The conversion number your ad platform reports is probably wrong

The report said 697 conversions. Forty-five of them were real. The other 652 were a technical anti-spam goal that fired on almost every form touch, and Yandex Direct had been counting each one as a lead.

That is the short version. The long version is that nobody in that account did anything wrong on purpose. The goal selection was never touched, and Direct's default behaviour is to fold every Metrika goal it can see into one column called "conversions". Technical goals included. The dashboard was honest about what it was adding up. It just wasn't adding up leads.

What 697 versus 45 does to your cost per lead

Cost per lead in that account read about 2 BYN. After we cut the anti-spam goal out of the selection and left only the goals a human triggers by asking to buy, it read about 8 BYN.

Nothing changed in reality. The money had always been 8. The client had been running a campaign for months believing the unit economics worked, because the denominator was inflated four times over by a goal that measures whether a form validator ran.

This is the part people underestimate. A wrong conversion count is not a reporting inconvenience you can mentally divide by four. It is an input to every decision downstream: which campaign gets more budget, which keyword gets paused, which landing page "wins" the test. All of it was being decided on a number that mostly described bot filtering.

And if you run automatic strategies, it is worse than reporting. You told the algorithm what a conversion looks like. It went and found more people who look like that. Feed it an anti-spam trigger and it will get very good at buying traffic that triggers anti-spam.

Proxy goals: the ones that correlate with money but are not money

A proxy goal is any goal that tends to happen near a sale without being one. Metrika accounts accumulate them the way a garage accumulates paint cans. Nobody deletes goals, and every contractor who ever touched the account left a few behind.

The usual suspects:

• Time on site, or scroll depth. Measures patience, not intent.
• Click on a phone number. On mobile it half means something. On desktop the person just selected the text.
• Click on a messenger button. The chat opened. Whether anyone typed a word is a separate question nobody asks.
• Technical goals: anti-spam, form validation, a test goal from 2022 named test_2.
• A thank-you page that also loads on refresh, so one lead counts twice when the customer hits F5.

Names do not help you here. The 652-conversion goal in that account had a perfectly sensible name. What it actually fired on was written in its condition, and the condition is the only thing worth reading.

How to check your own account in about twenty minutes

1. Open the conversion settings of each campaign and write down which goals are actually selected. Not which exist in Metrika. Which are selected.
2. For every selected goal, open its definition in Metrika and read the condition. Ignore the name entirely.
3. Pull a report broken down by goal, one row per goal, for the last full month. You are looking for the goal that carries most of the volume. In a healthy account that is the form submission. In a sick one it is something you did not know was there. Note that Direct caps you at 10 goals per report, so if the account has more goals than that you will be doing this in batches.
4. Keep only goals a person cannot reach without deliberately asking to be contacted. Form sent, call started, order placed.
5. Recalculate cost per lead. Expect the number to get worse, and expect that to be the most useful thing that happens to you this month.

Step 5 is where most of the resistance lives. Someone has been reporting the good number upward, and fixing the goals means walking into a room and saying the leads cost four times more than everyone thought. Do it anyway. The alternative is doing it in six months when the sales team finally asks where all these leads went.

The count is usually the first problem, not the only one

Once "lead" means lead, the rest of the account stops hiding. A different account we went through scored 79.82 out of 100 on a fixed checklist, and once real conversions were in place two things were obvious: 92 % of the spend was going to search queries that had never produced a single application, and all 27 active ads were running without a second headline. Neither of those is exotic. Both were invisible while the conversion column was noise, because everything looked like it was working.

That is the pattern I keep running into. Goal selection is not a small analytics detail off to the side. It is the lens. Get it wrong and every other diagnosis you make through it is wrong in the same direction, confidently.

So: go read your goal conditions. Not the goal names, the conditions. Then look at which goals your campaigns actually optimise toward, fix the selection, and give the strategies a couple of weeks to relearn on honest data before you judge anything else in the account. Of all the checks I run, this is the one that changes the picture most often.