Half the audience isn't people #
Imperva's 2025 Bad Bot Report found that automated traffic was 51% of all web traffic in 2024, and that bots built to do harm were 37% of it on their own. Imperva puts the jump down to AI and large language models, which made a working bot something almost anyone can build.
Advertisers pay for a large share of that traffic. Lunio's Global Invalid Traffic Report 2026 measured an average invalid traffic rate of 8.5% across ad channels and put the cost at $63 billion. Lunio expects the number to grow this year, and names the reason: AI agents. On one platform built for professionals, fraud0 measured 19.88% of LinkedIn paid traffic as invalid in 2025. That's about one paid click in five on the network most B2B budgets treat as the safe one.
Why the old checks stopped working #
Most fraud filtering looks at behavior. Did the visitor scroll, how long did they stay, did the mouse move like a hand, did they pass the attention check. Those were good signals when a bot was a script that loaded a page and left. An agent built on a language model reads the page, waits a human amount of time, moves the cursor and types an answer that makes sense.
Survey research found this out first. In a 2025 paper in PNAS, Sean Westwood of Dartmouth built what he called an autonomous synthetic respondent. It took on a demographic persona, simulated reading time and mouse movement, and typed open answers one keystroke at a time. Coverage of the paper reports it got past standard attention checks and bot tests 99.8% of the time. Critics fairly point out that this shows what is possible in a lab, and that nobody has documented it at scale in real panels yet. The checks it beat are the same behavioral kind that ad fraud filters lean on.
There is a plainer problem underneath. On most ad platforms, the company that counts your clicks is the company you pay for them. Nobody on that side of the table loses money when an agent clicks.
What a company can gain #
When half the audience can be faked, a confirmed person paying real attention is worth more than it used to be, and that is where the gain is. A company that changes what it is billed for gets two things back: the budget it was losing, and data it can trust again. The only change needed is to make the billable event something an agent can't produce cheaply. Three things are hard to fake together:
- A real, checked identity. A government ID and a live selfie cost a fraud operation far more than a fresh browser profile does.
- Proof of understanding. A quiz on the specific content, with no partial credit, means the person actually read the content they were shown.
- A written answer that someone checks before any money moves, with the money returned when it fails.
That is what we built Nexertise to do. Every professional who earns on it has passed Stripe Identity (a government photo ID plus a live selfie) and confirms a device passkey such as Touch ID before each engagement. They read the brief for at least 25 seconds, pass a three-question quiz with all three answers right, and write three answers of 20 to 100 words each. Every submission is graded before the advertiser is charged. If it fails, the budget goes back to the campaign. Pricing runs from $5 to $75 per approved engagement, depending on the professional's tier.
Each engagement costs more than a click does, because each one has passed every check above before it is billed.
What this does not claim #
We don't claim zero bots, and we don't claim any single check can't be beaten. An identity check proves a real person is behind the account. It does not prove that person didn't ask a chatbot for help with their answer, which is why every response also goes through AI-text detection and a quality review. We also don't verify job titles; that field is self-entered, and campaigns target by industry and reputation tier. The volume is far smaller than an ad network's. The full list of checks, and what each one does not do, is on how verification works.
Sources #
- Imperva, 2025 Bad Bot Report (2024 data): 51% automated traffic, 37% bad bots.
- Lunio, Global Invalid Traffic Report 2026, as reported by the ANA (January 21, 2026) : 8.5% average invalid traffic rate, $63 billion.
- fraud0, 2025: 19.88% of LinkedIn paid traffic and impressions invalid.
- Sean J. Westwood, “The potential existential threat of large language models to online survey research,” PNAS, 2025; 99.8% figure as reported by 404 Media .