- You cannot buy keywords here. The product data decides which auctions you enter, and the title is its richest field.
- Titles run 1 to 150 characters in Google’s spec. Front-load the part that survives truncation. [1]
- Brand, product type, then distinguishing attributes. Product type is the token most missing titles lack.
- Every generative-AI title must ship in
structured_titlewith digital source type set to trained algorithmic media. [1][2] - Send
titleas well and Google uses it instead, silently voiding the declaration. [1] - Five custom labels, numbered zero through four, are where your margin and stock logic becomes budget control. [3]
- Supplemental feeds override attributes by product ID, which makes them the safe place to test. [5]
- Google publishes no attribute weighting table. Every one you have seen is a vendor inference.
Someone types “merino base layer mens” into Google. Somewhere in your catalog sits the exact product they want. Whether it enters that auction was decided months ago, by whoever wrote the line of text in the title column of your export, and that person was almost certainly not a marketer.
This is the uncomfortable shape of retail search now. In a Shopping or Performance Max campaign there is no keyword to add, no audience that holds, and no placement you control in any durable way. Google reads the product data, decides what the thing is, and decides which queries it belongs in. The campaign is a budget and a target wrapped around that decision. The decision itself lives in a text file.
Where the match actually happens
The title is the densest text Google has about a product, and it does the heaviest lifting in working out what the product is. Google’s specification allows 1 to 150 characters for it. [1] Practitioners consistently report that only around the first 70 characters survive truncation in common placements, which is observation rather than published spec, but it is stable enough to plan against.
That gives you two jobs in one field. The first 70 characters have to carry the words a shopper would actually type. The remaining 80 exist to disambiguate the product from its siblings and to catch longer, more specific queries.
| Position | What belongs there | The failure you see instead |
|---|---|---|
| Opening | Brand, when the brand is searched | A style code, or a marketing name nobody knows |
| Immediately after | Product type, in the words a shopper uses | The category left implicit because the page shows a photo |
| Middle | The one or two attributes that decide the purchase | Every attribute, in the order the database stored them |
| Tail | Variant detail: size, color, capacity, pack count | Repeated brand, or the same words in a different order |
| Nowhere | Promotional language, all caps, shipping claims | “Best price”, “FREE DELIVERY”, “SALE NOW ON” |
The single most common defect is a title with no product type in it. A title reading “Aurora 400” is complete nonsense to a matching system. It is also completely normal, because internally everyone calls that product the Aurora 400 and nobody noticed the machine cannot. Put “insulated sleeping bag” in front of it and the product becomes eligible for a category of demand it was previously invisible to.
Resist the vendor tables that assign percentage weights to each attribute. They are inferences from private datasets. Google does not publish how it weights title against image against category, and treating one agency’s regression as a specification is how teams end up optimizing a number nobody outside that agency can see.
The attributes that decide eligibility, not ranking
There is a category of feed work that has nothing to do with performance and everything to do with whether the product is allowed to serve at all. Identifiers, category assignment, availability, price accuracy against the landing page, and for apparel the size, color, gender, and age group fields. These are pass or fail. A product failing them is not ranking badly. It is absent, and no bid change will ever surface it.
Google’s product data specification is the authority on which of these are required, conditionally required, or optional, and it varies by category and country. [7]
The useful mental split is that attributes govern eligibility and titles govern reach. Teams routinely spend a quarter tuning titles on a catalog where 18% of the products cannot serve, which is a bit like rewriting the menu of a restaurant that has not passed its fire inspection.
Custom labels are your real campaign structure
Google gives you five custom labels, numbered zero through four, and lets you decide entirely what each one means. Once set, they become product groups you can attach budgets, targets, and exclusions to. [3][4]
This is the lever people underuse, because it looks like a reporting convenience and is actually the only durable product-level control you have left. The scheme worth building is business logic the platform genuinely cannot infer:
- Contribution margin band, so the products you make money on can carry a different target from the ones you do not. This is where feed work meets what a conversion is worth.
- Stock depth, so a line with four units left stops absorbing prospecting budget.
- Price competitiveness against the market, so you stop paying to lose a comparison you already lost.
- Lifecycle stage, separating a launch that needs impressions from a run-out that needs clearing.
- Seasonality, so a decision made in October does not have to be remembered in March.
Note what is not on that list: anything Google can work out for itself. Labeling by category or by brand duplicates data the feed already carries and buys you nothing.
Supplemental feeds: the safe place to change your mind
Google recommends supplemental feeds for keeping price, availability, or sale information current when the primary feed makes that difficult, and they work by matching product IDs and overriding or adding attributes. [5]
The strategic value is larger than the documented use case. A supplemental feed is the one place you can change titles, labels, or category assignments for a subset of products without a release, without touching your store’s export, and without asking anyone. That makes it the natural home for title testing: override a pattern across one product type, leave the rest of the catalog alone, and roll the winner into the primary feed once it has earned it.
Where AI helps, and the rule almost nobody implemented
This is the part that changed recently, and it changed in a direction most teams have not noticed.
Google now requires that all titles created using generative AI be provided using the structured title attribute instead of the title attribute, with a digital source type sub-attribute declaring the origin: trained algorithmic media for AI-written text, default for everything else. [1][2] Google also ships its own endpoint for this work: a text suggestions service in the Merchant API whose one method takes product information and returns candidate titles and descriptions. [6]
Then comes the trap. Google’s documentation states that if you provide both the structured title and the title attributes, only the title attribute is used. [1] Most feed tools populate the legacy title field by default and will keep doing so while you add the structured one alongside it, which means the declaration you believe you made is inert and the field you believe is live is being ignored. If you are going to do this, the migration is a replacement, not an addition.
Give the model data, never a blank page
Pass the structured attributes you already hold: brand, product type, material, dimensions, pack size. The model’s job is to select and order facts, not to supply them.
Prevents inventionConstrain the shape, not the wording
Fix the pattern and the length band in the instruction. A model asked for a good title writes marketing copy. A model asked for brand, then product type, then two attributes, under 150 characters, writes a feed title.
Keeps controlDiff every output against its source row
Reject any title containing a number, unit, material, or capacity that does not appear in the source data. This one check catches the majority of real damage, and it is a string comparison, not a judgment call.
The gate that mattersShip into structured_title, and remove the old field
Write the AI-written text to structured_title with digital source type set to trained algorithmic media, and stop sending title for those products, or the declaration does nothing.
Read the top revenue decile by hand
Automate the tail, review the head. The products that carry the account are few enough that a human can read every one of their titles in an afternoon, and expensive enough that nobody should not.
Human in the loopWhat models are genuinely good at here
- Reformatting thousands of titles into one consistent pattern from fields you already have
- Pulling structured attributes out of messy free-text descriptions and spec sheets
- Mapping a catalog to a taxonomy, at draft quality, for a human to spot-check
- Writing the description nobody has had time to write since 2019
- Flagging titles that break your own rules, which is classification and cheap
What breaks, reliably
- Inferring attributes that were never in the data, stated with total confidence
- Replacing a specific product type with a broader one that reads better and matches worse
- Adding promotional phrasing and capitals that Merchant Center rejects outright
- Mapping into a neighboring taxonomy branch that is plausible and wrong
- Quietly dropping the variant detail that made the title unique across a size run
The pattern across both columns is the same. Models are strong at transforming information you possess and dangerous at supplying information you lack, and a feed is mostly a question of what you possess.
The feed QA pass
The checks we run before a bulk title or attribute change reaches live product data, in the order that catches the most damage soonest.
- What to export and record first, so a bad batch is reversible
- The eligibility checks that decide whether a product can serve at all
- The title rules, including the 70-character window and what may never appear
- The diff every AI-written title has to survive before it ships
- What to watch in the first 48 hours, and what not to judge yet
Knowing whether it worked
Feed changes are unusually hard to read, because the effect arrives as a change in which queries you match, not as a change in a rate you already track. Impressions on previously silent products, movement in the queries a product shows for, and product-level share are the leading signals. Conversion rate is the lagging one and it moves for twenty other reasons.
Give a bulk title change a couple of weeks before judging it, change one thing at a time across a product type rather than everywhere at once, and keep a control group of products on the old pattern. It is the same discipline as any other test, and it is skipped more often here because a feed change feels like maintenance rather than an experiment.
The reason all of this pays is structural. Bid strategy, budget, and campaign controls are converging across every advertiser in your category, because everyone is being handed the same automation. The product data is the last place where the work you do is still yours.
Sources
- Google Merchant Center Help · Title [title] and structured title [structured_title]1 to 150 characters, the requirement that generative-AI titles use structured_title, the digital_source_type values, and the rule that title overrides structured_title when both are sent
- Google Merchant Center Help · AI-generated content
- Google Ads Help · Use custom labels for Shopping adsfive custom labels, numbered 0 through 4, used to build product groups in Shopping and Performance Max
- Google Merchant Center Help · Custom label 0-4 [custom_label_0-4]
- Google for Developers · Content API for Shopping: best practicesthe recommendation to use supplemental feeds for price, availability, and sale updates
- Google · Merchant API product studio: TextSuggestionsServicethe GenerateProductTextSuggestions method, taking product information and returning suggested titles and descriptions
- Google Merchant Center Help · Product data specificationthe required, conditionally required, and optional attribute tiers, including apparel requirements
Frequently asked questions
Why does the product title matter so much in Shopping and Performance Max?
Because you cannot add keywords to a Shopping or Performance Max product listing. Google works out which queries a product belongs in from the product data itself, and the title is the richest text field in that data. Change the title and you change the set of auctions the product enters.
How long can a product title be?
Google’s specification allows 1 to 150 characters for the title attribute, and the same range for the content sub-attribute of a structured title. Practitioners commonly report that only about the first 70 characters survive truncation in common placements, so the useful working rule is to write for 150 and front-load the first 70.
What is the right structure for a product title?
The pattern most operators converge on is brand, then product type, then the attributes that distinguish this item from its siblings, such as color, size, material, or capacity. The reasoning is simple: a shopper searching for “merino base layer” needs those words present, and words early in the title carry more weight than words at the end.
Do I have to declare that a product title was written by AI?
Yes, if generative AI wrote it. Google states that all titles created using generative AI must be provided using the structured title attribute instead of the title attribute, with the digital source type sub-attribute set to trained algorithmic media. The default value marks a title that was not AI-generated.
What happens if I send both structured_title and title?
Google ignores the structured one. Its documentation is explicit: if you provide both the structured title and title attributes, only the title attribute is used. This is the trap in most implementations, because feed tools often keep populating the legacy field by default, which silently voids the declaration you thought you had made.
How many custom labels do I get, and what should they hold?
Five, numbered zero through four. You decide what each one means. The highest-leverage scheme is business logic the platform cannot infer: contribution margin band, stock depth, price competitiveness, and lifecycle stage. Those become the product groups your budgets and targets attach to.
What is a supplemental feed for?
Overriding or adding attributes for products that already exist in your primary feed, matched on the product ID. Google recommends it specifically when keeping price, availability, or sale information current in the primary feed is difficult. In practice it is also the safest place to test title patterns, because you can change a subset without touching your store’s export.
Can I let an AI model rewrite my whole catalog?
You can, and you will regret the part you do not check. The documented failure mode is semantic drift: the model infers attributes that were never in the source data, swaps the specific product type for a vaguer one, or adds promotional language that Merchant Center rejects. Use the model to fill and format from data you already hold, never to invent product facts.
Does Google publish how it weights feed attributes?
No. Vendors publish weighting tables and they can be useful as hypotheses, but Google does not disclose the relative importance of title, image, category, or landing-page signals. Any table that assigns a percentage to each attribute is somebody’s inference from their own dataset.
What is the fastest feed change with a real effect?
Putting the product type into titles that are missing it. A title reading “Aurora 400” tells Google nothing about what the object is, so it matches loose, low-intent queries or none at all. Adding “insulated sleeping bag” to the front of that title changes which auctions the product is even eligible for.
Mikołaj Salecki
Writes about media, tech, and AI business for people who actually run digital. Former agency lead. Skeptic of frameworks that read better than they perform.
