Google disapproves a product when the price in your feed and the price on your landing page disagree. There is no tolerance band: Google's documentation states that prices in the HTML "need to match exactly the prices uploaded in Merchant Center", and says a mismatch may escalate to account suspension. Agentic checkout raises the cost of that gap, because an agent does not merely display your price. It quotes it and tries to transact on it.
This guide is for e-commerce and growth teams who already send a product feed to an AI shopping surface, or are about to. By the end you will have an eight-step pipeline that keeps price, stock, eligibility and policy data consistent across your catalogue, your feed and your checkout, plus a field-level parity table, a refresh cadence matched to how fast your prices actually move, and alert thresholds that fire before an agent sells something you cannot ship at that price.
Why parity is a pipeline problem, not a feed problem
In an agentic setup at least three systems hold a price: your commerce database, the product feed you publish, and the checkout session the agent opens. Google's own UCP guidance tells merchants using scheduled feeds to "send supplemental price and inventory updates to avoid checkout price mismatches", which is an admission that the feed layer drifts. OpenAI's Agentic Commerce Protocol feed spec is blunter: submit current price and availability with every row, and note that sale-window, expiration and availability dates do not schedule anything. They describe state, not future state.
So parity is not a one-time mapping exercise. It is a job that runs, fails loudly, and gets watched.
Step 1: Name one source of truth and one write path
Pick the system that owns price and stock, usually the commerce platform or ERP, and write down which humans and jobs may change a price there. Every other store of that value becomes a derived copy with a timestamp. Teams that skip this step end up with promotion prices set in the merchandising tool, the feed built from a nightly database replica, and nobody able to say which one is right at 14:00 on a Friday.
Step 2: Map the fields that must agree, field by field
Do this against the actual specs, not from memory. The field names below come from the ACP file-upload product spec and Google's UCP Merchant Center prerequisites.
| What must agree | ACP feed field | Google / UCP side | Parity rule |
|---|---|---|---|
| Price | price, formatted 79.99 USD, no thousands separators, uppercase ISO 4217 | price in Merchant Center | Exact match with the price rendered in the landing page HTML. No tolerance. |
| Promotional price | sale_price, above zero, below price, same currency | sale_price | Must be withdrawn when the promotion ends. Dates do not end it for you. |
| Stock | availability: in_stock, out_of_stock, pre_order, backorder, unknown | Merchant Center availability | An unrecognised or empty value rejects the row outright. |
| Checkout eligibility | is_eligible_checkout, with seller_privacy_policy and seller_tos | native_commerce(checkout_eligibility), defaults to FALSE if absent | The native_commerce group is fully replaced on each upload, so send every flag every time. |
| Identity | item_id stable and never reused; group_id must differ from item_id | id must match your Checkout API product ID, or use merchant_item_id, which takes precedence | One ID map, versioned, owned by one team. |
| Returns | accepts_returns, return_deadline_in_days, return_policy URL | Merchant Center return policy stating cost, window and a link | Feed window, policy page and checkout screen must state the same number of days. |
| Shipping | shipping_price, non-negative | Merchant Center shipping settings | Omitting it means unknown, not free. Never leave it blank to imply free shipping. |
Step 3: Set a cadence your price volatility can survive
OpenAI's integration guidance describes an initial sample feed for validation followed by daily snapshots. Daily is a floor, not an answer. Work out your own number:
- Pull 30 days of price changes from the source of truth. Count changed SKUs per day and the largest single change.
- Measure the lag between a price change landing in the source of truth and appearing in the published feed.
- Multiply: changed SKUs per hour multiplied by lag hours is the number of rows that are wrong at any moment.
In Turkey that arithmetic bites harder than in euro or dollar markets. TÜİK put October 2026 consumer price inflation at 2.58% month on month and 29.77% year on year, so a catalogue that reprices to keep margin is touching rows continuously, not seasonally. If you reprice at 09:00 and your snapshot publishes at 03:00, every agent quote for 18 hours carries yesterday's price. The fix is not a faster snapshot. It is a snapshot plus a delta channel for price and availability only, which is exactly what Google recommends via the Merchant API alongside scheduled feeds.
Step 4: Decide what happens when they disagree
Write the policy before the incident. For each mismatch class, choose one of three behaviours: honour the agent's quoted price and absorb the margin, suppress the SKU from agentic surfaces until parity returns, or reject the checkout session. Honouring protects trust and is defensible for small deltas. Suppression is the right default for stock mismatches, since selling an out-of-stock item costs more than missing a sale. Rejection should be rare and reserved for cases where honouring is genuinely uneconomic, because an agent that gets a rejection may simply route the buyer to a competitor and not come back.
A useful shape: honour below a dH1 threshold you set per category, suppress between that and a hard ceiling, reject above it. Put the thresholds in config, not in code.
Step 5: Build the reconciliation job and its alerts
One scheduled job, three comparisons: published feed row against source of truth, published feed row against rendered landing page HTML, and feed row against what your checkout endpoint would quote for the same SKU. Sample rather than sweep if the catalogue is large, but stratify the sample so bestsellers and recently repriced SKUs are always included, since those are where mismatches actually happen.
Alert on rates, not on single rows: mismatch rate above your agreed threshold for two consecutive runs, any mismatch on a top-100 SKU, any row rejected for an invalid availability value, and a feed updated_at timestamp older than your cadence allows. Note that neither platform publishes a parity SLA. Google's help page says disapproved products are re-crawled "over the next few hours or days" and a requested site review can take up to 12 hours, so your internal monitoring is the only early warning you control.
Step 6: Flag what must never reach an agentic checkout
Google's UCP documentation lists product types that must have checkout_eligibility set to FALSE or left blank: subscriptions and installment plans, personalised, refurbished, used, final-sale and pre-order items, anything with bundled services, special shipping, gifting or in-store activation requirements, age-restricted and prohibited content, services, rentals, virtual items, and goods that require software installation. Derive this flag from product attributes you already hold, rather than maintaining a hand-curated exclusion list that goes stale.
Step 7: Run the pre-launch parity check
- Every required ACP field present on every row:
item_id,title,description,url,brand,seller_name,image_url,availability,price. - Price strings parse as
amount CURRENCYwith a decimal point and no separators. - Every
sale_priceis below itspriceand in the same currency. gtinvalues are 8, 12, 13 or 14 digits with a valid check digit.- Variant rows carry
group_id,listing_has_variationsandvariant_dict, withgroup_iddifferent fromitem_id. - Checkout-eligible rows carry
seller_privacy_policyandseller_tosURLs that resolve. - Return window in the feed equals the window on the policy page.
- Rendered landing page HTML price matches the feed for a stratified sample, tested with JavaScript disabled, since prices injected after load can trigger the mismatch error.
Step 8: Confirm your market is actually supported before you promise a date
This is where teams outside the United States lose a quarter. The ACP file-upload spec describes standard uploads as using target_countries: ["US"] and store_country: US, while the feeds API accepts a target_country ISO code and several attributes are marked as requiring feed-specific setup. Those two facts do not contradict each other, but they do mean a Turkish or EU merchant should confirm market support with the platform in writing before committing a roadmap. On the Google side, the Merchant Center UCP integration hub is labelled Pilot and MCP transport Beta, on a page last updated 29 October 2026 that references implementation docs versioned 2026-04-08.
Where this breaks down
This pipeline keeps three copies of a value consistent. It does not make the value correct. It will not catch a margin error, a mispriced bundle, or a currency conversion applied twice, because all three systems will agree on the wrong number.
It also stops at the catalogue boundary. Tax and VAT display logic, the checkout session's own pricing and tax calculation, reference-price duties under EU and Turkish price-indication rules, and marketplace offers where marketplace_seller requires separate feed setup are all out of scope here and need their own treatment. And because both specs are moving, any field name in this guide is worth re-reading against the source before you ship.
One widely repeated figure deserves a caveat. Adobe Analytics reported AI-referred traffic to US retail sites growing 138% year on year in May 2026, converting 54% better than non-AI traffic. That is vendor panel data, and the published reporting does not define how an AI referral is identified or how the sites are selected, so treat it as directional rather than as a planning input. [INTERNAL DATA NEEDED: Switas client-side figures on AI-assistant referral share and its conversion rate versus other channels, to replace the Adobe panel figure with first-party evidence.] [INTERNAL DATA NEEDED: observed feed-to-landing-page mismatch rates and typical repricing lag across Turkish e-commerce catalogues we have audited, to give readers a benchmark for step 3.]
Frequently asked questions
Is a daily feed snapshot enough?
Only if your prices change less often than daily and your stock never sells out mid-day. Otherwise pair the snapshot with a delta channel carrying price and availability, which is the pattern Google recommends alongside scheduled feeds.
What tolerance does Google allow on price?
None. The documentation calls for an exact match including currency, and where a landing page shows multiple currencies, the prominently displayed one must match the feed.
Can I use sale-price dates to schedule a promotion?
In the ACP feed, no. The spec says sale-window, expiration and availability dates do not schedule changes, so your own job has to add and remove the sale price.
What happens if availability is missing?
The row is rejected. Omitted, empty or unrecognised availability values are not treated as unknown by default, so send the explicit unknown value if that is genuinely the state.
Why did my checkout eligibility flag disappear after an upload?
Because the native_commerce attribute group is fully replaced on each upload. Partial submissions silently clear the flags you left out, and a missing checkout eligibility value defaults to FALSE.
Should I block agentic surfaces until parity is solid?
Suppressing specific SKUs is better than blocking the surface. Suppression is reversible per row and keeps your eligible catalogue earning while you fix the rest.
Who should own this pipeline?
Whoever owns the feed in production, with a named second owner in merchandising for the threshold config. Parity failures are usually process failures between those two functions rather than engineering bugs.
What to do next
Run steps 2, 3 and 7 against your own catalogue this week: the field map, the volatility arithmetic, and the stratified landing-page comparison with JavaScript disabled. That is a day of work and it tells you whether you have a parity problem or a pricing problem. If you would rather have the mapping, the reconciliation job and the alert thresholds built and monitored for you, that is the work we do, and Feedance is where we run product feed scoring and attribute optimisation at catalogue scale.
Sources
- OpenAI, Agentic Commerce Protocol: Products (file upload spec)
- OpenAI, Agentic Commerce Protocol: Key concepts
- OpenAI, Agentic Commerce Protocol: Feeds API
- Google, Universal Commerce Protocol: Prepare your Merchant Center account (updated 29 October 2026)
- Universal Commerce Protocol, specification overview
- Google Merchant Center Help, Mismatched product price
- Google Merchant Center Help, Inaccurate price due to inconsistency between feed and landing page
- Forbes Türkiye, TÜİK October 2026 consumer price index (3 November 2026)
- Digital Commerce 360, Adobe Analytics on AI-referred retail traffic (17 June 2026)







