How to Write Product Descriptions for AI Shopping Agents

If you sell on Shopify and your brand still gets skipped when a buyer asks ChatGPT for the best wool sweater, the missing piece is usually the product description. Most descriptions are written for a human skimming a page. AI shopping agents do not skim. They parse sentence by sentence and pull out the ones that answer the buyer's question. This guide shows what to write, what to cut, and what good copy looks like next to bad copy in three categories.

The three-sentence pattern that wins

Every product description should hit three jobs in three sentences:

1. Name the concern. What is the buyer worried about? Fit, durability, sound, scent, returns. Say it plainly.
2. Give the answer. What does this product do about that concern? One clear claim.
3. Back it with proof. A number, a material, a test result, or a named third party.

If the page does this in the first 100 words, AI agents can quote the page. If the page buries this under a brand origin story, the page gets retrieved and then dropped. A Search Engine Land study found that only 15% of pages ChatGPT retrieves end up in the cited answer. The pages that survive that second pass share the same shape.

Here it is on a real product, side by side.

Bad (apparel, merino base layer):

Crafted for the modern adventurer, our merino base layer combines old-world craft with a forward-thinking spirit. We believe in clothes that move with you, season after season. Sumptuously soft and ethically made, this is base layer reimagined.

Good (same product):

Worried about itch and odor on multi-day trips? This merino base layer is 100% Australian merino at 200 gsm, washed and tested for next-to-skin softness. Independent reviewers at Outdoor Gear Lab gave it a 4.6 out of 5 for warmth-to-weight and noted it does not smell after three days of hiking.

The first version has zero extractable facts. An AI agent reading it can't tell you what the fabric is, what it weighs, or who tested it. The second version has the fabric, the weight in grams per square meter, the country of origin, and a named third-party score. A buyer asking "best merino base layer for a four-day hike" gets the second one cited.

Why adjective stacks get skipped

The most common mistake on premium DTC pages is the adjective stack. "Buttery soft, sumptuously crafted, ethically sourced." Each word is doing the same job (telling you it's nice) and none of them give an agent something to extract.

An AI agent treats "buttery soft" as a brand opinion. It can't quote it as fact because every brand says it. Compare:

• "Buttery soft" -> not extractable.
• "Mid-weight 320 gsm cotton, brushed twice for softness" -> extractable.

The rule is simple: if a competitor could copy the same sentence and put it on their page, an AI agent will not cite it. Replace adjectives with the thing that produces the adjective. Soft becomes a fabric weight. Premium becomes a country of origin or a tested grade. Durable becomes a warranty or a tear-test number.

The numbers AI agents want to see

Agents read product descriptions for facts that map to fields in their internal model of a product. The fields that matter most:

• Weight in grams or pounds.
• Dimensions in cm or inches, including a key measurement most buyers care about (sleeve length, seat depth, driver size).
• Material with percentages if it's a blend.
• Fabric weight in gsm for textiles.
• Country of manufacture.
• Warranty length in years.
• Care instructions in one line.
• A named third-party fact (review score, certification, lab test, magazine award).

Every field above maps to a property in Schema.org's Product description spec or Google Merchant Center's product description spec, both of which AI agents lean on heavily. Google's own guidance is to put the most important details in the first 160 to 500 characters of the description. AI agents read the same shape.

Side-by-side: furniture

Bad (modular sofa):

Designed for modern living, our modular sofa redefines the way you relax at home. Configurable, comfortable, and crafted with care, it adapts to any space and any mood.

Good:

Worried about a sofa that looks great on day one and sags by year three? This modular sofa uses kiln-dried hardwood frames, 8-way hand-tied springs, and 1.8-density foam with a 10-year warranty on the frame. Each module is 35 inches deep with a 22-inch seat. Ships flat, assembles in 20 minutes with no tools. Apartment Therapy named it one of their top three modular sofas of 2025.

What changed. The bad version is three sentences of vibe with one config detail buried in the middle. The good version names the buyer's actual fear (sagging), then answers it with the frame material, the spring type, the foam density, and the warranty length. The dimensions are stated as inches, not "generous" or "deep." The assembly note replaces "easy to set up." The third-party fact replaces "designed for modern living."

A buyer asking "best modular sofa that does not sag" gets the second one in front of an AI agent because the page directly answers the sag question with a number and a warranty.

Side-by-side: audio

Bad (over-ear headphones):

Immerse yourself in sound. Our flagship over-ear headphones deliver studio-grade audio with a sleek, refined design. Engineered for the audiophile in all of us.

Good:

Looking for headphones that handle long mixing sessions without ear fatigue? These over-ear headphones use 50 mm beryllium-coated drivers, weigh 320 g, and clamp at 4.2 N for a snug but not tight fit. Frequency response is 5 Hz to 40 kHz. Reviewed at 4.5 out of 5 by RTINGS for neutral sound, with a noted bass lift around 80 Hz. 2-year warranty on drivers and headband.

What changed. The bad version is three sentences and zero facts. The good version names the use case (long mixing sessions), then answers with the driver size, weight, clamping force, frequency response, a named lab review, the bass note that lab reported, and the warranty.

A buyer asking "best over-ear headphones for mixing under $400" gets the second one cited because the description directly speaks to mixing and ships every spec a reviewer would compare.

What to cut from existing descriptions

Most premium DTC descriptions hide facts under three layers. Cut them in this order:

1. The brand origin story. "Founded in 2014 by two cousins in Brooklyn." It belongs on the About page, not the product description block. AI agents discount origin copy because every brand has one.
2. The mood paragraph. "For the modern adventurer who values craft." Pure vibe. Replace with the buyer concern.
3. The marketing-deck reuse. Anything copied from a press release or a deck slide. The cadence is too long and the facts are too soft.

If you cut these and find you have 30 words left, that is a sign your product description was 90% packaging. Rewrite using the three-sentence pattern, then add the fact list (weight, dimensions, material, warranty, third-party note).

How to handle variants without repeating yourself

A jacket in five colors should not have five copies of the same description. The shared facts (fabric, fit, country, warranty) live in the parent description. The variant-specific facts (color name, pantone-ish reference, season availability) live in the variant block. AI agents read both and prefer the parent for the core extract.

The shape that works:

• Parent description: 150 to 250 words covering concern, answer, proof, and the spec list shared across variants.
• Variant block: 1 to 2 sentences per variant covering color or size-specific notes, with the variant inventory and price exposed in your agent-readable product feed.

If your storefront repeats the parent description on every variant URL, AI agents see duplicate content and pick one to cite (often the wrong one). Make sure your product schema uses one canonical description and references variants through hasVariant or offers.

What stays in marketing voice

Not every word on a product page has to read like a spec sheet. The hero headline, the section breaks, the brand story further down the page, the photography captions: these can stay poetic. They serve a buyer who is already on the page and feeling the brand.

The description block is different. The description block is the part that ships in the merchant feed, gets pulled into the search results snippet, and lands in the AI agent's extract pass. That block has to be plain. Keep the brand voice for the parts of the page humans linger on.

The 60-minute rewrite workflow

You do not need a content sprint to ship this. The fastest path on a Friday afternoon:

1. Pick five PDPs. Start with your top three sellers and your two highest-traffic-but-low-conversion products.
2. Score each on the three-sentence pattern. Read the first 100 words. Does it name the concern, give the answer, and back it with a fact? If not, mark it for rewrite.
3. Rewrite the first 100 words. Use the patterns above. Concern, answer, proof. Then add the fact list.
4. Cut the origin story and adjective stacks. Move them to the brand or product story sections of the page.
5. Push live. Most Shopify themes let you edit the description block without a deploy.
6. Measure citation lift over 14 days. Ask ChatGPT, Claude, and Perplexity for "best [your category] under [your price]" once a week. Track which products get cited and how the descriptions get summarized.

You will not see a citation lift on every product. You will see it on the products where the rewrite was the missing piece. That tells you whether the gap was copy or signal somewhere else (schema, feed, third-party reviews).

Where this fits in the bigger picture

A clean product description is one piece of a working AEO setup for ecommerce. Schema, feed, and third-party reviews matter too. The description is the part you control today, with no developer time, and the part most brands have neglected the longest.

If your brand is invisible to ChatGPT and you are not sure where to start, the description rewrite is the cheapest test. The full citation playbook covers what to ship after the descriptions are clean. The recommendation guide covers what to expect once the page is extractable.

The buyer is asking the assistant. The assistant is reading the description. Either your sentences answer the question or someone else's do.