Autostrategies and bots: what the algorithm actually trains on

Conversions marked as fraud do not take part in training a Yandex Direct strategy. The query “autostrategies and bots” often sounds as if every empty click becomes a lesson for the bids. The antifraud page already says the opposite about what has been recognized: recognized fraud stays out of training. The strategy trains on goal completions that remained in the statistics. Ten completions a week is the threshold for that training, not a sign of fraud and not a sign of a wasted budget.

How the click price rises on the same budget is covered in why the click price rises. Below is which strategy price is not the report’s CPA, how many completions the help page asks for, and what happens to the weekly budget in the middle of the week.

What the strategy trains on

For a strategy to train on a goal, that goal needs at least 10 completions a week. If there are fewer, the help page advises taking a goal higher in the funnel, or changing the budget or the conversion price. Separately, you check the goal with your own conversion and look at whether Metrica counted it. The ten is about whether there are enough signals to train. It does not say that nine completions are “clean” and the eleventh is “a bot”.

A fraud conversion Direct has already recognized does not count toward those ten. A completion Metrica recorded, and that was not recognized as fraud, does. That is why “your own conversion, and did Metrica count it” comes before a strategy change. Training will follow what landed in the goal, minus recognized fraud, not what you consider a lead in the CRM.

“Maximize conversions” strategies with an average target, including an average conversion price or a share of advertising expenses, pick bids by the predicted conversion rate. The average is held across a calendar week. On single days it can sit higher or lower. The help page names a training minimum of 10 conversions a week, 70 for app promotion, and asks you to judge the result after 7–14 days. The 10, the 70, and the 7–14 days are training conditions, not a fraud threshold.

An adjustment changes the goal. At 200 ₽ per conversion and a group adjustment of +50%, the group’s average conversion price is 300 ₽. That is the help page’s arithmetic, not a suggestion to pile adjustments on. If you need to stay on the price you set, the page offers payment for conversions rather than for clicks.

The strategy price and the report CPA

The current conversion price is read in the report wizard, in the metrics CR, CPA, and “Conversions”. The price set in the “Maximize conversions” strategy is a limit for future conversions. The report’s actual CPA can land either above or below that limit. The help page does not promise the two numbers will match.

A worked example, not a report. The strategy limit is 500 ₽. In a week, spend is 20,000 ₽ and the report has 25 conversions. Actual CPA is 800 ₽, above the limit. Another week: 20,000 ₽ and 50 conversions, actual CPA 400 ₽, below the limit. The page allows both outcomes. Neither 800 nor 400 proves that fraud entered training. Recognized fraud conversions do not take part in training, and the actual price simply divides the spend that happened by the conversions that happened.

A cost-share limit asks for income

For “Maximize conversions” with a limit on the share of advertising expenses, you pass promotion income into Direct: the amount and the currency of the goal value. The help page lists the ways:

• the goals “Autocollect: order placed” and “Autocollect: purchase” for an online shop;

• a value in the strategy settings;

• goal income in the Metrica interface;

• reachGoal with a value;

• ecommerce with a price.

The page explicitly allows the value in the Metrica interface to be an average order or an approximate estimate, without exact financial figures. That is not an export of a payment fact from a CRM. A strategy limited by share then holds a share of the estimate you passed. A cash receipt is not inserted into training on its own.

The weekly budget and its restart

For strategies, the budget help recommends a weekly budget of at least 10 × the margin per conversion per week for “Maximize profit”, and at least 10 × the conversion price for “Maximize conversions”. The minimum weekly budget is 300 ₽. The factor of 10 is a recommendation for the strategy to work, not a profitability threshold and not a sign of waste.

The maximum spend in a day is counted on Moscow time. With a schedule of one day, up to 100% of the weekly budget and the carryover. With two days, up to 50%. With three or more, or with no schedule limits, up to 35%. That is the day’s ceiling, before a separate talk about a restart.

Direct distributes the weekly budget from Monday through Sunday and does not exceed it in the week, except on a restart. A restart happens when the weekly budget changes, when the payment model changes, and when the start date is moved into the future. Stopping and turning the campaign on again, changing the strategy, the type of limit, or the click or conversion price does not restart spend.

If the budget is changed in the middle of the week, what was already spent is not subtracted from the new limit. On pay-per-click, the restart day can take up to 35% of the old budget plus 35% of the new one. On pay-per-conversion, up to 100% of the old plus 100% of the new. The help page’s example: a budget of 7,000 ₽, 4,000 ₽ spent by Thursday, the budget changed to 5,000 ₽ — from Thursday through Sunday another 5,000 ₽ can be spent. Seven, four, and five thousand are the help page’s example, not a normal spend. The page would rather you change the budget at the start of the calendar week, so the new limit is not stacked on the tail already spent from the old one.