Product Content
Writing product descriptions that AI assistants can actually recommend
AI assistants recommend products whose descriptions answer the question being asked. Most Shopify descriptions answer nothing. Here is the structure that works, a before-and-after rewrite, and how to decide which products to fix first.
August 4, 2026 · 13 min read
An AI assistant recommends a product when the product page explicitly answers the constraint in the shopper's question — material, fit, use case, price, compatibility, care. Descriptions that list only a name, a size and an adjective give the assistant nothing to match against, so they are never retrieved regardless of how good the product is.
There is a specific moment where most Shopify stores lose an AI recommendation, and it is not the moment merchants expect. It is not the crawl. It is not the schema. It is the instant an assistant has your product page in front of it, is holding a shopper's question with three constraints in it, and finds that your page addresses none of them.
The assistant does not guess. It moves to the next page. This article is about making sure your page is not the one it moves past — what the assistant is actually doing when it reads a description, the structure that survives that reading, a worked rewrite, and how to prioritise when you have four hundred products and no appetite to rewrite all of them.
What the assistant is doing when it reads your page
Imagine the shopper asked: "a waterproof backpack for commuting with a laptop, under 100 dollars". The assistant now holds four constraints — waterproof, commuting, laptop capacity, price ceiling — and is looking at candidate pages to see which ones it can defend recommending.
Defending is the key word. An assistant that recommends a product and is wrong looks bad, so it is biased toward products whose pages state the relevant facts plainly. If your page says "water-resistant" it may or may not satisfy "waterproof". If your page says nothing about laptops, the assistant cannot claim your bag fits one. Ambiguity resolves against you every time.
The five questions every description must answer
Across categories, buying questions cluster into the same handful of constraint types. A description that covers these is retrievable for the overwhelming majority of realistic questions in its category.
| Constraint | The question behind it | What to state |
|---|---|---|
| What it is, precisely | "Is this the kind of thing I mean?" | Category, format, quantity, and what makes it different from the adjacent thing |
| Who it suits | "Is it for someone like me?" | Skin type, hair type, body type, skill level, room size — the qualifier that excludes as much as it includes |
| Measurable attributes | "Does it meet my hard requirement?" | Dimensions, capacity, duration, weight, materials, power — with numbers and units |
| Use case | "Will it work for what I'm doing?" | The specific situation: commuting, travel, sensitive skin, gifting, small kitchens |
| Practicalities | "What happens after I buy?" | Care, compatibility, warranty, returns, shipping constraints |
The second row is the one merchants resist. Saying who a product is for means saying who it is not for, and that feels like turning away business. It is the opposite: a product that suits everyone matches no specific question, and specific questions are where recommendations happen.
A worked rewrite
Here is a real pattern — a description that is not badly written, merely empty of retrievable facts — and the same product rewritten to answer questions.
Our signature Amber Oud. A luxurious blend crafted with care. Long lasting and unforgettable. 50ml.Before — 17 words, zero answerable constraints
Read it as an assistant would. Is it alcohol-free? Unknown. Does it suit sensitive skin? Unknown. How long does "long lasting" mean? Unknown. Is it unisex? Unknown. Every question a buyer might ask returns unknown, so the page can be cited for nothing.
A hand-poured amber oud built on warm resin, vanilla and a dry cedar base. It is alcohol-free and unscented at the base, which makes it suitable for sensitive skin and for wearing in close quarters. Expect 8+ hours of wear on skin, longer on fabric. Unisex. The 50ml bottle is refillable, and refills ship in recyclable pouches.After — 58 words, six answerable constraints
Count what the second version can now be retrieved for: alcohol-free fragrance, sensitive skin, office-appropriate scent, longevity questions, unisex gifting, refillable and low-waste packaging. Six distinct question families, none of which the first version could enter.
Structure: the first two sentences do the work
When an assistant quotes or paraphrases your page, it lifts a passage. That passage arrives in the answer without your product title above it, without your images, and without the paragraph before it. It has to make sense alone.
Which produces a concrete rule: the opening two sentences should independently identify the product and its single most differentiating fact. Not a brand story, not a scene-setting image — the thing that would make someone say "yes, that one".
- Sentence one: what it is, stated so a stranger could picture it. Category, format, key material or ingredient.
- Sentence two: the differentiator that answers the most common constraint in your category — the one buyers filter on before anything else.
- Body: measurable attributes with units. Dimensions, duration, capacity, weight, coverage.
- Body: who it suits and who it does not, named plainly.
- Close: practicalities — care, compatibility, what is in the box, what happens if it does not fit.
Bulleted specification lists are worth keeping alongside the prose rather than instead of it. The prose answers "is this right for me"; the list answers "does it meet my hard requirement". Assistants use both, and the list is where numbers and units belong unambiguously.
The failure patterns, and what to write instead
| Not retrievable | Retrievable | Why |
|---|---|---|
| Long lasting | 8+ hours of wear on skin | A number can satisfy a constraint; an adjective cannot |
| Premium quality materials | Full-grain leather, brass hardware | Names a material a buyer can filter on |
| Perfect for any occasion | Works for the office and for evening; too warm for gym use | Includes and excludes, so it matches specific questions |
| Fits most laptops | Fits laptops up to 16 inches; padded sleeve | A measurement, not a hope |
| Gentle formula | Fragrance-free, pH 5.5, no sulphates | Three separately searchable facts |
| Sustainably made | Refillable bottle; refills in recyclable pouches; made in Portugal | Verifiable specifics rather than an unbacked claim |
Which products to fix first
Rewriting four hundred descriptions is a project nobody finishes. Rewriting the right thirty is an afternoon with a measurable outcome. The ordering that matters:
- 1
Start with products that already sell
Rank by revenue over the last 90 days and take the top tier. These have proven demand; a thin description here loses recommendations for questions you would definitely have won.
- 2
Then products with traffic but weak conversion
High sessions, low conversion rate frequently means the page fails to answer the question that brought the visitor. The same gap that loses a human is the gap that loses an assistant, so this work pays twice.
- 3
Then the head of each category
For every category you want to be recommended in, fix the one or two products you would want an assistant to name. A category with no strong candidate cannot be won.
- 4
Fix the template, not just the products
If descriptions are thin because your product-creation process has no content step, rewriting today's catalogue only resets the clock. Add the five constraint types to whatever form or checklist your team fills in when a product is created.
- 5
Re-measure after publishing
Record which buying questions returned your domain before the rewrite, wait, and ask again. Without the before number you cannot tell improvement from noise — and assistant answers are noisy.
The long tail — products with no revenue and no traffic — can wait indefinitely. A thin description on a product nobody asks about costs you nothing, and treating the catalogue as a uniform obligation is how this work stalls.
Writing for two languages
For stores selling in Arabic and English, one trap is worth naming: machine-translating the English description produces text that is grammatically fine and commercially dead. Product terminology in Arabic is not a word-for-word mapping, and a translated description inherits the English sentence rhythm, which reads as foreign to a native shopper and reads as low-quality to an assistant weighing sources.
Write from the same facts rather than from the same sentences. The constraint list — material, fit, use case, care — is language-independent. The prose should be generated natively in each language with a consistent glossary of category terms, so that the same product attribute is called the same thing across your entire catalogue.
Frequently asked questions
How long should a product description be?
Long enough to state every fact a buyer filters on, and no longer. In practice that lands between 80 and 200 words for most physical products. Word-count targets produce padding, and padding dilutes the facts that matter.
Will AI-written descriptions hurt me?
The origin matters far less than the content. A model-written description full of specific, true attributes performs better than a human-written one full of adjectives. The risk with generated copy is not detection — it is fabrication, so every specific claim needs to be checked against what the product actually is.
Should I use the manufacturer's description?
Only as raw material. Identical text across dozens of retailers gives an assistant no reason to prefer your page over anyone else's, and duplicated content has always been a weak position. Keep the specifications, rewrite the prose, and add what you know that the manufacturer does not — how it performs, who returns it and why.
Do bullet points or paragraphs work better?
Both, doing different jobs. Prose carries the reasoning about who a product suits and why, which is what gets paraphrased into a recommendation. Bullets carry unambiguous specifications with units, which is what gets checked against a hard constraint.
Does this help conversion too, or only AI?
It helps conversion first. Every constraint an assistant needs answered is a constraint a human shopper also has, and unanswered questions on a product page are a well-documented cause of abandonment. AI visibility is largely the same work with a second beneficiary.
How do I know a rewrite worked?
Two measurements: whether assistants now return your domain for buying questions that mention the constraints you added, and whether the product's conversion rate moved against its own prior baseline. Both need a before figure recorded before you publish.
The verdict
Product content is the least glamorous and highest-leverage work in AI visibility. Crawler configuration is a one-afternoon fix with a binary outcome. Structured data is a template change. Descriptions are the part that requires knowing your products and saying true, specific things about them — which is precisely why so few stores do it, and why doing it works.
Take your top thirty products by revenue. For each description, mark every sentence with the buyer question it answers. The unmarked sentences are the ones costing you recommendations. That exercise takes an hour and tells you more than any score will.
More in Product Content