Negative keyword automation in Yandex Direct: where it stops working

Every advertiser has the same dream: a script that finds junk queries and adds them to the negative keywords while you sleep. Such scripts exist, and they do take work off your hands. The problem is what they are sold as: a finished solution you switch on and forget. The one decision that matters, block a word or leave it, an automat cannot make, and below I explain why.

In the first part I argued that negative keywords are a filter, not a list. This part is about the limits of that filter: what you cannot hand to a machine however much you want to, how a list rots from the inside, and what to do when negatives stop scaling and the junk keeps flowing anyway. For a company selling into Russia from abroad these limits bite harder, because the account lives in a language head office cannot read.

The script does not know your business

Any automat that cuts “bad” words cannot ring your sales department. And the answer to the question “is this traffic relevant or not” sits precisely there. No regular expression will compute any of the following: whether a segment converts into quality inquiries, what margin the product line carries, whether the item is even in stock, whether the landing page is ready to receive that query, how your sellers handle such calls, and whether another page on the site would serve the query better, so that impressions should be split by cross-negation rather than blocked.

The conclusion follows: the filter, in the end, is not the script. It is expertise, meaning an understanding of the niche, common sense and agreement with the client. A blunt test shows this clearly. Remove all negatives at once and inquiries jump sharply. They are junk, not hidden demand, and only the people who took the calls can confirm it.

That said, the technique can and should go to the machine: parsing search terms, lemmatising them, running frequency analysis, building cross-negation by formula. The automat lays the table. The person at the table makes the decision.

The Russian angle is about who that person is. When a European or Asian company sells into Russia, a Russian-speaking contractor sees the search terms, a distributor or local partner sees the customers, and head office sees a spreadsheet. Whoever approves negatives must be the party that hears Russian customers directly; the specialist brings them candidate words with numbers, not a finished list.

The list rots unless you prune it

The usual mistake is to only accumulate negatives. The list has to be reread in full every few months, because your understanding of the niche grows. At launch you do not know the context a word lives in, and checking it against live search results does not guarantee a correct decision the first time. Half a year later it turns out that a phrase you blocked as irrelevant is in fact hot demand.

The second source of rot is that the list was edited by different hands. You, colleagues, the client, the previous contractor. In audits of other people’s accounts I have found several hundred relevant words sitting in the negatives, and in one case the name of the advertised product itself. A relevant word blocked by accident at the start cuts inquiries for months, and no report shows it. What you see is “demand seems a bit weak”. So a review of the negatives is a separate calendar task, not a side effect of routine clean-up.

For a foreign advertiser there is an extra source of rot. Russian has rich morphology, and in Yandex Direct a negative keyword works on the lemma by default, blocking every form unless you fix one with an operator. That is how a blocked word reaches further than the person who added it intended, and nobody at head office can spot it in a list they cannot read. The review has to be done by a native speaker who understands the product. An account inherited from a previous agency deserves the same reading before you trust it.

What to cut first when space is limited

In Direct the volume of negative keywords is capped, so you cannot block everything and have to choose. Advertisers used to Google Ads shared lists of thousands of entries find the constraint unfamiliar. The way to choose without guessing is to break search terms down into individual words, or lemmas, attach each word’s impressions, clicks, spend and leads, and calculate cost per lead, conversion and click-through rates and click price for each.

A candidate for the negative list is visible in the numbers: dozens or hundreds of impressions with zero clicks, a click-through rate several times below average with no leads at all, a cost per lead a multiple of the mean, or a word plainly outside your topic. The order of urgency runs like this:

• first, words that have already eaten budget and produced no conversions;
• then words that convert, but at a price you cannot carry;
• then low click-through traffic with no leads;
• and separately, words that bring no traffic at all but keep collecting impressions, dragging down the click-through rate of their keywords and your position in the auction.

The method needs accumulated statistics. On a new campaign, frequency analysis by individual word is unreliable: there is too little data to judge a single lemma. A company that has just launched in Russia spends the first month blocking obvious junk by hand; the ranked cut comes once the account has something to count.

Where negatives hit the ceiling

There is a structural limit. A share of impressions always comes from unique queries that appeared in the account history exactly once. No negative list will ever gather them. So do not attempt to exhaust the full depth of single-word keywords: the list is unmanageable and still leaks. The real unit of collection is the two-word mask.

Beyond that point the tool is no longer negation. It is a change of match type. Short, high-volume keywords in broad match pull in the maximum number of synonyms, so do not run them in broad at all. Two-word phrases and high-volume three-word phrases go into exact match.

One Direct-specific detail trips up people coming from Google Ads. If the same keyword exists in both exact and broad match, pausing the exact version does not stop impressions for the full-match query. They simply move over to the broad keyword. To remove a specific query, you add it as a negative in its exact form at the level of that same broad keyword. In Google Ads an exact-match keyword takes priority when a query matches it exactly. In Direct there is no such priority, so “which keyword catches this query” has to be enforced by hand.

There is also a pairing that helps split impressions: keep the same keyword in exact and in broad match at the same time. Full matches between query and keyword then go mostly to the exact version, and the broad version receives mainly nested depth and synonyms. The split is never complete, since exact has no priority over broad in Direct, but the share of junk falls noticeably.

When moving an entire keyword into exact match feels wasteful, there is a compromise: fix the two most frequent word forms with the exact-match operator and consciously let the small tail of forms go. For example, the Russian phrase купить слона (“buy an elephant”) becomes two keywords with fixed forms, купить слона and купить слонов (singular and plural), and both stay in the same group so that statistics are not fragmented. In my own projects, two forms take the lion’s share of reach and cut off the bulk of synonyms. Check the share on your own data; it differs by niche.

This compromise has no equivalent in an English-language account, because English barely inflects. A specialist working by analogy with Google Ads either ignores word forms and lets broad match run wild, or locks every phrase into exact match and loses reach. Choosing the two forms is a judgment a native speaker makes in seconds and a script never makes at all.

What this means if you sell into Russia from abroad

The mechanics belong to tooling: exporting search terms, lemmatising them, computing metrics per stem, flagging candidates. A contractor who does this by hand is wasting your money. One who claims the script does the rest is wasting it differently.

The decisions belong to a weekly loop between the specialist and whoever hears Russian customers: a distributor, a local sales partner, a Russian-speaking support desk. Here are the words, here are their numbers, which of these are real demand for you. That exchange is what makes the list defensible when head office asks why a query was blocked.

The review belongs in the calendar, every few months, against the current product range. Stock changes, new lines launch, the sales team learns which calls it can close. And the ceiling belongs to match types, not to longer lists. How this fits into the wider setup of a Russian account is covered on the Yandex Ads page.

What to do with your negative keywords this week

• Reread the negatives in full. Find the relevant words that got there by mistake, whether yours, a colleague’s or a previous contractor’s, and give the demand back.
• Put numbers on priority. Break search terms into words, attach the metrics, and cut first what has consumed budget without converting.
• Take short high-volume keywords out of broad match. You cannot negate your way around them, so do not run them in broad at all.
• Move frequent phrases into exact match. Close two-word and heavy three-word keywords with the operator, not with an endless negative list.
• Stop waiting for the script. Clear disputed words with the client’s sales side in Russia, and leave the machine only the data preparation.

Negative keyword automation is like an autopilot. It holds the car on a straight road, and a human takes the junctions. A list nobody rereads slowly starts cutting into your own demand. And where negation has hit its ceiling, the rescue is not a longer list but a different match type. Junk is produced faster than you catch it, and the winner is not the advertiser with the longest list. It is the one who understands where catching it is pointless.

If you want to know how much relevant demand is locked inside your negatives and where the junk is still flowing past them, come to a free review: we will go through the list together.

Frequently asked questions

Sources

• Russian version of this article on belokrylovo.ru: Минус-слова: пределы автоматизации и ревизия