I’m a mathematician who ended up in advertising. PhD in mathematical modelling, eight years teaching at Kyiv Polytechnic, and then I found the one place where a model gets graded in cash every day: Google Ads. I’ve been doing that since 2006. Before I started my own company I spent two years as a research engineer at a UK firm building real time click fraud detection, which is where distrusting a reported number stopped being a personality trait and became a job title. I launched Doctor Ads in the UK in 2023, and over my career my clients have run more than $770 million in ad budgets. The difference is what I put under the microscope. Most ad audits grade your account against the platform’s own scoreboard, and the platform wrote that scoreboard. I grade it against your bank statement. In practice that means going line by line through spend that produced activity but never produced a customer. It’s dull work, and it’s where the money is. Analytics tells you what happened. Forensics tells you what you’re missing.

Referrals and speaking. Someone hears me at a conference, or someone who did sends me a name. Rankings help too. Being ranked number 6 in the 2026 Top 50 Most Influential PPC Experts put me in front of people who were already looking and had no idea I existed. The channel I actually want more of already exists, it’s just slow. I wrote a book on Google advertising, published in three languages, number 1 New Release on Amazon US, and the people who arrive from it are a different species. They’ve recognised their own account in somebody else’s numbers before they ever contact me. An introduction sends me a person who likes me. A number sends me a person who has already diagnosed themselves and wants a second opinion. Here’s the unglamorous part. Referrals don’t scale, and they make you lazy about proving anything. The month you stop publishing, the pipeline is still full, so you never feel the cost. You feel it a year later.

I handed over the first pass. Reading millions of rows of search terms, change history and auction data isn’t human work. Machines flag the anomalies and rank them by money at risk. That used to be a week of an analyst’s life, and the analyst did it worse. I took back the verdict. We tried letting the model write the conclusion from its own flags. The arithmetic was always right and the conclusion was often wrong, because the model reads the account and the answer lives in the business. A search term with zero conversions looks like pure waste until you learn it’s brand defence. I recognise that failure from the fraud detection years. The machine was never wrong about the pattern. It was wrong about what the pattern was worth. And it isn’t just my shop. I surveyed my own audience, 830 people, 500 of them business owners, and 73 percent said they already use AI actively and get no result from it.

Two things changed, and only one of them is measurable. The measurable one isn’t about me. A randomized field experiment this year, just over a thousand US searchers tracked in their own browsers, found that when an AI Overview appears, outbound organic clicks fall by roughly 40 percent and zero-click searches rise by about a third, while sponsored clicks don’t move at all. It’s a working paper, not gospel, but the direction matters. The AI layer is eating the free traffic, not the paid traffic, and most owners have that backwards. The other change is that people arrive pre-briefed. They quote my own numbers back at me on the first call, and they’ve ruled out three other options before I hear from them. It also means a weak answer doesn’t get a second chance, because the AI layer summarised you once and moved on. Are they coming from ChatGPT? Some, yes. I wouldn’t trust anyone who gives you a precise percentage, including me. That traffic mostly arrives with no referrer, so it shows up as direct, and our attribution for it is broken. What I can see is who gets recommended, and these systems reward depth. A repair business asked me why ChatGPT kept naming its competitor. It didn’t have a ranking problem. It had published nothing worth citing.

Before you open any ad platform, write down the maximum you can pay for a customer, and take that number from your own margin, not from a benchmark. That sounds generic until you see the spread. Inside the single Google Ads quiz category, across the accounts I work with, blended cost per click ran from about $0.03 to $0.36 over a six month window. A 10 to 13 times gap inside one category label, because free personality tests and paid subscription funnels are filed under the same word. Feed the bidding engine the category average and it’ll hit your number exactly, burn the budget, and you’ll walk away saying AI doesn’t work. The AI did its job. The number was wrong before it ever got there. What to ignore: the platform’s own report card. Google grades your ads, and in the 20 accounts where I could compare like for like, ads rated Excellent beat ads rated Average on click through rate in 11 of them. A coin flip with a certificate. Be precise about which machine you’re judging. The one that spends your money is extremely accurate, the one that grades your work is not, and most owners trust them equally. Then ignore anything sold as prompt training. Context beats prompts and it isn’t close. Automate the work, not the judgment.