The data model after go-live – why product data needs continuous governance

The work on the data model does not end with the go-live of a PIM system – it changes its nature. Why data models drift gradually without a decision framework, and how a tiered governance rhythm of 30, 90 and 365 days counteracts this.

Aaron Apelt PXM consultant
21 July 2026 · 12 min read

Summary

  • The go-live does not mark the completion of the data model, but the transition from conception to operation.
  • Without a defined change process, changes happen anyway – just uncoordinated. The result is a gradual model drift.
  • A tiered rhythm of change decisions (30 days), clean-up (90 days) and strategic review (365 days) keeps the model manageable.

In my first article, I discussed how a clean data foundation is a prerequisite for a good product experience, but does not replace it. In this article, I want to address the question of how this foundation actually remains stable after a PIM system has been introduced.

In most companies, introducing a PIM system is a clearly defined project. Ideally, requirements are gathered, attributes defined, product classes modelled, maintenance processes described and responsibilities assigned. The project has a beginning, a budget and an end point.

In reality, however, what follows after go-live is often a black box. PIM projects frequently run on tight budgets. One reason for this is that a PIM system often has no directly visible impact on business figures. Even if this is frequently due to the simple fact that the corresponding measuring points are missing, it still affects the course of such projects and has to be taken into account.

In the PIM context in particular, we are regularly confronted with conceptual questions that cost time and effort. Much like software projects, PIM projects tend to build up technical debt. The difference is that it does not arise in the code, but in the business structure of our data.1 In a great many cases, this debt consists of wanting to tackle certain topics later.

During the course of a PIM project, governance within the company has the least impact, since its processes and structures often only relate to operation after go-live. For this reason, governance is most frequently postponed to a phase after the actual PIM project.

As soon as this happens, the vast majority of PIM projects run into a massive problem. No matter how good the structures created for commissioning are – if there is no defined framework for their further development, the same unstructured data and processes emerge as before the PIM introduction.

The go-live is therefore not the end of the work on the data model. It is the moment from which the model has to prove itself in day-to-day operations and continue to evolve.

Data models begin to age after release.

None of this is new: if we once again compare our PIM domain with software development, this pattern has long been known. Decades ago, Meir M. Lehman described that a system in real-world use must be continuously adapted in order to remain useful, and that its complexity increases unless it is actively counteracted.2

Both statements can be transferred almost one-to-one to product data models. The occasions for adjustments are manifold and, as a rule, legitimate.

  • Assortment changes: New product groups bring properties with them that the existing model does not provide for.
  • Channel requirements: Marketplaces and retailer portals demand additional or differently structured information.
  • Regulation: New legal requirements create information obligations that previously did not have to be mapped.
  • Business needs: Marketing and sales require new content for campaigns, advice or personalisation.
  • Customer expectations: Comparability, filterability and transparency are becoming more important.

If there is no defined process for these occasions, changes happen in an uncoordinated way.

Let us look at a deliberately simplified example. At go-live, a retailer maintains the attribute "colour". A few months later, marketing needs a sales-oriented colour name such as "midnight blue" and creates the attribute "colour marketing" for it. Shortly afterwards, a marketplace requires a colour from a fixed list of values, and a third attribute, "colour marketplace", is created. And because the manufacturer's information does not quite fit into any of the three fields, the attribute "manufacturer colour" is eventually added as well.

Each of these decisions is understandable in its own right. In the end, however, there are four attributes for a single property, all of which have to be maintained differently. It is unclear which of them is authoritative for the colour filter in the online shop.

This is exactly what I call model drift. The actual structure of the model gradually moves away from its original conception, without this development ever being based on a conscious decision. Typical signs of this are:

  • similar attributes that exist multiple times with slightly different meanings
  • definitions that are interpreted differently by different roles
  • channel-specific requirements that have ended up in the central, actually channel-neutral model
  • product classes whose attribute assignment becomes increasingly confusing
  • a growing number of exceptions, special rules and workarounds
  • attributes for which nobody clearly bears responsibility for source and maintenance

The tricky part is that the system initially continues to work in this state. The costs do not show up as a failure, but as a gradually increasing effort: the model becomes harder to understand, its maintenance becomes more error-prone, onboarding new employees takes longer and every further change becomes riskier.

Data models rarely become outdated through one big wrong decision. They become outdated through many small changes without a shared decision framework.

Governance does not mean constant rebuilding.

This is where data governance comes into play. The DAMA-DMBOK defines data governance as the exercise of authority and control – that is, planning, monitoring and enforcement – over the management of data assets.3 Applied to our data model, this means that there is a binding framework within which it is decided whether, how and by whom the model is changed.

Continuous development is often equated with constant change. Good governance, however, ensures that not every new requirement automatically leads to a new attribute or a new special solution. Its most important task is therefore not adding, but weighing up.

Before any change to the model, at least the following questions should therefore be answered:

  • What specific business problem is to be solved?
  • Does the customer need this information, or does it merely arise from an internal wish?
  • Is there already an attribute that covers or could cover this need?
  • Is the requirement relevant for several product classes and channels, or only for a single case?
  • Who supplies the information initially and who is responsible for maintaining it permanently?
  • What effects arise for interfaces, exports, validations and downstream processes?

These questions alone noticeably change the discussion. Instead of talking directly about the design of a new field, the first step is to clarify whether a change to the data model is necessary at all.

To stay with the colour example: had marketing been asked with its first request whether a suitable attribute already exists and whether the requirement could not also be solved channel-specifically, four attributes would probably not have been created.

A tiered governance model: 30, 90 and 365 days

In many companies, governance is associated with a cumbersome administrative apparatus. For operating a product data model, however, such an apparatus is neither necessary nor helpful.

To enable companies to get started with a simple governance construct, a straightforward process is recommended. Important: this is a simplified model that is only intended as inspiration. In real-world use cases, the specific conditions and requirements should be incorporated into the design of such a concept.

RhythmSubjectGuiding question
30 daysChange decisionWill the model be changed – and if so, how?
90 daysClean-upWhere is the model more complex than necessary?
365 daysStrategic reviewDoes the model still fit the market, the company and customers?

Every 30 days: deciding on changes

At the operational level, a small committee assesses the open change requests. I call it the Product Data Council. Anyone familiar with the change advisory board from IT service management will recognise the idea – except that here it is about the business structure of the product data and not about changes to the IT infrastructure.

Only the roles that are actually needed should take part:

  • the person responsible for PIM or product data
  • the data steward or the person responsible for the data model
  • a representative of the requesting department
  • if required, e-commerce, marketing, sales or IT

For many companies, a monthly meeting of around 60 minutes is already sufficient. What matters is that every request is concluded with a clear decision. As a rule, five possible outcomes are enough:

  1. The requirement is included in the central data model.
  2. An existing attribute is used or extended.
  3. The requirement is solved exclusively channel-specifically.
  4. Further information is needed for a decision.
  5. The requirement is deliberately rejected.

The point is not to approve as many requirements as possible. The point is that every change is justified, documented and assigned to a responsible person. Only this traceability makes it possible to review decisions later and, if necessary, to revise them.

Every 90 days: cleaning up the model

Governance that deals exclusively with new requirements is one-sided. It has a process for adding, but none for removing. Under these conditions, the model can only grow.

At the quarterly level, the existing model is therefore specifically checked for unnecessary complexity. Questions such as these are helpful:

  • Which attributes are rarely or never maintained?
  • Where are the empty-value rates conspicuously high?
  • Which attributes are regularly confused during maintenance?
  • Where are there duplicate or overlapping definitions?
  • Which information is maintained but not used in any relevant channel?
  • Where is a clear data source or a responsible person missing?

This creates a defined process that not only extends structures, but also simplifies, merges or removes them. Removing attributes in particular is demanding in practice, since it is rarely clear who still depends on a piece of information. That makes it all the more important to ask this question regularly – and not only when a larger migration is due.

Every 365 days: checking the strategic fit

At least once a year, the data model should be considered from a strategic perspective. Unlike the other two levels, this is not about individual attributes, but about the fundamental suitability of the model:

  • Does the model still fit the current assortment strategy?
  • Does it support new sales channels and business models?
  • Does it adequately reflect customers' information needs?
  • Which regulatory requirements are foreseeably becoming relevant?
  • Where does the model already limit new product experiences today?

This annual perspective prevents a model from being technically well maintained while gradually drifting away from the market, the company and its customers. After all, a model can be entirely consistent in itself and still be unsuitable for its actual purpose.

Not every problem needs a new attribute.

Governance that only distinguishes between "include" and "reject" falls short. Many requirements can be solved without extending the central model.

Just because a marketplace demands an additional field does not mean it has to become a central product attribute. And just because two channels name the same information differently does not mean they necessarily need different data. In both cases, the cause lies not in the product information itself, but in its presentation in the respective channel.

Possible alternatives to extending the model are:

  • adjusting the assignment of existing attributes to product classes or channels
  • using channel-specific mappings that leave the central model unchanged
  • deriving information from existing data based on rules
  • preparing content only in the target channel
  • mapping temporary or one-off requirements outside the core model

A good data model is therefore not characterised by as many attributes as possible. It contains the right information in a form that is understandable, reusable and maintainable in the long term.

Governance is product experience, too.

Governance is often understood as a purely internal discipline – as a tool for order, efficiency and risk minimisation. This view is not wrong, but it falls short. Because the decisions we make about the data model have a direct effect on the later product experience.

Whether we as customers can compare products with one another depends on whether the decisive properties are available as structured attributes with consistent values. Whether we understand content depends on the precision of the underlying definitions. And which information is available to us in search, filtering, advice or personalisation at all is ultimately determined by how the product classes are modelled.

Governance should therefore not only ask: "Is the information correctly represented in the system?"

But also: "What better decision can a customer make with the help of this information?"

If the second question is missing, exactly the situation I described in my first article arises: the data model becomes ever more complete internally and ever less relevant externally.

The easiest way to get started

Getting started with continuous governance does not require setting up a comprehensive programme. A pragmatic start can consist of five steps:

  1. Collect change requests centrally. All requirements for the data model end up in one place, regardless of which department they come from.
  2. Document requests in a structured way. Each request describes at least the underlying problem, the expected benefit, the data source and the responsible person.
  3. Define decision criteria. The questions above are agreed as a binding framework.
  4. Introduce a monthly decision meeting. The Product Data Council meets in a fixed rhythm and concludes every request with a decision.
  5. Carry out the first health check after three months. The first quarterly clean-up creates a baseline for everything that follows.

This certainly does not solve every problem. It does, however, change the way the model evolves: individual ad-hoc requests become a process, and spontaneous adjustments become conscious decisions.

What decision-makers should take away

Regularity does not mean constantly tinkering with the data model. It means creating a reliable rhythm in which relevance is regularly reviewed and renewed – operationally every month, structurally every quarter and strategically every year.

For decision-makers, it should be clear that a data model is not finished at go-live, but is working under real conditions for the first time. From this moment on, changes cannot be avoided, but they can be steered. Those who forgo this will have to face the question of how their product data is structured sooner or later anyway – only under greater pressure and at higher cost.

  1. 1
    Ward Cunningham coined the technical debt metaphor in 1992 in his experience report "The WyCash Portfolio Management System" (OOPSLA '92). ↩︎
  2. 2
    Meir M. Lehman formulated the laws of software evolution from the 1970s onwards, among others in "Programs, Life Cycles, and Laws of Software Evolution", Proceedings of the IEEE, 1980. This refers in particular to the laws of "Continuing Change" and "Increasing Complexity". ↩︎
  3. 3
    DAMA International: DAMA-DMBOK – Data Management Body of Knowledge, 2nd edition, 2017. ↩︎