Why “it reads like AI” is not a verdict on the technology
Open any feed: “In today’s fast-paced world, an increasing number of businesses…” — the style gives itself away in the first line. But the problem is not the model, it is how it is used. An unconfigured model knows nothing about your product, your audience or your vocabulary, so it writes an average of the entire internet at once. Grammatically fine, and empty.
The difference between filler and working content is not the choice of model, it is the system built around it. That same model, given a brand voice, a fact base and hard rules, produces copy an editor accepts on the first or second pass. We rely on this ourselves: product cards, articles and proposals are all written here with the model-plus-configuration combination, and clients cannot tell them from copy written by hand.
Three reasons generation fails
1. The model does not know the facts about your product
Ask a model about your product and it will confidently invent the specifications. That is hallucination, and for a business it is the main risk: an invented power rating on a product card means returns and complaints. There is one cure — the model must write only from your fact base: specifications, case studies, internal rules. Where data is missing, it leaves a placeholder rather than making something up.
2. There is no brand voice
Every real company has its own way of speaking: formal and precise, or dry and ironic. Leave that unspecified and the model writes internet-average. Configuring the voice means loading your best pieces, a glossary of terms and a list of banned clichés. After that, the output starts sounding like you.
3. Nobody checks the result
Generation without quality control is a conveyor belt of defects. In a working system every piece passes automated checks: for filler and clichés, for originality, and for consistency with the facts on record. And for the first few months a human proofreads, until the system is calibrated.
What do search engines say about it?
The most common question is “will we be penalised for AI copy?” The published position of the major search engines is consistent: what is assessed is the usefulness of the content, not how it was produced. Google states plainly in its guidance that automation is acceptable when the content is made for people; what gets penalised is generated filler aimed at manipulating rankings. The logic elsewhere is the same — low-value content is filtered regardless of whether a person or a machine wrote it.
What a working generation system must do
- •Write from a fact base: specifications, case studies, prices — with no improvisation,
- •Hold the brand voice: glossary, tone, banned phrasing,
- •Adapt to the format: product card, long read, post, newsletter — different rules apply,
- •Check itself: originality, filler, clichés, discrepancies against the fact base,
- •Scale: a thousand product cards overnight is a normal task, not a heroic effort,
- •Work from search data: write for real queries rather than about something in general.
That last point is worth underlining: content nobody searches for is money burned. Before writing an article, check whether the topic gets searched at all. We use an AI SEO specialist for that: it gathers the real search volumes and hands the copywriter a brief — the query, the structure and the questions the piece must answer.
Confidentiality: forgotten in 90% of cases
When an employee pastes a contract into a public chatbot with “make this more formal”, the company has just sent a trade secret to somebody else’s server. Content is no different: draft strategies, internal numbers and unannounced products go into a cloud you do not control.
The answer is local generation: a language model running on your server, inside the company perimeter. Modern open models write well enough for the majority of business copy, and sensitive material physically never leaves the building. The full comparison is in on-premise LLM vs cloud AI and on the on-premise AI server page.
What it costs and when it pays back
| Option | Cost | What you get |
|---|---|---|
| Freelance copywriter | from 40,000 ₽/month | 10-20 pieces a month, quality depends on the person |
| Content agency | from 80,000 ₽/month | process overhead; getting to grips with your niche takes months |
| “Just ChatGPT” done by staff | from 2,000 ₽/month | filler and hallucinations, plus staff time spent rewriting |
| A configured AI copywriter | from 15,000 ₽/month | unlimited flow in the brand voice, facts from your base, self-checks |
For an online shop with a thousand product cards, or a company running an active blog, configured generation pays for itself in the first month on the difference against manual copywriting alone. Then there is speed: an article takes minutes rather than days.
Checklist: are you ready to deploy this
- •There is a regular need for copy — product cards, articles, posts, newsletters — at least 20 pieces a month,
- •There is a source of facts: specifications, price lists, case studies, in whatever shape,
- •There are three to five pieces you consider the benchmark for your style,
- •There is someone willing to review the output and give corrections for the first few weeks.
If all four are yes, deployment takes one to two weeks: we capture the brand voice, load the facts, calibrate on test pieces and put it into production. After that the system writes and you approve.