Why AI product descriptions are only as good as the data behind them
A language model knows nothing about your product. It knows patterns: how product descriptions in a category usually sound and which features tend to be mentioned there. Whatever it writes about an individual item comes either from the data it is given or from that probability. You cannot tell the difference from the finished text.
This leads to the most important mechanism. When a detail is missing, the model rarely leaves the gap empty. It fills it with whatever would be plausible for similar products. A face cream suddenly becomes “suitable for all skin types”, although nobody has checked. A pillowcase described as a “cotton blend” quickly turns into “pure cotton”. Specialists call this hallucination: an invented statement phrased convincingly.
The quality of generated copy is therefore decided not by the model but by the dataset. Thin data going in produces fluent text coming out, part of which is guesswork. The next question is how complete your own data needs to be.
How to tell whether your product data will hold up
A simple test helps before any decision. Take ten random items and check whether someone with no product knowledge could write a correct description from the data fields alone. If they could not, an AI cannot either.
Completeness. Material, dimensions, ingredients, usage, origin: every detail meant to appear in the text has to exist as a field. Knowledge that lives only in the buyer’s head does not reach the text. Or it gets replaced with something plausible.
Clarity. Free-text fields such as “notes” or “other” often contain contradictions, outdated information and internal remarks. A model cannot distinguish between “changing from autumn” and a valid product feature. Structured fields with fixed values are far more robust here.
Currency. If a formulation or a supplier changes, the change has to be in the data before any text is regenerated. Otherwise the AI keeps spreading the old state in clean prose.
Whether a spreadsheet is enough or a dedicated system is needed depends mainly on range size and the number of people involved. That is covered in the article on product data management. For generated copy the rule is the same either way: fields first, generation second. Even complete data, however, does not solve a second problem.
Why generated copy sounds so interchangeable
Without further instructions, a language model writes the most probable text. That is the strength of the technology and, in retail, also its problem. The most probable text about a scented candle is the one your competitors get from the same tool. Phrases such as “high-quality”, “perfect for” or “brings warmth to any room” therefore appear in many shops at once.
Brand language comes from decisions that depart from the average. Which words the brand never uses, how long its sentences are, whether it claims benefits or describes features. These decisions have to exist in writing before a model can follow them. A sentence such as “we sound friendly and competent” is not enough. It fits almost any brand.
What actually helps is examples and prohibitions. Three to five good descriptions written by hand serve as models. Add a list of unwanted words and guidance on length and structure. How to define a brand voice in the first place is covered in the article on brand voice. Without this foundation, the text is not wrong, but it belongs to no brand. More troublesome are the errors that only surface later.
Errors that only surface after publication
The expensive errors in generated product copy are rarely grammatical. They concern content and sound right. That is why they are usually not caught at sign-off but weeks later, and by people outside the team.
Wrong features. A customer orders because the text promises “fragrance-free”. She sends the product back because it contains fragrance. The return is the smaller problem. Bigger are the review that follows and the question of how many other texts contain the same error.
Inadmissible claims. In cosmetics, food or supplements especially, models tend towards efficacy claims that do not stand up legally. A generated “clinically proven to calm irritated skin” is quick to write and hard to substantiate. The limits that apply here are covered in the article on cosmetic claims in online retail.
Hidden duplication. When many similar items are generated from the same template, the texts often differ only in colour or size. To search engines such pages look almost identical. To customers it remains unclear how the variants actually differ.
What all three cases share is that the correction does not end with one text. Whoever finds an error has to search every text produced from the same data or the same template. That brings us to where the real cost lies.
Why reviewing takes longer than writing
The obvious assumption is that checking a finished text is quicker than writing it. With generated copy, that is often untrue. A person writing their own text leaves gaps visibly open or asks. A model closes every gap with a smooth sentence. Whoever reviews it therefore has to check every statement against the data, including the ones that sound self-evident.
On top of that comes a habituation effect. After twenty flawless texts, the twenty-first gets read faster and more loosely. That is exactly when the invented feature slips through. A review that relies on instinct gets worse with volume, not better.
A review built on fixed questions is more reliable. Does the text contain a detail that appears in no data field? Does it use a word from the prohibited list? Does it promise an effect? Part of this can be checked automatically, for instance by matching against the fields or searching for trigger words. Final approval stays with a person who knows the range.
A different calculation is therefore more realistic. Time spent writing drops considerably, while time spent on data and review rises. Whether anything is left over depends on the conditions in the next section.
When AI product descriptions pay off for a small shop, and when they do not
Some situations make the sums work clearly. Many similar items with good data. A range of hundreds of variants with well-maintained attributes is the most favourable case. The template is built carefully once, and review happens by sampling with fixed questions.
Second languages and further channels. If a reviewed German description exists, an AI translation or a shortened version for a marketplace is often more reliable than a text from scratch. The facts are already settled; the model only rephrases. A native speaker should still proofread it.
Rough drafts for people. Some teams have the AI produce a structure or first version, which someone with product knowledge then reworks. This saves less time than hoped. It does, however, help against the blank page.
The cases where it does not pay off are just as clear. Small ranges that need explaining. With thirty products that each have their own story, the text is part of the product. Here a person writes faster than they could build templates, data and review rules.
Regulated categories without expert review. Without anyone able to judge efficacy claims, every generated text carries a risk. Patchy data. As long as the ten-item test from the second section fails, the AI mainly produces correction work. In that case the money is better spent on the data.
What remains after introduction
Generated product copy is not a one-off project. New items arrive, formulations change, models are updated and write differently afterwards. Introducing generation means running a workflow: maintaining data, adjusting templates, sampling output, tracing every error found back to all affected texts. This needs one responsible person and a rule for when texts are regenerated.
A good first step is a small category with the best data in the range. Within a few weeks it shows how much time review and data maintenance really take. Only then is it worth deciding whether to extend the approach to further categories.
How we set up such workflows and look after them over time is described under AI Operations Care.