Rebecca Tushnet, including the court's account of what DoubleVerify's own complaint conceded about pre-bid GIVT filtration; DoubleVerify's published responses to the Adalytics March 2025 report; Adalytics' pre-bid bot filtration report and its citation of Shailin Dhar's earlier vendor testing; Senator Mark Warner's March 2025 letters to the Federal Trade Commission and the Department of Justice and his office's summary of them; and the MRC's definitions of general and sophisticated invalid traffic. We have found no public FTC or DOJ finding on the Warner referrals; absence of a public finding is not evidence of inaction. Nothing here is a finding that DoubleVerify, HUMAN Security, Integral Ad Science, Google or The Trade Desk carried, bid on or billed fraudulent traffic, or that any avoidance tier was negligently selected by any named party. Rich Kahn of Anura Solutions, quoted in part one, sells invalid traffic detection and competes with companies named here. Right of reply went to HUMAN, The Trade Desk, Integral Ad Science and the MRC earlier today, with answers requested by Monday; DoubleVerify is being contacted on this piece. ADOTAT is reader-supported. No paid placement in editorial.