Customers used to type "running shoes men 44" into search. Today they ask: "Which running shoes with good cushioning suit a beginner running on asphalt?" No results list answers that question any more — an AI system does, and to do it, it needs something many retailers have treated as an afterthought: structured, meaningful product data.
At Google Marketing Live 2026 Google introduced two building blocks for exactly this, both of which nearly disappeared in the noise around Universal Commerce Protocol and Ask Advisor: conversational attributes in the product feed and AI Performance Insights in the Merchant Center. Together they shift the rules of product visibility away from ad optimisation and towards data quality.
1. From keyword to purchase advice
AI Mode has been available in German-speaking markets since October 2025. Alongside the classic results list there is now a dialogue-based surface where users ask follow-up questions, compare and get advice. For retail that is more than a new presentation format: the purchase advice that used to happen in the shop or on the product detail page moves one level up — into the answer itself.
That shifts the decisive question. It is no longer only "do I rank for this keyword?" but "does an AI system understand my product well enough to name it in a recommendation?" And the answer depends not on ad budget but on which information about a product is machine-readable at all.
This is precisely the structural problem in many feeds. Title, price, availability and GTIN are enough to appear in a shopping ad. They are not enough to answer "are these headphones suitable for long flights?" Close that gap and you become quotable. Leave it open and you get passed over — with no error message, no rejection in the account, simply through not being mentioned.
2. The starting position in numbers
- 6 new conversational attributes are available in the Merchant Center, rolled out globally.
- 30 question-and-answer pairs are possible per product, and likewise up to 30 linked products.
- 14 % of shopping queries show an AI overview, according to an analysis of 20.9 million keywords.
Fourteen per cent sounds modest at first — the direction is what counts. Against the previous year that is a rise of roughly 5.6 times. Shopping queries are among the areas where AI answers are gaining ground fastest. Preparing today is not playing a niche channel; it is playing the normal case of the coming years.
3. Conversational attributes: the six new fields
Conversational attributes are optional data fields. They supplement the existing feed, do not replace core attributes and do not affect a product's approval status. They are submitted through supplemental feeds or the Merchant API. The rollout is global, so retailers in Germany can already work with them.
question_and_answer
In our view the most effective field: structured question-and-answer pairs about the product, up to thirty of them. Instead of hoping an AI system extracts the answer from running text, you supply it directly. For Bluetooth headphones: "How long does the battery last?" — "Up to 40 hours on one charge." Or: "Can two devices be connected at once?" — "Yes, multipoint Bluetooth is supported."
document_link
Here you deposit URLs to PDF documents: user manuals, assembly instructions, technical data sheets, certificates, buying guides. Up to five documents per product are possible, each up to 50 MB. That makes information accessible which previously sat outside the feed — and which often decides the matter for products that need explaining.
related_product
This field describes how products in the catalogue relate to each other, referenced by id or gtin. Google distinguishes six relationship types: `part_of_set` (part of a set, such as a chair in a dining group), `required_part` (needed for operation, such as a lamp's battery), `often_bought_with`, `substitute` (an equivalent alternative), `different_brand` (the same product under another brand) and `accessory`. Up to 30 links per product are possible.
item_group_title and variant_option
Both fields work alongside the existing `item_group_id`. The group title names the product family ("Men's T-shirt organic cotton"), while the individual product title describes the specific variant ("Men's T-shirt organic cotton — black, M"). The variant option names what distinguishes the variants: colour, size, material or storage capacity. That lets an AI system resolve a query such as "is the shirt available in black, size M?" unambiguously.
popularity_rank
A value that places a product relative to your own catalogue. A rank of 95.5 means the product performs better than 95.5 per cent of the range. This gives Google a signal for questions such as "what is your best-selling running shoe?" — questions that are simply unanswerable without the field.
In practice: start with `question_and_answer` for your twenty highest-revenue products. The content usually exists already — in FAQ sections, support tickets and questions from sales. This is about structuring, not new production.
4. AI Performance Insights: share of voice in AI search
The second novelty answers the question that most often stays open with AI visibility: how do I know whether I appear at all? The new report sits in the Merchant Center under Analytics → Products → the AI Performance tab, and shows how a brand and its products perform in AI Mode and in AI overviews.
The metrics
At the centre sits share of voice: your visibility compared with defined competitors. Two traps are worth knowing — with too few impressions the value is reported as 0, and with no competitors stored in the account it reads 100 per cent. Neither figure is a success message; both point to incomplete configuration. Share of voice is supplemented by query frequency (how often a query is asked) and query type (what intent sits behind it).
Three phases of the customer journey
The report assigns conversational queries to three phases: discovery (general orientation), evaluation (comparison and detail questions) and purchase (shortly before the transaction). Each phase gets its own share of voice. Because users often ask several things in a single prompt — price and comparison, say — queries can belong to more than one phase. The pattern is what tells you something: strong in discovery but weak in purchase means losing customers exactly where revenue happens.
Concrete optimisation hints
Beyond measurement, the dashboard shows frequently used AI shopping terms — formulations about properties or benefits, such as "maximum cushioning", that buyers prioritise in their queries. Retailers also see how many of their own products match those terms. The report additionally lists sought-after product attributes such as size, colour or material, making visible where technical details are missing from your own data sources.
Availability, and what it means for German retailers
An important limitation: AI Performance Insights are initially in a pilot phase for selected Merchant Center accounts in the US. An extension to Australia, Canada, India and New Zealand is announced; there is no date for Germany. The report also considers organic AI traffic only — paid ads do not feed into it, and there is no cross-category overall report so far. The conversational attributes, by contrast, are available globally. The sensible order is therefore set: the data basis first, the reporting afterwards.
5. The groundwork before the flourish
Conversational attributes only work on a load-bearing foundation. An AI system that can derive nothing from a title like "Art. no. 4471-SW" will not be convinced by thirty Q&A pairs either. Before filling the new fields, look at the basics.
Titles that answer a question. The product title should carry brand, product type and the two or three purchase-deciding features — in customers' language, not internal catalogue language. Conversational queries contain benefit formulations; if those appear nowhere in the title or description, the bridge between question and product is missing.
Descriptions as an answer, not ad copy. Google made the point plainly in its Marketing Live announcement: strong product descriptions are decisive for being found in the AI era. In practice: what is the product suitable for, what not, who is it made for, what is included? Superlatives without substance are worthless to an AI system; verifiable properties are not.
Complete technical attributes. Size, colour, material, weight, dimensions, energy efficiency — exactly these structured details are what the new report evaluates to show gaps. Missing attributes are not a cosmetic flaw but an exclusion criterion as soon as a user filters or asks for that very feature.
6. Relevant without an online shop too
The Merchant Center addresses retailers with a product catalogue. The principle behind it applies to any company that wants to appear in AI answers: AI systems recommend what they understand unambiguously in structured form. Service businesses, trades and B2B providers can achieve the same effect by other means.
Concretely: service pages that answer real customer questions instead of listing services. Structured data such as FAQPage, Service and Organization. Maintained company data with consistent details across every directory. And an llms.txt that shows AI systems which content is relevant to them.
7. Five mistakes that cost visibility
Marketing phrases instead of real answers. "Premium quality for the highest demands" answers no question. Q&A pairs that only repeat ad language give AI systems no additional information — and still cost maintenance effort.
Trying to roll out everything at once. Pulling six new fields across ten thousand articles fails on resources and data quality alike. The more reliable path is the reverse: bestsellers first, check results, then scale.
Contradictions between feed and website. If the feed says 40 hours of battery life and the product page says 30, both sources lose value. Consistency across channels is not tidiness, it is a trust signal.
Assigning relationship types arbitrarily. An accessory is not a spare part, and an alternative is not part of a set. Wrongly set relationship types produce unsuitable recommendations — and therefore returns rather than revenue.
Filling the fields once and leaving them. Popularity values, availability and product relationships change. A feed maintained only at project start produces misinformation after a few months — worse than a gap.
8. Six steps for the coming weeks
- Review existing content. Go through FAQ sections, manuals, guides and support enquiries. Most of the material for conversational attributes already exists, just not in structured form.
- Start with the bestsellers. Equip the 20 to 50 highest-revenue products with Q&A pairs first. That is where the leverage is greatest and the effort manageable.
- Map product relationships. Assign accessories, spare parts, frequently co-bought items and alternatives to the right relationship type — not blanket "similar products".
- Label variants unambiguously. Add group title and variant option so queries for a particular colour, size or storage capacity can be resolved cleanly.
- Complete technical attributes. Systematically fill in missing size, colour, material and dimension details. These are exactly the gaps the AI Performance report will later expose.
- Store competitors in the account. Without defined competitors, share of voice is worthless as a metric. This groundwork pays off before the report becomes available in Germany.
9. Product data is the new ranking signal
The 2026 Merchant Center changes look technical at first glance. Their consequence is strategic: visibility in retail shifts from ad optimisation to data quality. Know your customers' questions and answer them machine-readably, and you get recommended. Rely on ad budget and treat the feed as a duty, and you lose ground — slowly but steadily.
That is good news for smaller providers, because data quality cannot be bought, only earned. A specialist retailer with 300 well-described products has realistic chances against a large provider with 300,000 carelessly maintained ones.
Our assessment: the conversational attributes were the least spectacular and simultaneously most consequential announcement of Google Marketing Live 2026. They cost no budget, only care — and that is precisely why a large part of the market will ignore them at first. For everyone else, that is a window.
10. FAQ: the key questions
What are conversational attributes in the Google Merchant Center?
Conversational attributes are six optional product data fields Google introduced in 2026: `question_and_answer`, `document_link`, `related_product`, `item_group_title`, `variant_option` and `popularity_rank`. They supplement the existing product feed rather than replacing it, and do not affect a product's approval status. They are submitted through supplemental feeds or the Merchant API. Google's AI systems use this additional information to match products to conversational queries in AI Mode, Gemini and other AI surfaces.
What do the AI Performance Insights show?
They show how a brand and its products perform in Google's AI search. At the centre sits share of voice against defined competitors, supplemented by query frequency, query type and an assignment of queries to the discovery, evaluation and purchase phases. The report also lists frequently used AI shopping terms and missing product attributes. It considers organic AI traffic only; paid ads do not feed into it.
Are the new Merchant Center features available in Germany?
It differs. Google is rolling out conversational attributes globally, so German retailers can use them. AI Performance Insights are initially in a pilot phase for selected Merchant Center accounts in the US, with an announced extension to Australia, Canada, India and New Zealand. No date has been named for Germany — but the product data can be prepared now.
Is this worth it without an online shop?
The Merchant Center itself addresses retailers with a product catalogue. The principle applies to every company: AI systems answer questions from structured, unambiguous information. Service providers achieve the same effect through clean service pages with real customer questions and answers, structured data such as FAQPage and Service, maintained company data and an llms.txt. In both cases what matters is that the answer to a concrete user question exists in machine-readable form.
