How Google and Yandex detect AI content in 2026: 7 factors and what to do about it

In March 2024, Google integrated the Helpful Content System into core-ranking. In 2025, Yandex publicly stated that mass generation without editing is a violation. By the spring of 2026, domains that ignored these signals lost 70-95% of organic traffic. They didn’t “sag a little” - they collapsed. This article is about what exactly both search engines detect and how to write AI-assisted content that doesn’t get caught. Based on my experience with 11 published articles on dpustovalov.com, written in conjunction with Claude.

Disclaimer: “AI detection” is an incorrect term. Neither Google nor Yandex use a pure AI detector like ZeroGPT. They use signal packages that indirectly catch massify-generated content and punish it. This article is about signals, not about a mythical classifier.

AI-content in 2026 - not “whether it was written by a neural network”, but “does this content have human authorship and experience.” Search engines don't catch the first one. The second one is easy to catch.

What Google is doing in 2026

Helpful Content System in core-ranking

Until March 2024 Helpful Content was a separate “update”. Since March, this has been part of the core algorithm and works constantly. Signals it uses:

• People-first vs search-first. Content written “to rank”—H1/H2 “top 10 X” formulas, filling with synonyms, repeating structures—is penalized.
• First-hand experience. The signal is provided by specifics that are impossible to know without experience: personal metrics, specific errors, names of real projects.
• Author authority. Linking content to an author who has verifiable expertise. Person schema, sameAs, public reputation.
• Site purpose. Sites with clear specialization rank better. “Marketing + cooking + travel” on one domain is an anti-pattern.

E-E-A-T: experience, expertise, authoritativeness, trustworthiness

Not a separate algorithm, but a guideline for raters (real people who evaluate the quality of search results). Their estimates are used to train models. In 2024, Google added a second E (experience) - this is the first E, experience. This means that first-hand experience has become a clear factor.

For AI-assisted content this means:

• If content about marketing is written by AI without personal data of the site owner, it will not pass the experience criterion
• If the author exists only as a name in the footer without bio, schema and sameAs, authoritativeness will not work
• If the site does not have contacts, legal information, or clarity on “who wrote this and why”, trustworthiness will not work

SpamBrain and mass generation patterns

Google has publicly talked about SpamBrain, a neural network antispam component. In 2024–2025, he learned to detect “scaled content abuse” - the mass production of similar pages. Signs that SpamBrain catches:

• Publication rate - 5+ articles per day on one domain without an explicit editorial board
• Semantic monotony - articles with the same structure, differing only in the substitution of keywords
• No internal linking between publications
• Lack of external citations and links

What Yandex is doing in 2026

YATI 2 and Y1: Transformer Ranker

YATI (Yet Another Transformer with Improvements) is a Yandex neural network ranker, appeared in 2020, updated in 2023 (YATI 2). Y1 is the next iteration, presented in 2025. What do they do with AI-content:

• Semantic “freshness” of the text - how much it differs from template formulations in the trained sample
• Coupling with search intent - understands what the user was looking for and evaluates whether the page is responsive
• Correlation between text and behavioral signals - if the text is formulaic and users quickly leave, this is a double negative signal

Behavioral factors (the main filter in the Self)

Yandex's behavioral signals weigh significantly more than Google's. This is the difference that makes Yandex “tighter” in terms of AI-content. Signals:

• Time on the page. Generic AI-text holds the reader for 30–40 seconds, an article from a practitioner — 3–5 minutes
• Scroll depth. AI reading usually scrolls to the middle and closes
• Pogo-sticking. Returning to delivery after a visit is the main negative signal. If 70%+ of users return - Yandex understands that the page is not responding
• Returns. If 0% of users return to the site in 30 days, it is considered “one-time”, which is bad

ICS and Webmaster Quality Score

ICS - website quality index. Contains the share of “low-quality pages” in the total mass. If this percentage exceeds ~20%, the domain is demoted in the search results as a whole, and not page by page. This is the only real “ban for AI” in Yandex.

7 factors that scorch AI-content

1. Lexical stamps (Google + Yandex)

Words and phrases that models generate disproportionately often: “is”, “represents”, “in today’s world”, “it’s no secret that”, “significant”, “effective”, “worth noting”, “AI is a game changer”. The full list is in my voice-checklist (I post it separately).

How to treat: go through the text through the list, replace or delete. For 1500 words there are usually 0–2 “acceptable” tells left.

2. Structure symmetry (both)

All paragraphs are 4 lines long. All H2 “3 points of 3 adjectives”. Each section ends with “Thus...”. This is a structural AI-tell that rankings notice through statistical readability metrics.

How to treat: paragraphs of different lengths - from one line to seven. Different number of items in lists. Not every section is summarized.

3. Lack of original data (Google is critical)

AI doesn't know your specific metric since the 2024 project. AI doesn't know your customers' names. AI doesn't know your working price per hour. If there is nothing specific in the article, it will not pass the E-E-A-T experience criterion.

How to treat: each article must contain at least 3-5 unique personal data points - numbers, names, dates, specific situations.

4. Lack of author markup (Google)

Schema.org Person with author is associated with Article, sameAs profiles on LinkedIn / GitHub / hh.ru, short bio in the footer of the article. Without this, Google cannot attribute experience to the author.

How to treat: configure schema once, keep it up to date. Detail - in mine article about marketing sovereignty.

5. Pace of publication (Yandex is stricter)

7+ long articles a day is physically impossible for one person. Search engines know this. Sustainable pace for personal brand: 1 longread in 2–4 days, 2–3 per week maximum.

How to treat: Don't write faster than you can edit. The quality of readers is higher than the quantity of publications.

6. Behavioral indicators (Yandex critical)

If an article has a page time of 40 seconds and a pogo-stick of 75%, Yandex will demote it, no matter how much you cleared AI-tells. This means that the text must really hold the reader.

How to treat: write only what you have to say. Do not stretch the texts to length - cut out the weak parts.

7. Internal link network (both)

An AI factory typically publishes isolated articles without an internal network. This is a weak “scaled content” signal. Solution: each new article links to 3-5 old ones, contextually, not in the general “related” block.

How I wrote this very article about protection from detectors

Meta level: this article is an example. I’ll tell you what I did to make it pass too.

1. Gave Claude a brief + the entire voice folder with coordinates and AI-tell checklist. Without this, the draft would be generic.
2. I received a draft. I went through AI-tells - I found “is”, “worth noting”, “at the same time” - I threw it out.
3. Added personal data. Specific dates of Google updates, names of internal components (SpamBrain, YATI), my own experience with 11 articles.
4. Internal links. On article about marketing sovereignty, on about page, to other blog materials.
5. I read it aloud. I stumbled on three places - I rewrote it. Final editing took ~50% of the time spent writing the draft.
6. Schema checked. Article + Person + Breadcrumb. Author is linked to Person schema at /about.
7. Temp. This article is one of ten that I am releasing during May 2026. This is ~2–3 per week - a sustainable pace.

What NOT to be afraid of

• Use AI as a draft. This is not a fine. The penalty is to publish a draft without editing.
• Indicate “AI-assisted” in meta. In my layout it is meta name="ai-content-declaration" content="human-authored" - this is a signal to AI crawlers, not a classifier for Google.
• Use AI for translation or adaptation. Google has made it clear that this is normal use.
• Make meta articles about AI. This article is meta. It ranks the same as an article about any other technological phenomenon.

The stack where I keep anti-detection

• Voice library — 11 files with coordinates, pillars, checklists. Loaded into Claude before each draft.
• AI-tells checklist — 60 phrases and structural patterns. Find & Replace before publication.
• Internal linking map — Google Sheet, which marks which articles should link to which. Filled in when a new one is added.
• Schema.org validation — Rich Results Test before each deployment. Ya.Metrica for tracking behavior.
• Chief editor + Turgenev — for cleaning from office debris and water on the final pass.

If you have a website with AI-content and traffic has dropped

Symptoms of Helpful Content/YATI downgrade:

• Gradual (not overnight) drop in traffic by 30–80% over 2–3 weeks
• The drop appears in Search Console / Ya.Webmaster as a decrease in the average position, not the number of impressions
• Brand queries continue to work, information queries are the first to decline

What to do:

1. Audit of the top 50 pages using AI-tells, remove weak ones entirely (not “improve”, but delete)
2. Add author markup and personal data to the remaining ones
3. Reduce the rate of publications to a sustainable level
4. Wait 4–8 weeks - the system does not recalculate signals instantly

If you are currently working with AI-content and are not sure how protected it is from detectors, come for a 30-minute discovery. I’ll analyze your publications from the point of view of the 7 factors above, and show you what 3 edits to make in the first week to stop the drop in traffic.

Related materials: content plan template, CTR benchmarks in VK Ads, a complete guide to performance marketing, cleaning text from clutter and traces of AI.