Three months, 2224 articles, 27 Zen channels, 33 client websites. Publication is fully automated, and rankings were measured across 26 thousand checks. The analysis below is based on this project’s figures, and its main value is not in how to generate text, but in what happens to an article after publication: does it make it into Yandex’s top results, does Alice pick it up for a quick answer, does it bring people to the website? It also honestly lists what did not work—and a lot did not work.
Zen as a search platform, not a feed
SEO does not work inside Zen: it is a recommendation feed, not a search engine. What works is its connection with search. Zen articles are indexed by Yandex and Google, so you can write material for a specific search query and receive search traffic for months. Essentially, it is the same kind of SEO article you would publish on your own website, only on someone else’s domain with established trust.
This explains the discrepancy with how nine out of ten authors run their channels. A typical channel lives by its feed: impressions, read-throughs, subscribers, and viral topics. An article gets a peak within a few hours and then dies. A channel built for search works differently: an article is published not for the feed, but to secure a position in search results and work from there for weeks and months. The feed becomes a side effect.
The gap between the two pictures is visible in a sample of 1004 articles: 3112 total opens, of which 1696 impressions came from the Zen feed and 1416 from search. Material can be almost invisible within the platform and still perform excellently in search results. The practical takeaway for anyone running such a channel: do not look at Zen’s internal statistics at all. Look at Yandex rankings and visits to the website.
How a search-focused channel differs at the process level
- Topics are taken not from the author’s interests, but from keyword research with confirmed search volume. Not “I’ll write about land drainage,” but an answer to the query “how to install drainage on clay soil.” Collection, cleaning, and clustering are handled through the API of a search-volume checking service.
- The content format (instructions, comparisons, price analyses, problem breakdowns) is determined by the live search results for the query, not by the author’s genre preferences. Before generating the text, the model studies the top results and chooses the format.
- Facts instead of opinions: market prices, specific figures, and verifiable data. This also helps prevent invented facts—the model reads the search results first and writes afterward.
- A structure designed for extraction: a direct answer in the first paragraphs after the headline and a fact block. These are the sections Alice picks up.
- The metrics are different: positions in the top-10 and top-3, inclusion in AI answers, the share of reads from search, and visits to the website.
The share of visits to articles from search across all channels stays around 26–27% and grows as the channel matures. Over the entire period, 132 500 reads accumulated: 78 thousand from the feed, 34 thousand from Yandex and Google search, and 20 thousand from social networks and messengers, where articles arrive through readers’ reposts (482 reposts over three months).
An article almost never ranks for the query it was written for
This is the main result of the entire project, and it overturns the usual logic of turnkey work. The figures are as follows:
| Metric | Share of articles |
|---|---|
| Top-10 for at least one query from the cluster | 34% |
| Top-3 for at least one query | 25% |
| Inclusion in a Yandex AI answer | 35% |
| Top ranking for the target high-volume keyword the article was written for | 5% |
In absolute numbers: across 22 thousand queries, 1063 articles rank in the top-10, 824 in the top-3, and 749 appear in AI answers. The difference between 34% and 5% is the whole point of the long tail. An article ranks not for its own keyword, but for a range of related formulations around it.
Next comes the dependence on query length, calculated from 26 thousand ranking measurements and 20 330 unique queries. The difference in the probability of reaching the top between a two-word query and a nine-word query is a hundredfold. Alice’s block includes 15% of articles for queries of nine or more words, versus fractions of a percent for two-word queries.
The case explains this as follows: people increasingly speak rather than type, and voice queries are longer and almost always phrased as questions. In addition, search itself has shifted toward displaying a ready-made answer, and a ready-made answer can address a specific question, but not two words. This cause-and-effect relationship cannot be verified using public data, but the correlation itself was measured on a sufficiently large sample.
Two practical rules grew out of this. First, articles targeting longer formulations are published first, while short, high-volume queries wait at the end of the queue. Second, the FAQ section is built around real queries, and each question becomes a separate doorway from search. The article originally had three questions; now it has seven or eight with detailed answers, and these are also picked up in Alice’s block.
You still need to write for high-volume queries. You just should not expect the article to rank specifically for them.
The measurement method that initially ate half the result
At first, rankings were measured for a single target keyword—the one the article had been written for. With this method, much of the actual result is simply invisible: an article about 3D visualization appeared in Alice’s block for queries that had not been measured at all. Now up to five formulations are tracked for each article: the target keyword, similar semantic queries, and questions from the article’s own FAQ. The project acknowledges that this still understates the statistics, but the trend is visible.
The second observation about search results: exactly one Zen article appears in Yandex’s top-10. This was checked across two thousand queries; 0,2% were exceptions—rare cases where two articles appeared in the results. The task, therefore, is not to “make it into the top,” but to knock out whoever already occupies the slot. At the same time, a slot for Zen often exists even in blatantly commercial results where a website cannot break through into either organic search or the ad block.
The experiment that was not confirmed
The hypothesis was simple: question headlines often appear in the top results, so a question format must help. The test involved 145 articles older than ten days that were not in the top-10; only their headlines were manually changed to questions. The keyword, text, and images remained unchanged. After three weeks, the picture was as follows: changing the headline initially lowered the rankings, then reindexing took place and everything returned exactly to its original position. There was no gain for either the target keyword or related queries.
However, a parallel analysis produced a finding more valuable than the experiment itself. An exact keyword match appeared in the headline of 72% of the project’s own articles—but in only 7% of competing articles ranking in the top results. The average headline length also turned out to be shorter than that of competitors. In other words, optimizing for an exact match distinguished its materials from top-ranking ones, and in the wrong direction. The rule “the keyword must appear in full and at the beginning of the headline” had to be removed from the prompts.
The right approach: write the headline in natural language and name the topic the way a person would say it. The word order in the keyword can change freely, endings can be declined, and query clutter such as geo-modifiers should not be dragged into the headline.
Four mistakes that cost results
All of them surfaced in practice, and, tellingly, readers on Zen—not the analyst—found the first three. The platform’s internal audience is mature and sharp-tongued, and its comments served as a quality-control system.
The first issue to emerge was duplication across word forms. Non-obvious repetitions of the same query were not collapsed during semantic processing, and the channel received dozens of articles about the same thing with different headlines. The comments were predictable: “stop writing the same thing for two months.” After collapsing them, about 5000 keywords remained out of 9500, and the repetitive feed disappeared on its own.
Next, invented figures came to light. In technical niches with many models, sizes, and parameters, the generation produced nonexistent numbers. Readers reacted directly: “at least learn something about what you’re writing about.” The solution is that same consultation of search results before generation: first the model reads other materials and gathers real numbers, then it writes.
A similar problem arose with up-to-dateness: in niches tied to legislation, articles came out with last year’s rules. The solution is the same preliminary collection of source material.
The most expensive mistake turned out to be formatting, and it went unnoticed for three months. The assumption was that the articles were losing to the top in volume and text quality; the test compared a couple hundred of the project’s own articles outside the top with a couple hundred competing top-ranking articles for the same queries. Volume had nothing to do with it: 9% of competitors in the top had articles shorter than three thousand characters. The difference was formatting. Lists were used in 94% of the project’s own articles, but the HTML conversion function parsed the headings and collapsed list lines into an ordinary paragraph. Visually, everything looked like a list: “1.”, “2.”, text. According to Zen’s editor markup, it was not a list, and the platform received one continuous block of text. After the conversion was fixed, the number of images rose from two to five, and simple infographics replaced abstract illustrations.
Channel age matters more than the volume of publications
Dividing channels by type produces the most inconvenient figure in the entire case. The 17 channels built from scratch accumulated 73 500 reads, of which search accounted for only 9%. The 8 channels transferred by clients with an established history had 59 000 reads, and the share from search was 46%—with the same publishing discipline.
For a young channel, volume drives the feed, while search catches up noticeably later. In addition, a new channel has a kind of sandbox period: according to observations from the project, articles are blocked from indexing for anywhere from one to four weeks. There is no point planning for results before that period ends.
The estimate for reaching a meaningful volume of search traffic is sobering as well: from one to three articles are published per day, or 60–90 per month per channel, while the critical mass is approximately 500–1000 articles. That is a year of work.
The economics: text generation has long ceased to be the main expense
Two technical techniques reduce generation costs severalfold, and both are available to anyone using a model through an API.
Prompt caching. The unchanged part of the request (format rules, structure, style restrictions, requirements for lists and headings) is charged at a discount of up to 90% on repeated reads. In this project, the permanent part accounts for 86% of the input—about 13 thousand tokens; each request contains roughly 5 thousand unique characters: the keyword, source material, and links. Almost the entire prompt is cached.
Batch processing. The response does not arrive instantly but within a day, which reduces the price by 50%. For a pipeline in which articles are published on a schedule, the delay costs nothing. Batch processing saves more than caching, and together the two techniques provide a total discount on input tokens of up to 95%.
After this optimization, it became clear that text was no longer the main expense. An article cover costs three times as much as the text itself. This led to the decision to remove image generation from the article: infographics are rendered from HTML into PNG by an ordinary browser, which is free and, unlike a neural network, renders Russian text accurately. Only the cover is generated by the neural network. The main expense ultimately lies not in the articles, but in fixed overhead: servers and services.
What Zen does not provide
This is where the part that is usually omitted from case studies begins.
There is no link equity. Links in the body of the article are closed and pass no authority—just as on other trusted platforms of this type. A link in an article is meant for a real person to click, not for building link mass.
An increase in the client website’s visibility was not confirmed. A representative measurement: a website for electrical installation services with no SEO at all, whose channel has been run since the beginning of July. According to Webmaster, before and after: impressions 127 per day versus 118, clicks 14 per month versus 19, pages in search 54 before and 51 afterward. There was no effect over a month and a half. The same applies to the other websites—no change in the statistics is visible anywhere.
But on that same site, Zen brought in 397 people in August, while the site’s entire organic search brought in 23. In other words, the platform works as an independent source of traffic, not as a boost to the site’s rankings. These are different things, and they shouldn’t be confused.
Lead generation was not measured. Lead-generation goals are not configured for all clients, and conversions are visible only in isolated cases. For now, this is about brand awareness and additional touchpoints, not leads.
Why all the traffic figures are understated
Every article contains two anchor links with UTM tags, and the anchors are unique. Over the entire period, these tags generated 5283 visits to client websites, which means a conversion rate from reading to clicking of about 4%. But the metric does not classify some visits from Zen as clicks: the referrer is not passed, and the visitor ends up counted as a direct visit. Indirect confirmation: direct visits to client websites increased by 41% after the channels were launched, comparing equivalent periods before and after.
Some of this increase consists of bots, and some of it consists of real readers whom the tag-based report does not detect. The project does not draw an exact boundary, and that is more honest than attributing the entire increase to Zen.
Platform rules: what is allowed and what is risky
Scaling is built around the co-authoring functionality: an administrator can legally manage other people’s channels, which is something most similar platforms do not offer. That is precisely why the project ultimately focused on Zen, although it started out more broadly—with automation across several platforms at once. Supporting a zoo of editors and rules proved unmanageable, and on one of the platforms, a channel was regularly restricted for being a commercial account.
Two points are worth clarifying, because here the price of a mistake is a blocked channel.
About the number of channels. The case study mentions a rule of “up to five channels per Yandex account.” Zen’s current help documentation says something else: one channel is linked to a single Yandex ID, and a separate account is required for a second one. The scaling mechanism remains the same, however—it relies not on the number of channels under one login, but on access permissions and co-authoring. A channel can also be handed over to a client simply by changing the contact details in the email settings.
About generated content. The practitioner’s position is that the rules do not prohibit generation; there is quality-based moderation, and across 27 channels and 2224 articles, there were no bans—except for a channel in a topic that was prohibited from the outset. The platform’s rule history is more complicated: in March 2025, the rules explicitly allowed the publication of AI-generated content, but in July of the same year that permission was removed, returning to the previous wording. In other words, there is currently no formal permission for generated content, and a working pipeline operates in a gray area where text quality serves as the practical filter. It is possible to build a business around this, but only with your eyes open.
One more detail about bans: blocking one channel does not affect the owner’s other channels. This was verified on a personally owned channel blocked for a prohibited topic—the neighboring channels continued to operate.
How quality is measured when detectors lie
Running the texts through AI detectors produced the expected result: some services showed 99% machine-generated text, while others judged it human-written. The project abandoned this metric and replaced it with reader reactions. As long as the comments contain sarcasm and complaints about the author, something is wrong with the text. When the comments shift toward additions and substantive discussion, the quality threshold has been reached. Read-through rate provides indirect confirmation: 80% of those who open an article scroll to the end.
What to take away from this
- Measure an article’s rankings not by a single target keyword, but by a cluster of phrasings together with the questions from its FAQ. Otherwise, half the result simply remains invisible.
- Build the publication queue from longer phrasings to shorter ones, and compile the FAQ from real queries.
- Do not tailor the headline to an exact-match occurrence—almost nobody in the top results does that.
- Collapse inflected forms in the semantic analysis before launch; otherwise, the channel will be flooded with duplicates.
- Make the model read the search results before generating content: this eliminates made-up numbers and outdated rules in one stroke.
- Check that the markup in the output is actually markup. Lists that only look like lists to the eye do not count.
- Enable prompt caching and batch processing—together, they save up to 95% on input tokens.
- Be honest about timelines: a sandbox takes up to a month, while meaningful search traffic requires hundreds of articles and roughly a year.
And keep the main caveat in mind: a channel on someone else’s platform does not improve your website’s rankings. It brings people in directly. In the electrical-installation case, that meant 397 people in a month versus 23 from the site’s own organic traffic—even though the site’s visibility did not move by a single point.
