- The reason AI product descriptions read as generic is almost never the model. It is the input: a product name and a category cannot produce a sentence about fit, weight or compatibility, so the model fills the gap with adjectives.
- Baymard's benchmark finds 10 percent of top ecommerce sites still ship descriptions that do not meet shopper needs, and the missing pieces are consistent: materials, measurements and what the thing works with.
- Nielsen Norman Group traced 20 percent of task failures in ecommerce testing to incomplete or unclear product information, which is a conversion problem long before it is a writing problem.
- Google does not rank by how text was produced. It acts on scaled output with no added value, and its Merchant Center spec now carries separate structured fields for AI generated titles and descriptions.
- The step almost everyone skips is the verification pass. A model that invents a certification or a wrong socket size costs you a return, a dispute, and in some categories a regulator's attention.
- Feed the row, never the product name. If an attribute is empty in your data, the description gets no sentence about it. That single rule removes most hallucinated specifications.
Three hundred products, a spreadsheet exported from an old system, and a weekend. That is the shape of the job for most people who reach for AI here. Nobody is writing a sonnet about a tent peg. They are trying to get a catalogue live before the season starts, with descriptions good enough that shoppers do not bounce and Google does not treat the pages as filler. The filler worry is narrower than it feels, because the spam policy targets volume without added value rather than AI authorship.
This piece is the long version of how to do that. It is written to stay useful, so the specifics that will change are datestamped and the method is the part meant to last. Everything below was checked against live documentation in August 2026.
Why do AI product descriptions all sound the same?
Because the model was given nothing to work with. Ask any assistant for a description of "Merino crew neck sweater, navy" and it will produce four confident sentences about comfort and versatility, because that is the only thing derivable from those four words. The output is not lazy. It is a correct response to an empty brief.
Nielsen Norman Group tested this class of output directly and published its findings in the article on why product specific generative AI needs to write for the web. The recurring faults are worth memorising because they are exactly what you will be editing out: too long for the question asked, bullet lists that are formatted but not scannable, a structure that buries the important fact underneath supporting detail, and jargon delivered to a general audience. One participant's reaction to a yes or no question answered at length was that it was a lot of information for just yes or no.
The same researchers found that human revised versions consistently beat the AI originals on clarity and concision, and that broad assistants produced more readable output than the narrow AI features embedded inside retail products. Neither finding argues against using AI. Both argue that the generation step is the middle of the process, not the whole of it.
What a shopper is actually looking for
Not prose. Baymard Institute's benchmark of leading ecommerce sites found that 10 percent still carry product descriptions insufficient for users' needs, and its usability sessions show what happens next: people abandon the product page, sometimes the site, and they make wrong assumptions that come back later as returns. Participants spent one to three minutes hunting elsewhere on the page for a detail the description should have carried, and they avoided chat support rather than asking.
Three categories of missing information account for most of it.
- Materials and ingredients. Critical in beauty, health and apparel. Baymard reports that half of beauty shoppers need ingredient information before buying. Shoppers also value explicit exclusions, the BPA free style of statement, because absence is a claim a list cannot make.
- Dimensions. Labelled, with units, covering both the overall object and the parts people care about. A second measurement system helps if you sell across borders.
- Compatibility. For accessories, spares and consumables this is the purchase decision. Shoppers will not guess, and a missing compatibility line reads as a reason to leave.
Nielsen Norman Group's guidance on writing better product descriptions puts a number on the cost: 20 percent of overall task failures in their ecommerce studies came from incomplete or unclear product information. It also adds a requirement that most catalogues fail quietly, which is comparability. Two similar products described at different levels of detail are hard to choose between, and a shopper who cannot choose does not buy the cheaper one, they leave.
Start from your attributes, not from a prompt
Here is the method. It is five steps, and the order matters more than any individual prompt.
One: export what you already know. Every product you sell has facts attached to it somewhere, even if that somewhere is a supplier PDF and your own memory. Get them into columns: size, weight, material composition, care, country of origin, what it fits, what is in the box, warranty. This is unglamorous and it is the entire difference between a catalogue that reads like yours and one that reads like everyone else's. If the data is a mess, our walkthrough of moving product data between systems with AI covers the cleanup before you get here.
Two: write one description by hand. One. Pick a product you know well and write the version you would defend to a customer on the phone. This is not a prompt template, it is a pattern: it fixes the order of information, the length, the voice, and what you refuse to say. Everything downstream is measured against it.
Three: generate from the row. Pass the attribute row, not the product name. Instruct the model that any attribute that is blank produces no sentence. That one rule removes the majority of invented specifications, because a hallucinated fabric weight almost always arrives to fill a silence. Ask for the structure your hand written example uses, and ask for the key fact in the first line, which is where people read most and where Nielsen Norman Group's skim research says attention actually is.
Four: verify the claims. Covered in its own section below, because it is where the money is.
Five: measure the page, not the paragraph. Four weeks after publishing, look at returns, presale questions and add to cart rate for the products using the new template against those still on the old one. Description quality is measurable and almost nobody measures it.
Where each version of the description has to live
A description is not one artefact. The same product needs a page version, a feed version and a structured data version, and they have different rules. Getting this wrong is the most common reason a shop's products stop appearing in shopping surfaces without any obvious error message. The table below is assembled from Google's own specifications as they stood in August 2026.
| Destination | Limit | What is disallowed | What the AI step should output |
|---|---|---|---|
| Product page body | No hard limit | Nothing formal, but padding costs you | Lead line, scannable spec list, then detail |
| Merchant Center title | 150 characters | Promotional text, all capitals, gimmick characters | Product identity plus the variant that distinguishes it |
| Merchant Center description | 5,000 characters | HTML, links, sales information, accessory details | Plain text with line breaks and lists only |
| Structured data | Governed by required properties | Data that contradicts the visible page | Values that match the page exactly, fewer and accurate |
| Marketplace listing | Per marketplace | Varies, usually contact details and cross selling | A trimmed variant, not a copy paste |
Two of those rows deserve a note. Google's product data specification requires that the feed description match the description on your landing page and stay focused on the product itself, which rules out the promotional flourish an AI tool adds by default. It also now exposes structured alternatives to the title and description fields intended for AI generated content carrying the appropriate label. On the markup side, Google's introduction to product structured data is explicit that supplying fewer complete and accurate properties beats supplying every possible property badly, which is the opposite of what a generation script tends to do when told to fill every field.
If you sell on marketplaces as well as your own site, the mapping is its own job. We covered the parts worth overriding by hand in the piece on what to override in AI generated marketplace listings.
Does Google penalise AI written product descriptions?
No, and the confusion here costs shops real time. Google's documentation on using generative AI content on your website states that its systems reward quality rather than judging how content was produced. What it does act on is scaled output produced without adding value for users, which its spam policy calls scaled content abuse.
The practical line runs roughly here. Three hundred descriptions generated from three hundred genuinely different attribute rows, each carrying facts a shopper needs, is a catalogue. Three hundred pages spun from a template with the product name swapped in is the thing the policy exists to catch. The difference is not detectable by counting AI usage. It is detectable by asking whether the page answers a question that could only be answered by someone who has the product.
Two operational details follow from the same documentation. Accuracy obligations extend to the metadata, so an automatically generated title tag, meta description, structured data value or image alt text is held to the same standard as the body. And for ecommerce specifically, AI generated imagery is expected to carry IPTC DigitalSourceType metadata, with AI generated product data marked separately. If you are also generating your photography, the labelling question is covered in our piece on AI generated product images and where they are allowed.
Disclosure and detection are different questions. Nothing in Google's guidance asks you to put a badge on the page saying a model helped. What it asks is that the content earn its place. Our separate look at AI content detection for shops explains why detector scores are a poor proxy for either.
The verification pass, and why it is not optional
This is the step that separates a catalogue you can defend from one that generates disputes. A language model asked to describe a product will produce a plausible specification when the real one is absent, and plausible specifications are the expensive kind of wrong. Five claim types account for nearly all of the damage.
- Measurements. Wrong dimensions produce returns, and returns in apparel and furniture are where a small shop's margin goes to die. Check every number against the source row, not against the previous description.
- Materials and composition. A fabric blend or an ingredient list is a legal claim in several categories, not a descriptive flourish. If the supplier sheet says 70 percent, the page says 70 percent.
- Compatibility. The single most damaging invented fact for anyone selling parts or accessories, because the customer discovers it after installing something.
- Certifications and standards. Models are fluent in the names of standards. A generated line claiming a certification you do not hold is the one item on this list that can bring a regulator rather than a refund, and it sits next to the other legally required fields most product listings are missing.
- What is in the box. Cheap to get right, disproportionately common as a complaint, and it drives contact volume that eats your week.
The efficient way to run this is not to reread every description. It is to have the generation step emit, alongside the text, a list of the attribute values it used. Then you are comparing two short lists rather than proofreading prose, which is a task a person can actually do three hundred times.
How long should a product description be?
As long as the questions require, and no longer. There is no ranking threshold to hit, and the widely repeated word counts are inventions. The useful test is the one Nielsen Norman Group's skim research implies: a reader gets the most value from the beginning of the description and the beginning of each line, so the first line has to carry the fact that decides the purchase.
In practice a workable shape for most physical products is a lead line of one or two sentences, a scannable list of the specifications a shopper needs to compare, then any longer explanation for people who want it. A tent peg needs three lines. A mattress needs considerably more, because the questions are more numerous. Consistency inside a category matters more than absolute length, which is the comparability point from earlier.
Running this across three hundred products without losing the plot
Batch by template, not by alphabet. Products that share a shape of question, all the cables, all the candles, all the dresses, should be generated together against the same hand written example, because that is what produces the comparability shoppers need. It also makes the verification pass fast, since you are checking the same five fields repeatedly instead of context switching.
Variants are the trap. A colour variant does not need its own description, it needs the shared description plus the distinguishing attribute, and generating one per variant is how a shop ends up with hundreds of near duplicate pages. Decide the variant policy before generation, not after. Getting the underlying grouping right first is why we wrote about categorising a catalogue with AI: the structure decides how much writing there is to do.
Keep the source of truth in the data, not in the published text. When a supplier changes a specification, you want to edit one cell and regenerate, rather than hunt through prose. That is the argument for owning the pipeline, and it is the reason our own AI ecommerce store builder writes the generated code and content into a repository you control rather than locking the catalogue inside a hosted editor. When the data and the templates are yours, the second pass costs an afternoon instead of a rewrite.
On tool choice, there is no single correct answer and the market moves quarterly. We keep a working comparison of AI copywriting tools and what each is actually good at, and the honest summary is that the difference between tools is small next to the difference between a good attribute table and a bad one.
What belongs in the prompt, and what does not
Prompts are the part of this people over invest in, so keep the guidance short and spend the saved time on the attribute table. Four instructions do most of the work, and they are stable across whichever tool you use this year.
The first is the role of the data: state that the attached row is the only source of fact, and that no sentence may assert anything absent from it. The second is the shape: give your hand written example and ask for the same order of information rather than describing the order in words, because a sample is a far more precise instruction than an adjective. The third is the audience: name who buys this and what they already know, since Nielsen Norman Group's testing found unexplained jargon to be one of the standard failure modes when a model is left to guess. The fourth is what to refuse: list the words you never want, and if you sell in regulated categories, the claims you are not permitted to make.
What does not belong is the marketing brief. Asking for copy that is engaging, persuasive or that converts pushes the output toward exactly the atmospheric filler that costs you the first line of the description. Ask for accuracy and structure. Persuasion in a product description is mostly the absence of unanswered questions.
Length control is worth a specific instruction rather than a word count, because word counts produce padding when the data is thin and truncation when it is rich. Ask for one sentence per attribute group present in the row. A sparse row then yields a short description honestly, which is the correct behaviour.
Selling in more than one language
Machine translation of a finished description is the wrong order of operations, and it is the mistake that fills European catalogues with sentences that are grammatically fine and commercially wrong. Sizes, units, voltage, certifications and the legally required wording differ by market. If you translate the English output you carry the English assumptions with it.
Generate per market from the same attribute table instead, with the market's units and its required phrasing as part of the instruction. The attribute table is language neutral, which is precisely why building it first pays for itself twice. Our longer treatment of this sits in the piece on running a multilingual store with AI translation, including where a human review is still non negotiable.
One practical warning about feeds. Google's specification asks the feed description to match the landing page for that locale, so a market whose page you translated but whose feed you did not will produce a mismatch. That class of error tends to surface as quiet disapprovals rather than a visible failure, which is why it can persist for months.
When should you not use AI for this at all?
Three cases, and recognising them saves more than any prompt. The first is a catalogue under twenty products where you know each item personally. The setup cost of the method exceeds the writing, and your own sentences will be better. The second is anything where the description is the product: bespoke work, art, vintage items whose value is in provenance and condition. A model has no access to the thing that makes those sell. The third is a category where a wrong claim is a legal event rather than a return, such as supplements, medical devices or children's products. Use AI to draft the structure there if you like, but the claims come from the documentation and a person signs them off.
There is also a softer case. If your differentiator is voice, if people buy from you because of how you write, generating the catalogue will erode the asset you are actually selling. Use the method for the specification half and keep the voice half yours. That split works better than most people expect, because shoppers read the two halves for different reasons.
What does good look like after four weeks?
Three numbers, all of which you already have. Return rate for the products on the new template versus the old, because wrong or missing specifications show up there first. Presale contacts per hundred sessions, because a description that answers the question stops the email. And add to cart rate on the product page, which moves when the lead line does its job. If none of the three moves, the descriptions may be pleasant and still not be doing any work.
One more measurement, easy to forget. Search whether your own descriptions are being quoted back by AI answer engines when someone asks about a product like yours. The pages that get lifted are the ones with clear, factual, structured statements rather than atmosphere, which is the same shape of writing that converts. Our piece on what AI shoppers read on a product page goes through the specifics of that.
A short version, if you only do three things
Build the attribute table before you open any AI tool. Forbid the model from writing sentences about attributes you did not supply. Check the numbers before publishing. Everything else on this page is refinement of those three, and a shop that does only those will already be ahead of the tenth of the market Baymard measures as falling short.