✓TOKENS USED1,042,000,000
← Back to posts

The AI Agent Playbook for Google Merchant Center

June 10, 2026

Most ecommerce marketers treat Google Merchant Center like a closet they only open when something smells weird.

You know the one.

Product disapproved. Feed issue. Missing attribute. Shipping error. Landing page problem. Some random warning from Google that looks important but might also be complete noise. You click around for 20 minutes, fix the one thing that is on fire, close the tab, and go back to Google Ads because that is where the money looks like it is happening.

I get it.

Google Ads feels like the action. Budgets, ROAS, campaigns, asset groups, bidding, creative. Merchant Center feels like plumbing.

But for Shopping and PMax, Merchant Center is the raw material.

If the feed is messy, the campaign is messy. If the catalog has old products, bad titles, missing attributes, dead landing pages, duplicate product families, weird country targeting, and no useful labels, Google is making decisions on bad inputs.

And there is a lot of money running through those inputs. Tinuiti found Google Shopping spend increased 14% year over year in Q3 2025. Among retailers running both campaign types, Performance Max accounted for 68% of Shopping spend and 67% of Shopping-attributed sales.

Every one of those product ads starts with data from Merchant Center. The feed helps decide what can advertise, what Google shows, and which searches your products can match.

This is where Claude Code and Codex are stupidly useful.

Give the agent read/write access and context on the store. It can audit 10,000 products, find what is costing you money, then fix titles, attributes, custom labels, stale products, and broken listings through the API or the upstream feed.

You set the rules and approval thresholds. The agent manages the catalog.

Not another dashboard. An operating layer that can find and fix the mess.

Who this is for

This is for three people.

First, the growth lead at an ecommerce company who owns Google Shopping but does not have time to become the full-time feed janitor. You want to know what is actually hurting performance, what can wait, and what to tell the team to fix.

Second, the marketer who wants to look sharp to their boss. Not "I found 812 feed issues." Anyone can screenshot that. The useful version is "I separated the 812 issues into 3 revenue blockers, 40 cleanup items, and 769 warnings we can ignore for now."

Third, agency owners. If you manage five, ten, twenty ecommerce accounts, Merchant Center is where your margin quietly disappears. Every new client has a messy feed. Every client thinks their catalog is fine. Every audit turns into a weird archaeology project. Agents make that work scalable.

If you are doing this for one store, the agent saves you hours.

If you are doing this across many stores, the agent becomes the operating model.

The idea

Do not start by asking AI to "optimize my Merchant Center."

That prompt is too broad. It will give you the same boring checklist every feed audit gives you:

  • improve titles
  • fix disapprovals
  • add missing images
  • add GTINs
  • clean up product types
  • resolve shipping issues

Technically correct. Also not that useful.

The better frame is:

What is stopping Google Shopping from spending money on the right products?

That question forces the agent to think like an operator.

Not every feed issue matters. A missing optional image on a product that never gets clicks is not the same as a landing page error on a product family you are actively trying to scale. A shipping warning in a country you do not sell to is not the same as a US disapproval on a best seller. A title cleanup project across 8,000 products might be worth doing, but probably not before the 12 dead URLs blocking your main products.

There are a lot of things you can do.

What is the thing you should be doing?

That is the job of the agent.

The setup in plain English

You need five pieces.

1. A Google Cloud project

This is where the API access starts. Create a Google Cloud project for the store or for your agency. If you are an agency, I would not dump every client into one sloppy project. Keep it clean.

Turn on the Merchant API.

Google has been moving from the older Content API for Shopping toward the newer Merchant API. The exact API surface will keep changing because Google enjoys making everyone relearn the same thing with new names. Fun little hobby of theirs.

For this workflow, the principle matters more than the API branding:

Your agent needs a supported way to read products, product statuses, account settings, and reports from Merchant Center.

2. A service account

A service account is an application identity. It gets its own email address.

It is not a human login. It does not open the Merchant Center UI. It exists so your script, Claude Code workflow, Codex task, or scheduled audit can authenticate without you clicking OAuth screens every morning.

Create the service account in Google Cloud.

I use service-account impersonation, so there is no long-lived JSON key sitting on my laptop.

My Google account is allowed to mint a short-lived API token for the service account through Google Cloud IAM. It is cleaner, safer, and easier to revoke.

3. Merchant Center user access

This is the part people miss.

Creating the service account in Google Cloud is not enough. You also have to add the service account email as a user inside Merchant Center.

Go to Merchant Center, People and access, add the service account email, and give it permission to read and update the catalog.

The connection can support writes from day one. The workflow should still begin with an audit so you can verify the data and decide which changes the agent can make automatically.

Read first. Prove the diagnosis. Then let the agent apply the repeatable fixes while a human approves anything risky.

Same rule I use for ad accounts.

4. A local working folder

Create a folder for the store.

Something boring:

gmc-audits/
  store-name/
    README.md
    context.md
    credentials/
    scripts/
    exports/
    reports/
    staged-changes.md

The point is not the exact structure. The point is that the agent has a place to work where context can accumulate.

context.md should explain the business in normal language:

  • what the store sells
  • primary country
  • core product categories
  • margin notes if you have them
  • what products matter most
  • what products should not be pushed
  • where revenue truth lives, like Shopify, Triple Whale, GA4, Looker, or something else
  • who approves changes

Without this file, the agent will treat every product like it matters equally.

They do not.

5. Claude Code or Codex

You can do this in either.

Claude Code is great when you are sitting with the work and iterating. Codex is great when you want it to run through a repo, update files, create scripts, and leave behind artifacts you can review.

For this specific workflow, I like using them the same way I would use a smart junior operator:

  • read the docs
  • inspect the account
  • group the mess
  • make a recommendation
  • write it down clearly
  • do not change anything unless I explicitly approve it

That last line should be in the project instructions.

The first audit I would run

Do not start with the whole catalog.

Start with the money blockers.

Ask:

Audit Merchant Center for primary-country product issues.
Separate revenue blockers from cleanup work.
Group products by product family, URL, item group ID, status, country, and issue type.
Do not recommend feed changes yet. Show me what is actually happening.

The reason for "primary-country" is simple.

A store that sells only in the US might show hundreds of disapprovals in other countries because shipping is not configured there. That looks scary in Merchant Center. It may not matter at all.

Your first job is to separate:

  • products blocked in the country you sell to
  • products blocked in countries you do not care about
  • warnings that affect eligibility
  • warnings that are annoying but not urgent

This is where an agent helps because it can keep the whole picture in its head.

A human sees 900 issues and either panics or ignores the whole thing.

An agent should say:

"Most of this is international shipping noise. The actual urgent problem is a small set of primary-country products with landing page errors and missing required apparel attributes."

That is useful.

The second audit: duplicate product families

This is where Merchant Center gets weird.

Big Shopify catalogs collect history. Old product pages, seasonal drops, bundles, renamed products, variant changes, feed app quirks, products that were unpublished but still exist somewhere.

Over time you get a catalog that looks fine from 30,000 feet and completely cursed when you zoom in.

The prompt:

Find products with duplicate or overly generic titles.
Group them by source product ID, item group ID, URL handle, price range, country, and approval status.
Tell me if these are true variants of one product or separate product families sharing the same title.

This is one of the highest-value audits because bad titles make reporting muddy.

If 40 products all share some generic title like "Men's Shirt Online," you cannot tell what is actually working. Your team cannot talk about it clearly. Your agency cannot segment it cleanly. Your boss gets a report that sounds specific but is actually mush.

The agent should untangle it:

  • these are true size/color variants of one product
  • these are separate product families with lazy titles
  • these are old products still syncing
  • these are active products that need clearer naming
  • these are products that should be excluded from the feed

Do not merge everything just because the titles look similar.

That is how you create product mismatch problems.

The agent should check whether the "replacement" product already exists as its own approved listing before recommending a redirect or repoint.

That little check saves you from doing something dumb.

The third audit: landing pages

Landing page errors are boring until they are not.

If a product URL returns a 404, Google cannot reliably send traffic there. Obvious.

The less obvious issue is what to do next.

A normal feed cleanup would say "update the URL."

Sometimes that is right. Sometimes the old product should be removed because the new product already exists in the feed.

Prompt:

Find all products with landing page errors.
Check the current URL status.
If there is a likely replacement page, check whether that replacement already exists in Merchant Center.
Recommend one of: fix URL, redirect source page, exclude old product, remove stale product, or investigate.

This is the kind of decision tree you want.

Not "fix URL."

That is not a strategy. That is a button press.

The fourth audit: apparel attributes

If you sell apparel, this is where you get free matching quality.

Size, color, gender, age group, material, pattern. Boring fields. Important fields.

The agent should not just say "1,200 products missing color."

Great. Thank you. I am healed.

The useful version:

Build a missing apparel attribute matrix.
Prioritize products by primary country, approval status, clicks, spend, revenue importance, and product role.
Show the top 50 products to fix first.

This turns a giant feed project into a sane work queue.

If you have a team, this is what you hand them.

If you are an agency, this is what makes you look like you know what you are doing instead of dumping a scary CSV on the client and saying "feed needs work."

The fifth audit: product types and labels

Merchant Center cleanup only gets interesting when it changes how you manage Shopping.

Product types and custom labels are the bridge.

Most stores underuse them. They either leave them blank or use whatever the feed app spits out.

That is a waste.

Custom labels should reflect how you want to manage spend:

  • hero
  • core
  • seasonal
  • clearance
  • high_margin
  • low_margin
  • bundle
  • new_arrival
  • do_not_push
  • needs_review

Do not get cute. Use labels a normal person on your team will understand.

Prompt:

Propose a custom label system for Shopping management.
Use product role, margin, seasonality, priority, and action.
Keep it maintainable from Shopify or the feed app.
Do not invent labels that require manual upkeep no one will do.

That last sentence matters.

A clever label system that no one maintains is just future clutter with better branding.

What the final report should look like

The report should not be a giant essay.

I know, hilarious coming from the guy currently writing a giant essay.

But the audit output itself should be tight.

For a growth lead or agency owner, I want five sections:

1. What matters this week

The 3 to 10 issues that could affect spend, revenue, or reporting right now.

Example:

  • primary-country landing page errors on active products
  • high-click products with limited eligibility
  • duplicate titles splitting product family clarity
  • missing required attributes on products currently receiving traffic

2. What is noise

This section is underrated.

Your boss or client sees 800 Merchant Center warnings and thinks everything is broken. You need to explain what is not urgent.

Example:

  • unsupported-country shipping disapprovals
  • missing optional fields on products with no traffic
  • image suggestions on low-priority products

This is how you look calm and in control.

3. What to fix first

Ranked queue.

Not vibes. Not "best practices."

Priority, issue, product group, reason, owner, source of truth.

4. What to change in the feed structure

Titles, product types, custom labels, missing attributes, stale products.

This is usually a Shopify or feed app job, not a Merchant Center UI job.

Fix the source of truth.

5. What to monitor daily

The ongoing system:

  • new disapprovals in primary country
  • landing page errors
  • price/availability mismatches
  • products with spend but bad revenue
  • approved products that should be getting more impressions
  • feed changes that could affect active campaigns

That is the artifact.

If you can produce that every week across multiple stores, you have a real agency advantage.

How this helps you manage multiple stores

One store is a project.

Multiple stores is an operating system problem.

If you are managing ten Merchant Centers, you do not want ten different one-off audits living in ten random Google Docs. You want the same workflow every time:

  1. Connect store.
  2. Pull product data.
  3. Pull product status.
  4. Pull recent Shopping/PMax performance.
  5. Group issues.
  6. Rank by business impact.
  7. Stage recommendations.
  8. Review with client or team.
  9. Fix source of truth.
  10. Monitor daily.

That is where Claude Code and Codex are powerful.

They make the work repeatable.

The goal is not to become the person who manually fixes every feed. That is how you build yourself a little services prison. Ask me how I know.

The goal is to build the process once, then run it across every store.

Same prompts. Same report structure. Same staged changes format. Same human approval rule.

Different catalog.

That is leverage.

The source-of-truth rule

Write access is the point. The agent should be able to fix repeatable catalog problems, not just write another audit.

Merchant Center feeds usually have a source of truth upstream:

  • Shopify
  • feed app
  • custom feed
  • ERP
  • PIM
  • Google Sheets
  • API integration

If Shopify or a feed app owns the field, the agent should fix it there. If the API owns the product data, it can update Merchant Center directly. Otherwise the next sync will overwrite the work.

For risky or ambiguous changes, the agent should stage them for approval:

priority
product group
issue
evidence
recommended action
source of truth
risk
owner
approved?
completed?

Then a human approves.

Then the change happens in the right place.

This is how you avoid the AI doing something technically correct and operationally stupid.

The tools I would use

If I were setting this up for a store today, I would use:

  • Claude Code or Codex for the agent workflow
  • Google Cloud service account for Merchant Center access
  • 1Password for secrets
  • Python scripts for repeatable pulls and joins
  • Google Sheets for staged changes if a human team needs to approve
  • GitHub for the prompts, scripts, and client context files
  • Shopify or the feed app as the source-of-truth fix layer

You do not need a big SaaS platform to start.

Actually, starting with a big platform can make you lazy. You click the shiny dashboard, export the automated audit, and still do not understand what is happening.

Start simple.

Make the agent pull the data. Make it explain the mess. Make it prove which issues matter. Then decide if you need more tooling.

The beginner version

If you are doing this for the first time, do this in order.

Step 1: Connect read/write access

Get the service account connected and prove the agent can list products and product statuses.

Your first win is:

Show me all product issues in my primary country, grouped by issue type.

Step 2: Create the business context file

Write one page explaining the store.

The agent needs to know what matters.

If the store sells 8,000 products but 80 products drive the business, say that. If certain categories are low margin, say that. If the team only sells in one country, say that. If PMax is the main campaign type, say that.

The agent cannot read your mind. Annoying, I know. Would be convenient.

Step 3: Run a primary-country disapproval audit

Ignore the global mess at first.

Find what blocks the actual market.

Step 4: Run the duplicate title/product family audit

This is where you find the catalog weirdness.

Old products. Lazy titles. Multiple product families with the same name. Variants that should be grouped but are not. Products that should be excluded.

Step 5: Run the missing attribute audit

Prioritize by business impact.

Do not fix 2,000 products alphabetically. That is how souls leave bodies.

Fix the products that matter first.

Step 6: Build a label plan

Use labels to make Shopping easier to manage.

Think in campaign terms:

  • what should get budget?
  • what should be separated?
  • what should be watched?
  • what should be excluded?
  • what is seasonal?
  • what is core?

Step 7: Apply the safe fixes

Let the agent automatically handle changes you can validate with clear rules. Stage the judgment calls in a sheet or markdown table for approval.

Always update the real source of truth so the next feed sync does not undo the work.

Step 8: Turn it into a weekly operating rhythm

Once the first audit is done, the ongoing job is small:

  • what changed?
  • what broke?
  • what matters?
  • what should we fix next?

That is what your boss or client wants.

Not a 40-page feed audit.

They want confidence.

The agency angle

If I were starting an ads agency today, I would build the Merchant Center audit into onboarding.

Before touching campaign structure, I would run:

  • feed health audit
  • product family audit
  • primary-country disapproval audit
  • custom label audit
  • source-of-truth audit
  • Shopping/PMax performance join

Then I would show the client:

"Before we scale anything, we need to know whether the catalog is clean enough to scale."

That is a strong position.

Most agencies want to jump into campaigns because that is what the client thinks they bought.

But if the feed is broken, your campaign work is standing on mud.

The agency that can explain this clearly looks different. More senior. More in control. Less like a button-clicking vendor.

That is the move.

What not to do

Do not create a long-lived service account key if impersonation will do the job.

Do not let the agent auto-apply changes before you trust the workflow.

Do not treat every Merchant Center warning like it has the same priority.

Do not fix things in Merchant Center if Shopify or the feed app will overwrite them tomorrow.

Do not use client names or product details in public case studies unless you have permission. Use the shape of the problem. That is enough.

Do not let the AI think for you.

Use it to get through the mess faster so you can think better.

That distinction matters.

The bottom line

Google Merchant Center is not the sexy part of ecommerce growth.

Good. That is why there is opportunity there.

The sexy parts get crowded. Campaign launches, creative tests, AI ad generators, PMax hacks, whatever the LinkedIn crowd is yelling about this week.

The gritty catalog work is where a lot of the real leverage hides.

If you can take a store with thousands of products, connect an agent to Merchant Center, separate signal from noise, fix the source-of-truth problems, and turn the catalog into something Google Shopping can actually understand, you are not just "doing feed cleanup."

You are building the operating layer for paid ecommerce.

That is useful for one store.

It is a weapon across ten.

Connect read/write access. Build the context file. Run the first audit. Automate the safe fixes. Keep a human in the loop for the risky ones.

Then repeat it until managing a giant catalog feels boring.

That is when you know the system is working.

Sources checked while drafting: