01 · The task
Optimise your product feed: establish the facts before the format
A good product feed reliably describes which item you offer and in which variant. Titles, GTINs, categories and labels each serve a different purpose. Keeping those purposes separate lets you correct errors precisely and then check whether the submitted data is actually right.
The outcome is a verified product record
Imagine a shoe available in several sizes: the title sounds convincing, the product page offers a choice, and the item appears on Google. Yet the exported size may still differ from the selected offer. Conversely, a GTIN missing from the visible page text does not mean it is missing from the feed. Visibility, completeness and factual accuracy are three separate things to check.
This guide is for small and medium-sized shops that want to improve an existing product feed. You will check the product's identity, the description of its variant and its place within your assortment. The result is a Feed Quality Register: every relevant discrepancy receives an evidence-backed expected result, a named owner, a specific correction and a date for checking it again.
You need a current product export, access to the product data that has actually been processed, and reliable master data. A neatly completed spreadsheet is not enough. What matters is whether the same factual information survives all the way from its source to the selected offer. Even a successfully processed file does not prove that its product information is correct.
Current as of 21 September 2026. The shop examples and pilot rules below are editorial working models. They promise neither approvals nor better rankings nor any particular revenue. Assessing the economics of margin, returns and stock remains a separate task.
02 · Data provenance
Assign an authoritative source to every field
An error needs a defined place to be fixed
Start by mapping where the information comes from. Where is the product name created, who assigns the internal item number, which manufacturer file supplies the GTIN, and which application writes the categories? Many shops combine several paths: inventory management, the shop platform, a feed app and downstream rules. A correction made in the wrong place may disappear at the next import or unintentionally change offers for other countries.
For each field, record both the source of the product facts and the technical route used to transfer them. The factual source might establish, for example, that a particular variant is black and made of leather. The technical route explains which shop field produces the value for color or material from that information. Ownership here means that someone can verify the content and correct the mapping permanently; access to the advertising account alone is not enough.
If your data sources, connections or basic account have not yet been set up, start with the guide to setting up Google Merchant Center. The optimisation below builds on that foundation. It does not replace a complete initial setup or cover an appeal against an account suspension.
Before making changes, save an export with its date, data source name and affected market. Also record whether you are looking at a raw value or one that has already been transformed. 'Title corrected' is incomplete if it only refers to what appears in a preview. A useful record identifies the original value, the rule and the result after processing.
Avoid competing corrections: Agree which system or process has authority over each field. If you need a temporary override, document its purpose, scope and planned end date.
03 · Order of work
Prioritise defects by their actual consequences
Product identity and purchasing decisions come first
A missing word in a title is different from an identifier that points to another product. Prioritise work according to the potential harm and how certain you are of the finding. A suspicion without a verified source first needs investigation; a proven mix-up needs a high-priority correction.
| Priority | Observation | Risk | Evidence | Next action | Completion check |
|---|---|---|---|---|---|
| P1 · Identity | Identifier belongs to another item | Wrong product association | Manufacturer and variant data | Correct the association | Processed identifier is correct |
| P1 · Variant | Size conflicts with the destination offer | Wrong purchasing decision | Selection and order line | Repair the variant mapping | Entire path checked |
| P2 · Meaning | Title omits an essential variant detail | Offer remains ambiguous | Product data and assortment | Refine the title rule | All affected variants are correct |
| P2 · Classification | Unsuitable product category | Misclassification | Actual primary function | Identify a suitable category | Processed value checked |
| P3 · Organisation | Multiple spellings of one label | Unclear filters and groups | Agreed list of permitted values | Standardise the values | Use and assignment checked |
These priorities are an internal working model, not a Google classification. If the account shows an actual warning or disapproval, add its exact wording and the affected destination. It may change the order of work. An incorrect material also deserves particular attention when it matters to the purchase, even if the system reports no error.
Record the scope as the number of offers that are clearly affected. A faulty rule covering every size in a product line is one shared cause with several consequences. A single defective record, by contrast, may be a local exception. Making this distinction avoids manually editing a hundred records when the real task is to repair a central mapping.
04 · Offer identity
Keep the ID stable and the variant unambiguous
An internal identifier does not replace a manufacturer identifier
The id attribute identifies an offer in your product data. It has a different purpose from a GTIN. An internal item number can make a useful ID if it points permanently and unambiguously to the relevant variant. That does not turn it into a globally assigned product identifier.
First check that every offered variant retains its correct ID. Switching between sizes must not result in two different offers sharing one ID. A title change should not create a new identifier either. The rules for stable product IDs require you to retain the ID for the same product and not reuse old IDs for other products.
Keep a small reference table linking the inventory system, shop variant and exported offer. If a connector adds extra components to the identifier, document the complete value. Do not accidentally compare an internal SKU with an ID that has a different technical structure and hastily label the difference an error.
When migrating or changing feed apps, check this mapping before the first production export. Also establish whether the same product already uses consistent IDs across different languages or countries. Changes to that system are a separate task and should not happen incidentally during a title test.
Verification question: Can another person use the ID to find exactly the same variant without having to interpret its title? If not, you lack a reliable connection between the systems.
05 · Manufacturer identifier
Check the GTIN's source and its link to the product
A valid sequence of digits can still be the wrong identifier
Distinguish three GTIN states: a matching identifier has been verified; an identifier probably exists but is not yet known; or verified manufacturer information confirms that the product has no such identifier. These states require different actions. An empty export field alone does not tell you which one applies.
The Google specification for GTIN explains how to handle assigned product identifiers. Submit the correct identifier for the product or variant you actually offer. Missing or incorrect information can limit visibility and, depending on the case, lead to disapproval. An internal SKU, a guessed number or the GTIN of a similar item is not a substitute.
Use a source you can substantiate, such as a manufacturer file and the markings on the specific product. Check the model, version and packaging unit. A formally correct check digit confirms only part of the syntax; it confirms neither the manufacturer nor the association with your offer. Do not accept a number simply because a validation tool accepts it.
Treat identifiers as text in spreadsheets so that leading zeroes are preserved. Compare the complete value again after export. Correctly maintained inventory data is of little use if an intermediate step reformats the numbers. The register therefore contains both the verified source and the value that was actually processed.
A useful quality metric is the share of correctly submitted GTINs among products with a verified assigned GTIN. Do not count products that demonstrably have no GTIN as the same type of defect. Keep unresolved cases separate so that apparently high completeness is not achieved by entering unsupported values.
06 · Evidence rather than assumptions
Do not use identifier_exists as a shortcut to a fix
Unknown is different from absent
The identifier_exists attribute describes whether assigned unique product identifiers exist. It does not describe how fully your spreadsheet has been populated. Set it to false only when the absence of the relevant identifiers has actually been established. Waiting for a supplier to reply is not a sufficient basis for that decision.
| Case | Known facts | What to check | Owner | Appropriate action | Not a substitute |
|---|---|---|---|---|---|
| GTIN verified | Source matches the variant | Export and processing | Product data team | Submit the verified value | A neighbouring item's number |
| GTIN unresolved | Information is missing | Manufacturer or supplier | Purchasing | Obtain evidence | An invented GTIN or false |
| No GTIN assigned | Absence is documented | Brand and MPN | Product owner | Use existing identifiers correctly | Denying the existence of all identifiers |
| Own manufactured product | Shop is the manufacturer | Manufacturer role and identifier system | Assortment manager | Check the appropriate brand and own MPN | Automatic false for handmade products |
The rules for unique product identifiers, brand and MPN also distinguish manufacturers from resellers. A manufacturer that is the sole seller of its own product and has no official brand can use its shop name as the brand and its own MPN. This does not permit retailers to assign invented manufacturer data to other companies' products.
For exceptions, document who confirmed which information and when. 'Handmade' or 'private label' alone is not a complete justification. Specific evidence makes later questions easier to answer and prevents a legitimate exception from being copied across the entire assortment as a blanket rule.
07 · Variants
Connect each variant to the correct offer
The group, individual record and landing page must agree
A variant family consists of the same base product in versions that are actually offered, such as different sizes or colours. Each sellable variant retains its own ID. The shared item_group_id for variants connects these offers; this grouping is required for variants in the relevant Shopping ads targeting Germany and in free listings.
Check a family as one connected system. Do its shared characteristics agree? Are the distinguishing sizes and colours correct? Does each image belong to the relevant version? Related products do not become variants simply because they belong to the same collection. A parent model also does not need to be exported as an additional purchasable offer if it only organises the selection.
Then follow the submitted URL. The product landing page requirements call for a matching, specific offer. Google recommends preselecting the correct variant on the landing page. Check the selection from the initial page load, together with price, currency and availability. A later change by a script is not a reliable solution to an initially incorrect state.
Test at least different sizes and an alternative colour from your pilot assortment. Open the URL directly, without previously saving a selection in the shop. Then check that the order line still describes the same variant. This reveals errors that may remain hidden when you simply click through a product page that is already open.
Extended variant fields: Where relevant, document which additional group and variant details your exporter supports. New optional fields do not replace checks of the underlying ID, attribute and URL mappings.
08 · Product titles
Build titles from verified product characteristics
A title should distinguish the offer
Start by asking which details a buyer needs to distinguish this item from the other offers in your assortment. Product type, model, material, colour and size may all matter. Their order depends on the assortment. A universal formula intended to work equally well for shoes, prints and spare parts is not a useful quality target.
The requirements for title and structured_title allow up to 150 characters. This does not mean that the full title will appear everywhere. The often-mentioned focus on the first 70 or so characters concerns display, not a second field limit. Put important details early and avoid promotional additions and unsubstantiated characteristics.
Editorial template for a conditional example title: Brand + model + product type + verified material + colour + size. Include each component only when it is available, correct and relevant to the offer.
Verified source data might, for example, produce 'Example Brand Luma Women's Leather Trainers, Black, Size 39'. This is an invented wording example, not a statement about a real item. If you have no evidence for the material, do not use skilful wording to make 'leather' sound plausible; resolve the product information instead.
Read titles without the rest of the product page as well. Do they identify the version being sold? Do they repeat the same information unnecessarily? Do they contain fragments left by empty fields? A fluent title can still be factually wrong. Assess its claims first, followed by concision, order and readability.
09 · Generated text
Check both the content and submission method of AI titles
The origin information belongs in the export
The title specification uses the attribute structured_title for titles produced with generative AI. The text goes in content, and its origin is indicated through digital_source_type=trained_algorithmic_media in the submitted data. A subsequent manual review is useful, but does not remove the need to check this origin information.
Check how your tool actually submits the finished text. If it sends both title and structured_title, Google uses only title under the current specification. A structured column that looks correct in the spreadsheet is therefore insufficient evidence that the intended submission method is being used.
Define which input fields may be used to produce the text. The system may write from verified characteristics, but must not invent sustainability claims, suitability for particular health conditions or a special provenance, for example. Flag unclear information before generating text. An instruction such as 'make the title more attractive' is not precise enough for a production catalogue.
Check a sample that deliberately includes difficult cases: missing material, a very long model name, several similar variants and a product without an established brand. This shows whether the template handles gaps honestly or fills them with assumptions. Record rejected wording and the reason for rejection, and improve the rule before it generates titles for further items.
Approval therefore requires two separate pieces of evidence: the title describes the product correctly, and the export submits it using the appropriate structure. If either remains unresolved, the generation process is not yet ready for wider use.
10 · Google category
Classify the product by its main function
Google's taxonomy is not a free-text keyword field
Google classifies products automatically. Supplying google_product_category as an additional attribute is optional; it lets you specify a category from Google's predefined taxonomy within the documented use cases. These include category-specific requirements and the organisation of relevant campaigns. Use an appropriate predefined category, rather than inventing a sales description.
Choose the most specific category that accurately describes the product's actual main function. Submit either its ID or its full path, not both together. If a product is difficult to classify, first gather the factual details. A decorative design does not automatically turn a functional item into a decoration.
A list of product families with a few borderline cases is often enough for this review. Include more than just the bestsellers: accessories, sets and unusual versions are particularly useful for revealing where a blanket classification is too broad. Record why the selected branch fits the product so that someone can understand the decision later.
Also check whether an old rule is replacing the category you have verified. The shop's visible navigation is not sufficient evidence of the processed category. Its structure may be different for merchandising reasons. The immediate success of this correction is an accurate classification that is processed consistently. A better position or cheaper clicks cannot be assumed to follow automatically.
11 · Your own catalogue structure
Give product_type a clear hierarchy
Your taxonomy can differ from Google's
Use product_type to describe your own catalogue hierarchy. An illustrative example would be Shoes > Women > Trainers. This path does not claim to represent a branch with the same name in Google's taxonomy. It should fit a documented structure that your team also understands when filtering products and analysing results.
The field is optional and allows up to 750 characters. You can submit it repeatedly, up to five times; however, only the first value is used to organise bidding and reporting in Shopping. If you intend to use it for this purpose, decide deliberately which path comes first. Supplying additional values does not make that decision for you.
Avoid changing hierarchy levels, such as using “Trainers” in one place and “Women > Shoes > Trainers” in another, unless there is an explicit reason for doing so. Agree on spelling, separators and how to handle items that appear in several shop collections. A promotional collection should not automatically displace the product's established classification.
The broader structure of categories, product pages and indexing is covered in the guide to Shopify SEO and shop architecture. The feed task is narrower: check that the intended hierarchy path is exported and arrives in the right place. Merely supplying it does not create a campaign budget or a bidding strategy.
12 · Internal segmentation
Use Custom Labels with a fixed list of values
A label describes a group; the action associated with it is configured separately
The five optional fields custom_label_0 to custom_label_4 help you organise products internally. Each field holds one value per product. Within a Merchant Center account, each field supports up to 1,000 distinct values. Values beyond this limit are ignored for reporting and bidding; for that label, the affected products are treated as though no value had been supplied.
Assign a fixed purpose to each field. For example, one field might identify a clearly defined part of the catalogue, another a test group and a third a review status. Write down the permitted values and the conditions for using them. Creating a new label value for every timestamp or individual SKU produces unnecessary variation and makes the system harder to understand.
Example review status: review, verified. A change in value requires a documented review; it is not an automatic diagnosis provided by Google.
Example test group: pilot_a, baseline. Keep the assignment stable throughout the agreed observation period unless an urgent correction is needed.
A label called review does not pause an offer. If a group should be handled differently or excluded, the corresponding campaign settings must actually be configured and their effect checked. Record the data classification and the action separately.
Import groups based on financial criteria only from a process with clear responsibility for those decisions. This feed review checks whether the intended value is transferred correctly; it does not also calculate new margin thresholds and turn them into binding business decisions.
13 · Public product comparison
Use a footwear item to examine a concrete inconsistency
What a public product page establishes and what remains unknown
The Ankobags product page for a black women's trainer (in English) shows how publicly visible information can be turned into a review task. When checked on 21 September 2026, the page displayed an English product title, material information and a size selector covering EUR 36 to EUR 41. This is enough to compare the statements on the page, but not to judge the private product feed.
One specific inconsistency appeared in the size description: its final size mapping stated “11 (EUR40)”, whereas the selector displayed “11 (EUR41)”. The first step is therefore a question for the people responsible for the product: which mapping is correct for this model, and which places need updating once that has been established? Choosing either value without manufacturer or inventory records would be guesswork.
Apply this approach to your own shop. Record the values you observed, the time of the check and the affected variant. Have the correct size confirmed by someone with the relevant product knowledge. Then check the description, selector, exported variant value and preselected landing-page offer together. The correction is complete when all these places describe the confirmed version consistently.
A GTIN that does not appear in the visible page text is not thereby proven to be missing from the feed. The public page also cannot establish advertising performance, sales success or a particular implementation by an agency. The value of this example lies in the observable checking process and the precise task it produces.
14 · Working tool
Maintain a Feed Quality Register with clear sign-off
Each row connects the finding, responsibility and verification
The following table provides a working structure you can copy, with deliberately hypothetical entries. It does not show actual feed data from Ankobags or any other shop. Use one row per discrepancy and connect it to the complete set of affected products. After verification, enter the actual date, finding and decision in the verification field.
| Field and source | Observed value | Expected value and basis | Owner and priority | Action | Verification and status |
|---|---|---|---|---|---|
| GTIN · Supplier → Export | Blank; assignment unconfirmed | Confirmed variant identifier or documented exception | Purchasing · P1 for incorrect assignment | Obtain evidence; check mapping | Agree a date · Open |
| Title · Shop → Template | Material stated without evidence | Confirmed properties only | Product data team · P2 | Clarify source; correct template | Read processed title · Open |
| Category · Rule → Google | Broad catch-all group only | Appropriate main function within a permitted use case | Feed owner · P2 | Check taxonomy and rule | Sample after processing · Planned |
| Label · Test plan → Export | Several spellings | One value from the agreed list | Campaign owner · P3 | Standardise values; check usage | Close with date and finding · Planned |
Add the specific offer IDs, the saved source version and evidence of the finding to your working file. If the cause remains unclear, the action is “investigate”, not “approve”. Completing an enquiry does not mean the data correction is complete.
Agree on a clear completion rule: the responsible person confirms that the expected value is factually correct; the technical check shows it after processing; and, for changes involving variants, the landing-page offer also matches. This turns a task list into an auditable sign-off process that can be reused at the next import.
15 · Controlled change
Test the correction along the actual data path
A supplemental file and a rule both need verified matching
Choose a pilot set that includes both typical and difficult cases. One possible working example is twelve offers: several variants from one family, different product types and at least one verified exception. This is not a Google requirement. The size must suit your catalogue; a small test only provides protection if it covers the relevant ways errors can arise.
A supplemental product data source can enrich existing products, but cannot replace a standalone catalogue or add and remove products itself. For the usual ID-based matching, check the ID, feed label and language. The feed label is not one of the five Custom Labels. Also check that the advanced data source features required for this are enabled in the account.
With attribute rules in Merchant Center, the actual processing order matters. “The most recently edited place wins” is not a reliable general rule. The “Take latest” function applies to certain fields concerning price and availability, not to titles or GTINs in general. Save changes as a draft first, examine the test results and apply them only after sign-off.
- Back up: Record the original values, rules, affected IDs and version.
- Limit the scope: Correct one underlying cause and avoid unintended changes elsewhere.
- Compare: Check the input value, transformation and expected output for each pilot case.
- Approve: Apply the tested change and retrieve the processed values again.
A draft is not yet an active correction. After applying it, also check the next regular data run: does the value remain correct, or does another source overwrite it? If the change introduces a new factual discrepancy, stop expanding it and return to the documented, suitable previous version. Any incorrect statement already known to exist in that version does not automatically become acceptable; it still needs a specific solution.
16 · Verification
Separate data quality from later campaign performance
Verify the full data path before observing effects
After processing, check the same offers again. Do not compare arbitrary before-and-after samples. The following sequence shows what each check can establish and when the task must be reopened.
| Check | Evidence | Passes when | Reopen when | Record |
|---|---|---|---|---|
| Source | Confirmed master data | Expected value has a factual basis | Manufacturer information contradicts it | Source and date |
| Processing | Current item value | Expected transformation is visible | Another rule overwrites it | Version and finding |
| Landing-page offer | Variant URL opened directly | Version and offer match | Selection changes or conflicts | Checking context and result |
| Next data run | New export and result | Correction is preserved | Old error returns | Time and owner |
For the related review of the shop and purchase journey, use the checklist for preparing a Shopify shop for Google Shopping. The feed register records exactly which dataset was checked. A checkout that generally works does not replace a correct transfer of the selected variant.
Afterwards, if needed, monitor advertising performance and business metrics in a separate log. Record concurrent changes to prices, budget or the catalogue. A change in clicks alone does not prove an effect from the title correction. The immediately supportable result is more limited: the documented defect has been resolved, and the correction survives the next data run.
17 · Frequently asked questions
Questions about optimising an existing product feed
Does every item need to have a GTIN?
No. What matters is whether a GTIN has been assigned to the specific product. An existing identifier should be submitted correctly. An unknown identifier needs to be researched; verified absence is a different case. Do not invent a number to improve a completeness score.
Can I resolve missing GTINs with identifier_exists=false?
Only when you have established that no relevant assigned product identifiers exist. A blank GTIN field alone does not justify this value. Also check the brand and MPN, and your role as manufacturer or reseller.
Does the product title have to be exactly 70 characters long?
No. The maximum is 150 characters; the shorter guideline concerns display. Above all, check whether your customers can clearly identify the offer.
What additional checks are needed for AI-generated titles?
Check factual accuracy and the structured submission method, including the indication of origin. Ask to see an actual export and compare it with your documented approval. A text preview is not enough.
Are google_product_category and product_type the same thing?
No. The first uses Google's predefined taxonomy; the second uses your own hierarchy. Both must make sense for the product, but they serve different purposes. Do not therefore copy a shop collection into Google's category field without checking it.
Does a review label automatically stop advertising?
No. A Custom Label initially provides grouping information. Any intended exclusion or campaign behaviour must be configured and checked separately. The register should therefore record both the meaning of the label and the action that has actually been agreed.
Does a public product page prove there is a feed error?
It can reveal a contradiction between visible statements. Whether the same error exists in the submitted product record can only be established by examining that record. A GTIN that is not publicly displayed is likewise no proof that the identifier is missing from the feed.
When can I close an entry in the quality register?
When the expected value has been factually confirmed, is correct after processing and matches the affected landing-page variant. Also check that the next regular data run preserves the correction. A green upload status or a saved rule alone is not enough.
18 · Next step
Start with one product family and a clear cause
A properly completed pilot provides the basis for expanding the work
Choose a relevant product family now. Save the current export and identify the authoritative source for the ID, GTIN, title and classification. Compare the variants with their landing-page offers by opening the URLs directly. Then enter only observed discrepancies and explicitly unresolved questions in the Feed Quality Register.
Prioritise the error with the greatest concrete impact. Establish the expected value, correct the cause and check the entire data path again. Expand the change only once it remains stable in the pilot set and the next regular run. For recurring product families, you can turn this into a regular review procedure.
Product feed optimisation then becomes verifiable work on data. You know which statement changed, why the change is correct and who checked it. That clarity also helps when you add new variants or suppliers, or switch to a different exporter.
If you would like to review your sources, rules and product mappings together and put them on a reliable footing, Salestudia can help with Google Merchant Center setup and optimisation.