Google Ad Exchange and AdSense monetization · I research programmatic auction mechanics and invalid traffic detection using Google patents and antitrust case materials
Beyond the research I build ad formats of my own. The first is a scratch banner that raises the price of the ad next to it — a working prototype with the economics laid out.
Hello, friend
I have spent six years on Google Ad Exchange (AdX) and AdSense monetization; a thousand sites have passed through my hands. My answers do not come from second-hand retellings or support macros — they come from Google's own patents and from the materials of the antitrust case against it.
Don't take my word for it. Ask your own question and see what the model says: every answer arrives with the list of documents behind it and a screenshot of the exact page the quote came from.
Why did my CPM drop after I changed the floor price?↑I charge nothing and sell nothing. To keep the bots out, access is password-protected — message me on Telegram or LinkedIn and I will send it over.
Let me be blunt: most MCM partners do not understand it either. Their work comes down to poking at GAM settings — nudge a floor, watch, roll back. I know dozens of MCMs who lost their status because they onboarded publishers with bad traffic and could not see it coming.
It is not laziness. It is that Google's public documentation describes the rules, not the mechanics. It cannot tell you why your traffic was valued the way it was.
AI made it possible to stop reading other people's summaries and instead take the primary sources and build a system that answers from them.
At this volume any model starts inventing: it produces plausible percentages that are nowhere in the documents. I checked — they all do it. That is why every question goes through four passes by four different models.
gpt-oss-120bdeepseek-v4-flashqwen3-235byandexgpt-32kThe verifying model is deliberately from a different family than the writing one: they fail differently, and what one invents the other will not confirm against the excerpts. It regularly catches invented percentages and non-existent names — which is the best proof the system works as intended.
Not a link to somewhere out there, but the specific patent sheet or presentation slide the quote came from. The system finds the right page by matching the text of the excerpt against the text of every page in the document.
The main base answers questions about auctions, bids and invalid traffic. But there are adjacent subjects that cannot be dumped into it: the system pulls twenty excerpts per question, and that is a hard ceiling. Put CAPTCHA patents in there and they start winning space away from IVT patents on the very questions where IVT is what you need.
So each such subject gets its own store and its own agent, and the code decides which one to ask based on the wording of the question:
Ask about CAPTCHA and the specialist agent joins in with its own base, its material marked separately in the combined answer. Ask about auctions and it stays out of the way entirely.
How a recording that sat in the base as dead weight became one of the most cited documents.
Among the case exhibits there is a fourteen-minute video. The storage accepted it as video and extracted nothing: no audio, no text, no images. I checked all 878 excerpts the system had cited in its answers — this recording appeared not once.
Inside were two layers, and both valuable. On screen, Google slides about the move to a first-price auction, with the explicit line "no last look for programmatic". Below them, a transcript burned into the frame as subtitles, numbered by page and line. Real people are talking: the Exchange Bidding product manager explains the changes, publisher representatives push back.
"All that control — and control is the real keyword here — is going to be lifted from us, and we just have to hope Google is acting in our best interest. That is a lot to swallow" — from the publisher side, lines 40:4–40:7.
No speech recognition was needed: the text was already on screen, and it carried transcript page numbers that the audio does not have at all. I found the exact bounds of the subtitle strip, cut it out of frames every second and a half, and ran it through text recognition.
Two checks paid off. Deposition transcripts run twenty-five lines to a page, so lines numbered 33 and 34 were recognition errors and I dropped them. And I re-sampled the gaps in the numbering at half-second intervals: not a single line was added, which means those are genuine edit cuts. Only excerpts were entered into evidence, not the whole recording — now that is verified rather than assumed.
A recording that had never once taken part in an answer is now cited alongside the patents. Ask about per-buyer floor prices and the system answers with a reference to it — and shows you the page.
Four findings that change how you configure monetization. I verified each one myself in the source documents.
Out of a hundred advertiser dollars. Twenty go to the exchange. This is not an outside estimate — it is Google's own slide, entered into evidence.
In a second-price auction the winner pays the next bid down. If both top bids belong to the same buyer, that buyer sets its own price.
The patent describes not "good" and "bad" traffic but zones with a discount coefficient: 0.5, then 0.3, then zero. No notification is sent.
The person clicked through and came straight back. It is the share of such clicks that moves you into the discount zone. That is ad relevance and page speed, not bots.
"In auctions where GDN second-prices itself, Bernanke allows >14% margin capture" — the caption on Google's own slide, entered into evidence.
The research is open in two forms. One is a finished analysis you can simply read. The other is live access to the system, where you can ask your own question.
Twelve sections on how money and fraud actually work in Google advertising. Auction mechanics, where the fees go, how invalid traffic is caught, and why traffic gets discounted instead of rejected.
Ask your question and get an answer from the documents in about ten seconds. For a hard question there is deep analysis: it is split into parts, each searched separately, and the result is checked against the sources by a different model.
Password-protected access — message me and I will send it over.
I am not selling a way around anti-fraud — there isn't one, and attempts end with a lost account. I am offering something else: understanding what is happening to your traffic and why.
Once you know the mechanics you stop poking at settings blindly. You see which signals are collected about you, which of them you can change and which you cannot. And you know in advance how taking on the next traffic source will end.
Follow me on LinkedIn so our contacts are kept and you don't lose track of the research. I publish analyses as the base grows.
Follow on LinkedIn Message on Telegram