Optimizing your product feed for AI product advice: the technical guide 2026
AI product advice is only as good as the product feed behind it. This guide shows which attributes, variant relationships and update cycles a Shop Agent needs in 2026 and how to test whether your feed is ready for advice using real customer questions.

Optimizing your product feed for AI product advice: the technical guide 2026
TL;DR: the essentials in 30 seconds
- Key point: AI product advice can only recommend as well as your product feed is structured. If an attribute is missing, the agent can neither filter nor compare by it.
- Four data layers: a feed that is ready for advice contains identification (ID, GTIN, brand), advisory attributes (material, dimensions, intended use), relationships (variants, accessories, compatibility) and freshness (price, stock, delivery time).
- Structure beats volume of text: every property belongs in its own field with a consistent unit and fixed values. Facts that only appear in marketing copy can't be evaluated reliably by the AI.
- Lever for returns: in KPMG's Consumer Snapshot on return behaviour, 60.4% of online shoppers cite inaccurate descriptions or deviations from product images as a reason for returns.
- One feed, multiple channels: a clean feed serves the Shop Agent in your own shop, Google Merchant Center and AI shopping channels such as ChatGPT at the same time.
- Test instead of hoping: check the feed before go-live with real customer questions from chat, tickets and on-site search.
Definition:
A product feed for AI product advice is a structured file or interface that describes every product with a unique ID, advice-relevant attributes, variant relationships and current price and stock. An AI agent uses this data to match customer needs expressed in natural language to suitable products, compare products and answer questions based on facts.
You optimize a product feed for AI product advice by moving every advice-relevant property into its own, consistently filled field, linking variants and accessories cleanly and keeping price and stock up to date. The product description remains important for customers, but the AI also needs the same facts in structured form. This guide shows which attributes a feed needs in 2026, how to normalize values, how to connect data from Shopware and Shopify and how to test whether your feed is ready for advice using real customer questions.
Why does the product feed determine the quality of AI advice?
The product feed is the knowledge source from which an AI agent derives its product recommendations. Whatever is missing or contradictory there, the agent can't answer correctly.
Modern advisory agents work with Retrieval-Augmented Generation (RAG). With this approach, the system retrieves the matching records from the catalog for every customer question and formulates the answer on that basis. The language model provides language understanding and phrasing; the facts come from your feed. If a customer asks for a day-hike backpack under 1.2 kg with a compartment for a hydration bladder, the agent has to check three conditions: intended use, weight and features. If the weight only appears as "extra light" in the marketing copy, the match fails. The agent then recommends unsuitable products or stays vague.
The consequences reach all the way to your return rate. In the KPMG Consumer Snapshot on return behaviour in Germany from December 2025, 72.6% of online shoppers cite the wrong size and 60.4% cite inaccurate descriptions or deviations from product images as a reason for returns. Both causes depend directly on the quality of the product data. AI advice that reads sizes, dimensions and materials from clean fields addresses exactly that.
What distinguishes an advisory feed from a shopping feed?
A shopping feed is built for display in search engines, an advisory feed for the conversation with customers. The shopping feed mainly has to fulfil mandatory fields. The advisory feed additionally has to answer every question an experienced sales assistant in a store would answer.
The Google Merchant Center product data specification defines mandatory attributes such as ID, title, description, product link, image link, price and availability. Among the common issues, Google lists incorrect values for product category or GTIN and missing variant attributes such as item_group_id, color or size. This feed is a good starting point because it already contains unique IDs, prices and variants. For advice, it rarely suffices. It usually lacks exactly the attributes customers ask about: what is the product suitable for? Which device is it compatible with? Does it fit true to size?
The difference lies in the perspective. The shopping feed describes what a product is. The advisory feed also describes who it suits and what it is for. An AI agent understands needs, but it can only map them to products if the matching attributes are in the feed.
Which attributes does a product feed for AI product advice need?
A product feed that is ready for advice needs attributes on four layers: identification, advisory attributes, relationships and freshness. The graphic shows how typical marketing copy maps onto these four layers.
One product, four data layers
On the left, the product knowledge as many shops maintain it. On the right, the same content as a record that is ready for advice. Choose a layer to see where the information comes from.
Before: facts in marketing copy
The Ridge 30 by Talweg is our lightweight all-rounder for day hikes. At just 1.1 kg and 30 litres, it carries everything for a long day on the trail. The polyamide fabric is water-repellent, and the back has a sleeve for hydration bladders up to 2 litres. Pairs perfectly with our Rain Cover M. Available in blue and green. Now just €89.90, ready to ship.
After: structured record
Identification
Advisory attributes
Relationships
Freshness
"I'm looking for a backpack for day hikes, under 1.2 kg and with room for a hydration bladder."
Fictional example product. The values are for illustration only.
Identification: how does the AI recognize a product unambiguously?
Every product needs a stable, unique ID that doesn't change with updates. Add the brand, model name, manufacturer part number (MPN) and, where available, the GTIN. The GTIN (Global Trade Item Number, formerly EAN) is a globally unique article number whose allocation in Germany is managed by GS1 Germany. It helps the AI recognize identical products across channels, so duplicates don't show up as two different recommendations.
Advisory attributes: what do customers really ask about?
Advisory attributes are all properties that influence a purchase decision. They include material, dimensions, weight, size and fit, intended use, target group, experience level, care instructions and technical specifications. Which attributes matter depends on the product range. A bed shop needs mattress size, type of wood and load capacity. A bike parts shop needs wheel size, axle standard and brake system. A fashion shop needs cut, fit and sizing notes such as "runs small".
A dedicated field for intended use is particularly effective. Customers rarely describe technical values; they describe situations: "for beginners", "for winter", "for small apartments". If your feed lists such usage contexts as fixed values, the AI reliably bridges the gap from need to product.
Relationships: how are products connected?
Relationships link a product to its variants, its accessories and compatible products. Variants such as colors and sizes belong together via a shared group ID, which in the Google feed is item_group_id. This lets the agent recognize that a backpack in blue and the same backpack in green are one model, and ask specifically which variant the customer wants. Compatibility and accessories are the basis for meaningful cross-selling. Without these links, the AI suggests accessories based only on text similarity, which quickly leads to wrong recommendations.
Freshness: which data has to be updated daily?
Price, availability, stock level and delivery time change constantly. These fields need the shortest update cycles. If an agent recommends a sold-out product or quotes an old price, it loses customers' trust faster than through any imprecise wording.
How do you structure attributes so the AI can evaluate them reliably?
Structure attributes according to three rules: one value per field, one unit per numeric value and fixed value lists. This normalization decides whether the AI can filter, sort and compare.
The first rule concerns mixed fields. A "Details" field containing "30 l, 1.1 kg, polyamide, blue" is readable for people but useless for a comparison. Split such fields into volume, weight, material and color. The second rule concerns units. If the weight appears once in grams, once in kilograms and once as "approx. 1 kg", the agent can't sort reliably. Define one unit per attribute and store numbers as numbers. The third rule concerns value lists. If a material is called "polyamide" in one place, "nylon" in another and "PA" in a third, those are three different values for filter logic. Define a list of permitted values per attribute and maintain synonyms in one central place, ideally in your PIM system (Product Information Management, the central database for product information).
Also be honest about gaps. An empty field is better than an estimated value. An agent with a clean RAG connection then answers that no information is available on that point and, if in doubt, hands over to your team. An incorrectly filled field, by contrast, leads to a wrong recommendation that sounds convincing. Experts call this a hallucination, even though the error actually lies in the data.
How important are product titles and descriptions for the AI?
Titles and descriptions remain important, but they complement the attributes rather than replace them. The AI uses the text to understand benefits and context. For hard facts, it relies on the fields.
Write titles according to a fixed pattern of brand, product type, model and the most important distinguishing feature. Google limits titles in Merchant Center to 150 characters, a length that also works well for advice. Write descriptions so they explain facts rather than just repeat them: why is the material suitable for rain? What length of hike does the volume cover? Explanations like these help the AI convey the benefits to customers in its own words.
Contradictions are critical. If the field says "Weight: 1,100 g" and the text says "only 900 grams light", the agent has to guess. So check whether text and attributes state the same values. For large product ranges, an automated check that compares numbers in the body text with the structured fields pays off.
How do you connect the product feed to an AI agent?
There are three common ways to connect it: a direct interface to the shop system, a feed URL that is fetched regularly or a file upload as CSV. Which one fits depends on your shop system and how often your data changes.
The direct interface is the best choice for price and stock because the agent sees changes without an export run. Shopware 6 and Shopify provide APIs for this, through which products, variants and availability can be retrieved. A feed URL suits systems such as WooCommerce or Magento and shops that already generate a feed for Google. CSV upload is the simplest way to get started but only fits product ranges that rarely change. As a rule of thumb: price and stock as close to real time as possible, static attributes with every change, but at least daily.
Providers like melibo combine these approaches. The Shop Agent is connected via the Shopware integration or directly to Shopify; other shop systems deliver their product data via feed URL or CSV into the melibo Knowledge Hub. During setup, an AI assistant suggests which field in your feed corresponds to which advisory attribute, for example when a field is internally called "mat_1" instead of "Material". Other providers also offer advisory agents, for example moinAI with its AI product advisor or trbo with the trbo Agent. The same applies to all of them: the platform provides language understanding and conversation management, while your feed determines the quality of the advice.
Step by step: how to optimize your product feed in six steps
The following order deliberately starts with your customers' questions, not with your existing data structure.
1. Collect customer questions:
Analyze chat histories, tickets and search terms from your on-site search and note which properties customers ask about. This list is your target picture for the attributes.
2. Assess the current state:
Export your feed and determine the fill rate for each attribute. Flag fields with mixed values, inconsistent units or pure free text.
3. Define the attribute schema:
For each category, define the mandatory attributes, one unit per numeric value and fixed value lists. Create intended use as a separate field.
4. Enrich and normalize data:
Split mixed fields, standardize values and fill in missing information from manufacturer data or data sheets. Leave fields empty if a value isn't verified.
5. Link relationships:
Group variants via a group ID and store compatibility and matching accessories as references to product IDs.
6. Automate updates:
Decide which fields come in real time via the interface and which via feed export, and monitor the export for failures.
How do you test whether your feed is ready for advice?
Test your feed with a fixed set of real customer questions before the agent goes live. As a starting point, we recommend 50 to 100 questions covering all important categories and typical needs.
Assess every answer against three failure patterns. If the agent recommends an unsuitable product, an attribute is usually missing or a value is mapped incorrectly. If the answer stays vague, the attribute exists but is sparsely filled or hidden in body text. If the agent mentions a property the product doesn't have, there is almost always a contradiction between text and fields or an incorrectly mapped value. Each failure pattern leads you back to a specific spot in the feed.
After go-live, testing continues in day-to-day operation. Conversation analyses show which questions the agent fails on. In melibo, this is handled by the Bot Gym training area, where you evaluate conversations and refine advisory rules. Metrics such as the click-through rate on suggested products and the revenue generated through the chat show whether your feed improvements are working.
Is the effort also worth it for Google and ChatGPT?
Yes, because in 2026 the same structured feed serves several AI channels at once. Besides your own Shop Agent, external shopping assistants also read product data from feeds and structured data.
In its product feed specifications, OpenAI describes how merchants provide catalog data for product search and purchase in ChatGPT. Google supplies its shopping surfaces through Merchant Center. On the website itself, the schema.org/Product vocabulary ensures that machines can read product data directly from the HTML. For properties without their own schema.org field, there is the additionalProperty property; for identifiers, the gtin property. How AI agents use this data in the buying process is explained in the article on agentic commerce 2026.
The difference lies in control. In your own shop, you use advisory rules to decide which follow-up questions the agent asks and which products it prioritizes. In external AI channels, the platform decides. A good feed improves both sides, but only in your own shop do you control brand voice, advisory logic and customer data yourself. The results this delivers in practice are shown in the article on AI product advice in online shops.
What are the most common product feed mistakes shops make?
The most common mistakes are facts buried in body text, inconsistent values and outdated stock levels. They usually don't arise from carelessness but because the feed was originally built for a different purpose.
Many shops maintain product knowledge where it is easily visible to people: in descriptions, image captions or PDF data sheets. For the AI, this knowledge is hard to access. A second classic is a category tree that has grown over time, in which identical products are sorted differently. The agent then compares products that aren't comparable. A third mistake concerns variants that are listed as standalone products. Customers then get five nearly identical recommendations instead of one recommendation in the right size.
Ownership of the data is often underestimated. Without a clear owner, feed quality gradually deteriorates. So define who specifies new attributes, who closes gaps and who fixes errors surfaced by conversation analysis.
Free strategy consultation
How well can your product catalog already advise?
In 30 minutes, we'll work out together whether your product range benefits from active AI advice and which product data your Shop Agent needs for it. Our team will then support you in structuring your feed.
Frequently asked questions about product feeds for AI product advice
Which attributes matter most for AI product advice?
The most important ones are the attributes your customers actually ask about. In almost every product range, these include material, dimensions, size or fit, intended use, target group and compatibility. They are complemented by unique IDs, variant groups and current prices and stock levels.
Is my Google Shopping feed enough for an AI shop agent?
It's a good starting point but rarely enough on its own. The Google feed contains IDs, prices, availability and variants. For advice, it usually lacks attributes such as intended use, compatibility or fit notes.
How often should a product feed for AI advice be updated?
Price, stock and delivery time should be synchronized as close to real time as possible, ideally via a direct interface to the shop system. For static attributes such as material or dimensions, updating with every change in the shop or PIM system is enough, but at least once a day.
Can AI fill in missing product data by itself?
AI helps with preparation, for example by splitting mixed fields or mapping field names. It shouldn't invent missing facts. Values such as weight, material or compatibility belong in the feed from verified sources such as manufacturer data or data sheets.
How long does it take to make a product feed ready for advice?
That depends on the size of the product range and the starting quality. Shops with a well-maintained Google feed and a PIM system often only need to add advisory attributes. At melibo, the average time from contract signature to go-live of a Shop Agent is around 30 business days, and the team also supports you in structuring your feed.
Conclusion
AI product advice is no substitute for good product data; it makes it visible. The more clearly your feed covers identification, advisory attributes, relationships and freshness, the more precisely the agent recommends and the less often products come back as returns. Start with your customers' questions, build your attribute schema around them and test the advice before it goes live.




