Product-Feed Optimisation for LLM Recommendations (Beyond Google Merchant)

Your feed is no longer a Shopping-ads file. It is the primary authority AI agents read to decide what to recommend, trust and sell. Here is how to optimise it for every surface at once.

1. The feed stopped being an ads file

For most of the last decade the product feed had one job: power Google Shopping and Performance Max. You filled in the dozen attributes Merchant Center demanded, fixed disapprovals when they appeared, and otherwise left it alone. That era is over. In 2026 the same structured file is read by every AI shopping surface — and for ChatGPT it is not merely a signal, it is the primary authority on what you sell, what it costs and whether it is in stock.

The hub for this cluster — getting your products recommended by AI shopping agents — covers the five-layer AISLE framework. This spoke goes deep on the layer most merchants get wrong: Identified — the feed and structured data that make an agent certain enough to choose you. Get it wrong and no amount of off-site authority rescues you, because you are filtered out before ranking begins.

The shift is best understood as a collapse of the funnel. The old model was a sequence: a human searched, clicked, browsed a product page you controlled, and converted. The agentic model compresses that into a single transaction layer — the shopper states a preference, the agent evaluates inventory across merchants, and it recommends or buys. In that model your product page is no longer the interface; your feed is. It is read, compared and acted on, often without the shopper ever seeing your PDP. Which means the data quality you used to be able to hide behind good design and persuasive copy is now exposed, naked, as the thing the machine judges you on. Analysts frame this bluntly: bad product data used to hurt conversion; in agentic commerce it prevents selection.

2. One catalogue, many surfaces: the shared-field reality

The most important operational insight of 2026 is also the most reassuring: although five-plus checkout surfaces each launched on a different protocol with its own feed spec, the majority of required fields are shared. You are not building five feeds. You are governing one clean catalogue and distributing it. The protocol-behind-the-surface map tells you which single integration covers what.

Read that table as a sequencing plan, not a menu. One correct GMC feed or UCP manifest now covers Google’s cross-merchant Universal Cart and Copilot’s UCP path; one correct ACP feed covers ChatGPT and Copilot’s ACP path simultaneously. Use a feed-management tool (DataFeedWatch, Feedonomics, GoDataFeed and similar) to hold the single source of truth and syndicate — doing it by hand at thousands of SKUs is not viable, and variant-level richness is now expected.

Choosing where the source of truth lives

The architectural decision underneath all of this is where your canonical product data lives and how it flows to each channel. Three patterns are common in 2026. Smaller Shopify and Etsy merchants can lean on platform-native syndication — Shopify’s Agentic Storefronts push to ChatGPT, Perplexity and Copilot from one setup, so the platform is effectively the source of truth. Mid-market brands typically run a dedicated feed tool that ingests the raw catalogue, applies enrichment rules and mappings, and exports per-channel formats with a validation layer in front. Enterprise and headless stacks build a “universal agent gateway” — a single API the core commerce engine exposes that any protocol can query. Whichever pattern fits, the non-negotiable is the same: one governed dataset, many rendered outputs, never divergent hand-maintained files. The moment two channels disagree, an agent treats both as suspect.

3. The attribute checklist, in three tiers

Think of attributes in tiers. Tier 1 makes you eligible. Tier 2 makes you matchable to specific queries. Tier 3 makes you trusted. Most merchants fill Tier 1 and stop — which is exactly why the merchants who complete Tier 2 and 3 win the recommendation.

Tier 1 — Eligibility (non-negotiable)

• Identity: title (write it like a knowledgeable sales associate, not a keyword string), description, brand, and a GTIN/UPC. GTINs are the product’s cross-reference key — without them, products are auto-excluded from Performance Max and Gemini native commerce.
• Commercials: price with ISO 4217 currency code, sale price where applicable, real-time availability status, condition.
• Media + taxonomy: clean high-resolution primary image (white/neutral background, no watermarks or promo text), correct Google product category.

Tier 2 — Matchability (the semantic layer)

This is where conversational queries are won or lost. “Show me lightweight options” needs weight data; “will this match stainless-steel appliances?” needs a finish attribute; “a rug for a high-traffic hallway” needs a usage scenario. Fill the descriptive attributes most catalogues leave blank:

• material, colour, size, style, dimensions, weight, finish, construction_method
• usage_scenario / lifestyle_fit, care_instructions, compatibility, variants modelled explicitly (each colour/size with its own title, attributes and image)
• Be specific, not vague: “1200-thread-count Egyptian cotton” beats “high quality”; “20V brushless motor, 400 in-lbs torque” beats “powerful”.

Tier 3 — Trust + rich media

• Reviews in the feed: product review count and average rating. In Google these sit outside the feed; in ChatGPT’s ACP feed you can supply them directly, giving you more influence over perceived quality.
• Performance signals (ACP): popularity_score (0–5, reflect genuine sales velocity — don’t inflate) and return_rate (a lower percentage signals reliability).
• Rich media: video_link (publicly accessible HTTPS; YouTube is safest) and model_3d_link (GLB/GLTF) for furniture, electronics and anything where dimensionality matters.
• Trust pages: live, linkable return-policy and seller-info pages with accurate return windows. Agents check these as reliability signals.

Image and visual optimisation for multimodal agents

AI surfaces increasingly “see” your products. Multimodal models and visual-search systems analyse the actual image, not just its URL, so image quality is now an attribute in its own right. Your primary image should be clean and well-lit on a white or neutral background, high resolution, and free of watermarks or promotional text — overlaid “SALE” badges and busy lifestyle backdrops trip both image-quality scoring and the model’s ability to identify the product. Put lifestyle and in-context shots in the additional-images slots, where they add richness without confusing the hero recognition. For categories where dimensionality or texture matters, the 3D model and video links from Tier 3 give multimodal agents far more to match against than a single flat photo, and are still rare enough to be a genuine edge.

Variant modelling: the detail that wins specific queries

Conversational queries are frequently variant-specific — “mahogany desk, 48 inches wide”, “snapback cap in navy”, “mens UK 9”. If your sizes, colours and configurations are collapsed into a single generic product record, the agent cannot confidently match the exact variant the shopper asked for, and you are filtered out of precisely the high-intent queries you most want to win. Model each meaningful variant explicitly: its own title, its own attributes (colour, size, material), its own image and its own availability and price. ACP also allows custom variant attributes beyond the standard colour/size, so you can encode the intent-heavy dimensions specific to your category. At scale this is impossible by hand, which is the practical reason enrichment tooling — and disciplined rules rather than ad-hoc edits — has become standard.

Worked example: titles and descriptions that match intent

Titles and descriptions are parsed as natural language, so write them the way a knowledgeable sales associate would speak — specific, accurate, structured — not as keyword strings. “Running Shoes Men Running Shoe Athletic Running” confuses the model; “Acme Trailblazer X2 – lightweight 280g trail running shoe, mens UK 9” gives it the brand, model, key spec, category and variant in one legible line. Front-load the attributes shoppers actually ask about (weight, material, size, use-case) because those are the tokens an agent matches a conversational query against.

4. The ACP feed in practice

ChatGPT’s feed is merchant-controlled and push-only: you deliver a gzip-compressed file (.jsonl.gz, .csv.gz or .xml.gz; CSV/TSV/XML/JSON all accepted) over encrypted HTTPS to an OpenAI-provided, allow-listed endpoint — there is no passive crawl. Two flags govern visibility: enable_search and enable_checkout. Titles cap at 150 characters, descriptions at 5,000. Here is the shape of a single item (illustrative — adapt; do not paste verbatim):

Illustrative snippet — ACP feed item (JSONL)

Getting onboarded to ACP

If you’re on Shopify or Etsy, the heavy lifting is done for you: Shopify built the integration, so you apply at chatgpt.com/merchants, and once approved you enable the ChatGPT sales channel in your admin (a paid plan, Shopify Payments and US eligibility are the usual requirements); US Etsy sellers are included automatically through Offsite Ads. Everyone else implements the feed, an ACP-compliant checkout endpoint and a payment integration through a verified PSP such as Stripe, with TLS 1.2+ and documented privacy and compliance policies required before production approval. OpenAI validates an initial sample or full submission before you go live. The payment primitive worth understanding is the Stripe Shared Payment Token — scoped to a single merchant and cart total and time-limited, so ChatGPT never sees the buyer’s raw payment credentials and you remain merchant of record. Crucially, results are not ads: placement isn’t for sale, so feed quality, relevance and trust signals are the only levers you have.

5. Google: the native_commerce flag and UCP manifest

Google Merchant Center remains the highest-volume distribution channel for AI shopping, because Gemini draws directly from GMC inventory across Search, Maps and the Gemini app. Two moves matter beyond a healthy feed. First, activate native_commerce = true so eligible SKUs can show a Buy button inside Gemini and AI Mode without the shopper leaving the surface — vendor data puts the conversion uplift around 34% over recommendation-only visibility. Second, UCP compliance expects a machine-readable manifest at your domain’s /.well-known/ path declaring the capabilities you support, which agents discover automatically.

Google’s wider agentic-commerce moves are worth tracking because they all feed off the same catalogue. Alongside UCP and native checkout, Google has expanded catalogue feeds, introduced Direct Offers (making the offer object first-class data rather than a marketing afterthought), and a Business Agent that answers shopper questions from your product data — which is only as good as the conversational attributes and Q&A pairs you supply. The strategic implication is that enriching one Merchant Center catalogue now pays off across an expanding surface area: Search, Maps, Gemini, AI Mode, Universal Cart and the Business Agent all draw on it. That leverage is the strongest argument for treating feed enrichment as core infrastructure spend, not a marketing chore.

Keep GMC’s diagnostics dashboard clean as your fastest feedback loop: it scores data quality on completeness and surfaces price-mismatch and image-quality issues directly, so disapprovals are a to-do list, not background noise. A healthy GMC feed is also your UCP head start — the same data covers Universal Cart across Search, Gemini, YouTube and Gmail as that surface rolls out through 2026, meaning the cleanup you do for Google today is cleanup you never repeat for the surfaces that follow.

6. Align the feed to your on-page schema

The feed and your product-page structured data are two readings of the same truth, and agents cross-check them. If they disagree, you lose. Whatever you assert in the ACP/GMC feed — price, availability, rating, brand — must match the Product, Offer and AggregateRating schema on the PDP. Perplexity and ChatGPT’s browse mode read that on-page schema directly, so a page that is feed-consistent and schema-rich is legible to the surfaces that don’t take a feed at all. The hub article carries a full JSON-LD example; the rule here is simply: one source of truth, rendered identically in feed and page.

It is worth being honest about the limit of feed work, because it sets up everything else in this cluster. A perfect feed makes you eligible and legible — it cannot, on its own, make you preferred. When an agent has ten feed-clean options and must name three, it reaches for the same external evidence that has always driven authority: reviews, editorial round-ups, brand reputation and citable third-party content. So the feed is the price of admission, and earned trust is what wins the seat. That is precisely why this is a link-builder’s problem and not just a merchandiser’s — the disciplines in our link building strategies guide are what convert feed eligibility into actual recommendations.

Multi-market feeds: currency, VAT and regional stock

For UK and European merchants, the feed carries extra failure surface that US-centric guidance ignores — and agents are merciless about it. Region-specific pricing, VAT-inclusive display, currency codes (GBP vs EUR) and per-market availability all have to be correct for the market the shopper is in, or your products are filtered out of that market’s queries. ACP supports geo-targeted pricing and availability within the feed, and GMC handles multi-country feeds, but the discipline is yours: a record that is right for the US and wrong for the UK actively harms you in UK answers. The structural side of serving multiple markets — hreflang, regional architecture and the signals that tell an engine which catalogue serves which shopper — carries straight over from search; our work on international link building and link building for European markets applies directly to agentic feeds.

7. Per-surface specifics that change your build

The shared-field reality means one catalogue covers most of the work, but four surfaces have quirks worth building for deliberately.

The trap to avoid is reading each surface’s onboarding guide in isolation and concluding you need four separate builds. You do not — the required fields overlap heavily, and the differences are mostly at the edges: a handful of extra attributes one surface accepts, a flag another requires, a payment primitive a third mandates. Map those edges once, decide which single integration covers the most surfaces for your platform, and treat the per-surface extras as enrichment rather than parallel projects. The merchants who waste the most effort in 2026 are the ones who treated “ChatGPT”, “Gemini” and “Copilot” as three unrelated channels and rebuilt the same catalogue three times.

Microsoft Copilot — the dual-protocol bridge

Copilot is the only major surface that consumes both protocols: it launched checkout on ACP and added UCP-feed support in April 2026. That makes it a free rider on work you do for either Google or OpenAI — a correct ACP feed reaches ChatGPT and Copilot’s ACP path at once; a correct Merchant Center feed reaches Universal Cart and Copilot’s UCP path. Microsoft’s own figures (vendor-reported) claim materially higher purchase rates for Copilot-assisted journeys, so it is rarely worth a bespoke integration — just make sure your existing feeds are clean.

Perplexity — Merchant Program plus citable pages

Perplexity runs a Merchant Program for structured catalogue submission, but it also reads on-page Product schema directly and weights recency heavily. The practical move is twofold: submit the catalogue where the programme allows, and make sure your PDPs carry fresh, schema-rich, citable data — a visible “last updated” date and a clean Offer block do real work here that they don’t elsewhere.

Alexa for Shopping (formerly Rufus) — no feed at all

Amazon retired the Rufus brand on 13 May 2026 and there is no Alexa-for-Shopping feed to submit. The surface pulls from the existing Amazon product detail page, A+ content, reviews and Q&A. Your levers are listing completeness, image quality, competitive pricing and answered questions — entirely on-Amazon work, disconnected from your open-web feed strategy. Treat it as a separate workstream.

Google Universal Cart — one feed, many entry points

Google’s cross-merchant Universal Cart spans Search, Gemini, YouTube and Gmail, rolling out from US Search and Gemini first through 2026. The operationally important point: a single SKU needs native_commerce: true to carry a Buy button across all of those surfaces at once. That is unusual leverage — one feed flag, many discovery entry points — and it rewards merchants who get their Merchant Center data complete early rather than channel by channel. It also reinforces the core thesis of this article: the work is in the catalogue, not in chasing each surface individually.

8. Five feed mistakes that get you excluded

Each of these quietly removes products from consideration before any ranking happens:

1. Missing GTINs. Auto-exclusion from Gemini native commerce and Performance Max. The single highest-impact gap in most catalogues.
2. Feed/PDP price or stock drift. Reads as catalogue-wide unreliability and is the leading rejection reason across surfaces.
3. Templated, keyword-stuffed titles. Confuse the natural-language parser and lose specific-query matches to better-written competitors.
4. Empty Tier 2 attributes. No weight, finish or usage data means you simply don’t match “lightweight”, “stainless-steel-friendly” or “high-traffic” queries.
5. Poor primary images. Watermarks, promo text or busy backgrounds trip image-quality scoring and multimodal analysis. Keep the hero shot clean; put lifestyle shots in additional images.

9. Measuring whether your feed is actually working

Feed health is one of the few parts of agentic commerce you can measure relatively directly — use that. Three signals tell you if the work is landing:

• Platform diagnostics. GMC’s data-quality score and disapproval list, and the ACP feed’s validation/acceptance status, are your fastest feedback loop. Drive disapprovals to zero before chasing anything subtler.
• Recommendation sampling. Ask each engine the real buying questions your products should answer, and log whether you appear and how you’re described. A wrong spec in the answer usually traces straight back to a feed field.
• Agent-attributed orders. ACP/UCP webhooks and your platform’s agentic-channel reporting tell you what sold and via which surface — capture this from day one so you can tie feed changes to revenue.

This is the Evaluated layer of the AISLE framework applied narrowly to data quality. Treat it as a standing dashboard, not a one-off audit — prices, stock and models drift constantly, and a feed that was clean last quarter rarely still is.

Treat feed governance as a standing discipline

The brands that hold AI visibility are not the ones that ran a one-time clean-up; they are the ones who built feed governance into operations. That means a few unglamorous habits. Assign clear ownership — product data fragmented across marketing, merchandising and IT is the root cause of most mismatches, so consolidate to a single accountable source. Put validation in front of every export, not after, so a bad price or missing GTIN is caught before it reaches an agent rather than diagnosed after recommendation share drops. Monitor the platform diagnostics on a cadence and treat disapprovals as incidents, not background noise. And re-audit attribute completeness quarterly, because catalogues grow and new SKUs rarely launch at 95% completion without a rule forcing it. AI feed-management agents are emerging that can run a full spec-compliance audit, group disapprovals by cause and even check on-page structured data against the feed automatically — useful leverage, but they automate the discipline, they don’t replace the ownership.

The pay-off for getting governance right compounds. Every surface that comes online — and the surface area expanded from one experiment to five-plus live checkouts in twelve months — reads the same clean catalogue. The merchant who governs data well onboards each new agent in an afternoon; the one who doesn’t re-litigates the same mismatches on every new platform.

10. Composite case study: a homeware retailer’s feed clean-up

The following is an anonymised composite, drawn from patterns across several mid-market retailers in 2026; illustrative, not a single account.

A UK homeware brand with roughly 8,000 SKUs had a “healthy” Google Shopping feed by traditional standards — it passed Merchant Center — yet was almost invisible in ChatGPT and Gemini shopping answers. An audit found the usual three problems. Around a quarter of SKUs lacked GTINs, silently excluding them from native commerce. Tier 2 attributes (material, dimensions, finish, usage scenario) were blank on most of the catalogue, so the products never matched conversational queries like “stain-resistant rug for a busy hallway”. And a pricing-engine lag meant about 7% of SKUs showed a feed price that differed from the PDP — enough to depress trust catalogue-wide.

The fix sequence mattered more than any single change. GTIN backfill and the price-consistency guardrail came first (restoring eligibility and trust), then a focused enrichment push drove the top 50 SKUs to 95% attribute completion, then the ACP feed went live with reviews and popularity signals included. Over roughly three months, sampled recommendation share for their hero products rose from near-zero to appearing in a clear majority of relevant category queries across ChatGPT and Gemini, and agent-attributed orders moved from nothing to a small but compounding line. The lesson echoes the hub: fix accessibility and data integrity first, enrich second, distribute third.

Two details from the project are worth carrying into your own. First, the order of operations was not negotiable: enriching attributes before fixing the price mismatches would have poured effort into products the agents were already discounting as unreliable. The integrity work has to clear before the enrichment work pays. Second, the wins concentrated on hero SKUs precisely because completion was concentrated there — the long-tail products that stayed at 40–50% attribute coverage saw little movement, which validated the 95% rule rather than contradicting it. The honest read is that feed work behaves like compounding maintenance, not a campaign: the retailer’s biggest ongoing risk now is regression, as new SKUs launch incomplete and the pricing engine drifts again, which is exactly why governance — not a second clean-up — is the follow-on investment.

11. Frequently asked questions

Do I need separate feeds for each AI surface?

No. Most required fields are shared. Maintain one clean source of truth and distribute it — a correct ACP feed covers ChatGPT and Copilot’s ACP path; a correct Merchant Center feed covers Gemini, Universal Cart and Copilot’s UCP path. A feed-management tool handles the syndication.

How often should I refresh my feed?

As often as your data changes, up to the surface’s limit (15 minutes for ACP). In practice, push price/stock deltas at high cadence and a full feed nightly — constant full pushes are wasteful and unnecessary.

Are GTINs really mandatory?

Functionally, yes, for AI commerce. Products without GTINs are excluded from Gemini native commerce and Performance Max, and lose the cross-reference that lets agents verify and trust your product. Backfilling them is usually the highest-ROI feed task you can do.

Can I influence Amazon’s Alexa for Shopping with my feed?

No — there is no external feed. Alexa for Shopping draws from your Amazon listing, A+ content, reviews and Q&A. If Amazon matters to you, treat it as a separate on-Amazon optimisation workstream.

Will a perfect feed get me recommended on its own?

It gets you eligible and legible, not preferred. A clean, complete, consistent feed clears the bar to be considered; when an agent has several feed-clean options and must pick a few, it leans on external evidence — reviews, editorial round-ups and brand reputation — to choose. Do the feed work first because nothing else counts without it, then invest in the earned-trust signals that turn eligibility into the actual recommendation.

12. Your Monday-morning action plan

If you want a sequenced 30-day version: week one is integrity — GTIN audit and the price/stock guardrail, because nothing downstream counts until the data is trusted. Week two is the hero enrichment push to 95% on your top 50 SKUs, Tier 2 attributes first. Week three is distribution — register for the feeds that take one and switch on native_commerce. Week four is instrumentation — stand up diagnostics monitoring and a recommendation-sampling routine so month two starts with a baseline. Then the individual actions below become your repeatable checklist.

• Audit GTIN coverage across the catalogue. Every missing GTIN is a product excluded from Gemini native commerce and Performance Max.
• Run a feed/PDP price-and-stock consistency check and ship one automated guardrail that flags any mismatch.
• Drive your top 50 SKUs to 95% attribute completion — prioritise the Tier 2 descriptive fields competitors leave blank.
• Register for the feeds that take one: apply at chatgpt.com/merchants for ACP; confirm GMC health and switch on native_commerce; check Perplexity’s Merchant Program.
• Set a delta-push cadence (15-min for changes, nightly full) rather than blindly pushing everything constantly.
• Reconcile feed values to on-page schema so the surfaces that read pages, not feeds, see the same numbers.
• For Amazon, work the listing: there is no Alexa-for-Shopping feed — the levers are PDP completeness, A+ content, image quality, reviews and answered Q&A.