Implementation
Audit Shopify Products for Google AI Mode Shopping
Confirm Merchant Center sync, repair feed mismatches, and test conversational queries so Shopify products can compete in Google AI Mode.

A product can look finished on your Shopify storefront and still be invisible in Google AI Mode. The theme is not the catalog Google reads. AI Mode draws product answers from Google Merchant Center and the Shopping Graph. If the feed is thin, stale, or missing a variant the shopper named, the conversation recommends someone else.
That is the failure this audit is for. A merchandiser sees the backpack on the site, sees it in ChatGPT, and cannot find it when they ask Google AI Mode for a 26-litre clamshell bag that fits a 16-inch laptop. The instinct is to rewrite the theme or “do Google SEO.” Neither move touches the record Google actually uses.
This is a different job from ChatGPT Shopping on Shopify Catalog. Catalog is the ChatGPT path. Google AI Mode and Gemini take the Google & YouTube sales channel into Merchant Center. Treat the two audits as siblings, not duplicates. If you skip that split, you will repair the wrong file and call the channel broken.
The work has five stages, in this order, because each stage only makes sense after the previous one is true:
- Confirm the Google data path is actually on.
- Check that Merchant Center and the live product describe the same item.
- Make titles, attributes, and variants answer conversational constraints.
- Add the optional conversational fields Google documented for AI Mode.
- Test queries and measure channel results without treating one chat as a ranking report.
Shopify's Q1 2026 commerce data is the reason this audit is worth a week of catalog time. AI chatbot referrals grew more than 8x year over year. AI-referred orders grew nearly 13x. On product pages, AI-referred visitors converted at nearly 50% higher rates than organic search visitors, and those orders carried about 14% higher average order value. Organic search still sent more sessions than all tracked AI platforms combined. The channel is growing. It is not a replacement for search. Use those figures as context for why the work is worth staffing. Do not paste them into a board deck as your store's results.
Separate Google AI Mode from ChatGPT and from Shopping ads
Before you open a product, name the surface. Shopify's agentic storefronts documentation lists Google AI Mode and Gemini beside ChatGPT, Microsoft Copilot, and Meta. The data path is not shared. A Catalog Mapping change will not repair a Merchant Center disapproval. A Shopping-ads bid will not invent a missing size.
| Surface | Product data Google or the agent actually uses | Checkout | Common mistake |
|---|---|---|---|
| Google AI Mode and Gemini | Merchant Center via Google & YouTube | Can use Shopify-powered direct checkout when that setting is on | Rewriting the theme and ignoring the feed |
| ChatGPT Shopping | Shopify Catalog | Merchant checkout in ChatGPT or a new tab | Treating Catalog Mapping as a Merchant Center fix |
| Google Shopping ads | Merchant Center plus campaign settings | Store checkout or ads destination | Bidding to fix a missing attribute |
| Free Shopping listings | Merchant Center, when the listing is approved | Store or Google surfaces | Assuming ads sync equals AI Mode readiness |
Read the table as a diagnostic, not a glossary. If the backpack is missing only in AI Mode and already approved for Shopping ads, the pipe is probably open and the record is thin. If the backpack is missing from ads and AI Mode, start with approval and identifiers. If ChatGPT shows it and Google does not, you are in this article, not the Catalog worksheet.
Shopify says you manage AI-channel availability in Sales channels → Agentic. For Google, also confirm the Google & YouTube channel is connected to a Merchant Center account and that products intended for the online store are syncing. “Agentic is on” is not the same sentence as “Merchant Center has this SKU.”
Do not call this “Google SEO.” A missing result can be a disconnected channel, a disapproved item, a variant mismatch, or a relevance miss. Each failure has a different repair. Stage 1 isolates the first of those failures: whether Google can see the product at all.
Stage 1. Confirm the route before rewriting copy
Start in Shopify, then open Merchant Center. The audit fails if those two systems disagree. Copy work done on a disconnected channel is theatre. You will feel busy and change nothing Google can use.
Shopify checks
- [ ] Google & YouTube is installed and connected to the correct Google account.
- [ ] Merchant Center is the account you actually operate, not an old agency leftover.
- [ ] Products meant for sale are available to the online store, so they can sync.
- [ ] Google AI Mode and Gemini are enabled under Sales channels → Agentic.
- [ ] You know whether Shopify-powered direct checkout is on for Google.
- [ ] Shipping, returns, and refund policies are complete in Shopify settings.
The leftover-account box is the one stores skip. An agency connected Google & YouTube two years ago. You log into a different Merchant Center on Monday morning and wonder why Shopify “is not syncing.” It is syncing. It is syncing to an account you do not open.
Shopify documents that eligible US stores can appear in free Shopping-tab listings after a successful Merchant Center sync. AI Mode is not the same surface as the Shopping tab, but it reads the same product record. A product that never reaches Merchant Center cannot appear in either place. That is why Stage 1 ends in Merchant Center, not in the theme editor.
Merchant Center checks
- [ ] The item is approved, not disapproved or pending for a policy you have not read.
- [ ] Availability and price match the live product page at the time you check.
- [ ] The landing page is reachable, indexable, and purchasable without a wholesale login.
- [ ] Identifiers (GTIN, brand, MPN) are present where Google expects them.
- [ ] Variants share a stable item group and do not appear as unrelated products.
If a box fails, stop. A better description will not repair a disconnected channel or a disapproved SKU. Write the failure on a card: “route” or “approval.” Then leave copy alone until that card is closed.
When every box is green, Google can use the product. The next lie is quieter. Shopify, the live page, and Merchant Center can still describe three different items.
Stage 2. Compare three representations of the same product
Pick 20 commercially important products. Include more than bestsellers. Bestsellers hide architecture problems because someone already babysits them. You need at least one simple product, one size or colour family, one metafield-heavy product, one sale item, and one awkward case such as a preorder or a bundle. Those five patterns are where feeds usually split from the storefront.
For each product, open three views on the same day:
- Shopify admin
- The live product page
- The Merchant Center product record
Write down title, description, price, availability, image, GTIN, brand, colour or size, and landing-page URL. If any field disagrees, the feed is lying to Google. AI Mode will answer from the lie, not from the prettier theme.
Typical breaks, and what they cost:
- The storefront title is merchandising copy. The feed title is an internal SKU name. A shopper who names the category never matches the record.
- Inventory is in stock on the site and
out_of_stockin Merchant Center because the sync lagged or a location was excluded. The conversation treats a live SKU as unavailable. - The primary image is a lifestyle crop. The feed image is a leftover packshot of the wrong colour. The model “shows” a product the shopper will not receive.
- Size 8 is sold out. The parent still reads as available. The shopper is promised a size you cannot fulfill.
- A metafield holds material and capacity. The feed never received those fields. The conversation cannot answer the constraint the shopper actually used.
Work one row all the way through before you “fix the catalog.” If admin says The Weekender, the PDP says Carry-on travel backpack, and Merchant Center says SKU-4419, you do not have a content problem yet. You have three sources of truth. Pick one canonical title and force the other two to match. Then repeat for price, stock, and image.
Shopify can keep price and availability current when the channel is healthy. That is not the same as sending every merchandising metafield. If a buying constraint lives only in a theme block, AI Mode cannot use it. Stage 3 exists to put those constraints into fields Google already knows how to read.
Stage 3. Write the record for a conversation, not a keyword
Google says AI Mode queries are longer and more specific than classic Shopping searches. A shopper does not type backpack. They say they need a carry-on bag that fits a 16-inch laptop, stays under a budget, and does not have a separate shoe compartment. Stage 2 made the three representations agree. Stage 3 makes that agreed record answer those constraints.
The product data specification is still the backbone. Conversational attributes come later. If title, description, price, availability, and identifiers are wrong, optional fields will not save the listing. Fill the backbone first. The worksheet for extras is a later article.
Title
Name the product a shopper would ask for. Include the category and the variant that changes the purchase when those facts are true. Do not write a slogan.
Weak: The Weekender — premium travel essential Useful: Carry-on travel backpack 26L, clamshell, 16-inch laptop sleeve, black
The weak title can win a branded search. It loses every constraint prompt. The useful title is not “stuffed.” It is the sentence a careful shopper already used.
Description
State what the product is, who it is for, materials, dimensions, compatibility, care, inclusions, and limits. Keep a short opening sentence, then specifications. A lifestyle paragraph that never names volume or laptop size forces the model to guess. Guessing is how you get a confident wrong answer.
The product-content workflow is the bulk process for a large catalog. This audit is the quality gate. Do not generate 400 new descriptions until the 20-product sample matches across admin, PDP, and Merchant Center.
Variants
If a shopper can buy it as a distinct SKU, it needs a distinct, consistent row. Use item_group_id so colours and sizes stay one product family. Human-readable option names beat internal codes. A shopper who says “navy” cannot match NV1.
| Failure | Shopper prompt | Repair |
|---|---|---|
| Parent in stock, size 8 empty | “Do you have this in women's 8?” | Variant-level availability |
Colour stored as NV1 | “Show the navy one” | Readable colour value |
| Sale price stale | “Find this under $80” | Current price and sale dates |
| Image does not match the selected colour | “Show the green version” | Variant image link |
| Three colours uploaded as three unrelated products | “Which colours does this jacket come in?” | Shared item group |
Each row is a different prompt type. That is why you do not “write better copy” as a first move on a variant miss. You repair the field the prompt actually touched.
Once the core offer can survive those prompts, you may add the optional fields Google built for follow-up questions. Not before. Stage 4 is an overlay, not a rescue.
Stage 4. Add conversational attributes only after the core feed is clean
Google's conversational attributes guide is explicit. These fields are optional. They complement the primary specification. Adding them does not change existing product-approval status. Google recommends a supplemental data source, or the Merchant API, rather than stuffing everything into one fragile primary file.
Do not copy the same sentence into description, product_highlight, product_detail, and question_and_answer. Google tells you not to duplicate. Duplication is how a feed looks “complete” in a spreadsheet and still teaches the model nothing new.
Use this order, because each field answers a different leftover question:
- Core offer fields are accurate. You proved that in Stages 1 to 3.
product_highlightandproduct_detailcarry facts that do not already live in the description.item_group_titleandvariant_optionmake the family readable when the shopper asks “which colours.”question_and_answeranswers real buying questions the other fields do not cover.document_linkandrelated_productonly when the file or accessory is real.popularity_rankonly if you can defend the number from your own inventory performance.
If you are staring at character limits or TSV escaping, you have left the audit. The companion article on Merchant Center AI Mode attributes is the field-by-field worksheet. Come back here when the supplemental file is uploaded and you need to know whether the conversation changed.
Stage 5. Test with a constraint matrix, then measure business results
A clean feed is a claim. Stage 5 is the check. Do not search only the product name. A branded hit proves recognition. It does not prove discovery. The shopper who already knows The Weekender was never your problem.
For each priority family, run six prompt types in Google AI Mode. Use the same country, language, and date. Record whether your product appears, which merchant is shown, and which fact is wrong. Wrong facts matter as much as absence. A hit that promises the wrong laptop size is a refund in waiting.
| Prompt type | Pattern | Example |
|---|---|---|
| Category | best [category] for [audience] | best travel backpack for a remote worker |
| Attribute | [category] with [attribute] | carry-on backpack with clamshell opening |
| Use case | [category] for [situation] | backpack for a three-day business trip |
| Budget | [category] under [price] | laptop backpack under $120 |
| Compatibility | [category] compatible with [item] | backpack that fits a 16-inch MacBook Pro |
| Exclusion | [category] without [undesired trait] | travel backpack without a shoe compartment |
One observation is not a trend. Repeat the set after a feed change and again the following month. If you changed titles and Q&A in the same week, you cannot say which edit moved the result. Change one class of field per cycle.
Then look at Shopify's Agentic reporting for the Google channel. Shopify documents sales, orders, online-store sessions, and online-store conversion by AI channel, with channel or referrer attribution on the order. Direct-checkout activity and referred storefront sessions are not the same event. Track them separately. A store that turned on Shopify-powered checkout inside Google will under-count “sessions” if it only watches the online store.
| Stage | Metric | Question it answers |
|---|---|---|
| Route | Approved, in-stock items in Merchant Center | Can Google use the product at all? |
| Representation | Admin, PDP, and feed agree | Is the record truthful? |
| Visibility | Priority prompts where the product appears | Does the data match the conversation? |
| Visit | Sessions attributed to Google AI surfaces | Did exposure become a store visit? |
| Purchase | Orders, sales, conversion | Did checkout finish? |
Read the table from the top. Do not jump to purchase and invent a title story. Do not credit a description rewrite for an order that appeared once after you changed three other things. Price, stock, competitors, and the shopper's context move at the same time. The measurement guide is the method for that restraint.
Fix problems in this order
When a prompt misses, do not start at Stage 4. Walk the stack you just built:
- Access. Google & YouTube, Merchant Center, Agentic toggle, policies.
- Approval and offer integrity. Disapprovals, broken URLs, stale price, wrong stock.
- Identity. Brand, GTIN, item group, variant labels, duplicate listings.
- Decision data. Materials, dimensions, compatibility, limits.
- Conversational fields. Highlights, details, Q&A, documents, related products.
- Query log and channel results. One hypothesis per cycle.
The 30-day plan below is this stack stretched across a month so a small team can staff it without boiling the catalog.
A 30-day plan
Week 1. Verify the route
Connect or repair Google & YouTube. Confirm Agentic settings for Google AI Mode and Gemini. Export Merchant Center disapprovals. Choose 20 representative products. Do not write copy this week. If the route is wrong, next week's prose will never arrive.
Week 2. Repair the records
Align titles, identifiers, prices, stock, images, and groups on those 20 SKUs. Do not write new prose until the three representations match. A changelog of which field moved on which SKU is the only way Week 4 can be honest.
Week 3. Add constraints and conversational fields
Rewrite descriptions that hide materials or fit. Add highlights and Q&A only where the core fields are silent. Keep the changelog. If you cannot point to the field you changed, you cannot test it.
Week 4. Test and choose the next fix
Run the six-prompt matrix. Review Google-channel sessions and orders in Agentic reporting. Pick one catalog hypothesis for the next month. Expand beyond the 20-SKU sample only if the sample is clean.
You cannot guarantee a Google AI Mode placement. You can guarantee that the product record answers the same questions a careful shopper asks. Eligibility opens the pipe. Specific, current, purchasable data is the reason a conversation has anything true to say about your store.