Agent takes over a site's SEO: Yandex impressions climb from 2.8K to 55K after an audit
The author of "Dots Over AI" handed the SEO for his camp's website to an agent: guides, metadata, links between pages, submitting pages to the index. In the first month, before the audit, Yandex showed the site in search results about 2.8K times. A month after the audit, that figure was 55K. These are impressions, meaning how many times the link was seen in search, not clicks, and there has been only one audit. So the author is adding a reviewer agent that double-checks the first one.
AI-processed from TG: @TochkiNadAI (Точки над ИИ); edited by Hamidun News
Picture a site that the search engine barely knows exists: in three months, Yandex sent it just 10 clicks. The author of the Telegram channel "Dots Over AI" took on the SEO for his training camp's site in May and handed the routine work to an agent. A month after the first audit, Yandex showed the site in search results 55K times, up from about 2.
8K a month earlier. An "impression" here means that a link to the site appeared in search results. It is not yet a visit or a lead.
But from a starting point like this, the growth is visible right away, and it raises a new question: who checks the agent's work when a human no longer has time to?
What the agent does
Right now the agent writes guides based on the camp's live sessions and lessons. It also handles page metadata (the titles and descriptions the search engine sees), internal linking, meaning links between pages of the site, and submitting pages to the index. From time to time the author asks it to run an audit. The agent pulls data from Yandex Metrica, Search Console and server logs on its own, and turns it into a work plan with checkboxes.
For a developer or product manager, this is a typical agent task: dozens of small operations, data spread across several systems, a clear outcome. The human remains the one who sets the goal and makes the decisions.
What the numbers showed, and the caveats
In the first month, before the audit, Yandex showed the site about 2.8K times. A month after the audit, impressions reached 55K. The author himself stresses that this is only the first period after SEO was switched on, and just one audit.
The numbers should be read with caveats:
- These are impressions, not clicks or leads. The post doesn't show how many people actually came to the site and bought anything.
- There is no control group. The growth could have come from the new guides, the internal linking or indexing, not necessarily from the audit.
- The site started from almost zero, and for such a site percentage growth looks especially impressive.
So this is better seen as a sign that the human-plus-agent setup works than as proof of the effect of any single step. The practical value lies elsewhere: one person with an agent covers a volume of work that would normally go to a small department.
Who checks the agent
The more the agent does on its own, the less time is left for manual double-checking. The author came across a post by Ilya Iserson describing a critic agent. It kicks in on its own for serious tasks and recalculates numbers with a script rather than "by eye". It then runs the result through eleven "lenses", that is, eleven different points of view for the review. To avoid looping forever, the critic makes no more than three rounds.
For anyone building agent systems, three rules follow:
- Hand number-checking to code. A script redoes the calculation from scratch and won't repeat an error the model made during generation.
- Set a hard cap on the number of review rounds. Three rounds limit both time and token spend, since each review round is a new set of model calls. Without a limit, the executor and the critic can argue indefinitely.
- Turn on review only for serious tasks. That way small edits don't inflate the bill, and the check fires where a mistake is expensive.
The post has no data on what such a loop costs in money and tokens, and it doesn't list the eleven lenses themselves. To assess the economics, you will have to read Iserson's original or run your own measurements.
What this means
The author ends on a note of self-irony: his SEO department has just one employee so far, and that employee is overconfident. Behind the joke is a working setup: an executor agent, a reviewer agent, and a human who decides what to believe. If you run an agent on SEO, content or analytics, build in verification from the start, not after the first mistake. It's good when the agent pulls data from Metrica, Search Console and logs by itself. But it matters even more that someone else recalculates its conclusions before you start making decisions based on them.
- Nov 21, 2026Amazon Bedrock научили считать по целым Excel-таблицам, а не по кусочкам из поиска
- Nov 21, 2026Amazon Bedrock добавил явное кеширование промптов для открытых моделей
- Nov 21, 2026Claude Code wrote all the code for an ESP32 rig and debunked a false 'discovery'
- Nov 21, 2026Nectar launches AI agent that finds influencers among 300 million profiles
Key Questions Answered
What the agent does?
Right now the agent writes guides based on the camp's live sessions and lessons. It also handles page metadata (the titles and descriptions the search engine sees), internal linking, meaning links between pages of the site, and submitting pages to the index. From time to time the author asks it to run an audit. The agent pulls data from Yandex Metrica, Search Console and server logs on its own, and turns it into a work plan with checkboxes.
What the numbers showed, and the caveats?
In the first month, before the audit, Yandex showed the site about 2.8K times. A month after the audit, impressions reached 55K. The author himself stresses that this is only the first period after SEO was switched on, and just one audit. The numbers should be read with caveats: These are impressions, not clicks or leads. The post doesn't show how many people actually came to the site and bought anything. There is no control group.
Who checks the agent?
The more the agent does on its own, the less time is left for manual double-checking. The author came across a post by Ilya Iserson describing a critic agent. It kicks in on its own for serious tasks and recalculates numbers with a script rather than "by eye". It then runs the result through eleven "lenses", that is, eleven different points of view for the review. To avoid looping forever, the critic makes no more than three rounds.
Ship AI Faster with Founders, Engineers & Weekly Architecture Reviews
Inside Vibe Club (hamidun.com/club) — weekly live masterminds with Zhemal Khamidun, Tech Radar, full LMS access, and 370+ production Claude Code skills.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.