Bots in Google advertising: The efficiency trap of advertising costs

Bots in Google advertising: The economics of advertising spending and virtual installation

60% of the installations purchased for $220 are robots. This is not an accident, but the result of Google's advertising system working properly according to the rules it has been designed for.

I have run this experiment: a tool app that relies solely on Google UAC (Universal App Campaign) running volume and has a target CPI of $1.5. After burning US$220 in three days, 147 installations were reported in the background. Connected to AppsFlyer for review, 2% for the next day, 0.3% for the seventh day, and almost zero event triggers. Check the device ID and find a large number of "pixel perfect users" with the same subnet, the same User-Agent, and even the battery power locked at 100%. This is not that the brushing team is lazy, this is that they have accurately met UAC's conversion definition -click, download, open, and report the first-time opening event. The four steps are completed, with no more than one second in between.

The underlying logic of UAC is a greedy optimizer: it does not recognize "users", only "conversion events." Whatever signal you feed it, it will buy traffic wherever it goes. When you set "opening for the first time" as the only optimization goal, the system will crazily tilt towards traffic that can trigger the event at low cost-click farms, simulator group control, SDK counterfeiting tools, they are better at running this funnel than real people. Real people have to think, hesitate, and may delete seconds after downloading; robots don't need to, and they will send you back install and session_start after the script is finished, and the cost is only one-third of that of real people.

Opponents will say: Google has "invalid traffic filtering", Play Protect, a "post-installation attribution" deduction mechanism, and a complete refund process. This is true. I did anti-cheating in the ByteDance commercialization team in 2022, and I know how much big factories invest in risk control. But you have to see clearly the incentive structure: Google's risk control goal is to "keep the reputation of the platform from collapsing and big customers from running away", not to "your ROAS is positive." It filters traffic that is "obvious, large-scale, and will be exposed by the media." Those "high-quality robots" that are scattered in tens of thousands of equipment farms, only run 3-5 times per device a day, perfectly simulate human behavior trajectories, and have passed SafetyNet certification are "long-tail real users" in the risk control model. Deduction? They will be deducted, but usually 60 days after the end of the settlement cycle, when your cash flow is exhausted and your budget for the next quarter is set, you will be refunded a 15% voucher. Everyone will be decent.

This matter is essentially the price of efficiency. If you regard the "cost of getting customers" as the North Star, the platform will help you push the "cost of getting customers" to the physical limit-even if the "customer" does not exist. Just like the factory owner only evaluates "output", and the assembly line starts to build mobile phones with only shells and no chips; when assessing "lines of code", the programmer will write 50,000 lines console.log for you. Once an indicator becomes a goal, it ceases to be a good indicator (Goodhart's Law). UAC turned "installation" into the only currency, and the market naturally issued counterfeit banknotes.

Last year, I met Lao Chen, the founder of a sea tool manufacturer. He has a team of 20 people and a monthly advertising budget of US$300,000. He showed me their "growth meeting": everyone stared at the "installation cost" column backstage, dropped from 1.42 to 1.38, and everyone cheered and the bonuses were distributed. No one asked,"Will these people put on it still be around next week?" Half a year later, the cash flow was cut off, and the LTV/CAC models in the financing PPT were all calculated based on 1.38, and the real recovery ratio was 1:0.4. He complained to me: "We are obviously 'optimizing' every day, but how can we optimize it, get closer to death?" Because you are optimizing a lie.

The platform will not help you stop losses. Whether it is Meta or TikTok, the same goes for it. They sell "a certain sense of scale", not "uncertain business growth." If you want to buy increments, you have to take apart the "installation" yourself: stuff the first day of payment, the seven-day retention, the core event completion rate, and even the device fingerprint entropy into the conversion goal, even if the cost rises to $5. Or simply don't buy UAC, buy search terms, KOL, and brand terms-at least there is a person behind those traffic, and even if he scolds you a few words, he will be a real person.

The advertising system does not produce users, it only produces indicators. Indicators look like users, not users.