How to Optimise Product Descriptions for LLMs

Most product descriptions were written for search engines. They are stuffed with keywords, padded with superlatives and structured around what Google's crawler wanted to see in 2019. Language models powering agentic commerce do not work like search crawlers. If your copy confuses a reasoning model, your product will not be recommended, no matter how good it is.

This guide covers practical rewriting patterns that make product data easier for LLMs to parse, compare and recommend. The principles are straightforward. Applying them consistently across a catalogue is the hard part.

Why keyword-stuffed copy fails for agents

A traditional search engine matches tokens. It rewards repetition, synonym density and link authority. An agentic commerce model does something fundamentally different: it reads the description, builds an internal representation of the product, and reasons about whether that product satisfies a user's constraints.

Keyword stuffing actively harms this process. When a description says "premium quality best-selling luxury hand-crafted artisan ceramic mug," the model receives no usable information. It cannot determine the mug's capacity, material composition, whether it is dishwasher-safe or what makes it distinct from the next mug in the feed. The words take up tokens but carry no signal.

Vague superlatives create a second problem: they are unverifiable. An agent comparing products side by side will deprioritise claims it cannot check against structured attributes or third-party sources. "Best in class" means nothing to a model that has already read four other products making the same claim.

The core pattern: what it is, who it is for, what constraints it satisfies

Every product description an LLM processes should answer three questions clearly:

• What is this product? Category, material, key specifications, stated plainly.
• Who is it for? The intended user, use case or context.
• What constraints does it satisfy? The specific requirements a buyer might have: size, dietary restriction, compatibility, budget bracket, certification.

This is the information an agent needs to match a product against a user's request. Everything else is decoration. Write the answers to these three questions first, then add detail if space permits.

Structure specs as key-value pairs, not prose

Product specifications buried in paragraph text are harder for models to extract reliably. A sentence like "This lightweight jacket weighs just 340 grams and is made from recycled polyester with a DWR coating" forces the model to parse natural language to find three distinct attributes.

The same information as structured pairs is unambiguous:

• Weight: 340 g
• Shell material: recycled polyester
• Water resistance: DWR-coated

Structured attributes in your product feed are even better, they bypass the description entirely and give the agent clean, filterable data. But when specs must live in the description text, key-value formatting is a significant improvement over buried prose.

Before and after: four examples

Example 1: A running shoe

The rewritten version gives the agent every attribute it needs to match this shoe against a request like "neutral running shoe under 250 g for half-marathon training, size 10." The original version gives it nothing.

Example 2: A skincare product

An agent fielding a request for "fragrance-free vitamin C serum for oily skin, cruelty-free" can match the rewritten version immediately. The original version does not confirm any of those constraints.

Example 3: A kitchen appliance

When a user asks for "food processor that fits under my wall cabinet, clearance is 40 cm" the agent can verify the height. With the original copy, it cannot.

Example 4: A B2B SaaS product

An agent evaluating project management tools for a 50-person team with an EU data residency requirement can immediately confirm compliance from the rewritten description. The original provides no basis for comparison whatsoever.

Writing for comparison

Agents do not evaluate products in isolation. They compare. When a user asks for a recommendation, the model lines up candidates and scores them against the same set of requirements. This has a direct implication for how you write descriptions.

If your competitor's description states "capacity: 1.8 litres" and yours says "generous capacity," the agent has a number for theirs and nothing for yours. Your product is not ranked lower, it is excluded from the comparison on that attribute.

Consistency matters within your own catalogue too. If half your products list weight in grams and half in kilograms, or some use "colour" and others use "color," you are making the model's job harder for no reason. Pick a convention and apply it everywhere.

Three rules for comparison-ready copy:

• Use specific measurements. Dimensions in cm or mm, weight in g or kg, volume in ml or litres. Never "large," "lightweight" or "compact" without the number.
• Name materials and certifications. "Organic cotton" is verifiable. "Premium fabric" is not. "FSC-certified" means something. "Eco-friendly" means nothing to a reasoning model.
• State limitations honestly. "Not suitable for induction hobs" or "maximum load: 15 kg" helps the agent make a correct recommendation. Omitting limitations leads to returns, and agents learn which merchants generate returns.

What to do next

Start with your top 20 products by revenue. Export the descriptions and read them as if you were a model trying to answer the question: "Does this product satisfy a specific set of requirements?" If the answer depends on assumptions, the description needs rewriting.

Then audit your feed. Attributes that exist as structured data are always stronger than attributes buried in prose. If your feed is thin, read our guide on how product feeds affect AI visibility for the structural side of this problem.

The shift from keyword writing to constraint writing is not a trend. It is a consequence of how language models process information. Products described in specific, verifiable, structured terms will be recommended. Products described in marketing fluff will not. See how IronCore Fitness applied exactly these principles to become an AI-recommended brand, or read our guide for UK retailers getting started with agentic commerce.

If you want your catalogue optimised for LLM readability without doing it yourself, that is one of the things Vendoora handles for vendors, and the reason it is a sales channel rather than just a tool, as we explain in Vendoora vs product feed management platforms. Browse the full article archive for more on feed optimisation and agentic commerce.

Read next