Product Data Quality and Governance: What It Is and How to Fix It

Product data quality is the first thing to break after a PIM goes live, and the last thing most teams measure. The tool stores the record. Quality and governance decide whether the record is fit to publish. This guide defines data quality, shows how to measure it, explains who should own it, covers metadata and taxonomy, gives a seven-step sequence to improve it, and ends with what changes when AI touches the data.

What is data quality?

Data quality is a product record that can still do the job you collected it for. A data set is high quality when it meets the requirements of its intended use. For product information that use is concrete: a channel can publish the item, a buyer can trust the specification, and a colleague can take the same row and reach the same conclusion.

Every extra place someone types a value is another place an error enters. A weight corrected in the ERP and left wrong in the marketplace feed is a quality failure, even when both systems look well kept. If you do not block bad entry and do not repair the records you already have, channels publish the wrong product.

The requirements change by marketplace, by country and by brand. The test stays the same. You cannot claim high quality without a picture of what good looks like for each channel.

The five dimensions of product data quality

The five dimensions are the usual way to make that picture explicit. They overlap in practice. Score them separately anyway: a complete record can be wrong, and a correct record can be missing the one field the channel demands.

• Accuracy. The record states facts about the product you ship. A wrong material, a stale price or a dimension in the wrong unit means the listing describes a different item from the one in the warehouse.
• Completeness. The record holds the fields the job needs, and they are current. A marketplace that rejects the row for a missing attribute has already given you its definition of complete.
• Consistency. People, exports and reports read the data the same way. The webshop, the price list and the portal can each be tidy and still contradict each other.
• Integrity. The record stays intact from the moment you receive it, through enrichment, to the version a channel publishes. A silent overwrite during an import is an integrity failure: you still have a row, but not the row you approved.
• Relevancy. Accurate data the business does not publish can still be the wrong data to maintain. Extra fields cost time and hide the attributes a report or channel was built to show.

Where quality breaks

Problems show up when databases are merged or systems are integrated and fields that should match do not, because the schema or the format differs. The same break appears after an acquisition, or when siloed files land in a warehouse for the first time. The ERP and the shop can both be treated as the system of record and still disagree, because each accepted a different version of the item.

The work is rationalization and validation. You decide which source wins for which field, and you check that the winner still matches the product. Basic principles of master data management puts this in the wider frame: product data is one domain of master data, and the rules for ownership apply beyond the catalog. Where ERP stops and PIM starts shows which system should own which attribute.

What to measure first

• Pick the channel that rejects you. Its required attributes are the completeness list. Start there.
• Pick the fields a customer can compare. Size, material, price, the claim on the label. Those are the accuracy checks that matter this quarter.
• Name the winning source per field. Supplier file, ERP or the PIM. Two winners means the next import undoes the cleanup.
• Re-check after publish. Integrity is the value that left the PIM, not the value still on the screen of the person who edited it.

The trade-off of a short list: you will ignore fields a future regulation wants. Add those when they have a reader. A quality program that tries to perfect every column before any channel is clean does not ship.

What is data management?

Data management is the administrative process of acquiring, validating, storing, protecting and processing data so that the people who need it can get a reliable version in time. PIM, DAM and MDM each manage a slice of the records a company uses. Data management is the work underneath all three.

Data is the facts in and around the organization: customer records, accounting entries, orders, product documentation, measurements from production. Structured data is organized and traceable in a database. Unstructured data still sits in files, documents, messages and images. A pile of values becomes information when it has structure, context and meaning, so someone can interpret it. Data management gets that information to the right person before the decision, not after.

The day-to-day work is maintenance: check files for changes, correct them, and add new values from external sources through links, so an update does not depend on retyping. Consistency matters as much as currency. If the same company, person or address is stored in more than one database, the values have to match. A spelling difference is enough for two systems to treat one company as two. Enterprise data management is the same work scoped to the whole organization.

Who owns product data quality? Governance

A PIM is only as good as the governance around it. Governance here means the policies, processes and structures that decide how product information is managed. The tool stores the record. Governance decides who may change a value, what good means and which version is allowed to go live.

In practice, data governance is the frame an organization uses to manage data for its own needs: policies for how data is handled and named roles for who does the work. Six things belong in it.

Policy and compliance

Write the lifecycle down: creation, publication, archiving. A policy people can follow keeps internal records consistent and keeps the company inside the legal requirements and industry standards that apply to the product. A policy nobody can follow will not stop an editor pasting a claim into a channel at the end of the day.

Quality standard

Set the standard for entry, the validation that runs on it and the audit that checks it. Use the five dimensions above and say which of them each channel requires. If the policy does not say, the audit has nothing to fail.

Roles

Name who owns which part of the record. Overlaps and gaps are where errors and rework come from. Put a person's name on the field. A shared mailbox will not approve a row.

Two roles matter most. Data owners have the authority to make binding decisions about a data domain. Data stewards follow consistent processes day to day. A master data team that works as a manual fix-it crew at the end of the pipeline is an entry team without process authority, not a governance function. In the PIMvendors Single Source episode with Nicola Askham, who coaches on data governance, the root cause of decay after go-live was almost never the tool: nobody was formally accountable across the full data lifecycle, spread over purchasing, category management, e-commerce and regional sales.

Workflow and approval

New or changed information is reviewed and approved before it is published. A workflow people follow stops errors reaching a channel. Six approval steps that everyone bypasses add delay and do not govern the publish. Pragmatism is part of governance: when every new product needs one person's sign-off and that person is on holiday, people route around the system. A framework that allows the rush job and the incomplete entry at intake, and then has a clear corrective path, holds better than one that cannot flex.

Security and access

Access rights keep sensitive product information away from people who should not see or change it. It matters in an ordinary catalog the moment the record holds a cost price, an unreleased claim or a supplier contract term.

Improvement and training

Review the process when the business or the tooling changes. Train people on the PIM, on the practices you expect and on new functions as they ship. A new marketplace attribute is a new lesson.

Why governance pays

• A record you can decide from. The value on screen is the value the channel will show, because a person approved it.
• Fewer loops before publish. Clear roles and a real workflow mean a change does not wait on a meeting to find out who may say yes.
• A compliance position you can explain. The lifecycle and approval are what you show when someone asks which claim was live on a given date.
• Sensitive fields protected. Cost, unreleased products and contract terms stay with the people who should see them.

Experian's 2018 Data Management survey of more than 1,000 data practitioners found that 55% said unreliable data has a substantial disruptive effect on the organization and 69% said incorrect data undermines their ability to deliver an excellent customer experience. Without a working governance structure, competitiveness and the ability to operate within the law both take the hit. Regulation adds pressure: the Digital Product Passport and PPWR both need structured, verified product data, which is what a mature governance framework produces.

The trade-off: governance adds an approval, and an approval feels like friction to the person who used to edit the sheet directly. Take the friction away and nothing stands between a draft claim and the channel.

Setting up a data management team

Once a PIM or MDM project is approved, the next question is who runs it. Work through it in this order:

1. Name why the organization needs the team and which business problems it owns. Write the problem down before the job titles.
2. Define the scope: which data sets, which business functions, which stakeholders.
3. Assess the current state: practices, processes and technologies, and the gaps.
4. Develop the plan: objectives, timeline, budget, metrics, plus a governance framework, policies, procedures and standards.
5. Build the team: analysts, engineers, architects and governance specialists as the scope requires.
6. Choose the tools last, following the scope: integration, data quality and security.
7. Communicate and train every stakeholder who creates or uses the data, then monitor against the metrics.

Askham's advice for starting is to begin with a specific business problem, not with "we need data governance." Start small and assign one person to coordinate.

Metadata and taxonomy

What is metadata?

Metadata is data that describes other data (Merriam-Webster). It is the information you attach to a page, document or file so the thing inside can be found, reused and judged relevant. For a document: author, file size, creation date and keywords. For a product image: file type, size, rights and the product it belongs to.

Metadata comes in a few kinds:

• Descriptive: title, subject, author, creation date.
• Rights: copyright status, rights holder, license terms.
• Technical: file type, size, compression. This is what digital-object management and interoperability use.
• Structural: an item's place in a hierarchy or sequence.

Instruments for managing it

A metadata approach starts with a metadata set: what each element is, why you use it and which information it applies to. These instruments build on each other:

• Controlled vocabulary. A fixed selection list, such as the twelve months or a list of audiences. It stops typos and lets you tag quickly.
• Taxonomy. A hierarchy of nodes, used when a list grows into main and subcategories. It makes results relevant and gives you something to filter on.
• Facets. Groups of metadata fields used together, mainly to filter on a website.
• Thesaurus. Terms organized by meaning: synonyms, related terms, a preferred term and a definition.
• Ontology. Entities connected by whatever relationships you need. Powerful and complex, and it takes a specialist to build.

Use standards that already exist and that other organizations already apply. You avoid inventing a private scheme, and the information stays findable because the other side uses the same terms.

Product taxonomy

A taxonomy is an authoritative, restricted list of terms or categories, where each term stands for one unambiguous concept (Encyclopedia of Knowledge Organization). A shopper looking for swim shorts walks a path such as Men's clothing, Summer clothing, Swimwear. That path supports indexing, tagging, categorizing and metadata, and it makes products findable when someone browses or searches.

A working taxonomy changes several things:

• Attributes follow the branch. A properly set up PIM attaches the right attributes to the right category. Coffee cups do not inherit the measurements of extension cords, so buyers are not shown specifications that belong to another product.
• Data is collected in a normalized way, so you can check for redundancy and keep a tight check on what you publish.
• Search and campaigns improve. A miscategorized product can show up on queries that do not match, which raises clicks and lowers conversion in shopping campaigns.
• Supplier content is easier to take in. A PIM that follows standards such as GS1 or ETIM can take in product content with relatively little rework. See product classification with GS1, ETIM and ECLASS. Classification models such as UNSPSC and GS1 GPC serve the same purpose for expanding catalogs.

Seven steps to efficient product data management

Efficient data management starts with KPIs that show whether the PIM process works, plus named owners and rules people follow. These seven areas show where you are mature, where you can improve and where you need to act.

1. Measure the PIM with KPIs

   Companies often go live without a finish line. Without numbers you cannot tell whether the system supports the business or where to improve. Set KPIs at the start, using the dimensions above. They are what later decisions hang on, including whether the software still fits.

2. Tie the data program to company goals

   Look at the PIM inside the company, not as a tool on its own. A high return rate or a damaged public image often traces back to inconsistent or incomplete product data on the output channels. A data strategy that supports company goals limits those risks.

3. Name who owns what

   Check the organization: which roles touch the process, and whether you have the people to hit the goals. Train the people in the system. Data management works when responsibilities, tasks and competencies are defined at each level.

4. Keep quality by following the rules

   Processing pays off only when standards hold: product descriptions, units of measure, dates, article numbers and the legal framework you must meet. Duplicates and incomplete or wrong address lines are signs that information management is slipping. Specifications and data models have to exist first, so product data can be delivered complete and in one form.

5. Harmonize processes so a change can move fast

   Several departments usually work in the PIM, so their processes have to line up on one database. Approval flows for products under legal rules, such as food or medical devices, need a clear structure and a record. A recall is the stress test: the article leaves the range, disappears from the shop or shelf, and the cash register systems get the update. Spell out which scenario needs which action.

6. Keep the data architecture understandable

   A large article count gets expensive to maintain and opaque for the people in the PIM. Catalog the data models and check them for consistency, or the next round of maintenance turns into clutter. A clear structure is how people find the fact they need.

7. Balance automation and manual work

   For some companies the online shop is the center. For others it is the print catalog and technical data sheets. That choice drives the setup. Individual, high-quality information means more manual maintenance. For large volumes in several languages, check how far the work can be standardized or automated. Recheck the objectives after go-live and adjust them. Data management is that recurring check.

The product data hierarchy of needs

A second way to sequence the work is as a pyramid. Skip a layer and the layer above is guessing.

• Collect. How product data enters the company: logging of clicks, views and purchases, external data, and user-generated content such as reviews.
• Move and store. A reliable flow between systems, storage that can grow, and pipelines that standardize, clean and organize before anyone reports.
• Explore and transform. Cleaning inconsistencies, errors and duplicates, spotting anomalies, and adding the attributes and metadata analysis needs.
• Aggregate and label. Analytics, the KPIs you will judge decisions by, and segments by region or purchase history.
• Learn and optimize. A/B tests, machine learning for recommendations and forecasting. AI belongs here, after the lower layers exist.

Work it from the bottom: collect and store before you clean, clean before you aggregate, aggregate before you train a model.

Data quality in the AI era

AI does not clean product data for you

AI processes whatever file you hand it, faster than a person with a spreadsheet. If the names, units and supplier labels already disagree, the disagreement shows up in more systems, sooner. In a PIMvendors webinar, Susan Walsh of The Classification Guru described a normalization module that merged three unrelated businesses into one supplier group (a hairdresser, a cleaning company and a law firm) because they shared a word. A repeated token is not a shared identity. Nobody spots that by eye on a sheet of a million rows.

Her practical points:

• Write the entry rules down. Naming, units, which supplier field wins, who may overwrite a value. Tools change every few years, and rules for how a value is entered still apply. A cleaned file with no rules is dirty again within months.
• Point AI at checkable work. Spreadsheet formulas, lookups, duplicate and pattern detection. Keep a person on classification that changes a commercial claim.
• Start from a pivot, not a platform. Put the export in a pivot table, sort description against product code and look at the inconsistencies a raw export hides. Fix by business impact.
• Fix the intake. If the supplier file is the source of the bad rows, cleaning the PIM and leaving the inbox alone spends the budget twice.

Training data matters too. In a Partium demo on parts data, the point was that AI needs quality data to train on, that synthetic data carries significant risk, and that companies should assess how trustworthy their data sources are. See what AI inside a PIM actually does and where it breaks.

Risk, bias and chaos

Models trained on messy data are as good as their worst input. Weak links between data sets lead to misleading insights. Loose access control lets AI reach data it should not. Unknown data origin makes outputs hard to trust, and historical bias in the data is replicated. Governance is the foundation: know where data comes from, control who and what can see it, and demand traceability from AI decisions. AI can also support governance, for example by finding duplicates and inaccuracies, monitoring access and tracing data lineage, as long as a person owns the rules. The idea draws on Tina Salvage's article in CDO Magazine.

Incomplete data drops out of AI answers

Product data that is missing, conflicting or buried in marketing copy is easy for an AI shopping assistant to skip. It moves on to a listing that looks complete enough to trust. Salsify's AEO 101 piece describes the exclusion pattern: missing or inconsistent attributes, conflicting dimensions across your site and retailers, vague copy that says nothing about material or fit, and no structured markup. The fix is the quality work above: complete the attributes that used to feel optional (material, size with units, compatibility, care), remove contradictions between channels, and add structure. See AI-ready product data and agentic commerce.

Governing AI agents that write to the PIM

AI agents now enrich product data through APIs and the Model Context Protocol (MCP), an open standard that lets an AI client call a system's tools. Before you connect another agent, decide who it answers to. Documentation from Akeneo and Bluestone shows the same pattern: the agent borrows an existing identity and carries that user's rights. Both recommend a dedicated least-privilege user for agents, and nothing forces it.

Inriver, drawing on outside research, suggests treating an agent like a new hire: a unique identity, explicit scope and task-specific rights. An agent enriching descriptions needs access to content fields, not to pricing approvals or channel credentials. Inriver cites IBM research that 35% of executives cannot say who holds authority to override an AI recommendation. The policy has to name that person per attribute group, and the PIM workflow has to route the conflict to them. Stibo Systems runs its agents inside its roles, permissions and controls, and Pimcore runs its agent bundle through its workflow engine, so agent work passes the same approvals as human work.

Three checks when you buy:

1. Give every agent its own identity and a named business owner. A shared API key leaves no trail of which agent did what.
2. Scope rights down to attribute group, locale and catalog, with propose-only rights on regulated attributes. Check that approval rules also fire on an MCP write.
3. Make every agent write traceable and reversible. Ask how you reverse a bulk change on thousands of records. Bluestone warns that change history should not be treated as a guaranteed undo function.

Related reading

Quality is measured per channel against five dimensions, governance gives it an owner, metadata and taxonomy give it structure, and the seven steps keep the work recurring. For the systems that run this work, see what is PIM, PIM features and capabilities and how to implement a PIM.

Where to go from here

Both tools are free and independent of any vendor.

Not sure you need a PIM yet?

A short check that shows whether a dedicated PIM helps your catalog, or whether spreadsheets and ERP still do the job.

Check if a PIM fits

Ready to compare vendors?

Answer questions about your scale, sector and must-haves and get a ranked shortlist.

Build my vendor shortlist