Does ChatGPT Recommend Your Products? What Your Product Feed Says About You (2026 Playbook)

You know your Google Shopping impression share. You know your Meta catalog's ROAS to two decimal places. You can tell me, right now, which of your products got disapproved in Merchant Center this morning.

Now answer this one: when a shopper opens ChatGPT and types "what's a good induction-compatible frying pan under $50?" — does your product come up?

Most brands can't answer that. And the reflex response — open ChatGPT, type your product name, screenshot the result — answers a different question entirely. It tells you whether the model has memorised your catalogue. It tells you nothing about whether you get recommended to someone who hasn't heard of you.

Here's the part that should interest you more than the anxiety: the thing that decides the answer is not your website. It's your product feed. Which means it's an asset you already own, already maintain, and can already change this week.

This is a guide to what those surfaces actually read, what they reward, and where most catalogues — including very well-run ones — quietly fall short.

Part 1: The feed became the ranking layer

Three major AI shopping surfaces converged on the same architecture within about a year, and they converged on formats you already produce.

ChatGPT (Agentic Commerce Protocol)

The ACP product feed specification defines what OpenAI ingests to index and display products inside ChatGPT. It accepts TSV, CSV, XML or JSON, and it accepts updates every 15 minutes.

The required fields are unremarkable if you've ever built a Google feed: item_id, title, description, price, availability, url, seller_name, seller_url — plus seller_privacy_policy and seller_tos once checkout is enabled. The recommended fields lean hard into variants: offer_id, color, size, size_system, gender, shipping, delivery_estimate in ISO 8601, and up to three custom variant dimensions.

The spec is explicit that freshness is a matching input, not just hygiene: frequent updates "improve match quality and reduce out-of-stock or price-mismatch scenarios."

Google AI Mode

Across 2026 Google has been adding Merchant Center attributes aimed squarely at conversational surfaces. Eight of them carry most of the weight:

Google's own documentation says several of these are "primarily intended for use in conversational experiences such as AI Mode in Google Search." That sentence is worth reading twice. Google is telling you, inside its attribute reference, which fields feed the answer engine.

Perplexity

The Perplexity Merchant Program is free — no fees, no commissions, no listing costs. Feeds go in as CSV over SFTP following Google Merchant Center specifications. GTIN accuracy is non-negotiable, and the ranking signals Perplexity surfaces are conspicuously feed-shaped: product data completeness, review sentiment, pricing transparency. Products typically start appearing four to six weeks after setup completes. Schema.org Product JSON-LD on the product page is part of the requirement set, not a nice-to-have.

The pattern

The dependency shows up in the data as well as the specs. In Search Engine Land's analysis of AI shopping recommendations, 83% of the products in ChatGPT's shopping carousel matched Google Shopping results, and Google's Shopping Graph now holds around 60 billion product listings.

The answer engines are not crawling their way to your products. They are reading feeds.

Related reading: we covered the transaction side of this shift — how UCP and ACP actually move an order — separately. This article is about the layer before the transaction: whether you get recommended at all.

Part 2: Your ad feed is not automatically an AI feed

This is the part that catches out sophisticated advertisers, and it's worth being precise about why.

A feed built for paid shopping is optimised for matching a keyword and surviving truncation. Titles get front-loaded with the brand and the head term because the rest gets cut off in a shopping unit. Descriptions get treated as a formality because almost nothing reads them. Attributes get filled in as far as the disapproval warnings demand and no further.

A feed read by a conversational surface is optimised for something close to the opposite: satisfying a constraint and answering a question.

Consider what the shopper actually typed:

"a good induction-compatible frying pan under $50, ideally dishwasher safe"

That's three constraints and a price band. To be a candidate, the machine needs to establish that your pan works on induction, survives a dishwasher, and costs under $50. Two of those live in your feed — price certainly, and material maybe. The other one probably lives in a bullet point on your product page, or in a PDF spec sheet, or only in the head of whoever wrote the packaging copy.

Here's the same product expressed both ways:

Ad-shaped

title:       BrandName Granite Frying Pan 24cm Black
description: High quality granite frying pan. Fast shipping.

Conversation-shaped

title:            BrandName Granite Non-Stick Frying Pan 24cm — Induction Compatible
product_highlight: Works on induction, gas, electric and ceramic hobs
product_highlight: Dishwasher safe up to 70°C
product_highlight: PFOA-free granite coating, 5mm forged base
material:          granite / forged aluminium
product_detail:    [Compatibility] Induction: yes · Oven safe: 180°C

Neither version is wrong. The first is a perfectly good ad feed. The second is a findable one. And critically, the second isn't new information — it's information the business already had, sitting in the wrong place.

This is the single most common gap we see: the constraint exists in the product, and in the marketing, but not in the feed.

Part 3: What the surfaces actually reward

Working through the specs, seven things carry disproportionate weight.

1. Constraint language, in structured fields

product_highlight gives you 150 characters that Google explicitly earmarked for conversational surfaces. Spend them on constraints — compatibility, capacity, care, certification, use case — not slogans. "Premium quality since 1985" cannot satisfy a query. "Fits standard 60cm cabinets" can.

product_detail is where the spec sheet belongs, organised into sections. If your category is technical, this field is the difference between being a candidate and being invisible.

2. Answered questions

Google's Q&A attribute takes up to 30 question-answer pairs per product, 1,000 characters each. That is an enormous amount of surface area, and almost nobody is using it.

The questions to fill in are not mysterious. They're the ones your customer service team answers all day, the ones in your on-site reviews, the ones in returns reasons. "Does this work on an induction hob" gets asked to a human somewhere in your organisation weekly. Put the answer in the feed.

3. Identifier integrity

GTIN accuracy is explicitly non-negotiable for Perplexity and shapes matching everywhere else. Identifiers are how a model establishes that the thing it's recommending and the thing you're selling are the same object. Missing or invented GTINs don't just lose you a ranking signal — they break the join.

4. Freshness

ACP accepts updates every 15 minutes and says plainly that frequent updates reduce out-of-stock and price-mismatch scenarios. If your feed refreshes overnight, the model is quoting yesterday's price to today's shopper. In a paid channel a stale price costs you a disapproval. In a conversational surface it costs you the recommendation and, worse, produces a confidently wrong answer with your brand attached to it. Feed update frequency stops being a technical setting here and becomes a visibility input.

5. Variant integrity

item_group_title, the new variant option attribute, and ACP's three custom variant dimensions all point the same direction: these surfaces want to understand a product family, not a wall of near-duplicate rows.

Variant structure has always been the hardest part of feed management, and it gets harder when the consumer is a machine trying to summarise. A parent that fragments into forty unlinked children reads as forty weak products rather than one strong one — and if the in-stock variants are buried among out-of-stock ones, the model may conclude the product isn't available at all. Our complete guide to product variant feeds covers the mechanics; what's changed in 2026 is the cost of getting it wrong.

6. Relationships and documents

Related product supports six relationship types — accessories, substitutes, often-bought-with, required parts and so on. Document link takes PDF manuals and spec sheets.

Both exist because conversational shopping is rarely a single question. "Will this fit my machine?" and "what else do I need?" are follow-up questions, and a catalogue that can answer them structurally gets to stay in the conversation.

7. Social proof and popularity

Perplexity weighs review sentiment. Google's popularity rank takes a 0–100 sales ranking. These are the signals you influence least directly through the feed — but they're also the ones most often simply left empty when they could be populated.

Part 4: How to tell whether you have a problem

You don't need a measurement platform to spot the symptoms. You need to know what they look like.

Symptom 1 — Strong in Shopping, absent in conversation. You have solid impression share for your category's head terms, but when the same need is expressed as a sentence, competitors appear and you don't. This usually means the constraints in the sentence aren't represented anywhere in your data — a product_highlight and product_detail gap.

Symptom 2 — Mentioned but never linked. The model names your brand and doesn't cite your site. Mentions are pleasant. Citations are traffic. This pattern typically points at page-level structured data and feed completeness rather than at the feed's descriptive fields.

Symptom 3 — Right product, wrong facts. The recommendation is yours but the price is stale, or it says out of stock when it isn't. A freshness problem, and a reputational one — a wrong answer in a confident voice does more damage than no answer.

Symptom 4 — Only your hero SKUs exist. Your three famous products come up; the other thirty-nine thousand never do. Usually a taxonomy problem before it's a content problem — if product_type and google_product_category are vague or mis-mapped, no downstream attribute can rescue the long tail. Category mapping is the foundation, not a formality.

Symptom 5 — Your competitor's spec sheet is better than yours. Read the answer, not just your position in it. If the model can state a competitor's dishwasher rating and can't state yours, you've just been handed a specific, cheap, fixable gap.

Part 5: Why occasional spot-checking isn't enough

Opening a chat window and asking a few questions is a genuinely useful first move. It costs an afternoon and it will tell you more than a quarter of speculation. Do it.

But understand its limits before you build a strategy on it:

Answers are stochastic. The same question, the same model, the same day can return different brands. A single reading is an anecdote. Any conclusion drawn from one screenshot is a coin flip presented as a finding.

Answers are personalised and localised. What you see is shaped by your account, your history and your location. Your view is not your customer's view — and it's certainly not your customer's view in another market.

The surface moves. Model versions change without announcement. A drop you notice in October may have nothing to do with anything you did in September. Without a fixed reference point, you can't tell a real decline from a provider's release note.

Coverage is the whole game. Your category isn't one question. It's hundreds of ways of expressing the same need, and your visibility varies enormously across them. Checking the five you happen to think of tells you about those five.

The practical consequence: manual checks are how you decide whether this matters. They're not how you manage it. Once it matters, you need consistent, repeated measurement against a fixed set of questions, tracked against competitors over time — which is exactly the problem Feedance's AI Visibility module is built to solve, using the catalogue we already hold for you.

Part 6: A feed readiness checklist

Everything below is inside your control and none of it requires waiting for anything.

If you'd like this run against your live feed rather than by hand, our free product feed audit tool covers most of it in a few minutes.

Part 7: Where to start

This week — find your worst gap. Take your top-selling category and read ten product records as a machine would. Not the page, the feed row. Can you tell, from that row alone, whether the product satisfies the three constraints your customers ask about most? If not, you've found the work.

This month — fill the constraint fields. product_highlight, product_detail, material and the variant attributes, for one category first. Enrichment at this level is a data exercise, not a rewrite of your catalogue — most of the information already exists somewhere in the business. Rules-based feed enrichment is how you avoid doing it product by product.

This quarter — fix freshness and structure. Move refresh cadence from nightly to intraday, and clean up the variant grouping that has been "good enough" for paid channels for years. These are the two changes with the longest lead time and the widest effect: they improve paid performance and conversational visibility simultaneously.

Note the shape of that list. Every item on it is something a well-run feed operation should be doing anyway — which is the genuinely good news here. Unlike most platform shifts, this one doesn't ask for a new team or a replatform. It asks you to finish the feed.

FAQ

Is AI visibility the same thing as SEO? No. Classic SEO optimises a page to rank in a list of links. AI visibility depends on structured product data being complete enough for a machine to establish that your product satisfies a stated need. The overlap is real — structured data and content quality help both — but the primary asset is different. In AI shopping, the feed does much of the work the landing page used to do.

We already have great feeds. Doesn't that mean we're fine? Not necessarily. Feed quality for advertising and feed quality for conversational surfaces optimise for different things. An ad feed rewards a tight, keyword-dense, truncation-proof title. A conversational surface rewards constraints, specifications and answered questions. Plenty of catalogues with flawless Merchant Center diagnostics are close to invisible in constraint-led queries.

Do I need to submit separate feeds to ChatGPT, Google and Perplexity? In practice you maintain one enriched source and export a compliant variant per destination. Perplexity follows Google Merchant Center specifications over SFTP; ACP accepts several formats with its own required fields. The differences are in field naming and required sets, not in the underlying data — which is why solving this at the source rather than per-channel is the sane approach. That's what channel exports are for.

How long until changes show up? Perplexity's own guidance is that products typically surface four to six weeks after setup completes. Treat anything faster as a bonus and don't judge a change after a week — between ingestion lag and the variance in the answers themselves, short-window readings are unreliable.

Which surface should I prioritise? Start where your category's buyers already are. If you sell anything technical or comparison-heavy, ChatGPT and Perplexity are where constraint-led questions get asked. If you already run Merchant Center well, Google's conversational attributes are the cheapest win available to you — the feed is connected, the fields are simply empty.

Brands that dominate paid shopping in their category often discover they're barely present when the same need is expressed conversationally. That's not a reason to postpone looking. It's the reason to look now, while the surface is still forming and the fixes are still cheap — because in almost every case the fix is a feed change, not a rebuild.

Start with the free product feed audit to see what your catalogue is missing. If you want to know where you actually stand against your competitors in AI answers — consistently, over time, rather than one screenshot at a time — talk to us.