Optimising Google Merchant Center Titles with AI, at Scale

A Google Merchant Center title is not a place to pack search terms. It has to describe the product on the landing page accurately, and put the deciding attributes where a shopper sees them before truncation. AI can restructure titles at volume, but if you ignore Google's requirements for generative-AI-created product data, the process is wrong from the first upload.

Four rules to hold on to

From Merchant Center's current title specification:

1. a regular title can be up to 150 characters;
2. many surfaces show only the first 70 characters or fewer, so important attributes go first;
3. titles must not include promotional text, gratuitous capitals, or attributes that don't match the landing page;
4. titles created by generative AI should be submitted as structured_title with digital_source_type set to trained_algorithmic_media.

Rule 4 is the one "bulk-rewrite your titles with AI" tutorials skip. Even when the content is accurate, the generation method has to be declared in the data.

There is no single good title formula

Different categories have different deciding attributes:

• apparel: brand, category, gender/age, material, colour, size;
• electronics: brand, model, product type, key spec, capacity, colour;
• home: brand, product type, material, dimensions, colour or room;
• customised goods: product type, customisation method, material, size or recipient.

Set the attribute priority per category before generating anything. Don't run one formula across the whole catalogue.

Step 1: prepare a clean feed worksheet

Keep at least these columns:

Fix missing attributes and broken variants first. On dirty source data, AI just writes the errors more fluently.

Step 2: generate per category, not across the whole sheet

A prompt that works:

You are a Google Merchant Center product data editor.
Use only the input fields. Do not add materials, features, certifications or target users that aren't supplied.
Category: home storage
Attribute priority: brand > product type > material > size > colour
Requirements: 3 English title candidates; max 150 characters; most important information within the
first 70 characters; no price, discount, "Free Shipping", all-caps or exclamation marks.
Output: title, character count, and which source fields were used.
Product data: ...

Making the model list its source fields is the fastest way to catch invented information.

Step 3: variant titles must be distinguishable

When the same product exists in red, blue or several sizes, the title should reflect the variant and the corresponding attributes must be submitted separately. Identical titles across child items mean shoppers click and land on something other than what they expected.

Check the landing page too. Google looks at whether price, availability and product information are consistent between the feed and the page; renaming "Midnight Sky" to "Dark Blue" for search purposes when the page never says "Dark Blue" creates exactly that inconsistency.

Step 4: pick the field that matches how the title was made

If a title was created by generative AI, the current requirement is the structured field:

structured_title.content = the AI-generated title
structured_title.digital_source_type = trained_algorithmic_media

Google's documentation notes that if you submit both title and structured_title, the plain title takes precedence. So don't put contradicting versions in the two fields. Whether your feed tool or Shopify app supports structured titles has to be confirmed in the actual export mapping.

Step 5: run a small experiment before the bulk edit

Take 20–50 items from one high-impression, attribute-complete category:

• keep the original titles and the change date;
• verify the first 70 characters carry the core attributes;
• confirm there's no promotional wording, keyword stacking or off-page information;
• after upload, watch diagnostics and item review status;
• watch clicks, conversions and query matching, not impressions alone;
• expand to more SKUs only once results hold.

The five most common mistakes

1. Pasting a keyword list into the title. Readability drops and it often includes synonyms that don't appear on the page.
2. One attribute order for every category. Shoppers weigh apparel and electronics differently.
3. Ignoring the first 70 characters. You may submit 150, but what's at the end may never be seen.
4. Letting AI guess attributes. It looks like leather in the photo is not a basis for writing "leather".
5. Not declaring AI-generated titles. Structured titles and the digital source type are part of the current specification.

Launch checklist

• title matches the landing page and the specific variant;
• key attributes sit within the first 70 characters;
• total length under 150 characters;
• no price, shipping, promotional or exaggerated wording;
• brand, colour, size and material come from real feed fields;
• AI-generated content uses the structured fields the current spec requires;
• pre-change data is recoverable; validate small, then scale.

Worked example: one storage basket from feed row to title

Take a fictional home brand, Hollowpine, and one row from its feed worksheet: brand Hollowpine, product_type storage basket, material seagrass, size 30 x 30 x 28 cm, colour natural, key_feature foldable handles. The landing page shows all of these.

Run the Step 2 prompt with the home storage priority, and a typical batch of three candidates looks like this:

1. Hollowpine Seagrass Storage Basket, 30 x 30 x 28 cm, Natural, Foldable Handles   (78)
2. Hollowpine Eco-Friendly Seagrass Basket for Living Room, Bathroom, Nursery and
   Laundry, Best Gift Idea   (102)
3. Hollowpine Storage Basket Seagrass Natural 30cm   (47)

What a reviewer does with them:

• Candidate 2 is rejected. "Eco-Friendly" and the list of rooms are not in any source field, and "Best Gift Idea" is promotional. This is exactly what the "which source fields were used" column exposes.
• Candidate 3 is too thin. It drops the full size, and "30cm" alone could be read as any dimension.
• Candidate 1 is kept. Its first 70 characters end at "Foldable", so brand, product type, material, size and colour all sit before the cut-off. Only the least important word falls past it, which is the order you want.

Because candidate 1 came from the model, it goes into structured_title with the digital source type set, not into the plain title field.

Prompt for checking titles you already have

Before generating anything new, it often pays to audit the current feed:

You are a Google Merchant Center feed reviewer. For each row, compare current_title with the
supplied fields and landing_page_title. Flag: attributes not found in the fields, promotional
wording, all-caps words, missing core attributes in the first 70 characters, and identical titles
across different variant ids. Do not rewrite. Output id, issue, and the field that proves it.
Rows: [PASTE FEED ROWS]

That list tells you which category to start the Step 5 experiment with.

FAQ

I have 150 characters. Can I use the spare space for extra keywords? You can use it for real attributes that are on the landing page. A keyword list at the end is still a keyword list, and anything past the first 70 characters may never be seen.

Should every variant get its own AI-written title? Every variant needs a title that reflects its own colour or size, but that is usually a template job from the feed fields rather than a fresh generation per child item.

My feed app doesn't show a structured_title option. What now? Check the actual export mapping, as Step 4 says. Until you know the field and its source type reach Google, don't upload AI titles through a path you can't declare correctly.

If you need product facts and copy templates first, combine this with turning one product photo into a full Amazon and Etsy listing, but check feed titles separately against the product data specification. For translated feeds, see DeepL vs ChatGPT for product translation.

Sources checked 2026-07-12: Google title and structured title specification, product data optimisation guidance, AI-generated product data update.