The 2026 technical playbook for turning your product feed into the context layer AI agents reason over — the OpenAI, Perplexity and UCP specs, the fields that decide recommendations, and why feed hygiene alone is no longer enough.

Most ecommerce teams still treat the product feed as an advertising-operations chore: a file you keep clean enough to pass Google Merchant Center’s compliance checks and avoid disapprovals. In 2026 that framing is quietly costing them visibility on the fastest-growing discovery surface in retail. ChatGPT now processes around 50 million shopping queries a day; its product recommendations are drawn first and foremost from a structured feed — and that feed is no longer a copy of your Google Shopping export.

The shift is structural, not cosmetic. Your feed has stopped being a list of products to advertise and become the context layer an AI agent reasons over — the machine-readable evidence it uses to understand what you sell, compare you against rivals, and decide whether to recommend you to a buyer. A listing that reads “Blue Backpack – £49.99” will not survive that environment. The same product described with weight, material, compartment dimensions, compatible use-cases, certifications, GTIN and live inventory will be recommended ahead of it on every major platform.

This guide is the technical companion to our hub on getting your products recommended by AI shopping agents. The hub makes the strategic case; this article is the build sheet. We cover the OpenAI, Perplexity and UCP feed specifications field by field, the attributes that actually move recommendations, the multi-surface distribution model, and — the part every competitor’s checklist omits — why feed completeness gets you into the selection pool but earned authority is what gets you picked out of it.

The thesis of this guide: Feed hygiene is necessary and finite. Every competent merchant will eventually have a clean, complete, multi-surface feed, because the specs are public and the tooling is commoditising. The durable advantage is in the signals a feed alone cannot manufacture — verified review depth, cross-source consistency, and the citation history that makes an agent trust your data. The feed carries those signals; it does not create them. That is precisely where a link building function earns its place in an AI-commerce programme.

What this guide covers

Why the AI feed is a reasoning layer, not an ad feed — and what that changes
The OpenAI ChatGPT product feed spec: required, recommended and AI-only fields
Perplexity’s Catalog API, Google’s UCP feed, and the Shopify auto-eligibility shortcut
The 12 core attributes, the 20–30 enrichment attributes, and the fields that decide recommendations
Feed-to-schema consistency, freshness mechanics, and the most common rejection causes
The off-feed signals — reviews, consistency, citation age — that out-rank a perfect feed
A multi-surface implementation sequence and a measurement framework

Why the AI feed is a reasoning layer, not an advertising feed

On Google and Meta, the system decides what to show from the attributes you send and the bids you set; the feed is one input among several. In agentic commerce the feed is closer to the whole game on the eligibility side. Agents do not browse category pages or read your hero copy. They query structured data, evaluate machine-readable catalogues, and recommend — or purchase — on the buyer’s behalf. The feed is the primary evidence they reason over.

Two consequences follow. First, semantic density beats marketing polish. The agent is matching a long, natural-language query — “a quiet 48-inch mahogany desk with cable management for a small home office” — against attributes, not slogans. Klaviyo’s 2026 data found 30% of AI shopping queries run to eight words or more and most carry personal or emotional context. Feeds rich in factual, intent-aligned attributes match those queries; feeds full of “premium quality” filler do not. Second, freshness is a trust signal, not a nicety. Engines connected to live search will detect a discrepancy between your feed price and your site price and mark your data unreliable — quietly dropping you from consideration.

The reframe for link builders: “Beyond Google Merchant” is not just about extra platforms. It is about the feed graduating from a compliance artefact to a credibility artefact. Credibility is the link builder’s home turf — and several of the feed’s most decision-relevant fields (review count, average rating, return rate) are populated by signals you earn off the feed, not values you type into it.

The OpenAI ChatGPT product feed spec, field by field

The single most important thing to understand: the ChatGPT feed is pushed, not crawled. Merchants send a structured file to a secure OpenAI endpoint, which ingests, validates, indexes and — in OpenAI’s own documentation, repeatedly — ranks it. Formats accepted are CSV, TSV, XML and JSON, and the feed can be refreshed as often as every 15 minutes. It is a distinct specification with its own required fields, control flags and ranking inputs.

Control flags: enable_search and enable_checkout

OpenAI’s spec introduces two control flags found in no Google feed. enable_search governs whether a product can appear in ChatGPT shopping results at all; enable_checkout governs whether it can be purchased in-flow where checkout is available. These do not affect how the product looks on your own site — they simply switch the ChatGPT integrations on or off, which lets you suppress seasonal lines, gate premium SKUs, or A/B test visibility at the SKU level. Sync them with your existing Shopify or WooCommerce product-status workflows so discontinued items never linger in recommendations.

Required, recommended and AI-only fields

The spec groups attributes by object — basic data, media, pricing, variants, fulfilment, seller — and marks each Required, Recommended or Optional. The required core overlaps with Google, but the high-value differentiation lives in fields Google never had. The table below is the practical hierarchy: ship the required core first, then the AI-only enrichment fields that competitors leave blank.

Field group	Key fields	Why it matters for recommendations
Basic (required)	id (unique), title, description, product_category, brand, gtin/mpn, condition, product URL	The eligibility floor. Non-unique IDs and missing identifiers are top rejection causes
Description (AI-only)	Accepts plain text, HTML and markdown for the same product	Structured formatting and bullets let the agent parse attributes during conversational matching — a real edge over plain-text-only feeds
Pricing	price, sale_price, currency (ISO 4217), availability	Must match your site and schema exactly; mismatches drop you from consideration
Media	image_link, additional_image_link, video_link, model_3d_link (GLB/GLTF)	Rich cards require image completeness; video and 3D are barely adopted, so early movers gain visibility
Variants	item_group_id, colour, size, gender, size_system, custom variants	Group SKUs under one canonical group_id; custom variants match intent-heavy queries
Trust / performance	product_review_count, average_rating, return_rate, popularity_score (0–5)	Explicit ranking inputs — and the fields most tied to off-feed, earned signals
Fulfilment & seller	shipping, delivery_estimate, return_policy, return_window, seller_name, policy URLs	Trust signals agents surface directly; policy URLs become required when checkout is enabled

The field competitors ignore most: popularity_score and return_rate are explicit ranking signals with no Google equivalent. They are also the fields you cannot honestly fabricate — a high return_rate or an inflated popularity_score the agent can cross-check against external sentiment will erode trust. These fields reward brands whose real-world performance and external reputation back the numbers up, which is exactly why feed work cannot be divorced from reputation work.

Category, taxonomy and the multi-path question

The spec requires a product_category following a path such as Apparel & Accessories > Shoes. Where Google permits a single taxonomy node, ChatGPT’s documentation shows a single path but does not forbid more — and early evidence suggests multiple category paths may be supported, which would widen the queries you match. Until confirmed, choose the category most aligned with conversational intent rather than the one that fits your internal merchandising. For verticals, the spec supports category-specific attributes: apparel needs size_system, size_type and fit; electronics need compatibility and spec-sheet fields; food needs nutrition_facts and ingredients. Ship the minimum and you appear as a plain text mention; ship attribute-complete and you earn a rich product card, which drives materially higher click-through.

A worked example: the same product, two feeds

Abstraction obscures how large the gap is between a compliance feed and a reasoning feed. Take a single SKU — a £49.99 laptop backpack — and look at what each version gives an agent to work with.

Feed A: the Google-export compliance version

Title: “Backpack – Blue.” Description: “Premium quality blue backpack. Great for everyday use. Buy now.” Category mapped to a single broad node. Price and availability present. No GTIN. No dimensions, no material, no laptop-size compatibility, no review fields, plain-text description only. It passes Merchant Center checks. To an agent fielding the query “a backpack that fits a 16-inch laptop and a water bottle for daily commuting,” Feed A is almost mute: it cannot confirm laptop fit, cannot confirm the bottle pocket, has no reviews to build confidence, and no identifier to cross-reference. The agent has little reason to recommend it over a rival that answers the question directly.

Feed B: the reasoning version

Title: “Aerodesk Commuter 22L Laptop Backpack – Fits 16\” Laptop, Water-Resistant.” Description (markdown): a short factual summary plus bullet points for capacity (22L), laptop compartment (up to 16\”), material (recycled water-resistant polyester), external bottle pocket, weight (790g), warranty (2 years). GTIN and MPN present. item_group_id grouping the three colourways. average_rating 4.6 and product_review_count 318 populated from a live review system. return_window 30 days, free_shipping_indicator true. additional_image_link showing the laptop compartment in use. Category chosen for conversational intent.

Against the same query, Feed B matches on every constraint — laptop size, bottle pocket, commuting use-case — surfaces as a rich card rather than a text mention, and carries the review depth and policy signals that let the agent recommend it confidently. Same product, same price, radically different outcome. The difference is entirely attribute density and earned trust fields, and it is reproducible across every SKU in your catalogue.

Read the example carefully: Everything that made Feed B win is either a factual attribute you can author today or an earned signal (the 318 reviews, the 4.6 rating) you have to build. The authoring is a one-week project. The earned signals are the moat — and they are the part a feed-management tool cannot generate for you.

Beyond OpenAI: Perplexity, UCP and the Shopify shortcut

“Beyond Google Merchant” means at least four destinations in 2026, each with its own ingestion model. The good news for the 4.6 million-plus Shopify merchants is that the landscape is far simpler than it looks, because Shopify Catalog provides the same schema fields automatically for every store.

Destination	How it ingests product data	Practical onboarding
ChatGPT (OpenAI / ACP)	Pushed structured feed via merchant portal or SFTP/API	Direct submission, or automatic via Shopify Catalog (US merchants auto-discoverable since March 2026)
Perplexity	Accepts Google Shopping-format feeds; Catalog API ingests title, price, stock, attributes in real time	Separate Perplexity Merchant Program enrolment; minimal effort
Google AI Mode / Gemini (UCP)	Draws from Merchant Center; set nativecommerce=true to enable agentic purchasing	Merchant Center is the highest-volume channel; Shopify simplifies the connection
Microsoft Copilot	ACP + UCP; Shopify Payments is a Copilot Checkout partner	Shopify merchants auto-enrolled with opt-out only — no application

The strategic takeaway is that discovery eligibility is increasingly a configuration step, not an engineering project. Shopify Catalog exposes a compliant catalogue to ChatGPT, Copilot, Gemini and Google AI Mode largely by default; Universal Cart, launched at Google I/O on 19 May 2026, is extending the buy-in-chat surface further. If the eligibility layer is becoming a checkbox everyone ticks, the differentiation has to come from somewhere the checkbox does not reach. It does — and that is the back half of this guide.

One technical caveat that trips up international sellers: agents cannot simulate geolocation. A single URL with IP-based dynamic pricing does not work for them. Build distinct per-market product URLs with market-specific schema, ISO 4217 currency codes and region-appropriate shipping on each page, so the agent sees one unambiguous price per market rather than a price that shifts under it.

The attribute hierarchy: what to ship, in what order

Order matters more than completeness. Ship the core, then enrich, then align schema, then distribute. Skipping the first step makes the rest invisible; over-investing in exotic fields before the core is complete wastes effort. Here is the sequence.

Layer 1 — the 12 core attributes (the eligibility floor)

Complete on every SKU, no exceptions: title, description, brand, GTIN, MPN, category (Google Product Taxonomy), price, sale price, availability, condition, image URL, product URL. Most merchants shipping to Google have most of these; agents need all of them. Fill rate is itself a signal — merchants with 95%+ fill on core attributes see dramatically higher AI visibility, because incompleteness reads as unreliability.

Layer 2 — 20–30 enrichment attributes (the differentiation)

This is where you out-describe competitors. Material, dimensions and weight; compatible use-cases; certifications and sustainability claims; variant detail; the AI-only media (video, 3D); and the trust/performance fields. The principle is semantic density: every factual, query-relevant attribute is another natural-language question your product can answer. Write descriptions in the buyer’s language and around use-cases, not feature lists — that is what conversational matching rewards.

Writing descriptions an agent can reason over

The description field deserves special attention because ChatGPT’s spec accepts plain text, HTML and markdown for the same product — a genuine advantage most merchants waste. Structure earns parsing. A short factual lead sentence establishes what the product is; bulleted attributes give the agent discrete, matchable facts; and use-case phrasing connects the product to how buyers actually ask. Three rules separate descriptions that match queries from descriptions that do not:

Lead with the noun and the differentiators, not an adjective. “22L water-resistant commuter backpack with a 16-inch laptop sleeve” matches more queries than “premium everyday carry.”
State attributes as facts the agent can verify, not claims it must trust. Numbers, materials, dimensions and compatibilities are matchable; “best-in-class” is not.
Mirror buyer language and situations. If buyers ask about “hand luggage that fits under a Ryanair seat,” the phrase belongs in the description — long-tail, situational queries are where AI shopping concentrates.

This is the same use-case-driven, intent-aligned writing that earns links and citations elsewhere; here it is pointed at the feed. The discipline transfers directly from editorial content work to feed copy, which is why teams with a content and link building function tend to write better AI feeds than teams that treat the feed as pure operations.

Layer 3 — schema alignment (consistency is the signal)

Your submitted feed and your on-page Schema.org markup must agree. Agents parse Product, Offer (price, priceCurrency, availability in ISO 4217), AggregateRating and Review schema as standalone data, and they cross-reference it against the feed and against other sources on the web. A contradiction between feed, schema and site is read as unreliable data and suppresses recommendations. Render JSON-LD server-side: GPTBot, ClaudeBot and PerplexityBot download JavaScript but do not execute it, so client-injected schema is invisible to them. The full markup specification lives in our companion guide on structured product data for AI commerce, and the crawler-access mechanics in AI bot crawl optimisation.

Layer 4 — distribution and freshness

Push to every surface your buyers use, and keep it fresh. OpenAI supports 15-minute refreshes; treat freshness as a ranking factor for dynamic inventory and pricing. The metric to manage is data-freshness latency — the lag between a change in your store and the change reaching each agent. Stale price or stock data does not just cause a bad experience; it gets your brand flagged as unreliable across surfaces simultaneously.

Choosing a feed format and refresh model

OpenAI accepts CSV, TSV, XML and JSON. For small, static catalogues a scheduled file push is fine; for large or fast-moving catalogues, prefer JSON over an API or SFTP path that supports frequent partial updates, so a price change propagates in minutes rather than waiting for a nightly full-file regeneration. The emerging pattern for the most dynamic merchants is a real-time sync — an MCP server or live API that exposes current inventory and pricing on demand — which eliminates latency entirely but costs more to build and maintain. Match the model to your volatility: a furniture brand with stable prices does not need the infrastructure a flash-sale fashion retailer does. Whatever the model, monitor latency as a first-class metric, because it is the freshness signal agents actually respond to.

A practical sequencing note: do not chase real-time sync before the core attributes are complete. A perfectly fresh feed of thin records loses to a once-daily feed of rich ones. Completeness first, freshness second, real-time last — and only where volatility justifies it.

Vertical-specific feed requirements

The core and enrichment layers are universal, but conversational queries in each category lean on category-specific attributes the agent expects to find. Shipping the generic core and omitting the vertical fields is a common reason a technically valid feed still loses to a competitor’s in a specific category. The patterns below are the ones that most affect matching.

Vertical	Attributes agents expect	Query types they unlock
Apparel & footwear	size_system, size_type, fit, material, gender, care, colourway	“true to size?”, “breathable summer fabric”, “wide-fit trainers”
Consumer electronics	compatibility, spec sheet, connectivity, power, dimensions, warranty	“works with iPhone 17?”, “USB-C, under 200g”, “fits a 27-inch monitor”
Home & furniture	dimensions, weight, material, assembly, model_3d_link, room/use	“48-inch desk for a small room”, “flat-pack, fits through a standard door”
Food & beverage	nutrition_facts, ingredients, allergens, dietary flags, origin	“gluten-free, high protein”, “vegan and nut-free”
Beauty & personal care	ingredients, skin/hair type, volume, certifications, cruelty-free	“fragrance-free for sensitive skin”, “reef-safe SPF”

The principle generalises: identify the constraints buyers in your category actually specify, then make sure every one of them is a discrete, machine-readable attribute in the feed and a matching claim in your schema. The fastest way to find those constraints is to mine your own reviews, support tickets and the long-tail queries from your probe-query log — the same buyer language that should drive your descriptions. A feed built from real buyer constraints matches real buyer queries; a feed built from your internal product taxonomy matches the questions nobody asks.

The trust and performance fields, in depth

The fields most tied to recommendation outcomes — and least understood — are the trust and performance group, because they are the bridge between the feed and the off-feed reputation work the rest of this guide argues for. Treat them carefully; they reward honesty and punish manipulation, because agents can cross-check them against external sentiment.

product_review_count and average_rating

These are explicit inputs to whether an agent recommends and how confidently. They are also the clearest case of a feed field whose value is earned elsewhere: you transmit the number, but a review-generation system produces it. The strategic error is to optimise everything else in the feed and leave these empty — a complete feed with zero reviews still reads as a risk. Prioritise getting genuine reviews flowing and piped into the feed before you chase exotic media fields, because review signals do more work in the selection step than a 3D model does.

popularity_score and return_rate

OpenAI flags popularity_score (a 0–5 scale, or merchant-defined) as a ranking signal, and the spec exposes return_rate. Both are double-edged. A merchant-set popularity_score creates an obvious inflation temptation — but an agent that can compare your claimed popularity against external sentiment, review volume and citation frequency will discount a score that the wider web does not support, and may treat the discrepancy itself as an unreliability signal. return_rate is even less forgiving: a low, honest return rate is a powerful trust signal an agent will surface; a misstated one collides with the reality buyers report in reviews and on third-party sites. The lesson is consistent with everything else here: these fields reward brands whose real-world performance and external reputation corroborate the numbers, which is another reason feed work and reputation work cannot be run in separate silos.

The pattern across every trust field: The feed is a transmission layer for reputation, not a substitute for it. You can type a number into return_rate or popularity_score, but the agent’s confidence in that number comes from whether the rest of the web agrees. Build the reputation, report it honestly, and the trust fields become an asset; fake it and they become a liability the agent learns to distrust.

The most common rejection and suppression causes

Eligibility failures fall into a predictable set. Audit against this list before submission and after every major catalogue change.

Non-unique product IDs. Duplicate or unstable IDs break ingestion and matching. Use stable, unique identifiers.
Missing required fields or identifiers. Absent GTIN/MPN breaks cross-source matching and can exclude you from carousels entirely.
Marketing fluff instead of factual descriptions. “Premium quality” tells an agent nothing; attributes do. Factual, structured descriptions are explicitly favoured.
Price and availability mismatches. Feed-vs-site contradictions are a hard disqualifier on live-search engines.
Client-side-only schema. If JSON-LD loads via JavaScript, AI crawlers never see it. Render server-side.
Blocking the wrong crawler. Disallowing OAI-SearchBot removes you from ChatGPT recommendations regardless of feed quality.
Geolocation-dependent single URLs. IP-based pricing on one URL confuses agents; use per-market URLs.

Why this list matters strategically: Notice that every item here is a one-time or low-maintenance fix. None of it differentiates you, because every competitor can fix the same list from the same public specs. Clearing it is the price of entry, not a competitive moat. The moat is built in the next section.

What a perfect feed cannot do: the off-feed signals that decide recommendations

Here is the sentence the tool-vendor checklists bury or omit: getting into the selection pool is not the same as being selected from it. Even the researchers behind the most-cited 2026 feed study conceded that product sentiment, brand mentions in contextual sources, and product-specific signals likely drive the final selection and ranking within the carousel. In other words, once two feeds are equally complete, the agent reaches for signals that live entirely outside the feed.

It is worth being precise about why this happens, because it is structural rather than incidental. A feed is self-reported data: it is what the merchant says about its own products. An agent making a recommendation it will stake its usefulness on cannot rely on self-report alone, any more than a careful buyer would trust a product solely on the seller’s own description. So it triangulates — checking the merchant’s claims against independent reviews, third-party coverage, sentiment and citation patterns it has indexed over time. The feed establishes the candidate and its facts; the external corroboration establishes whether those facts can be trusted and whether the product is genuinely well regarded. The better-corroborated product wins, and corroboration is, by definition, something you cannot put in your own feed.

Three of those signals are the link builder’s responsibility, and none can be typed into a feed file.

1. Earned review depth and velocity (the fields you cannot fake)

product_review_count and average_rating are feed fields, but the values are earned, not authored. Three in five shoppers hesitate to buy a product with no reviews (ChannelEngine), and agents inherit that caution. Velocity matters as much as volume: AI surfaces reward freshness, so a steady monthly inflow of recent reviews signals current relevance in a way a large but stale count cannot. The feed transmits the number; a sustained review-generation programme produces it. Pipe the result back into the feed so the signal is legible to every agent.

2. Cross-source consistency (the web-wide audit agents run)

Agents cross-reference your feed against your schema, your site, marketplaces and third-party sources. Consistency across all of them — price, naming, specifications, policy — reads as reliability; contradiction reads as risk. Much of this consistency lives on properties you do not own: a marketplace listing with a different price, an outdated spec on a retailer’s page, an old brand name in a directory. Reconciling those is reputation work, not feed work, and it directly protects the trust an agent places in your data.

Run the audit the way an agent would. Take a representative SKU and gather every public statement of its price, title, key specs and policies — your site, your feed output, your schema, each marketplace listing, and any retailer or affiliate carrying it. Line them up and look for contradictions: a £49.99 feed price against a £54.99 marketplace listing, a “16-inch laptop” claim on your site against “15.6-inch” on a reseller’s page, a discontinued colourway still live on a third-party site. Each contradiction is a small reason for an agent to lower its confidence, and they compound. The fixes split into two buckets: the ones you control (feed, schema, site) you correct directly; the ones you do not (marketplaces, resellers, directories) you resolve through outreach and listing management — again, reputation work rather than feed engineering. Schedule the audit quarterly and after any price or spec change, because consistency decays silently as the web ages around a product.

3. Citation age and authority (the signal no spec contains)

The most important off-feed signal appears in no feed specification at all: how long, and how consistently, your products have been cited across AI-indexed sources. Agents do not treat all structured data as equal. Sources that have been referenced, linked and validated over time carry more weight than a newly optimised feed, however technically perfect. A brand named in buying guides for two years, cited by reviewers, and corroborated across the web is a confident recommendation; a brand with a flawless feed and no external footprint is a gamble the agent would rather not take.

This has a blunt operational implication: the citation work you start today is worth more the longer it runs, so the cost of delay is real and compounding. A competitor who began earning category citations a year ago holds an advantage you cannot close by perfecting your feed this quarter — you can only start closing it by beginning your own citation programme now. The asset that compounds fastest is a genuinely useful, linkable buying-decision tool that reviewers and agents reference repeatedly; we cover how to build those in building linkable buying-decision tools that AI agents cite. Treat citation age the way you treat domain age in classic SEO: a slow-building moat that rewards starting early and punishes waiting.

This is the whole argument for link building in AI commerce: A feed is replicable in a sprint. Citation history, earned reviews and web-wide consistency take quarters to build and cannot be bought in a feed-management tool. They are the durable, compounding advantage — and they are built with the link building and digital-PR toolkit. The full ranking-factor breakdown is in our guide to how agents choose which products to recommend; the citation-earning tactics are in our buying-guide citations playbook.

For the platform-by-platform weighting of these signals, see how ChatGPT, Perplexity and Gemini choose which products to recommend. For the outreach mechanics that earn the citations, see earning citations in AI buying guides and comparison answers. And because the marketplace surfaces weigh feed-vs-reputation differently, the trade-off is covered in marketplace vs owned-site visibility in AI shopping.

A multi-surface implementation sequence

Phase	Focus	Actions
Week 1–2	Core eligibility	Complete the 12 core attributes to 95%+ fill; fix unique IDs and identifiers; confirm OAI-SearchBot and PerplexityBot access; server-render JSON-LD
Week 3–4	Enrichment & flags	Add 20–30 enrichment attributes; write use-case descriptions; set enable_search/enable_checkout; add video/3D where relevant
Week 5–6	Distribution	Submit to OpenAI; enrol Perplexity Merchant Program; set nativecommerce=true in Merchant Center; confirm Shopify Catalog/Copilot enrolment; build per-market URLs
Week 7–8	Consistency & freshness	Reconcile feed vs schema vs site vs marketplaces; set refresh cadence; minimise data-freshness latency
Ongoing	Off-feed authority	Stand up review-velocity system; begin citation-earning outreach; monitor cross-source consistency

The shape mirrors the strategic point: phases 1–4 are a finite, eight-week engineering and operations project. The “ongoing” row is the part that never finishes and the part that actually wins recommendations once everyone’s feed is clean. Resource it accordingly — most teams over-staff the eight weeks and under-staff the forever.

Measuring feed performance in an opaque channel

Agentic analytics remain primitive, so measure directionally across three layers and accept imperfect attribution. Move on velocity, not certainty.

Feed health — track fill rate on core and enrichment attributes, rejection counts, and data-freshness latency. This is the only fully deterministic layer; keep it green.
Visibility — run weekly probe queries across ChatGPT, Perplexity, Gemini and Google AI Mode in your buyers’ language; log whether you appear as a rich card or a text mention, your position, and which competitors win. Card-vs-text is a direct readout of attribute completeness.
Outcome — segment AI-referred traffic from day one (it often lands as direct or referral from chat domains in GA4) and connect it to conversions. Recall that AI referrals convert markedly higher and return less, so modest AI volume can outperform its size.

For category benchmarks and the data points cited throughout, see our living link building statistics for 2026, and keep the technical SEO foundation sound, since an uncrawlable or slow page undermines every feed signal you submit.

Feed-readiness quick reference

Use this as a pre-submission and post-change checklist. If every line is true, your feed is eligible everywhere and carrying the trust signals the selection step rewards.

All 12 core attributes present on every SKU at 95%+ fill, with unique stable IDs and valid GTIN/MPN
Descriptions are factual, structured (markdown/bullets), use-case-led and written in buyer language
20–30 enrichment attributes and any vertical-specific fields populated
product_review_count and average_rating populated from a live review system; popularity_score and return_rate honest and corroborated
enable_search/enable_checkout set deliberately and synced to product status
Feed, schema and site agree on price, title, specs and policy; JSON-LD server-rendered
Submitted to OpenAI, Perplexity, Merchant Center (nativecommerce=true) and Copilot; per-market URLs where you sell internationally
Refresh cadence matched to volatility; data-freshness latency monitored
Cross-source consistency audited; off-feed review-velocity and citation programmes running

Common questions about LLM feed optimisation

Can I just reuse my Google Shopping feed?

Partly. Perplexity accepts Google Shopping-format feeds, and Shopify Catalog carries core fields automatically. But the ChatGPT spec has unique fields (enable_search, enable_checkout, popularity_score, return_rate, video_link, model_3d_link) and favours structured, conversational descriptions. Most merchants run a Google feed plus an adapted, enriched version for AI surfaces rather than a single shared file.

How often should I refresh?

As often as your data changes. OpenAI supports 15-minute refreshes. Stable catalogues are fine on a daily push; dynamic pricing or inventory needs near-real-time, because a feed-vs-live discrepancy gets your data flagged as unreliable. Complete first, fresh second.

Is the feed enough on its own to get recommended?

No, and this is the central point of this guide. A complete feed makes you eligible to be recommended; it does not make you the chosen recommendation. Once competing feeds are equally complete, agents select on off-feed signals — earned reviews, cross-source consistency and citation history — which are built with the link building and reputation toolkit, not in a feed-management tool.

What is the single highest-ROI feed change?

For most merchants, getting genuine reviews flowing and piped into product_review_count and average_rating, paired with factual, attribute-dense descriptions. Reviews do disproportionate work in the selection step, and attribute density is what lets you match the long, specific queries where AI shopping concentrates.

Where this leaves your feed strategy

The product feed has been promoted. It is no longer the advertising department’s compliance file; it is the evidence base an AI agent uses to decide whether your product is worth a buyer’s attention. That promotion rewards a specific kind of work: factual, attribute-dense, multi-surface, server-rendered, consistent and fresh. Get that right and you are eligible everywhere an agent shops.

But eligibility is the floor, and the floor is rising toward universal. The specs are public, the tooling is commoditising, and Shopify is turning much of the eligibility layer into a default. When everyone’s feed is clean, the agent falls back on the signals a feed cannot manufacture — earned reviews at velocity, consistency across sources you do not control, and the citation history that makes your data trustworthy. Those are built over quarters, with the link building toolkit, and they are what separate the brands agents recommend from the brands that merely qualify.

Build the feed as the precise technical instrument this guide describes — then build the off-feed authority that gives it weight. Start from the strategic frame in our hub on getting recommended by AI shopping agents, layer in the tactics from the 15 link building strategies hub, and treat the feed and the authority programme as two halves of the same job.