AI for writing articles: what Google and Yandex accept and the checks a draft needs

Search engines grade the article, not the tool that drafted it

People who look up AI for writing articles usually want to know two things: whether a model can fill the company blog and whether Google will punish the result. In a February 2023 post on Google Search Central, the Search Quality team promised to reward high-quality content however it is produced, meaning original content that shows what Google calls E-E-A-T: expertise, experience, authoritativeness and trustworthiness. The FAQ of that post adds that appropriate use of AI or automation is not against Google's guidelines, unless it is used primarily to manipulate rankings. Using AI gives content no special gains either: in Google's words, it is just content.

The line Google draws sits in its spam policies. Scaled content abuse means many pages generated mainly to manipulate rankings rather than help users, however they are made, and the first example on the list is generative AI used to produce many pages without adding value. Sites that violate the policies, Google warns, may rank lower or not appear at all. Papers for a journal or a university follow their own rules, such as the recommendations of the International Committee of Medical Journal Editors (ICMJE), which ask authors to disclose AI use, so check those first; this story is about articles for a business site or blog.

Where an AI draft helps and where it invents facts

A model is good at the parts of writing that follow a pattern: an outline, a first pass of plain explanations, transitions, a shorter version of a long paragraph. It is fast and rarely leaves a section empty. For a small business that has the knowledge but not the hours, that removes the blank page. It does not remove the work that makes an article worth reading.

The weak points are just as predictable. A draft states numbers, dates and rules with equal confidence whether they are true or not, so every fact needs a source a person actually opened. Yandex's help for webmasters lists the typical signs of low-value content: circling the same reasoning, sentences added for volume, shallow reasoning built on the page title, and unverified facts presented as reliable. It adds that such text is often made with neural networks, machine translation or keyword templates without further analysis or editing, and that a formally original text can still be of little use.

A drafting workflow with a fact check and a named sign-off

A workflow that holds starts before the model is opened. First the question: one search phrase or buyer question per article, checked against the pages the site already has. Then the sources: official documentation, the regulator, the standard, read by a person, with each address kept. Then the draft, a fact check that compares every number, date and rule with its page, and an edit that removes generic sentences, repetition and leftover chatbot phrases.

Google's guide to helpful, people-first content offers a final test in three questions: Who, How and Why. Who: Google strongly encourages accurate authorship information, such as bylines, where readers might expect it. How: it asks whether the use of automation, including AI, is self-evident to visitors through disclosures or in other ways, and the 2023 post adds that an AI byline is probably not the best way to show it. Why: the content should exist primarily to help people. The same guide says Google has no preferred word count.

Why AI articles end up alike and how shingles catch it

The quiet risk of AI articles is sameness. Drafts on neighbouring topics come out with the same structure, examples and sentences, because a model reaches for the most likely wording each time. Such pages then compete for the same queries. Google's spam policy lists a related practice: scraping content and passing it through automated transformations such as synonymizing or translating, where little value is added. Swapping words does not make a page new. Google's documentation also warns that identical or similar descriptions on every page of a site do not help.

Near-duplicates can be measured with a method older than any language model. A text is cut into shingles, runs of consecutive words: with four-word shingles, “the corner bakery opens at eight” gives “the corner bakery opens”, “corner bakery opens at” and “bakery opens at eight”. Two texts are then compared with the Jaccard index, which Wikipedia defines as the size of the intersection divided by the size of the union of two sets, from 0 for nothing shared to 1 for identical. Introduction to Information Retrieval, the Cambridge University Press textbook by Manning, Raghavan and Schütze, uses this test with 0.9 as an example threshold.

VJOURNAL, the studio's own newsroom built on AI-assisted drafts

VJOURNAL shows what these checks look like in code. It is the newsroom of VITON13 Studio and the studio's own product, not a client commission, published at journal.viton13.com in English, Russian, Spanish and French. On 19 September 2026 its four editions carried 3,754 stories: 938 in English, 940 in Russian and 938 each in Spanish and French. They are filed in 14 categories and 8 region desks and signed by 12 desk bylines.

The case fits this topic because of how the stories are made. Drafts are AI-assisted, and three of the four editions come from machine translation, so a style guide alone would not hold. Sourcing, the length of titles and descriptions, and repetition are checked by code. The journal's public editorial policy, whose counts are computed when the site is built, carries an AI disclosure: the How in Google's terms.

The gates every VJOURNAL story passes before and after drafting

The first check runs before a story is written. A publication gate compares the plan with the archive and returns CREATE, SUPPORT_EXISTING, UPDATE_EXISTING or REJECT_DUPLICATE. When an existing guide to a capsule wardrobe on a budget was proposed again, the gate returned REJECT_DUPLICATE with exit code 1. It is the cheapest step: an unneeded article costs nothing to refuse at the planning stage.

At merge, a story that reports facts needs sources on at least two independent outside domains, and its primary source must be on that list. First-party analysis is exempt from this rule. All 938 English stories cite at least two distinct source domains, 630 domains in all. A deterministic quality gate then holds SEO titles to 35–60 characters, descriptions to 90–155 and excerpts to 120–260, and rejects placeholder text and leftover chatbot phrases. Google itself sets no length limit for titles or meta descriptions and truncates both in results as needed, typically to fit the device width, so these ranges are the journal's own rule.

Every release also runs a uniqueness report over all 3,754 stories. It compares five-word shingles and treats a Jaccard similarity of 0.6 or more as a near-duplicate, a lower bar than the textbook example, so more pairs get flagged. Separately, tests forbid any eight-word phrase from repeating within an article body in the 42 repaired Design story groups, in all four editions; all 8 tests passed on 19 September. None of these checks writes a sentence: each only refuses or flags.

What the VJOURNAL checks do not prove about quality

The limits of the case matter more than its gates. Drafts are AI-assisted and the Russian, Spanish and French editions are machine translations. Stories are signed by desks, with no named reporters and no original interviews, and nothing in the case shows that a person reviewed each translation. Since Yandex names unedited machine translation among the common ways low-value content is made, this is the open question of the case. The eight-word test also covers only the 42 Design story groups, not the whole archive.

The uniqueness report is advisory: on 19 September one cluster of paraphrased paragraphs still ran through 54 stories in each edition. The case makes no claim about readership, traffic or search rankings, so it says nothing about a payoff in search. Its lesson is narrower. Code catches lengths, repeats, missing sources and duplicate plans cheaply, but it cannot judge whether an argument is right or a translation reads well, and that still takes an editor.

Writing with AI in-house or ordering a content system with checks

Before you start, prepare three things. A list of the questions your buyers actually ask, from calls, messages and search data, one article per question. A source policy in one paragraph: which pages count as a source and how many independent ones a factual article needs. And a name for who checks the facts and signs off, because the bylines Google recommends need a real person behind them.

Writing with AI yourself works when the volume is small and the person who knows the subject has an hour to check each draft. A content system makes sense when the volume outgrows one editor, when the site runs in several languages, or when you need gates like the ones described above. VITON13 Studio builds such systems and repurposes existing material with AI; the linked guide and service page carry the terms. Either way, AI for writing articles stays a drafting tool, and the editor answers for the text.