AI product enrichment: improve your catalogue without inventing claims

AI product enrichment covers tasks such as standardising supplied measurements, locating relevant specifications in approved documents and drafting explanations of verified features. Done well, it gives reviewers a clear chain from source fact to proposed change, and it makes sparse supplier information more useful without pretending missing evidence exists. The aim is a better catalogue and a manageable approval process, not simply more words per product.

Enrichment sits within a wider product data optimisation process. Correcting an attribute, resolving a contradiction and publishing an approved record matter just as much as writing a fluent description.

The need for it shows from the shopper's side. SO-ME, the consumer shopping platform Sighted powers, takes requests in natural language, and finding the retailer product that fits one is harder than matching a keyword, because conversational discovery asks for more depth about a product than a keyword search ever did, and the record has to hold it.

Decide which task you are asking AI to perform. Extracting “outer: 100% cotton” from a supplier specification is different from predicting a garment's composition from its photograph. The first has an identifiable source; the second does not establish the fact.

Give the enrichment system:

• the product and variant identifiers;
• the approved specifications;
• the existing copy;
• the permitted image references;
• the brand-writing rules.

Identify source versions and dates so reviewers know which information each draft used.

Separate facts from suggestions. A supplier's “water-resistant outer” belongs in the evidence packet with its scope intact. “Ideal for heavy rain” is a proposed suitability claim that needs additional support.

Add an explicit list of unknowns. “Laptop compartment dimensions unavailable” is useful input: it tells the drafting process which answer cannot yet be supplied. Without that list, the system may try to complete the story instead of preserving the uncertainty.

This is an illustrative product record, not a client result or a test of any model. The approved specification for a navy overshirt confirms a 100% cotton outer, two chest pockets and a machine-wash instruction of 30°C. It supplies no shrinkage testing, organic certification or fit measurements.

An AI draft proposes: “An organic cotton overshirt with a guaranteed shrink-free finish and a relaxed fit, made for effortless layering.”

The reviewer should not treat that sentence as one yes-or-no decision. Break it into claims:

• Cotton composition is supported for the outer.
• Organic status and shrink-free performance have no evidence.
• A relaxed fit has not been established.
• “Effortless layering” is an interpretation, and it should not imply dimensions or a fit nobody has verified.

An approved alternative might read: “A navy overshirt with a 100% cotton outer and two chest pockets. Machine wash at 30°C, following the garment's care label.” It is less dramatic and much easier to substantiate.

The missing fit information becomes a supplier question. If measurements are later approved, add them to the relevant attributes and the visible size information before drafting a stronger fit explanation.

Approved

The value matches the evidence, applies to the correct item and is expressed without changing its scope. Record the source and the reviewer, then route the change to its intended destination.

Revise

The underlying fact is supported, but the wording overstates it or conflicts with your tone. “Cotton outer” must not become “entirely cotton” when the lining and trims are unspecified.

Evidence required

The proposed detail could help a customer, but the source packet cannot establish it. Assign the question to a person who can obtain the answer, rather than leaving it to cycle through new AI drafts.

Rejected

The claim is incorrect, irrelevant or outside the permitted evidence. Keep the reason so the same suggestion does not reappear on the next run.

These four outcomes make review actionable. A generic confidence percentage rarely tells a merchandiser whether a material claim came from a current specification or an unsupported inference.

Brand voice controls vocabulary, rhythm and emphasis. It does not authorise extra capabilities. A playful brand can describe verified features playfully; it cannot invent a certification because the phrase sounds more compelling.

Provide short examples of approved copy, banned phrases and the level of detail your customers need. Ask reviewers to assess facts first and voice second, so polished wording is less likely to distract from a wrong attribute. The brand voice glossary entry explains that distinction.

Human approval is one checkpoint; the destination's requirements are another. Google Merchant Center requires AI-generated product data to be identified as such (as of October 2026): an AI-generated title goes in the structured_title attribute and an AI-generated description in structured_description, each with its digital_source_type set to trained_algorithmic_media. Build that handling into your publication process rather than assuming ordinary field mapping is enough.

For website content, Google's Search Central guidance on generative AI asks site owners to prioritise accuracy, quality and relevance, and to fact-check AI-generated text before publishing, metadata such as title elements, meta descriptions, structured data and alt text included.

Using AI to generate many pages that do not help users may breach Google's spam policy on scaled content abuse, so producing more pages is no substitute for making them useful.

Keep the approved facts aligned across visible pages and the destinations you support. The guide to feeds, schema and product pages covers that architecture.

Start with a varied small batch: straightforward records, incomplete records and records with conflicting sources. Measure how long a reviewer takes to reach an approval, how many claims need correcting and how often rejected claims return.

A workflow that drafts 1,000 descriptions but leaves 900 unapproved has not necessarily increased usable output. Count approved, verified improvements and track the reasons work is held.

Two of Sighted's pilots ran enrichment on a chosen set of products, the batch this guide recommends, and each was read against something. With The Game Collection, 200 products were enriched and 17 collections created; AI citation coverage went from 0% to 32% of the priority products tested, and the store's organic conversion rate rose 64%. With The Bean Coffee Company, the pilot products were enriched; their revenue rose 31% against 10% across the rest of the catalogue, and search clicks went from 121 to 366.

Both figures sit in their pilot's context: a defined set of priority products in one case, the rest of the same catalogue in the other. That is the comparison an enrichment programme needs before it claims anything, and the testing guide sets out how to build it.

Keep previous values and source references. When a supplier changes a specification, you need to identify which descriptions and attributes depended on it. That record is what makes enrichment maintainable rather than another one-off copy project.

Can AI infer missing specifications from an image?

It can suggest observations for investigation. Do not treat an image-based prediction as verified evidence of composition, certification, compatibility or performance.

Does every AI change need a manual review?

Choose controls according to the type of change and your evidence. A pre-approved formatting transformation differs from a new technical claim. Establish and validate the rules before automating approvals, then audit the exceptions and the outputs.

How do you avoid generic product descriptions?

Supply specific facts, useful customer questions and clear voice examples. If the source packet contains little distinctive information, rewriting it repeatedly will not create that information.

1. Decide which task you are asking AI to perform: extracting a stated fact from a named source, or predicting one.
2. Build an evidence packet for each product, with source versions and dates and an explicit list of unknowns.
3. Review each draft claim by claim and give every suggestion one of the four outcomes, keeping the reason for each rejection.
4. Give reviewers approved copy, banned phrases and the level of detail your customers need, and have them assess facts first and voice second.
5. Mark AI-generated titles and descriptions for Merchant Center, and fact-check website text, metadata included, before publishing.
6. Start with a varied small batch and measure review time, corrections and how often rejected claims return.
7. Keep previous values and source references so a changed specification can be traced to the descriptions that depended on it.

Sighted scores every product on the data shopping assistants read and drafts the missing fields from your own descriptions, options and photos. Your brand voice shapes the wording, not the facts.

• A photo is read only for what it shows, such as a colour, a visible zip or text printed on a label; a material or a fit is never guessed from how the product looks.
• The factual claims in a draft are checked against that data: a specification Sighted cannot verify is held out of what publishes.
• A description carrying an unverified claim goes to your review queue, even when you have chosen to publish automatically.
• On Shopify, each value Sighted writes goes into the category attribute the store already has for it and is recorded with the one it replaced, so you can revert it.
• The free AI visibility audit runs the scoring step on up to 50 of your product pages, and the enrichment is the technology behind SO-ME.