E-commerce
August 19, 2026
This page aligns nineteen cards under a single sentence: synthesizing reviews, specifications, and fit advice for a restricted list. This is not a ranking. These are not nineteen search engines. Rufus and Sparky carry "review synthesis" in a walled garden (Amazon, Walmart). Klarna is a buy-now-pay-later payment option. VendeeX (governance) and Karma (coupons) are in the filter, but their cards are not labeled as product search. Macaron calls itself "the world's first personal agent": landing page phrasing. No gross merchandise volume (GMV), no review accuracy rate, no fit adjustment rate. This article doesn't have them either. It categorizes what a machine can synthesize, not logos.
Google Merchant Center documents conversational attributes as optional fields, complementing the main product description, for AI agents and surfaces (AI Mode, and others) as much as traditional search. Questions and answers, PDF links, related products (accessory, required part, frequently bought together), group title, variant options (shoe width, size), self-declared popularity rank. Recommended supplemental source, main feed, or Merchant API. This does not change approval. Google Search Central documents Product markup: ratings, reviews, pros and cons, availability, shipping. Two families (snippets and merchant listings). A JPEG header image is not a FAQ. An image-based size guide is not [variant_option]. Lumping all nineteen into an "AI search" indicator means budgeting for a Rufus you don't have, or a FAQ you already have in a PDF.
If the idea of a software buyer is new: what agentic commerce is. The 2026 calendar: what has been delivered. The four filters: buying agents by task. Official totals remain human shopping journeys: 2026 e-commerce statistics.
Summary
Three professions in one sentence, nineteen logos
The introductory sentence already has three jobs. The cards do not cover them one by one.
Job | What Google documents | Without the field |
|---|---|---|
Specifications | Q&A, product_detail, document_link (PDF), related_product | The agent invents the headphone jack or stays silent |
Fit and variants | [variant_option] plus [item_group_title] (e.g. width, size) | A JPEG size guide is invisible |
Reviews | Product markup: ratings, reviews, pros and cons (especially snippets) | Rufus synthesizes the Amazon garden. Your Magento, no |
Pasting a shortlist and searching makes sense. A buyer who is "gathering information" mixes specifications, reviews, and size. Reading nineteen agents like nineteen FAQ engines, no. Daydream is a fashion graph. Amazon Buy for Me checks out after an Amazon payment page. ChatGPT and Gemini wear the label because a LLM can summarize. Summarizing is not having read your feed.
"more than 40,000 sites" (Phia), "more than 10,000 shops" (Daydream), "world's first personal agent" (Macaron): card phrases. Amazon, Walmart, Revolve, Zara: directory examples, not sourced partnerships here.
A merchant who "chooses a search agent" like a PIM has read the title, not the fields. You don't install this page. You maintain attributes, structured reviews, variants. The agents then read, or hallucinate.
Neighboring categories steal the word search: price comparison (an offer), bargain hunting (a code), checkout funnel (a payment). An FAQ is not [price]. A review is not a mandate.
Convert over 2,000 customers on average per month with Qstomy.
The world’s 1st Shopify AI dedicated to customer conversion
Empowering 200+ e-commerce merchants
The filter is not the label
Read the card labels, not just the filter. VendeeX is on this page. Its card says spend governance, returns, audit trail. Not product search. Karma says price, deals, coupons. Not search. The filter grouped them. This is not proof that they synthesize your reviews.
Klarna carries the search label. It is a payment schedule. An installment payment does not write an FAQ about materials. Macaron only appears in the task filters of this site here (search, gift search). Treat it like a directory listing, not like a PIM.
The "19" counter on the task page is a filter. It will change. It does not say how many of your customers are asking Gemini for an adjustment, nor if the answer comes from your [question_and_answer]. A Verified listing does not buy a badge.
Claude and Operator often read the product page, not the additional feed. A Merchant Center attribute only helps a headless browser if it is also on the page. Vocabulary (attribute, variant, review): agentic commerce glossary.
Conversational attributes: optional for approval, not for attribution
Conversational attributes, says Google, help AI systems understand the nuances of a product. They are entirely optional. They complement the main description. Submitting them has no impact on the approval of products that are already there. In other words: your listing can be approved and silent in AI Mode.
question_and_answer: FAQ pairs. Google's example: is there a jack port, is Bluetooth present. This is spoken characteristic, not keyword stuffing.
document_link: PDF (manual, assembly), multiple URLs separated by commas. If the FAQ is already in the PDF, do not duplicate it elsewhere: Google says to extract.
related_product: three sub-fields: relation type, identifier type, identifier. Accessory in GTIN, required part in internal ID, often_bought_with. Without this, "what does this go with" remains a hallucination.
item_group_title and variant_option: family plus variants. Example: organic t-shirt; or XL screen, 512 GB memory, color. Adjustment: narrow shoe width, size 8. A size chart in JPEG is not this.
popularity_rank: self-reported percentage within your inventory (example 95.5). This is not a Google ranking. Do not recycle it as a directory indicator.
Google recommends an additional source rather than inflating the primary feed. Merchant API possible. Repeated note: if it is already in description, product_highlight, or product_detail, do not copy it into conversational attributes. Two copies, two possible truths. The fundamental work (identifier, material, variant): what a machine needs to compare.
The long help center example is a Pixel 9 Pro (identifiers, GTIN, jack and Bluetooth FAQ, highlights, PDF). This is a documentation template. It is not a measured customer case. Do not present it in committee as "Google boosted Pixel thanks to attributes".
These fields feed AI Mode "and more", says the help, and also traditional search. This is not a Gemini payment tunnel integration. This is data. A checkout journey is chosen elsewhere: what platforms need to change.
Reviews are markup, not a conversation
The introductory sentence puts reviews on the same level as features. Conversational Merchant Center, in the cited page, does not specify a "customer review summary" field. Google documents ratings and reviews on the Search side: Product markup.
Two classes. Snippets: pages where one does not necessarily buy, more options for review info (pros and cons of an editorial review). Merchant listings: pages where people buy from you (sizes, shipping, returns). Providing merchant listing properties often makes you eligible for snippets. Both have their enhancements. Give as much rich info as possible, Google says: experiences change.
What a buyer sometimes sees: ratings, reviews, price, availability in a text result; Knowledge panel. What a directory agent summarizes: we don't measure that here. Rufus labels review summary in Amazon. Sparky in Walmart. Your Magento does not inherit their corpora. If your reviews only exist in a JavaScript block, the headless browser and the crawler see a blank page: a Shopify storefront readable by a LLM.
Do not confuse an AggregateRating score with a paragraph that ChatGPT summarized from three blogs. One is a field. The other is a conversation. The customer service for a misread size is yours. Breakages: where agentic commerce still breaks down.
The pros and cons of an editorial review are not your five-star product ratings. Google separates them. An agent that mixes the two invents a feature. Don't give it the excuse: keep product reviews and editorial content within their respective markups.
Enclosed garden against all trades
Four walled gardens under the same search filter.
Rufus and Amazon: questions, reviews, comparisons in the shop. You must manage the Amazon listing, not the [question_and_answer] Merchant Center.
Sparky and Walmart: same, Walmart garden, extension into other chats according to the directory. Your Magento FAQ does not get in there on its own.
Daydream and Shop: discovery (affiliate mode, Shopify channel). A style passport is not a Bluetooth FAQ.
All merchants (Claude, Gemini, Genspark, Manus…): promise to read the web. Without fields, they read the header visual, or a PDF if linked, or nothing.
Amazon Buy for Me carries the search label and remains an agentic payment tunnel. Search is the Amazon path before checkout. It is not your corpus of reviews.
A pitch like "we'll get you into search agents" can sell an extra feed, review JSON-LD, a more readable theme, or a directory listing. Four deliverables. One word "search".
Operator labels Instacart and DoorDash. Useful as an example of tasks. Insufficient as proof that your fashion reference has a synthesized fit.
What the agent was able to cite
The only merchandising question: what was the agent able to cite? The snippet doesn't say.
Buyer question | Useful field | What doesn't work |
|---|---|---|
What does this go with? | related_product (accessory, required_part, GTIN or id) | A "frequently bought together" block in JavaScript only |
What width, what size? | variant_option plus item_group | A scanned table, a copied sentence in the description |
Does it have Bluetooth? | question_and_answer or product_detail, not both copied | An icon without text |
Are people happy with it? | Review and rating markup | A block that renders after the fact |
Don't put an outdated price in an FAQ. The price has its own attribute. Search has its own. An agent reading "€199 until Sunday" in an FAQ synthesizes a dead promotion. Google, on the overview page, already says not to copy description, highlight, or detail.
If you judge "AI search" at the last click, Gemini in the chat and a Rufus click do not have the same pixel. Decide the attribution before shouting that "no one reads our guides."
A manual PDF via document_link can be better than a poorly written FAQ. Google says it extracts. Test what actually comes out in AI Mode, not what the sales pitch promises.
What this page does not measure
This page measures nothing of what matters for a shortlist.
Synthesis accuracy: did the agent invent material? No public rate.
FAQ coverage: how many references have [question_and_answer]. Help does not ask for it for approval. Neither does the directory.
Adjustment: size returns. Chargeback reason code: missing. Your ordering system, perhaps.
Market: nineteen listings is not the market. It is a filter at a date.
Do not budget this list as a channel with a known cost per acquisition. We do not publish one.
Product search, on cards that have it, is a usage label. It is not an FAQ certificate. It is not a review synthesis. We do not recycle the word as an indicator.
popularity_rank at 95.5 in the Google example is not your Search score. It is a field you fill in. An agent citing it as social proof is citing your marketing.
Three gestures, not nineteen steps
This week, field by field, not by logo.
Spoken features: ten real customer service questions on five core references, in [question_and_answer] or in a linked PDF. No prices in them. No duplicate descriptions.
Fit and variants: [variant_option] aligned within the group. Width, size, material. Get rid of the JPEG table.
Reviews: ratings and reviews markup visible without JavaScript. Marketplace garden (Amazon, Walmart) kept separate.
The ultimate test: would this work still be useful if Macaron disappeared tomorrow? A real FAQ and typed variants, yes. An exclusive connector "for the search filter", no.
Name the owners: PIM or merchandising for attributes and fit, technical SEO for review markup, customer service for FAQs, marketplace for Rufus. A single "AI search owner" stops at the first JPEG.
If you only have one hour: take five references, ask if a human can answer headphone jack, size, and compatibility without opening an image. If not, none of the nineteen will synthesize better than your internal search.
You do not turn on the search agents
You do not turn on "search agents" like a conversational agent. Buyers open Rufus, Gemini, Claude, Daydream. You hold fields. The cards on this page are not buttons.
If a seller promises to place you "in product search agents," ask for: additional feed, JSON-LD reviews, readable theme, or directory listing. Four deliverables. One word.
Conversational attributes are done now: they are optional for approval, not optional for being citable. A checkout funnel rail is chosen late. Postponing the UCP is not postponing an empty FAQ.
After the shortlist, the click, or the mandate, it is your after-sales service if the adjustment was in a JPEG. The directory did not collect the return. Do not attribute a commercial gesture to it.
The product directory, if you want the grid by logo: AI buying agents.
A field index, not a prize list
Search is fragmenting. Google did not "kill" Rufus. A Merchant Center FAQ has not "replaced" your Amazon reviews. This page is an index of poorly labeled synthesizers under a single word. The only card you hold: spoken features, variants, structured reviews.
Treat it as a starting point: what garden, what fields, what exists only in an image. Three questions. Not a search ranking.
You can ignore Macaron and still fill out [variant_option]. You can decline Claude and still maintain review markup. You can be off Amazon. The directory does not decide. Your fields, however, do.
Next, open two profiles, not nineteen. A garden where you already sell (Rufus or Sparky). An agent for all merchants (Gemini or Claude). Ask each time: what corpus, what field, what image it will not read.
Frequently Asked Questions
What is the best product search agent in 2026?
This page does not say. No public data on synthesis accuracy, fine-tuning, or reviews says so either. Nineteen maps, several gardens. A ranking would be an invention.
Rufus and Gemini, is it the same search?
No. Rufus synthesizes within Amazon (review synthesis label). Gemini reads what Google has of your feed and pages, including conversational attributes if you submit them. Two corpora. One word: "search".
Why are VendeeX and Karma on this list?
Because the page filter included them. Their cards are not labeled product search. Spend governance, coupons. Do not budget them as FAQs.
Are conversational attributes required to be approved?
No. Google says they are optional, with no impact on approval. Yes if you want an agent to cite a nuance (jack, Bluetooth, accessory, width) instead of inventing it.
Is a size guide in PDF format enough?
document_link exists for that. [variant_option] exists for the variant. A JPEG in the visual catalog, no. Test what AI Mode renders. We do not have rates.
What to do this week?
Five header references: FAQ or PDF, typed variants, reviews without JavaScript, related_product on a real accessory. No duplicate descriptions. No prices in the FAQ. Stop reading the nineteen logos as long as the header visual speaks for itself.
Is Macaron the world's first personal agent?
It's a listing phrase. Not a fact. We are not making a ranking or a protocol out of it.
Sources
Google Merchant Center: Conversational attributes: FAQ, PDF, related_product, variants (optional, AI Mode).
Google Search Central: Product Markup: ratings, reviews, pros and cons, snippets, and merchant listings.
Enzo
August 19, 2026
