Automated bidding works by buying more traffic that resembles your past converters. So a single fake conversion does not cost you once. It enters the model as a training signal, and the algorithm spends the next days hunting down more traffic that looks like the source that produced it. The click was the cheap part. The recurring spend the fake conversion triggers is the expensive part, and it keeps running until something corrects the record.
Bots now make up the majority of web traffic — 51% in 2024, with malicious bots at 37% (Imperva) — and they don’t stop at the click. A share of completed conversions are automated rather than human, and every one of those that your bidder learned from is still shaping where your budget goes.
What automated bidding is actually optimising for
Target CPA, Target ROAS, Maximise Conversions, Advantage+. The products differ at the edges, but underneath they share one mechanism: look at which clicks turned into conversions, then buy more clicks that resemble those. The model reads your conversion feed as ground truth and chases the patterns inside it.
That is usually a good mechanism. It is also completely dependent on the conversion feed being honest. The algorithm cannot ask whether a conversion was a person or a script that filled four fields in 90 milliseconds. It treats every counted conversion as a real one and learns from it. Feed it a fake, and it learns the fake just as faithfully as a true sale.
The doubling cost, concretely
Walk one fraudulent conversion through the system. A bot clicks your ad — that is the first cost, one click’s worth of budget. The bot then completes your lead form, and your pixel fires. Now your reporting shows a conversion attributed to a campaign, an ad group, an audience, a placement.
The bidder reads that conversion the next time it adjusts. It sees a source that produced a win and shifts budget toward more traffic with the same shape: the same placement, the same audience signals, the same geo, the same device profile the bot presented. That is the second cost, and it does not happen once. It recurs every optimisation cycle, every day, for as long as the fake conversion sits in the training data as a success. One $4 fraudulent click can steer hundreds of dollars of subsequent spend toward the bot’s lookalikes.
This is why fraud aimed at conversions is more damaging than fraud aimed at clicks. The ANA found only $439 of every $1,000 of programmatic ad spend reaches a real consumer, so the click waste alone is large. The conversion waste compounds on top, because it teaches the machine to keep paying.
Why blocking the click later does nothing about the second cost
Most fraud tooling acts at the click. It detects an invalid click and refunds it, or excludes the IP so it stops clicking. That recovers the first cost. It does nothing about the second.
By the time you block a click, the conversion it produced has already fired and already trained the algorithm. The bidder set its targets against the inflated number. It moved budget on the strength of a win that never happened. Excluding the IP stops new clicks from that address, but the damage to the model is already booked, and the model will keep chasing the pattern the fake conversion taught it. You have closed one door after the spend already walked through it.
Grading and retracting the conversion, keyed by click id
The fix has to act where the second cost is created, which is the conversion, not the click. ClickLens grades each conversion on its provenance — whether the submit was browser-trusted, whether a real pointer path preceded it, the time from page load to submission, paste-without-focus events, and the entropy of the field-fill order — plus funnel coherence. It reads how the form was completed, never what was typed into it.
A conversion that grades as fraudulent does not just get flagged in a report. On Google and Microsoft Ads, ClickLens writes the retraction back to the platform keyed by click id — gclid on Google, msclkid on Microsoft. The platform sees the conversion restated or removed against the exact click that produced it, so the bidder stops treating that source as a winner. When your CRM later marks a lead disqualified, that outcome can turn an earlier pass into a retraction candidate too, closing the loop through a signed webhook.
The platform write-back rolls out under shadow-mode measurement with a 10% holdout, so the correction is checked against live traffic before it moves real budget. Run a free audit to see what your conversions are teaching your bidder, or read how the grading is built and measured.
Sources
- Imperva (Thales), “2025 Imperva Bad Bot Report”, 2025. Accessed 24 June 2026. https://www.imperva.com/blog/2025-imperva-bad-bot-report-how-ai-is-supercharging-the-bot-threat/
- Association of National Advertisers (via WFA), “ANA’s 2024 Programmatic Benchmark Study”, January 2025. Accessed 24 June 2026. https://wfanet.org/knowledge/item/2025/01/21/ana-s-2024-programmatic-benchmark-study-progress-but-challenges-remain