You ask AI to write a product or service description for your website and get a page of generic words: “high quality”, “individual approach”, “the perfect solution”. Lots of text, little use. Often the cause is not the model but one phrase in the prompt: “at least 800 words”. Below: why that phrase hurts, what to write instead and how to check the finished text yourself.
We checked the facts against primary sources on 27 October 2026. All sources are named right in the text.
1. What went wrong: a minimum word count with a minimum of facts
A clear example: Horoshop, a Ukrainian platform for online stores, described its own mistake in an article of 28 August 2026. The prompt for product cards required “a minimum length of 800 words”, and that applied to every product, however much data the supplier provided. For products with few verified facts, a longer text did not become more useful. According to the authors, the model started to repeat itself, add generic phrases and exaggerate benefits. The author’s conclusion: the goal of a card is not to reach a word count but to give the buyer enough information to decide.
The mechanics are simple. You have three facts about a product, and the model must produce eight hundred words. There is nowhere to take them from, so it fills the space with repetition, clichés and invented benefits. This is one company’s example, not a study, but it matches what we see at work: the fewer facts in the prompt, the more filler in the output.
2. Can you set length at all: two sides
Opinions differ here, and we show both honestly.
Practitioners who advise setting a length. Length is part of the task: “a summary of up to 200 words” sets a frame and keeps AI from stretching the text. This works when you set an upper limit.
Research on length. In August 2025 two researchers (Xie, Lee, arXiv preprint 2508.13805) showed that models often miss a requested length because they do not keep their own word counter. A special “countdown” technique inside the prompt raised the share of exact hits on GPT-4.1 from under 30% to over 95% (one test). So a plain “at least 800 words” does not guarantee length, and an exact hit needs a separate technique.
What follows. You can set an upper limit (“up to two hundred words”): it only cuts. It is better not to set a lower limit: it forces repetition when facts are scarce. We found no separate study specifically about a “minimum”, so this is a rule of practice, not a proven law. Length should be dictated by what you actually have to say.
3. What to write in the prompt instead of a minimum
Remove “at least N words”. Then do three things.
Give real facts. Prices, specifications, deadlines, delivery and warranty terms, who it is for. Anthropic’s guidance for its own models says it directly: the more precisely you explain what you want, the better the result, and context or the reason behind an instruction helps the model give more targeted answers. No facts in the prompt means no facts in the text.
Ask it to answer four questions. This is Horoshop’s criterion for card quality; the article states it as a criterion, and the four questions in this form are our wording:
- Who is this product or service for?
- What customer problem does it solve?
- How is it used?
- What are its limits that the customer should know about?
Set an upper limit if needed. “No more than two hundred words.” If a hundred is enough to answer the four questions, let it be a hundred.
Sample prompt: “Here are the facts about the product: [your data]. Write a website description of no more than 200 words. It must answer four questions: who the product is for, what problem it solves, how to use it, what its limits are. Do not add anything that is not in the facts.”
The last sentence matters: it forbids inventing outright. We wrote separately about what happens when AI fills in the gaps for you: how AI hallucinates and what it can make up about your business.
4. What to check yourself after the answer
Even with a good prompt you must read the text. Google’s guidance on AI content (page updated 1 October 2026, available only in English) says models do not retrieve facts from a database, so their text must be manually checked for accuracy before publishing. The same goes for the page title, the description in search results and image captions.
A ten-minute check:
- compare prices, deadlines and terms with your price list and contract;
- cross out phrases that could sit on any competitor’s site (“high quality”, “individual approach”);
- make sure each of the four questions has a real answer in the text, not a generic phrase;
- if even one question is unanswered, add facts to the prompt and try again.
A separate point on volume. In its spam policies Google calls scaled content abuse the situation where many pages are created without benefit to people. A hundred “filler” texts in one evening is exactly that risk. More in our article on what Google advises checking by hand in AI-written texts.
Why filler hurts you in particular
Filler costs money twice. First, a customer who lands on the page from your ad or search does not find the answer to “is this for me”, and leaves: you paid for the click and got no lead. Second, in the same guidance Google writes that search quality raters look at whether content was made with “little to no effort, little to no originality, and little to no added value”. Generic phrases that could stand on any site fit that description. Google does not specify exactly how this affects your site’s rankings, so we do not claim more than it writes.
Example: what changes in practice
This is an illustration, not a test result. Suppose you sell an electric kettle and know three things about it: 1.7 litre capacity, glass body, switches itself off after boiling.
Prompt with a “minimum”: “Write a description of an electric kettle, at least 800 words.” The model has nothing to write from, so you get “unrivalled quality”, “stylish design for a modern kitchen”, “brings comfort every day” and invented benefits like “boils in 30 seconds” that you never confirmed.
Prompt with facts and questions: “Facts: 1.7 l capacity, glass body, automatic switch-off after boiling. Write a description of no more than 120 words. Answer four questions: who it is for, what problem it solves, how to use it, what its limits are. Do not add anything that is not in the facts.” The result will be shorter, but it will say who the kettle is for (families who drink tea and coffee several times a day), that it switches itself off, how to use it, and, honestly, that our facts say nothing about boiling speed, so the text says nothing either. If the model still added something invented, you see it at once because the list of facts is right there.
From our practice
This is our experience, not statistics. The weakest AI texts almost always come where the prompt has no data and only wishes about length and tone. The best come where the owner gave a few dry facts and asked for brief answers to specific buyer questions. The same logic applies to ad copy: we wrote about choosing variants in 80% of what AI writes is better thrown away.
Summary
- The phrase “at least 800 words” makes AI write filler when facts are scarce.
- You can set an upper limit (“up to two hundred words”), not a lower one. Models are poor at hitting an exact length.
- Give real facts and ask for answers to four questions: who it is for, what problem it solves, how to use it, what its limits are.
- Read the finished text yourself: AI may invent prices and deadlines.
See how we set up and evaluate advertising in the Google Ads for business section.
Want a free audit of your Google Ads account? Send us a request: we’ll look at your account and tell you where money is being lost.