Platform

Solutions

Ressources

Pricing

Menu

Menu

Menu

Menu

Nov 4, 2026

How to optimize your product feed, titles, descriptions and structured data for AI shopping assistants

A shopper asks an assistant for "a lightweight rain jacket for cycling to work, under $150, packable". Your jacket fits every word. The assistant recommends two competitors, because your feed says "Jacket - Blue - M" and nothing about cycling, weight or packing. The feed you built for Google Shopping got the product indexed. It didn't give the assistant anything to match.

Disclosure: GetMint publishes this guide and makes one of the tools mentioned. GetMint does not edit feeds: this guide covers feed work you do in your own stack and names where our tracking fits. Platform specs were checked on vendor pages in September 2026.

This guide is the spec page for that problem. It covers how AI shopping systems read a feed, which fields decide eligibility and which decide recommendation, what each platform requires as of September 2026, an audit with pass and fail criteria, title and description rewrites, schema.org markup for products and variants, a freshness workflow for large catalogs, the tools for each layer of the work, and a scoring model for deciding which fix comes first.

TL;DR

  • Assistants match meaning. A title that opens with the product entity and its two deciding attributes beats a keyword-stuffed one on every engine.

  • A handful of fields decide whether a product is a candidate at all: title, description, brand, price, availability, link, image and a valid GTIN (or brand plus MPN). OpenAI's feed requires nine of them and Merchant Center seven, so get them right for every SKU before anything else.

  • Feed claims have to match the product page. When they disagree, the assistant picks one and quotes it, and nothing guarantees it picks yours. Google already requires the feed price to match the landing page.

  • Variants fail in two directions. Split badly and no single SKU carries the reviews or the specs. Merged badly and the assistant recommends the wrong size. One group ID per parent, one GTIN per child.

  • Schema.org Product, Offer and ProductGroup on the page should carry the same values as the feed, down to the unit.

  • One canonical catalog, enriched once, feeds every assistant. Separate feeds per platform multiply the errors.

  • A clean feed is necessary but not sufficient. Measure the outcome with a tool that runs a fixed set of shopping prompts many times per engine and tracks which SKUs appear, in what position, from which retailer. GetMint's AI Shopping module does this at SKU level on ChatGPT Shopping, Google AI Mode, Perplexity, Copilot and Alexa for Shopping.

BUILT FOR ENTERPRISE BRANDS

BUILT FOR ENTERPRISE BRANDS

Turn AI Search into growth channel
Turn AI Search into growth channel

Track how ChatGPT, Alexa for shopping, Gemini, Perplexity, Claude, and AI Overviews mention and recommend your brand, strategize your next moves with a GEO agent and our experts, then publish content AI cites.

Track how ChatGPT, Alexa for shopping, Gemini, Perplexity, Claude, and AI Overviews mention and recommend your brand, strategize your next moves with a GEO agent and our experts, then publish content AI cites.

What is AI shopping feed optimization and how does it differ from classic feed optimization?

Classic feed optimization answers to a channel spec. Google Merchant Center wants certain attributes in certain formats, rewards titles with the keywords shoppers type, and disapproves items that break the rules. The work is compliance plus keyword density, and it stops at "approved and served".

AI shopping feed optimization answers to a reader. ChatGPT Shopping, Perplexity, Google AI Mode, the Gemini app and Copilot take the shopper's sentence, which carries entities and attributes ("rain jacket", "cycling", "packable", "under $150"), and look for products whose data carries the same ones. A feed approved by Merchant Center that matches no attribute in the sentence gets passed over, with no error message attached.

What about Claude? Claude has no shopping surface yet: no product cards, no merchant program, no checkout. It can name products in an answer, but there is no shopping result to track, so there is no feed to optimize for it.

Three practical differences follow. Assistants lean on the description and the attributes, where classic Shopping results lean heavily on the title. Assistants also read the product page and third-party pages (reviews, roundups, retailer listings) alongside the feed, so what those pages say shapes the recommendation too; Google already compares your feed with your page, but only to check price and availability. And assistants answer natural-language qualifiers (use case, constraint, comparison) where classic engines answer a query string. The feed has to carry the qualifiers.

How do AI shopping systems read and interpret product feeds?

No assistant publishes its ranking logic, but the working model that fits how answers behave is this: the assistant maps the title, description and attribute values to concepts, and scores the product against the concepts in the prompt. Google's January 2026 Merchant Center update, which adds attributes for common product questions, compatible accessories and substitutes, points the same way.

In that model, "waterproof" and "rain jacket" resolve to the same requirement. "Packable" resolves to a use case. "Cycling" resolves to a scenario with its own attribute expectations (reflective details, a dropped hem, a helmet-compatible hood). A product whose data names those attributes is a candidate. A product whose data says "premium outerwear" is not.

Taxonomy decides candidacy before content does. The assistants lean on the category systems they know: Google's product taxonomy (optional in Merchant Center, which assigns one if you don't) and the product_category path in OpenAI's product feed. Schema.org's Product "category" property on the page should carry the same path. A jacket filed under "Apparel & Accessories" with nothing below it competes with every garment in the store. Filed under "Apparel & Accessories > Clothing > Outerwear > Coats & Jackets" (Google ID 5598), with a product_type of "Rain Jackets > Cycling", it competes with jackets and tells the assistant which kind. Map every product to the deepest accurate node and keep that mapping identical in the feed and on the page.

Having a feed is not the same as being readable by an agent. Readability means consistent field names on every row of the catalog, one format per attribute (grams, not "light"), values from a controlled list where the platform publishes one (color, size, material, condition), and an endpoint or file the assistant's ingestion pipeline can fetch on schedule. A feed with "Blue", "blue", "BLU" and "navy-ish" in the color column is four products to a human and noise to a parser.

The field-priority matrix: which feed attributes matter most for AI visibility?

Tier

Attributes

What they decide

Failure mode

Minimum viable

id, title, description, brand, price, availability, link, image_link, GTIN (or brand + MPN, or identifier_exists = no)

Whether the product is indexed and matched to an entity at all

Product rejected, or accepted and never matched

High effect

variant attributes (size, color, material, pattern), google_product_category and product_type, review aggregates (star_rating and review_count in OpenAI's feed, the product ratings feed in Merchant Center), shipping and return fields, additional images

Whether the product gets recommended for a specific qualifier

Product appears on brand prompts and never on use-case prompts

Enrichment

compatibility lists, use-case tags, certifications, sustainability claims, dimensions and weight, care instructions, bundle contents

Whether the product wins against a comparable competitor

Product loses "best for" prompts to a rival with the same specs and better data

Work the tiers in order, and finish a tier for the whole catalog before starting the next. A common pattern: a hundred products enriched down to the care instructions, and a thousand with no GTIN. Google's own GTIN guidance warns that products with missing or incorrect GTINs may have limited visibility, so the thousand are the bigger problem.

What each AI shopping platform actually requires

Platform

Program and format

Required fields

Variant fields

Notes (September 2026)

Google (AI Mode, AI Overviews, Shopping)

Merchant Center; XML, TSV or API

id, title, description, link, image_link, availability, price; brand for new items; GTIN strongly recommended

item_group_id, required in the US, UK, DE, FR, JP and BR

New conversational attributes (answers to product questions, compatible accessories, substitutes) rolling out since January 2026

OpenAI (ChatGPT)

Product feed, apply at chatgpt.com/merchants; JSONL, or a Google-compatible feed once OpenAI confirms it

item_id, title, description, url, brand, seller_name, image_url, availability, price

group_id, listing_has_variations, variant_dict (IDs must stay stable)

is_eligible_search and is_eligible_checkout flags. Instant Checkout was retired in March 2026; checkout happens on your site

Perplexity

Merchant Program; Google Shopping spec, CSV or XML

Same as Google

item_group_id

Free for merchants

Microsoft Copilot

Copilot Merchant Program via Microsoft Merchant Center; Google-compatible

As Microsoft Merchant Center

item_group_id

Copilot Checkout works with Shopify, PayPal and Stripe merchants

Titles cap at 150 characters and descriptions at 5,000 in Merchant Center (product data spec); OpenAI's spec gives the same figures as a target ("aim for at most 150 and 5,000 characters"). The four programs share most of their core fields, which is why one canonical record with a per-platform export works.

How to audit your product feed for AI shopping gaps

Run the checklist per SKU, pass or fail. Export the fails by tier.

Check

Pass

Fail

Example failure

Identifier

Valid GTIN (8, 12, 13 or 14 digits, correct check digit), or brand + MPN, or identifier_exists = no for custom goods

Missing, placeholder, invented, or failed check digit

"N/A" in the GTIN column for 300 SKUs, or one GTIN reused across every size

Title

Entity first, two deciding attributes, under 150 characters

Keyword list, or a variant code

"Jacket - Blue - M - SKU4471"

Description

Original, attribute-bearing, one use case named

Manufacturer copy, or under 30 words

Same 40 words as 60 other retailers

Category

Deepest accurate taxonomy node

Top-level node, or wrong branch

"Apparel" for a rain jacket

Availability

Matches inventory today

Stale, or "in stock" on a sold-out variant

Feed refreshed weekly, stock changes daily

Price

Matches the page and the schema

Differs from the page

Sale price on the page, list price in the feed

Variants

One item_group_id (group_id in OpenAI's feed) per parent, each child with its own GTIN and distinct color/size values, mirrored by ProductGroup/hasVariant on the page

Duplicated parents, missing group ID, or children with identical titles

Six "Rain Jacket Blue" items with different GTINs and no group ID

Page agreement

Feed values present in page schema and copy

Page says something else

Feed weight 340 g, page says 400 g

Images

Main image on a clean background at 500 x 500 px or larger (Google's minimum from January 31, 2027), variant-specific image per child, lifestyle and detail shots in additional_image_link

One image for all variants, or no additional images

Blue image on the red variant

A price failure is the one shoppers notice first. When an assistant already shows your product at the wrong price, our guide to checking ChatGPT Shopping and AI Mode walks through the fix from the answer side.

Then read the answers from the shopper's side: ten natural-language prompts for your category on ChatGPT, Perplexity and Google AI Mode, reading which of your products appear, which attributes the answer uses to justify its picks, and which competitor products take the prompts you lose. This is diagnosis, not a score: it shows what the assistants say, not how often. The full reading method is in that same guide; a product that passes every feed check and never shows up in the answers you read likely has a content or review problem, not a feed problem.

Variant fragmentation. It takes two shapes. Splitting: each size and color lives as its own product with its own title, so no single item collects the reviews or the specs, and the assistant sees twelve thin products instead of one strong one. Merging: one product with no variant attributes, so the assistant recommends "the jacket" and the shopper lands on the wrong size. Diagnose it by prompting for a specific variant ("in a women's medium, navy") and reading which item the assistant returns.

How to rewrite titles and descriptions for natural-language AI retrieval

The rule for titles: product entity first, then the two attributes a shopper uses to decide, then the differentiator. Brand where it helps recognition, variant attributes at the end. Under 150 characters, no promotional words. The brands below are illustrative.

Query intent

Before

After

Use case

"Jacket - Blue - M"

"Packable rain jacket for cycling, 340 g, reflective trim, Acme Commute, navy, M"

Comparison

"Acme X200 Vacuum PRO Best Cordless 2026"

"Acme X200 cordless stick vacuum, 45-minute runtime, pet hair brush, 2.5 kg"

Compatibility

"Phone case black"

"Slim case for iPhone 16 Pro, MagSafe compatible, matte black, Acme"

Gift

"Candle set 3pk"

"Soy candle gift set, three 200 g jars, lavender, cedar and fig, boxed"

Front-loading works because product cards and carousels truncate long titles, so the words that decide the match have to survive the cut. Google's own guidance says it plainly: "Put the most important details first. Users will usually notice only the first 70 or fewer characters of your title" (Google Merchant Center). It also works because a title that opens with the entity parses cleanly into product, attribute and variant. Keyword stacks ("Best Cheap Vacuum Cordless Stick Vacuum Cleaner 2026") give the parser one entity repeated and no attributes, and read as spam to a model as they do to a human.

Descriptions follow the same logic in three sentences. Sentence one: the product and the buyer it suits. Sentence two: the two or three attributes a shopper compares, with numbers. Sentence three: the limit, what it isn't for, which earns trust and prevents a wrong recommendation. Then a spec block with the attributes as label and value pairs, mirrored exactly in the feed and in the page schema.

Alignment is the part teams skip. The feed says 340 g, the page says "ultralight", the schema says nothing. The assistant reads three sources and has no reason to trust any of them. Generate feed, page copy and schema from one product record so the same value appears in all three, down to the unit.

How to mark up product pages with schema.org Product, Offer and ProductGroup

Product and Offer markup carries price, availability, GTIN and brand on the page itself. ProductGroup ties the variants to one parent, so the page tells the same story as the feed's group ID. Markup on its own does not buy a recommendation. Its job here is agreement: Google reads it to correct your feed, and any assistant that visits the page finds the same values it got from the feed.

Schema.org property

Where it goes

Must match feed field

name

Product

title

description

Product or ProductGroup

description

brand.name

Product or ProductGroup

brand

gtin (or gtin13, gtin12)

Each Product variant

gtin

sku

Each Product variant

id (item_id in OpenAI's feed)

offers.price and offers.priceCurrency

Offer inside each variant

price

offers.availability

Offer inside each variant

availability

aggregateRating

Product or ProductGroup

review fields (star_rating, review_count)

productGroupID

ProductGroup

item_group_id (Google), group_id (OpenAI)

variesBy

ProductGroup

the variant attributes (size, color)

hasVariant, isVariantOf

ProductGroup lists its variants; a variant on its own URL points back to its parent

one row per child in the feed

Here is the rain jacket as a ProductGroup with two variants, navy M and navy L, each with its own GTIN and Offer (illustrative brand, sample GTINs). It follows Google's product variants documentation; paste it into a script tag of type application/ld+json.

{ "@context": "https://schema.org",

"@type": "ProductGroup",

"name": "Acme Commute packable rain jacket for cycling",

"description": "Packable rain jacket for cycling commutes, 340 g, reflective trim.",

"brand": { "@type": "Brand", "name": "Acme" },

"productGroupID": "ACME-COMMUTE-RJ",

"variesBy": ["https://schema.org/size", "https://schema.org/color"],

"hasVariant": [

{ "@type": "Product", "sku": "ACME-RJ-NAVY-M", "gtin": "0012345678905",

"name": "Acme Commute rain jacket, navy, M", "size": "M", "color": "Navy",

"offers": { "@type": "Offer", "url": "https://www.example.com/rain-jacket?size=M",

"price": "139.00", "priceCurrency": "USD",

"availability": "https://schema.org/InStock" } },

{ "@type": "Product", "sku": "ACME-RJ-NAVY-L", "gtin": "0012345678912",

"name": "Acme Commute rain jacket, navy, L", "size": "L", "color": "Navy",

"offers": { "@type": "Offer", "url": "https://www.example.com/rain-jacket?size=L",

"price": "139.00", "priceCurrency": "USD",

"availability": "https://schema.org/OutOfStock" } }

]

}


Validate every template with Google's Rich Results Test, then check Merchant Center diagnostics for price and availability mismatches between the feed and the page.

How to keep feed data current and consistent at scale

Cadence by field. Price and availability: at least daily, hourly for fast-moving stock and in promotion periods. Promotions: pushed with a start and end so the sale price expires in the assistant's view. Discontinued products: removed from the feed the day they leave the catalog, with the page returning a 410 or redirecting to the replacement. Descriptions and attributes: whenever they're edited, with a quarterly review per category. In Merchant Center, keep automatic item updates on: Google then corrects price and availability from your page's structured data when the feed lags, which is one more reason the schema has to be right. In OpenAI's feed, push the change when a sale starts or ends, as the spec asks.

Enrichment for large catalogs. Category-level templates generate the attribute set each product type needs and flag the empty cells. AI-assisted extraction pulls attributes from existing product pages and manufacturer PDFs to fill them, followed by a QA checkpoint: a human samples a slice of each category plus every item where the extractor's confidence is low. An exception queue holds the items that fail QA until a merchandiser resolves them. The workflow scales to tens of thousands of SKUs because humans read exceptions, not rows.

Feed-to-page mismatch detection. A nightly job fetches each landing page's schema markup and compares price, availability, GTIN and title with the feed row. Mismatches go to the exception queue with the two values side by side. Expect the queue to be long in the first month, when page templates, promo tools and the feed still disagree, and to shrink once all three read from the same product record.

Ownership. One owner per tier: the feed engineer for the minimum viable fields and the sync, merchandising for the high-effect fields, content for enrichment. One shared rule: the product record is the source of truth, and nobody edits the feed by hand.

One catalog, every assistant. Build one canonical, enriched product record. Export it to Merchant Center, to OpenAI's product feed, to Perplexity's and Copilot's merchant programs, and to your page schema, with per-platform formatting handled by the export and not by separate content. Separate feeds per platform drift apart over time and produce the price disagreements the assistants then quote.

Which tools help optimize product feeds, titles and structured data for AI shopping?

No single tool does all of it. The stack has four layers, and most retailers already own one or two of them.

Layer

What it does

Examples

When you need it

1. Feed management

Rules, title templates, per-channel export to Merchant Center, OpenAI, Perplexity and Microsoft

Channable, Feedonomics, Productsup, DataFeedWatch, GoDataFeed

You export to more than one channel, or fix titles by rule across thousands of SKUs

2. PIM (product record)

Single source of truth for attributes, enrichment workflows, approvals

Akeneo, Salsify, Plytix

Attributes live in spreadsheets, ERPs and supplier files that disagree

3. Structured data on the page

Product, Offer and ProductGroup JSON-LD generated from the product record

Shopify and BigCommerce themes, Schema App, Yoast WooCommerce SEO

Your templates output incomplete markup or none for variants

4. AI shopping visibility tracking

Whether assistants recommend the SKU, in what position, from which retailer

GetMint AI Shopping; Merchant Center AI performance insights for Google surfaces

You need to know whether layers 1 to 3 moved recommendations

Layers 1 to 3 change your data. Layer 4 measures what the assistants do with it. Google's AI performance insights report covers AI Mode, AI Overviews and the Gemini app (Google) for English-language queries in a handful of markets, the US included, and flags attributes missing from your product data. GetMint's AI Shopping module covers Alexa for Shopping (formerly Rufus), ChatGPT Shopping, AI Mode, AI Overviews, Copilot and Perplexity, and shows each product's appearance rate, position and retailer against competitors. GetMint's Advanced plan tracks 3 AI models, Enterprise & Agency a custom set. Check each vendor's site for current features before you buy. For a side-by-side of the tracking tools with prices, see our AI shopping visibility tools comparison.

Which sources the assistants read per category

The feed is one input. Knowing the others tells you where a fix outside the feed pays. The table below is a starting hypothesis by category, based on the kinds of pages shopping answers tend to cite. Verify it for your own category and market before you plan around it.

Category

Source types that tend to matter

Consequence for the feed team

Electronics

Manufacturer spec pages, review roundups, marketplace listings

Specs in the feed have to match the manufacturer page to the unit

Beauty

Beauty retailer pages, ingredient databases, review platforms

Ingredient and skin-type attributes in the feed, mirrored on the page

Home and furniture

Retailer pages, buying guides, forums, Reddit among them

Dimensions, material and assembly fields filled; a guide page per use case

Apparel

Brand pages, review platforms, size-guide pages

Size and fit attributes, a size guide one click from the product page

Grocery

Retailer pages, nutrition databases

Pack size, price per unit, availability synced daily

Sports and outdoor

Specialist review sites, brand pages, forums

Weight, dimensions, use-case attributes; the spec table on the page

Two things follow. When a retailer page is the source an assistant cites, the retailer's copy of your data is what shoppers hear, so the feed you send the retailer matters as much as your own. And when review platforms dominate a category, no feed fix moves the recommendation until the reviews exist. Run the review campaign in parallel with wave two below; our Perplexity and Copilot guide covers how reviews weigh in those answers.

The mix differs by market and moves month to month. To see it for your own category, pull the domains and URLs cited on your shopping prompts. GetMint's Sources Explorer lists them per prompt, engine and market, keeping pages that cite you separate from pages that only mention you.

How to prioritize AI feed fixes by effect and effort

Score every fix on three axes, 1 to 3 each: expected effect on AI shopping visibility, expected effect on Google Shopping performance, and implementation effort (3 is easiest). Multiply the three. Highest total goes first. The scores below are a worked example for a typical catalog; score your own.

Fix

AI effect

Google Shopping effect

Effort (3 = easy)

Score

Add missing GTINs

3

3

3

27

Fix stale availability

3

3

3

27

Correct taxonomy to the deepest node

3

2

3

18

Rewrite titles, entity first

3

2

2

12

Restructure variants

3

2

1

6

Enrichment attributes

2

1

2

4

Write original descriptions

3

1

1

3

The quarterly plan follows the scores in three waves. Wave one, structural: identifiers, availability sync, taxonomy, mismatch detection. Wave two, content: titles, descriptions, variant structure and page schema for your top-selling SKUs. Wave three, automation: category templates, extraction, the exception queue, and enrichment for the long tail. Teams starting with wave three buy tooling to enrich products the assistants cannot see.

Measuring each wave means a fixed set of shopping prompts, each run several times per assistant, over several weeks, with a tool that samples every prompt repeatedly.

Why a hand count misleads. AI answers are sampled, not fixed. In a SparkToro and Gumshoe study, 600 volunteers ran 12 identical prompts nearly 3,000 times on ChatGPT, Claude and Google's AI, and the odds of getting the same list of brands twice were under 1 in 100 (SparkToro, January 2026). What holds steady is how often a brand appears across many runs, and measuring that takes volume. On 75 answers (25 prompts run three times), a 50% mention rate carries a margin of about ±11 points at 95% confidence; on 15 answers, about ±25 points. A hand count can't tell a real move from chance. Read answers by hand to understand them; count them with a tool that runs every prompt many times.

GetMint's AI Shopping module runs your shopping prompts daily and shows which SKUs appear, at what position and on which retailer, so your feed team sees the mention rate move on the SKUs they fixed. Plans are sized on a call, on a 12-month commitment. The number to report is the share of prompts where your product appears with the right price, before and after each wave, as a trend across several weeks of runs.

FAQ

What are the minimum product feed fields for AI shopping?

id, title, description, brand, price, availability, link and image, plus a valid GTIN or brand + MPN. OpenAI's product feed requires nine fields and Merchant Center seven. Merchant Center caps titles at 150 characters, and OpenAI recommends the same limit.

Do I need a separate feed for ChatGPT?

No. OpenAI accepts a Google-compatible feed once it confirms your registration, and Perplexity and Copilot use the Google spec. Keep one record and export per platform.

Does schema markup matter if I already have a feed?

Yes. Google uses page structured data to auto-update prices and availability, assistants that visit the page find what the markup says, and a feed that disagrees with the page loses trust. Use ProductGroup with hasVariant so the page mirrors the feed's group ID.

What tool helps me add product structured data?

Most e-commerce platforms output Product markup from their themes (Shopify and BigCommerce among them). Schema App and Yoast WooCommerce SEO add or extend it, including variants. Whatever generates it, feed it from the same product record as your feed.

How often should I update price and availability?

At least daily, and hourly during promotions or for fast-moving stock.

How do I know if my feed fixes worked?

Track a fixed set of shopping prompts with a tool that runs each one repeatedly on each assistant over several weeks, and follow the share of answers where your product appears at the right price. Reading a few answers by hand tells you why a product is missing, not whether the fixes moved the rate.

Conclusion

The feed decides whether an assistant can consider your product. The page schema decides whether what it finds agrees with the feed. Everything past that is measurement. Fix the minimum viable fields, make feed and page agree from one record, then watch which SKUs the assistants start recommending.

See which of your SKUs AI shopping assistants recommend, in what position and from which retailer: book a GetMint demo.

par

par

Jean-Noël Escande

Jean-Noël Escande

BUILT FOR ENTERPRISE BRANDS

Your competitors are already visible in AI. Are you?

See what ChatGPT, Gemini and Perplexity say about your brand. Set up in 5 minutes.

FOLLOW US

English

Stay ahead of the AI search curve.

Weekly insights on GEO, AI visibility trends, and strategies from 200+ brand case studies. No spam, just signal.

Join 3,000+ marketers. Unsubscribe anytime.

Subscribe

Subscribe

© 2026 GetMint. All rights reserved.

Stay ahead of the AI search curve.

Weekly insights on GEO, AI visibility trends, and strategies from 200+ brand case studies. No spam, just signal.

Join 3,000+ marketers. Unsubscribe anytime.

Subscribe

Subscribe

FOLLOW US

English

Stay ahead of the AI search curve.

Weekly insights on GEO, AI visibility trends, and strategies from 200+ brand case studies. No spam, just signal.

Join 3,000+ marketers. Unsubscribe anytime.

Subscribe

Subscribe