Insights

Google Opened a Window Into AI Shopping. Here Is What It Shows, and What It Does Not.

A senior point of view on Google's new Merchant Center AI performance insights, its conversational attributes, and the personal shopping agents arriving on top of both. Written for VPs and Directors of ecommerce, digital shelf, and ecommerce content at enterprise consumer brands.

Genrise Editorial18 min read

Recently, Google announced two additions to Merchant Center that matter more to consumer brands than anything else it has shipped for shopping this year. The first is AI performance insights, a native report that shows how a brand's products surface across AI Mode, AI Overviews, and the Gemini app, benchmarked against competitors. The second is conversational attributes, a set of new structured fields a brand can submit so that Google's AI systems understand its products the way a shopper asks about them. The report entered a limited U.S. pilot in July. The attributes are rolling out globally.

Since then, the context around both has changed. In September, Meta launched Muse, a personal AI agent that browses, fills in forms, and checks out on a shopper's behalf. Shopify opened every one of its merchants to Muse within two weeks. Amazon blocked it. An invite-only agent called Instinct, which shoppers simply text or call, passed 100,000 users.

Read together, three things shift materially:

  1. AI shopping visibility is no longer something a brand has to buy or build to see. Google now gives it away, inside the tool most retail feeds already run through.
  2. Product content has a new destination. Alongside the retailer product page and brand.com, a brand's structured data in Merchant Center is now a direct input to what Gemini says about its products.
  3. The shopper asking the question is increasingly an agent, not a person, and agents read structured data rather than skimming a page.

None of this lowers the bar. It moves where the bar sits. This piece walks through what Google actually shipped, what it deliberately left out, and what it means for the next twelve months of digital shelf strategy.

What Google actually shipped

It is easy to read the coverage and come away believing Google now publishes the prompts shoppers type into Gemini. It does not. What Google shipped is two separate instruments, and they do different jobs.

A measurement report. AI performance insights lives in Merchant Center under Analytics, then Products, then a new AI performance tab. Its headline metric is share of voice: a brand's AI impressions divided by the total AI impressions across the brand and its defined competitors for related queries. Next to it sits the competitors' average share, so a brand can see whether it is above or below its own market. Underneath, the report breaks visibility down by stage of the shopping journey (discovery, evaluation, purchase), by the conversational product terms shoppers use most, by the type of query being asked, and by the product attributes shoppers search on, with a completeness score showing which of the brand's products are missing them.

An input channel. Conversational attributes are new optional fields in the Merchant Center product data specification. Google describes them as a way to help AI systems and conversational agents understand the specific nuances of a product. They sit on top of a brand's existing feed, are best submitted through a supplemental data source, and do not affect the approval status of existing products.

One measures presence. The other shapes it. A brand that treats them as the same thing will misread both.

Google also shipped the plumbing that sits around them: the Universal Commerce Protocol for agent checkout, native purchase inside AI Mode and Gemini, and a new interface that lets AI agents read Merchant Center data directly. The checkout side of that story, and what Universal Cart changes for brands, is covered in the Universal Cart and agentic commerce piece. This piece stays on the data.

What the report shows, and what it withholds

The report is genuinely new. Until this year, the only way a brand could see how it appeared in Google's AI answers was to run prompts through a third-party tool or audit them by hand. Now the platform generating the answer reports on it directly, from the same Shopping Graph it uses to ground those answers. That is a meaningful change in where the evidence comes from.

It is also narrow, by design. Four limits matter for how a brand reads it.

It reports aggregates, not prompts. Google classifies individual conversational queries into terms, query types, journey stages, and attributes. It does not show any shopper's actual question. Google's own documentation notes that a single complex query can map to more than one category, which is a reminder that the categories are Google's interpretation, not the shopper's words. When a stakeholder says "Google is now showing us the prompts," the precise answer is that Google is showing the patterns it sees in them.

It measures impressions, not outcomes. The report has no clicks, conversions, or revenue. It answers "was the product shown?" It does not answer "did being shown sell anything?" A rising share of voice is a leading indicator at best.

Its benchmark is only as good as the competitor set. Share of voice is calculated against the competitors configured in Merchant Center. Define the set carelessly and the number is misleading. Define no set at all and a brand can look dominant against nobody.

It covers one platform. The report measures Google's own AI surfaces. It says nothing about Amazon Rufus, Walmart Sparky, ChatGPT, Perplexity, or the Gemini-powered assistants that retailers such as Kroger run on their own sites. For most consumer brands, those surfaces carry a large share of the AI-assisted shopping in their category.

None of this makes the report less useful. It makes it one input. The answer engine optimization piece sets out the measurement hierarchy the cluster uses: share of conversation at the topic layer, share of answer at the individual-prompt layer, and share of voice as the roll-up. Google's report delivers a version of the top layer for Google's surfaces. The layers underneath, which prompt, which answer, which claim caused the citation, and which competitor claim won instead, are still the part that tells a team what to change. A share-of-voice number that cannot be traced to a cause is a thermometer, not a diagnosis.

Microsoft reached the same place first. Bing Webmaster Tools added an AI performance report in February 2026, and it, too, stops short of showing individual prompts or answers. The pattern is becoming the default shape of platform-native AI data: a visibility score, a classification of demand, and no view of the conversation itself.

The six conversational attributes, up close

When Google first previewed conversational attributes at NRF in January, it described "dozens" of new fields. What shipped in the specification is six. Each has its own detailed page in the Merchant Center help center.

6
Conversational attributes shipped
Versus "dozens" previewed at NRF
Roughly half
Relevant AI Mode recommendations
Incorporated brand-submitted attributes in Google's lululemon test
AttributeWhat it carriesWhy it matters for AI answers
Question and answer [question_and_answer]Product questions and their answers, submitted as pairsLets the assistant answer the shopper's actual question from brand-supplied content rather than inference
Related product [related_product]Relationships to other products: accessory, required part, substitute, often bought withSupports the multi-step, "what else do I need" journeys conversational shopping produces
Document link [document_link]Links to product PDFs such as spec sheets, manuals, or sizing guidesGives the assistant a crawlable, authoritative source for detail too long for a feed field
Item group title [item_group_title]A single title for a product family with multiple variantsHelps the assistant treat variants as one product rather than competing near-duplicates
Variant option [variant_option]The properties that distinguish variants, such as size, flavor, or countLets the assistant match a shopper's stated need to the right variant
Popularity rank [popularity_rank]A product's popularity as a percentile of the brand's own inventorySignals which products are the brand's proven performers

Two details in Google's guidance deserve more attention than they have received.

First, Google says a brand that already submits specific detail in its description, product highlight, or product detail fields does not need to repeat it in the conversational attributes. That means the conversational layer is not a separate content universe. It is an extension of the same structured record, and the quality of the existing fields still carries most of the weight.

Second, the one published proof point is modest and specific. In a Google-published test with lululemon, brand-submitted conversational attributes were incorporated in roughly half of relevant AI Mode recommendations. That is evidence the content gets used. It is not yet evidence of what using it does to sales, and it comes from a single brand in a single category.

The pattern should look familiar. When Amazon capped titles at 75 characters in July and introduced a searchable Item Highlights field, it separated a product's identity from its indexable claims. The Amazon listing piece walks through that split. Google's conversational attributes do the same thing one layer up: the primary feed carries identity, price, and availability, and a supplemental structured layer carries the questions, relationships, and distinctions an assistant needs. Two platforms, the same underlying shift. Claims and answers are moving out of paragraphs and into their own addressable fields.

The content those fields want is also familiar. Q&A coverage, persona-aligned use cases, specific and citable claims, and consistency across surfaces are the dimensions of the convergent rubric set out in the AI shopping assistants field guide. Google has not invented a new content discipline. It has opened a new place to deliver one.

The agent layer arriving on top

Google's data would be significant on its own. What makes it urgent is what arrived in September.

Meta's Muse. Meta introduced Muse on September 8 as a personal agent that can open a browser, fill out forms, and complete a purchase once the user approves it. On September 21, Shopify announced a partnership that makes Shopify merchants discoverable and purchasable inside Muse by default, with product data flowing through Shopify's catalog and checkout running through Shop Pay over the Universal Commerce Protocol.

Amazon's wall. The same weekend, Amazon blocked Muse from shopping Amazon.com. Amazon said Meta had not disclosed that Muse would access its store, that the agent does not identify itself when it browses, and that it appears to capture and store customer credentials. Meta disputes the credential claim. The move follows the court order Amazon won against Perplexity's Comet browser in November 2025, and it is consistent with Amazon's long-standing position that shopping on Amazon happens through Amazon's own assistant.

Instinct. Instinct, built by Spear Street Technology, has no app. Users text or call it, and it connects to their email, messages, and location to handle tasks on their behalf, shopping among them. It is invite-only and passed 100,000 users within weeks of its August launch.

100,000
Instinct users
Passed within weeks of its August launch

Both postures have coherent commercial logic. Google and Shopify are betting that open catalogs and shared checkout protocols let merchants be found wherever the shopper's agent happens to live. Amazon is betting that the customer relationship, the account, and the transaction are worth protecting inside its own environment, where Alexa for Shopping and Rufus already serve hundreds of millions of shoppers. Neither is more correct. For a brand, the implication is the same either way: its products now need to be legible to agents inside open ecosystems and inside closed ones, which means the same governed product truth has to reach both.

What this means for each shopper persona

The cluster's three-persona model holds, and each persona is affected differently.

01

The human shopper

about 85% of traffic
Needs
A clear, complete product page with the specifications shoppers scan for.
Disqualifier
Missing product detail that makes comparison harder.
02

The AI-assisted human

roughly 10 to 15%
Needs
Conversational terms, intents, and direct answers that match how the shopper phrases a need.
Disqualifier
Content that cannot answer the recommendation question.
03

The autonomous agent

under 1%, rising
Needs
Identifiers, attributes, relationships, availability, and answers.
Disqualifier
Missing or contradictory structured data.

The human shopper (about 85% of traffic). Still the majority, and still best served by a clear, complete product page. Google's attribute completeness data is useful here too: the specifications shoppers search on in AI conversations are, in most categories, the same ones they scan for on a page.

The AI-assisted human (roughly 10 to 15%). This is the persona Google's report is built to measure: the shopper asking AI Mode or Gemini for a recommendation and choosing from what comes back. The frequently used terms and top intents in the report are the clearest signal yet of how this shopper phrases a need. Conversational attributes are the most direct way to answer it.

The autonomous agent (under 1%, rising). Muse and Instinct are early and their purchase volume is not yet measurable. But they change the reader. An agent acting on a standing instruction ("keep us stocked on the dog food she likes, under $40") never sees the hero image or the brand story. It reads identifiers, attributes, relationships, availability, and answers, and it moves on to the next eligible product if any of them are missing or contradictory. The generative AI in ecommerce piece describes structured-attribute completeness as an eligibility requirement for this persona. September made that concrete.

The four shifts to plan against in the next 12 months

Shift 01

1. Visibility becomes a commodity. Prioritization becomes the job.

Share of voice is now free from Google and from Bing, and more platforms will follow. As measurement proliferates, the scarce thing stops being the number and becomes the decision: which terms, which claims, which attributes, on which SKUs, this quarter. More dashboards produce more signal and more noise at the same time. The brands that benefit are the ones that can turn many inputs into a short, ranked list of changes, and then check whether those changes moved anything.

Shift 02

2. Product content now has three destinations.

For most of the last decade, "product content" meant the retailer product page. It now means at least three places, each read by different systems.

DestinationWhat it carriesWho reads it
Retailer product pagesTitles, Item Highlights, bullets, descriptions, enhanced contentRetailer search, Rufus, Sparky, retailer-owned assistants
Merchant Center feedCore product data plus conversational attributesAI Mode, AI Overviews, Gemini, Google's Business Agent
Brand.comPage content and matching structured product dataOpen-web assistants and agents that visit the site

Content written once and syndicated once a year cannot keep three destinations current. The work has to be continuous, and it has to draw on a single source of brand truth so the three do not drift apart. Contradictions between destinations are exactly what an agent treats as a reason to skip a product.

Shift 03

3. Brand.com becomes machine-readable, visibly.

The cluster has argued since the AI shopping assistants field guide that the brand site matters again, because open-web assistants read it. Agents make that more pressing. The temptation will be to publish a layer of content meant only for machines, hidden from people. That is the wrong move. Google's structured data guidelines are explicit that markup must not describe content that is not visible to readers of the page, and that violations can cause it to be treated as spam. The durable approach is structured product data, in schema.org markup, that mirrors a visible page carrying the same answers, specifications, and claims. Machine-readable and human-visible are not in tension. They are the same content, expressed twice.

Shift 04

4. Open and closed ecosystems both need the same governed truth.

With Google and Shopify opening catalogs to any agent and Amazon closing its store to agents it has not authorized, brands will be served by some assistants they can feed directly and some they can only influence through retailer content. Bespoke work per surface does not scale across that split. What scales is one governed source of claims and product facts, approved once, rendered in each destination's format. As agents begin repeating brand claims directly to shoppers, that governance stops being a legal nicety and becomes the thing that keeps the brand's story consistent wherever it is told.

The CPG wrinkle: who owns the feed

Most of the early commentary on these tools is written for retailers and direct-to-consumer merchants. Enterprise consumer brands face a complication that commentary tends to skip.

Merchant Center reports on, and accepts data for, the products a merchant submits. A consumer brand that sells mainly through retailers often does not own most of its Google listings. Those listings frequently come from the retailers' feeds. A brand's direct levers are its own brand.com feed, where it has one, and manufacturer-level product data, which Google handles through Manufacturer Center. How fully the new AI reporting and conversational attributes extend to that manufacturer layer is worth confirming account by account before building a baseline on it.

Two practical consequences follow. First, a brand's view of its own AI share of voice on Google may be partial, and the gaps may sit precisely where retailers carry most of the volume. Second, the content a brand supplies to its retail partners and syndication channels becomes an indirect input to Google's AI surfaces, which is one more reason for the three destinations to draw on the same source.

Where this fits in your 2027 content strategy

Google's new data does not change the cluster's argument. It sharpens it.

The measurement hierarchy in the answer engine optimization piece anticipated exactly this: platform visibility scores would arrive, and they would be precise about presence and silent about cause. The field-by-field logic of the Amazon listing piece now applies to Google's feed as well. And the grading approach in the PDP audit framework, which asks whether a page answers real shopper questions with specific, citable claims, is the same test Google's attribute completeness score now applies from the outside.

That is also how Genrise is built. The platform is deliberately agnostic about where visibility and demand data come from. Google's share of voice, conversational terms, intents, and attribute gaps become inputs alongside prompt-level testing across Rufus, Sparky, ChatGPT, and retailer-owned assistants, and alongside retailer search data, each with its source preserved. Genrise does not need to own the measurement to use it. Its value sits downstream: scoring every SKU on the AI Shelf Readiness Index across Content Foundation, SEO Performance, AI Shelf Visibility, and Brand's Right to Win, ranking the highest-leverage changes by expected incremental revenue, and routing them into content that humans approve. That content is increasingly delivered not only as retailer copy but as the structured context Google's feed and a brand's own site now ask for, all drawn from one approved claims library.

The last step is the one platform reports leave open. Google's report tells a brand whether it was seen. Genrise validates changes against organic outcome. A/B-tested campaigns across consumer healthcare brands consistently show 0.7% to 6% conversion uplift per SKU, and across a catalog, continuously improving content quality compounds into 2% to 5% incremental annual revenue growth.

0.7% to 6%
Conversion uplift per SKU
Across A/B-tested consumer healthcare campaigns
2% to 5%
Incremental annual revenue growth
From continuously improving content quality across a catalog
Visibility is the leading indicator. Revenue is the point.

References

Google primary sources

Agent and platform reporting

  • PYMNTS, Shopify Brings Shop Pay Checkout Solution to Meta's Muse AI Agent (September 21, 2026).
  • American Banker, Shopify adds Meta Muse to agentic AI strategy (September 2026).
  • GeekWire, Amazon blocks Meta's Muse AI assistant in new standoff over agentic shopping (September 20, 2026).
  • Bloomberg, Amazon Blocks Meta's Muse AI Agent From Its Retail Site (September 21, 2026).
  • PYMNTS, Instinct AI Assistant Targets $10 Billion Valuation (September 2026; user base above 100,000).
  • Search Engine Land, Google launches AI Performance Insights and Conversational Attributes in Merchant Center (May 2026).

Related Genrise insights

Frequently asked questions

AI shopping readiness

See how your catalog reads to AI Mode, Rufus, Sparky,
and the agents arriving behind them.

See which changes would move it most with a tailored walkthrough of Genrise.