Clients often arrive with the same complaint: the leads are coming in, but the cost per lead swings between 800 and 3,000 rubles for no clear reason. The semantic core is built and the goals are set. The first instinct is to blame the contractor. The problem usually lies elsewhere: the algorithm optimises for a lead, not for money.
Dmitry, head of advertising and promotion at Aspro, runs online stores and develops the Aspro.Cloud business management system. He lists the settings mistakes that leak budget in Yandex Direct.
Goals, structure and bids
The most common mistake is training the ad account on the wrong goal. Set 'form submission' as the target and the system will hunt for people who submit forms. You pay for them, and there is no guarantee they turn into buyers. If the business needs sales rather than leads, quality data has to be fed back into the system, or the algorithm cannot tell a random click from a genuine enquiry.
A second leak is a semantic core that is collected but not structured. When queries with different intent sit in one group, no ad can be relevant to all of them. CTR falls and clicks get more expensive. The fix is to split groups by meaning: hot queries apart from informational ones, with a separate ad for each group.
Bids also ignore unit economics. The algorithm does not know your margin. A 'maximum clicks' goal brings cheap traffic that does not convert; 'maximum conversions' without a CPA cap can drift into expensive segments. The acceptable cost per lead should be calculated in advance and used as the anchor for settings. Otherwise it is a lottery: 800 rubles today, 3,000 tomorrow.
Two more omissions do quiet damage. Without negative keywords, the ad keeps serving on queries that will never produce a lead, and the search query report needs regular checks, every couple of days in the first weeks. Without device and geo splits, mobile and desktop behaviour and conversion rates get averaged into one bid, and you overpay for the segment that does not pay back.
The swings in cost per lead are not a verdict on the contractor. Check the goals, the group structure, the bids and the negative keywords. Start with the goals: if the algorithm is learning from the wrong signal, nothing else will save it.
