Industries / Technology / Software and SaaS / Ad and marketing tech
Ad fraud
Ad fraud is selling fake views, clicks and app installs as real ones; this domain covers the schemes and the software that finds and filters them.
6
Milestones achieved, from 2006 to 2025.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
2006
Click fraud settled in court
Click fraud is clicking pay-per-click ads, by hand or by software, with no wish to buy, so the advertiser pays for nothing. On 27 July 2006 Judge Joe Griffin of Miller County, Arkansas, approved Google's $90 million settlement with advertisers led by Lane's Gifts and Collectibles: advertising credits worth about $3.80 for every $1,000 spent over four and a half years, and $30 million for the lawyers. Google did not admit liability. An expert report for the case by Alexander Tuzhilin of New York University's Stern School found Google's efforts against click fraud "reasonable", but concluded that no definition of an invalid click "can be operationalized", turned into a test software can run, because each depends on the clicker's intent.
LessonSoftware can judge behaviour but not intent, so Tuzhilin set out working definitions instead: rules written by experts, deviations from normal traffic, and trained classifiers.
Credit
Miller County Circuit Court; Alexander Tuzhilin
2015
Two grades of fake traffic
The Media Rating Council (MRC), the US industry body that audits audience measurement, issued its Invalid Traffic Detection and Filtration Guidelines in October 2015 and updated them in June 2020. They split invalid traffic (IVT) into two kinds. General invalid traffic (GIVT) is caught with lists and routine checks, such as known data-centre addresses and self-declared bots. Sophisticated invalid traffic (SIVT) needs "advanced analytics, multi-point corroboration/coordination, significant human intervention", and includes hijacked devices, emulators and paid human clicking. Measurement firms are audited against these rules for MRC accreditation.
Credit
Media Rating Council
2019
Every reseller on file
ads.txt (2017) lets a website list who may sell its ad space. On 31 July 2019 IAB Tech Lab's OpenRTB working group finalized two more specifications: sellers.json, a public file in which an exchange lists the accounts that sell through it and whether each is the owner of the space or a reseller, and the SupplyChain object, which travels with each bid request and names every company that sold or resold it. Google publishes a sellers.json file; the copy downloaded for this page on 20 November 2026 lists 885,466 sellers, of which 616,436, or 70%, are marked confidential, with no name shown, and the file still says "This file is a beta and is unverified."
LessonThe files help only as far as sellers are named and checked; Yash Vekaria, Rishab Nithyanand and Zubair Shafiq found misinformation sites pooling their ad space with unrelated sites through a few major exchanges to get past advertisers' brand-safety filters.
Credit
IAB Tech Lab
2020
Uber wins money back
Uber said its mobile ad campaign, run by the agency Fetch Media from 2014 to early 2017, was "fraud-infested": it had "overpaid Fetch and mobile advertising providers due to fraudulent attribution" of installs of the Uber app. Attribution is the record of which ad gets the credit, and the payment, for an install. In 2019 Uber added racketeering claims against the ad firm Phunware and its chief executive. On 9 October 2020 Phunware agreed to pay Uber $4.5 million and its insurer $1.5 million more; every party denied wrongdoing.
LessonPaying for credited installs pays whoever can claim the credit, including networks that claim installs they did not cause.
Credit
Uber; the settlement is reported in Phunware's filings with the US Securities and Exchange Commission
2021
Prison for a bot farm
A US jury convicted Aleksandr Zhukov in May 2021 of wire fraud and money laundering, and in November 2021 he was sentenced to 10 years in prison. Prosecutors said his "Media Methane" ad network, part of the scheme known as Methbot, ran thousands of fake websites on more than 2,000 rented data-centre servers with no real visitors, loaded ads on them, used software to imitate people viewing them, and leased or hijacked internet addresses to hide where the traffic came from. The Record reports that it took more than $7 million from 2014 to 2018.
Credit
the US Attorney's Office for the Eastern District of New York and the FBI, with Google and White Ops (now HUMAN)
2025
Fraud built into TV boxes
BADBOX 2.0 is malware that comes preinstalled on, or is downloaded to, cheap Android-based TV streaming boxes, projectors, tablets and car screens. On 5 June 2025 the FBI warned that the botnet, a network of infected devices controlled from outside, "consists of millions of infected devices" and sells access to home internet connections. Google told a federal court in New York that it held more than 10 million devices as of April 2025 and ran ad fraud through hidden ads and through games that load ads in hidden browsers; the court granted a temporary restraining order on 30 May 2025 and a preliminary injunction on 27 June 2025, finding Google likely to succeed.
LessonFraud moved from rented servers into devices in people's homes, so its traffic now comes from real home internet addresses.
Credit
HUMAN's Satori research team, Trend Micro and the Shadowserver Foundation with Google; FBI
Watch for
AI bots click ads
DoubleVerify, an ad verification firm, says that "in unprotected media, AI bots accounted for up to 15% of all clicks", that in some tests an ad click was "10X more likely to be from an AI bot than from a human", and that its new Agent ID Measurement sorts declared AI agents into GIVT and undeclared ones into SIVT.
Achieved whenthe MRC accredits a firm's detection of AI agents, or an independent test published with its method measures the share of ad clicks made by AI agents.
Claimed by
DoubleVerify, December 2025
Spotting AI browsing agents
AI browsing agents are AI models that run a real web browser to do tasks for a person, which makes them hard to tell from people. In a preprint, a paper not yet peer reviewed, Ethan Wang, Zubair Shafiq and Yash Vekaria report that their classifier, FP-Agent, picked out all seven agents they tested from humans by how they type, scroll and move the mouse, while Cloudflare's bot detection caught only one.
Achieved whenthe paper passes peer review, or another team repeats the result on other agents and on ad traffic.
Claimed by
the authors, May 2026
Open
01
Score ad filters on agents
A public test in which AI browsing agents do ordinary tasks on test pages that carry house ads (a site's ads for itself, so no advertiser pays), recording how often each agent's ad views and clicks pass as human or are filtered as invalid by free detectors and by any verification firm that agrees to take part.
The briefCredit · unclaimed02
Signed agents in bid requests
A check on a publisher's web server for Web Bot Auth signatures, a cryptographic proof a bot or AI agent attaches to its web requests to say who sent it, plus a Prebid module that copies the result into every bid request, so that buyers learn a visit came from a declared agent before they bid.
The briefCredit · unclaimed
Open 01 · the brief
Score ad filters on agents
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test in which AI browsing agents do ordinary tasks on test pages that carry house ads (a site's ads for itself, so no advertiser pays), recording how often each agent's ad views and clicks pass as human or are filtered as invalid by free detectors and by any verification firm that agrees to take part.
Why it is still open
The MRC's latest invalid-traffic update, of April 2024, names no category for AI agents. The FP-Agent study tested a general bot defence, not ad filters. DoubleVerify reports AI bots making up to 15% of clicks in unprotected media but has not published its data or method. Web searches in November 2026 found no published test of ad filters against named AI agents.
Built with what exists
Revive Adserver (GPL 2.0 licence) to serve the house ads; Perplexity's Comet browser and agents built on Browser Use (see the Note "Deleting 80% to start over"); FingerprintJS's BotD (MIT licence), an open-source bot detector that runs in the page; Playwright scripts as a plain-bot comparison.
First step
Install Revive Adserver on a test site with three house ads, add BotD, and log one agent through 20 tasks.
Done looks like
At least five agents, each run through 100 scripted tasks, with a published table of the share of their ad views and clicks that each detector labels invalid, beside the same share for plain Playwright scripts, released with the test pages and code.
Open 02 · the brief
Signed agents in bid requests
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A check on a publisher's web server for Web Bot Auth signatures, a cryptographic proof a bot or AI agent attaches to its web requests to say who sent it, plus a Prebid module that copies the result into every bid request, so that buyers learn a visit came from a declared agent before they bid.
Why it is still open
Web Bot Auth is still a draft: the IETF working group of that name published its protocol draft on 29 October 2026, and the draft says a valid signature tells nothing about "whether the Agent is benign", which it leaves to each website's policy. Cloudflare documents it as a way for websites to verify bots and agents, not for ad systems. A web search in November 2026 found no ad tech product or Prebid module that reads it. OpenRTB does not require an exchange to pass on a visitor's user agent, the browser's description of itself, which AdExchanger's sources gave in March 2025 as a likely reason ads were served to a crawler that declared itself as a bot.
Built with what exists
Cloudflare's published guide and the IETF drafts it cites, which build on HTTP Message Signatures (RFC 9421); Prebid.js and Prebid Server (Apache 2.0 licence); a self-built test agent that signs its requests with a key in a published key directory, following the guide's steps. No real person's traffic is involved.
First step
Put a test page behind a server that verifies Web Bot Auth signatures, and log whether a signing test agent's signature arrives with the page request.
Done looks like
A published module and test site on which signed test agents are marked as agents in at least 99% of 1,000 bid requests and unsigned Playwright scripts in none, with the proposed field name sent to IAB Tech Lab's OpenRTB working group.