The Post-Merchant-Center Product Feed: Structured Intent for Agents

TL;DR. A product feed used to be an advertising artifact — a file you submitted to one console to become eligible for Google Shopping, then optimised for coverage: more attributes, longer titles, richer copy. In agentic commerce the feed changes jobs. It becomes a continuously verified truth contract that engines ingest natively and agents check at transaction time. Its value migrates from coverage to fidelity-under-verification.

A feed makes three kinds of claim, and they run on different rules. Identity claims (GTIN, brand, canonical title) resolve you into the graph — the join key; get them wrong and you are invisible. State claims (price, availability, delivery) are verified at execution and against your own site; a caught mismatch is a failed transaction plus a reliability penalty that persists into future recommendation rounds. Quality claims (descriptions, superlatives, self-declared ratings) are self-asserted, so the agent discounts them and decides selection from earned corroboration off the feed.

The reflex is exactly inverted: teams pour budget into the quality column the agent discounts, treat identity as a checkbox, and under-invest in the state column that both gates the sale and carries the only durable penalty. Two instruments below — the Feed Claim Ledger and the Verification-Gap Test — sort your feed spend by which column actually pays.

1. The feed changed jobs

For fifteen years, a product feed had one reader and one purpose. You built it, submitted it to Google Merchant Center, and it made your catalogue eligible for Shopping ads and free listings. The feed was matched against keyword-ish queries, rendered in a visual grid, and its failure mode was a disapproval — a flag you cleared and moved on. Everything the industry learned about feeds it learned in that world: enrich attributes, lengthen titles, fill every optional field, because completeness bought you impression share. That instinct is now a liability, because the reader changed.

In 2026 the feed acquired new consumers with their own native specifications. OpenAI’s Agentic Commerce Protocol (ACP, the open standard connecting merchants to ChatGPT’s shopping surface) ingests a merchant-pushed feed as the single source of truth for titles, prices, stock, media and logistics. Product data are not crawled; the merchant pushes a structured file (CSV, TSV, XML or JSON) to a secure endpoint, refreshed as often as every fifteen minutes, and OpenAI states plainly that ChatGPT’s shopping results are not ads and not influenced by paid placement — they rely entirely on feed quality. Google’s Universal Commerce Protocol (UCP, the coalition standard for agentic discovery-to-checkout), launched at NRF in January 2026 with Shopify, Walmart, Target and twenty-plus partners, extends Merchant Center from a catalogue console into a full conversational and checkout layer across AI Mode, Gemini and, since May 2026, YouTube.

This is why the honest name for the artifact is a post-Merchant-Center feed. Merchant Center has not disappeared — a clean feed there still gets you discovered in AI Mode, and OpenAI even falls back to a Google-compatible feed profile when it cannot parse a native one. But the mental model built around it has expired. The feed is no longer one file for one Google surface governed by one disapproval queue. It is cross-engine transaction infrastructure, read by AI browsers and shopping agents that do not render your grid, do not read your marketing the way a shopper does, and increasingly try to buy against your data rather than merely display it. Merkle’s testing put it bluntly: when they asked agents to recommend a television, all the research and pricing happened inside the model and the brand websites never loaded at all.

The specific danger is that the old reflex is not merely unhelpful now — it is pointed at the wrong target. A decade of Performance Max and Shopping optimisation trained teams to treat the feed as an input to an auction: maximise attribute coverage, lengthen titles, fill every optional field, because completeness bought matches and impression share. Agents do not run that auction. They run a filter over facts, so the same energy poured into longer copy produces almost nothing, while the fields an agent actually gates on — a resolvable identity, a price that is true at checkout — sit neglected because the ad console never penalised you for getting them slightly stale. Google’s own agentic guidance even introduces a machine-readable capabilities layer alongside the feed — a manifest an agent can query to learn how to transact with you — which only underlines the shift: the feed is being read by something that wants to act, not something that wants to display.

Is Google Merchant Center still relevant in 2027?

Yes, but as a layer rather than the destination. Merchant Center remains the catalogue data layer — a clean feed there gets you discovered in AI Mode, and it doubles as a compatibility profile other engines can fall back on — while UCP sits on top as the conversational and checkout layer. What has ended is the era when Merchant Center was the single console for a single Google surface. The feed now flows to several engines with their own native specs and their own reliability bookkeeping, so “my Merchant Center is healthy” is a necessary check, no longer a sufficient one.

What is an agentic product feed?

An agentic product feed is a machine-readable catalogue an AI agent reads to decide whether it can find, trust and transact against your products without a human ever seeing your storefront. It differs from a Shopping feed in what it is optimised for: not visual impression share against a query, but verifiable, current, resolvable claims an agent can act on at machine speed. The same file that once fed an ad auction now feeds a purchase decision made programmatically, upstream of the click.

2. The three claims a feed makes

Every field in a feed is a claim, but not all claims are the same kind of claim. Reaching an agent, being described accurately by it, and being believed by it are separate problems; this article is narrower still — whether the transaction the agent tries to execute against your data actually completes, and what it costs you when it does not. To reason about that, split the feed into three classes by who can check the claim, and how. That single question — verifiability — is what makes the classes behave so differently.

Identity: the claims that resolve you

Identity claims say what the product is: the GTIN (Global Trade Item Number, the barcode-level product identifier), brand, MPN (manufacturer part number), canonical title and category. An agent verifies these against external references — GS1’s registry, other retailers carrying the same item, the knowledge graph — so it does not have to trust you to check them. They barely change. And they are the join key: if your GTIN is missing, duplicated or wrong, the agent cannot resolve your listing to a real-world product, and an item it cannot resolve is an item it can neither recommend nor buy. Identity is a gate, and the gate is binary.

State: the claims that are checked at execution

State claims say what is true right now: price, availability, delivery window, shipping cost, returns window. They are the most volatile fields in the feed and the most consequential, because the agent verifies them at the worst possible moment for you — at the point of transaction, and cross-checked against your own schema and product page. This is the column where a mismatch is not cosmetic. It is a broken promise, caught in the act.

Quality: the claims you assert about yourself

Quality claims are the ones only you make: the prose description, the superlatives (“the best pan you will ever own”), self-declared ratings, marketing copy. Nothing external corroborates them from the feed alone, so a rational reader that cannot phone your references treats them as a sales pitch — informative about what you want to be true, not evidence that it is. This is the column the old playbook told you to maximise, and it is the column the agent trusts least.

The reason this taxonomy earns its keep is that it predicts behaviour the old feed categories cannot. Sort fields by “required versus optional” and price and description land in the same bucket. Sort them by verifiability and they fly apart: price is a promise checked at execution, description is a pitch checked against nothing — two fields with identical formatting rules and opposite strategic weight. Any framework that groups the feed by data type, the way a validation tool does, will keep reporting both as “complete” while one quietly fails transactions and the other is quietly ignored. The classes below track what the agent does with a claim, not what the schema demands of it.

3. The Feed Claim Ledger

Put the three classes side by side and the investment logic falls out of the table rather than out of habit. The Feed Claim Ledger sorts every feed field by how the agent handles it — and the dividing line, running down the verifiable-how column, is who can confirm the claim without taking your word for it.

Read the table by its columns, not its rows. The volatility column tells you where the recurring work is (state). The mismatch column tells you where the durable penalty is (state again, with identity as the on/off switch). And the verification column tells you the uncomfortable part: the one class you fully control the wording of — quality — is the one class the agent will not take on trust.

Key takeaway. A feed field is worth what an agent can verify about it. Identity is a binary gate, state is a penalised promise, quality is a discounted pitch. The reflex from the Shopping era — pour effort into the copy you write about yourself — funds the only column with no verification behind it.

4. State fidelity: the promise the agent checks

In the Shopping world a stale price was an accuracy problem with a bounded cost: Google flagged the product, you fixed the feed, the listing came back. Nobody had acted on the wrong number in the meantime. Agentic commerce removes that buffer. The agent does not look at your price and pause; it parses intent into constraints, queries your state, and proceeds toward a purchase. When the number it committed to at recommendation does not survive to checkout, the transaction fails in the buyer’s hands — the “hallucination gap” that erodes user trust and merchant revenue at the same instant.

The static feed is structurally unable to keep that promise. A nightly CSV, the industry standard for a decade, leaves you wrong all day in any category with promotions or volatile stock; Google’s Shopping Graph now refreshes on the order of two billion listings an hour, and AI Mode prefers live data. The operational bar in 2026 is a lag window measured in minutes, not hours — Content API push updates, webhook feeds, or fetches on a fifteen-minute floor, with a live product-query endpoint for time-sensitive SKUs so the agent can bypass the cached file entirely. State fidelity stopped being feed hygiene and became a technical, data-pipeline problem.

What happens when an agent finds my feed price is wrong?

Three things, in ascending order of damage. The transaction fails, so you lose that sale. The engine records the failure against a merchant reliability score, and merchants with frequent out-of-stock or price mismatches get ranked lower in future recommendation rounds — even after the item is restocked or the price corrected. And, as of 2026, the mismatch is evidence: New York’s Department of Financial Services settled with two agent operators for 4.2 million dollars in May 2026, and the pattern running through the wave of cases is a demand to prove your listing was accurate and that the price the consumer saw matched the price charged. The consistent instruction from those cases is to treat your structured data as legal evidence, not marketing copy, so an agent cannot serve a stale answer in your name.

The penalty is what makes state different from every other column, and it compounds through the one property agents have that shoppers do not: memory. A human who hits a wrong price is annoyed once. An engine that hits a wrong price writes it down and consults that record the next time your category comes up. For example, a homeware retailer running a weekend flash sale on a nightly feed can be correct 95% of the week and still teach two engines that its promotional prices cannot be trusted — precisely on the high-intent, discount-seeking queries it most wants to win. Cross-source consistency raises the bar again: the price in your feed, your schema and your product page must agree at all times, because a disagreement between any two of them signals unreliable data and can remove the product from recommendation eligibility before any customer is even involved. That check costs the agent nothing and does not require trusting you — which is exactly why it is decisive.

Notice the asymmetry that makes this column so dangerous to neglect. Getting state right earns you no special credit — an accurate price is simply the price, the baseline expectation of any counterparty. But getting it wrong is recorded, remembered and priced into every future round, and the recovery is slower than the fall. That is the exact shape of a reputation penalty rather than a display error: the upside of correctness is bounded and the downside of a mismatch is open-ended and compounding. For example, a retailer that resyncs its feed correctly after a bruising promotional weekend does not instantly regain the standing it lost; the engine has already down-weighted its promotional lines and will re-test them cautiously before trusting them again. Feed accuracy, in other words, has become far more like a trust balance you draw down on every mismatch than a checkbox you tick once — which is why the pipeline that keeps it true is worth more than any amount of copy that describes what it once was.

5. Quality copy: the pitch the agent discounts

The counterpart to the penalised promise is the discounted pitch. Marketing prose is written to move a human — tone, aspiration, a superlative or two. An agent reading for constraints extracts none of that as evidence, because a signal only separates a good seller from a bad one if it is costly to fake, and writing “award-winning” costs the honest merchant and the liar exactly the same. So the model reads your description for the facts it can pull out and check, and treats the persuasion around them as noise. Merkle’s finding is the same point from the merchant’s side: marketing copy simply does not contain the technical specifications agents need, and the brands that win are the ones whose data can answer a precise question rather than the ones whose copy sounds best.

Does a richer product description get me recommended by AI agents?

Partly, and the part that works is not the part you think. A description helps to the exact extent it contains verifiable factual attributes — “induction-compatible: yes,” “dishwasher-safe: yes,” “3-litre capacity” — because those let the agent match your product to a constraint and can be checked against the spec. The subjective wrapper — “the finest cookware for serious home cooks” — adds nothing an agent will act on. So the honest answer re-sorts the field: the winning half of a “rich description” is really a bundle of factual state-and-identity claims wearing prose clothing, and it belongs in the verifiable columns. The prose that remains after you extract the facts is the pitch, and the pitch is discounted.

This is where selection actually lives, and why the feed cannot buy it. Among the products an agent has resolved and found reliable, the tiebreaker is not which one described itself most warmly but which one independent sources corroborate — the reviews, editorial coverage, forum and comparison signals the agent reads off your feed. Google’s own framing of the I/O 2026 announcements is that brands will influence agents through the structured data they expose and the off-page signals agents actually read — reviews, forums, editorial — not through on-site merchandising or paid placement at the point of decision. Which means the feed gets you eligible and the independent corroboration you earn gets you chosen. A read of which sources the engines already cite in your category tells you more about winnable selection than any amount of description polishing, and the same logic behind the factors that decide which product an agent recommends governs why a third party vouching for you outranks you vouching for yourself.

The magnitudes are not subtle. Ahrefs’ study of 75,000 brands found brand mentions correlating with AI visibility at 0.664 against 0.218 for raw backlinks, and Muck Rack’s May 2026 analysis of more than 25 million citations found 84% of what models cite is earned and 0.3% is paid. Read those two numbers together and the strategic point is unavoidable: the feed is how you become eligible and reliable, and the overwhelming majority of the signal deciding who gets named is generated by other people talking about you, not by you talking about yourself. No feed field lives in that majority. It is why a merchant with an immaculate catalogue and no independent standing gets found, quoted and then passed over for the resolvable, reliable rival the sources already vouch for.

6. The Verification-Gap Test

The ledger classifies; the Verification-Gap Test routes a budget. Run any feed field — or any line item in a feed-enrichment proposal — through three questions, in order, and it tells you not just what to do but what kind of spend it is.

1. Can an agent check this claim without trusting me? Against an external catalogue, at execution, or across my own sources — yes; only from my own assertion — no. A yes means it is a promise and fidelity is the whole game. A no means it is a pitch, and no amount of polish changes the discount; the leverage is elsewhere.
2. What happens the moment it is caught wrong? Nothing, a disapproval, or a failed transaction plus a reliability penalty that follows me into the next round. Rank every field by its downside, and you will find the money and the danger sitting in the same few state fields, not spread evenly across the feed.
3. Does keeping it true need a one-time fix or a live pipeline? Identity is a one-time fix — resolve it correctly once and it holds. State needs a live pipeline — this is the recurring spend, so buy the sync and the monitoring, not another manual audit. Quality is neither: stop trying to keep a pitch “true” and move the budget to earning corroboration off the feed.

The routing is the output: fix once and stop (identity), build the pipeline (state), stop gaming and earn elsewhere (quality). Most feed-optimisation proposals, scored this way, turn out to spend the majority of their budget in the third bucket while the first two go unaddressed — which is the exact shape of the problem the next section walks through.

Key takeaway. Verifiability, downside and cadence decide where a feed pound belongs. Identity fixes are cheap and final; state is the recurring pipeline spend that carries the penalty; quality is not a feed problem at all — it is an earned-corroboration problem wearing a feed costume.

7. Worked example: Brackenhouse

Brackenhouse is an invented but concrete case: a UK premium cookware retailer, roughly 45 million pounds in revenue, about 9,000 SKUs on Shopify, with a 2027 agentic-readiness budget of 180,000 pounds. Its board arrives pre-committed to a plan, because a decade of Shopping-feed optimisation and a paid-media agency taught everyone that enrichment is the lever: rewrite 9,000 titles past 150 characters, push descriptions past 500, add conversational attributes, load more images. The coverage play. Run the Feed Claim Ledger across the catalogue first, and a different picture appears.

Diagnosis. On identity, a 2024 catalogue migration left roughly 18% of SKUs with missing or duplicated GTINs — those items do not resolve on two of the three engines Brackenhouse tested, so no description, however long, can make them recommendable. On state, the feed is a nightly CSV in a category defined by weekend promotions and bursty stock; the two engines already drop Brackenhouse’s discounted lines mid-transaction because the promo price the agent quoted has expired by checkout. On quality, the descriptions are perfectly adequate — and it is the one column the agency wants the money in.

The naïve plan spends 89% of the budget in the column the agent discounts and nothing in the column that carries the penalty. Watch the two plans diverge over six months.

The naïve plan is not chosen out of foolishness; it is chosen because it photographs well. Rewritten titles and richer descriptions produce a visible before-and-after the board can see in a slide, a dashboard that turns green, a sense of motion. Fixing GTIN resolution and swapping a nightly CSV for a real-time pipeline produces almost nothing you can screenshot — right up until the moment it is the only reason your promotional lines still transact. This is the quiet trap of the whole category: the work that matters to an agent is mostly invisible to the humans approving the budget, and the work that reassures the humans is mostly invisible to the agent. A plan built to survive the next board review and a plan built to survive the next agent’s checkout attempt are rarely the same plan, and only one of them shows up in the reliability score.

• Month 0. Naïve treats thin copy as the problem. Ledger runs the Verification-Gap Test and diagnoses broken identity plus stale state — neither of which is a copy problem.
• Months 1–2. Naïve ships beautiful 500-character descriptions; found-rate barely moves because 18% of SKUs still do not resolve and discounted lines still fail at checkout. Ledger fixes GTINs — resolution jumps — and the state pipeline goes live, so the mid-transaction failures on promotional lines stop.
• Month 3. Naïve’s reliability score on the two engines has degraded further: the enrichment never touched the stale-price failures, which kept happening while the team polished copy. Ledger’s score recovers and its first independent reviews land, so it starts being named, not merely found.
• Month 6. Naïve is “cited but not chosen,” quietly down-ranked on the promo-heavy queries it most wanted, and carrying a reliability penalty that persists even after a belated sync fix. Ledger is resolvable, reliable and increasingly corroborated across engines — and because its state is now live, a flash sale finally helps it (an accurate live promo price is a reason to be picked) instead of breaking it.

The lesson is not that descriptions are worthless; it is that Brackenhouse’s constraint was never in the quality column, and a budget built on the old reflex spent its whole force where the agent was not looking. A degraded reliability score, much like a damaged link-velocity profile, is slower and costlier to recover than to protect in the first place — which is the case for buying the pipeline before you buy the prose, and for funding a genuinely citable asset such as an original data tool other sites reference over another round of copy.

8. The objection the ledger has to survive

The strongest counter is not “descriptions do not matter.” It is this: rich attributes demonstrably lift AI-Mode recommendation eligibility. Google itself, in its NRF 2026 guidance, calls missing attributes “lost recommendation opportunities” and tells merchants that when a shopper asks AI Mode “will this fade in sunlight?” the AI reads structured product data to answer — and recommends a competitor whose feed can answer if yours cannot. On that evidence, quality is not discounted at all; it is decisive. The objection is real and it has Google’s name on it, so concede it fully before answering.

Then bound it three ways. First, look at what the lift actually measures: it decides whether you are in the candidate set for a constraint the agent is filtering on — an eligibility-and-coverage function. And the attributes doing the work are factual and checkable — “fade-resistant: yes,” a capacity, a compatibility flag — which the agent verifies against the spec. Examined closely, the objection does not rescue the quality column; it re-sorts the winning attributes out of it and into the verifiable columns, which is precisely what the ledger says to do. The discount still falls exactly where it fell: on the subjective, superlative, un-checkable prose. Second, completeness saturates and is available to every rival, so it raises the whole field’s floor rather than deciding selection among the eligible — table stakes, not an edge. Third, even a maximally described product does not survive a state failure: the most beautifully attributed out-of-stock item still fails the transaction and still earns the penalty. Taken at its strongest, the objection sharpens the boundary between a factual attribute and a marketing sentence — and leaves the ledger standing.

What would actually break this model? A clean falsifier: if a major engine began treating self-declared quality claims as decision-grade without external verification — ranking primarily on description richness or a merchant-declared quality score, with no consistency or reliability weighting — then the pitch column would revive and the ledger would be wrong. The current evidence runs the other way: reliability scoring, cross-source consistency checks, the industry-wide finding that the overwhelming majority of what a model says about a brand is grounded in third-party sources, and the platforms’ own statements that structured data plus earned signals, not declarations, drive selection. Until that reverses, self-assertion stays discounted. It is worth watching the same way you would watch any shift in how a ranking system weighs a signal — that shift is the thing that rewrites the playbook.

9. What to do Monday

A concrete first week, ordered so the cheap final fixes come before the expensive recurring ones and the earned work starts in parallel:

• Run a resolution audit. Pull the share of SKUs with missing, duplicated or invalid GTINs and MPNs. This is your identity gate; nothing downstream matters for an item that will not resolve. Fix it once.
• Measure your state lag. How long, in the worst case, is your feed wrong on price and availability? If the answer is “until tonight’s sync,” you are failing transactions on every promotion. Move price and stock to a push or sub-15-minute cadence, starting with your highest-velocity and promo SKUs.
• Reconcile the three sources. Check that price and availability agree across feed, on-page schema and the product page itself. Any disagreement is a self-inflicted unreliability signal.
• Stress-test at checkout, not just at search. Have an agent actually attempt to buy your top promotional lines across ChatGPT, AI Mode and one more engine. Where it fails mid-transaction is where your reliability score is bleeding.
• Extract facts out of prose. Move the checkable claims buried in descriptions — materials, dimensions, compatibility — into structured attributes. Leave the superlatives; they are not doing the work you think.
• Fund the earned column separately. Selection among reliable, resolvable products is decided off the feed. Point real budget at independent reviews and editorial in your category — the newsjacking and editorial coverage, the journalist-sourcing platforms that earn genuine third-party corroboration, and, if you sell across borders, the per-market coverage across European and other markets agents read.
• Name an owner. Treat agentic-feed fidelity as a resourced discipline with a clear owner, not an add-on to whoever runs Shopping and paid strategy, and fold it into your wider technical and link-earning programme rather than leaving it stranded in the ads team.

The through-line is simple enough to put on a whiteboard: your feed stopped being copy the engine reads and became a promise the agent checks. Spend on the promises you can be caught breaking, resolve the identity that lets you be found at all, and earn the trust that decides the pick — the same durable logic behind every ranking system, from earning a mention on a high-signal community to the fundamentals of what link building is for, and visible across the 2026 link-building statistics. Whether you approach it as a cleanup discipline or a growth one, the machine is keeping score either way.