Two records can describe related books and still need separate lives

The attraction of a catalogue merge is easy to understand. A reader searches for a title and receives several similar results. Colleagues see duplicate-looking descriptions, duplicated maintenance and an unnecessary decision placed in front of the customer. One confident record appears to promise a cleaner shop. When we examined that proposition at Nerzavi, the difficulty was not whether repetition existed. It was whether the proposed rule could tell repetition from a meaningful difference in the publication being offered.

Our review covered 1,204 records in 217 proposed merge groups. We classified 840 source records as duplicate listings suitable for controlled consolidation and retained 364 as distinct editions or translations. The exercise cost ₹32 lakh. Those figures describe a bounded review, not an estimate for every book in the catalogue. They also do not tell a reader how many final product records remained after consolidation. The distinction matters because even a sound clean-up can acquire a larger success story than its actual evidence supports.

## Ask what the match permits someone to do

A similarity score is an observation about selected features. The business action that follows may be much stronger. Connecting two results says that a relationship may help the reader. Merging them may say that their separate identities are unnecessary. Substituting one offer for another can say that the customer will receive an equivalent product. These actions require different levels of confidence and different kinds of evidence. It is possible to have a useful match for discovery and an unsuitable match for substitution at the same time.

That was the distinction I had underestimated. I supported a rule because the repeated title and author looked persuasive. Editorial colleagues then opened examples and found abridgements, translations and revised texts. The underlying work could be related while the reading experience changed substantially. A person buying a particular translation might care about a difference the matching features treated as incidental. The existence of a high score did not remove that preference; it only made it easier for our process to overlook it.

The commercial argument for consolidation deserved a fair hearing. More records cost effort to maintain, and a confusing result page can make a good catalogue hard to use. A retailer cannot ask every customer to perform a bibliographic investigation. Keeping every historical listing separate would not have been an adequate answer. The task was to remove accidental repetition while retaining a comprehensible path through real alternatives. That required product and editorial judgement alongside the data work, rather than a contest between automation and manual care.

We found it helpful to discuss the reader's next action. If two entries were linked, could the person compare the edition details before selecting? If they were merged, which description and image would survive? If an unresolved relationship remained, could the shop still present useful choices without pretending the uncertainty was settled? These questions made the proposed data action concrete. They also exposed where a technically reversible merge could still create confusion in the customer's account of what had been offered.

## Preserve the reason for a relationship

Evidence for identity can come from different sources and can have different limitations. A supplier description may be incomplete. A publisher's edition statement may answer one question while leaving another open. A physical inspection may resolve an example without establishing a rule for every similar record. Our catalogue team therefore retained the basis of the review decisions, including the cases held for further examination. We wanted a later colleague to understand why a relationship existed, rather than simply inherit a confident flag.

That record of reasoning has a practical cost. It needs maintenance, and it can make a workflow slower than a broad automatic rule. The answer is to make the evidence proportionate to the action. A suggestion that helps someone discover another edition may tolerate ambiguity that an automatic replacement cannot. A process should not demand the same review effort everywhere merely because one difficult example exists. Equally, a large volume should not turn an unsupported identity claim into an acceptable one simply because careful review would be inconvenient.

Omar's editorial team helped us define a linked presentation. The reader could recognise a common work and then see meaningful edition choices. Technology colleagues implemented the migration and display, while buying colleagues retained the commercial offer. That division of responsibility mattered. My function could establish and explain relationships, but it could not decide every editorial interpretation or commercial substitution through a data rule. Naming those owners made the work more accountable without forcing the customer to see the internal organisation behind it.

We also retained unresolved cases. An unresolved record is not automatically a bad record; it may be an honest account of evidence still missing. It becomes a problem when its uncertainty is hidden or when nobody owns the next review. A useful catalogue needs a way to keep the book discoverable while avoiding a stronger claim than the evidence permits. That can be less visually tidy than one universal field, but the extra distinction is justified when it protects a choice that matters to the reader.

## Count the result at the level actually changed

The review changed catalogue treatment. It did not demonstrate an increase in sales, a reduction in every wrong-edition order or a universal improvement in customer satisfaction. Those outcomes would require their own evidence. I resist attaching them to the work simply because they are plausible benefits. A clearer product relationship can be valuable without a manufactured financial return, especially when the retailer can explain the specific confusion it has removed and the continuing work it has accepted.

The ₹32 lakh review expense also remains an expense. It does not disappear when the conclusion is to retain separate records. Nor should the 364 retained records be called errors prevented without a clear account of what would otherwise have happened. We know the proposed treatment changed. We do not know how every customer would have responded under the abandoned version. Keeping that boundary makes the account more credible and leaves room for later evaluation of the actual customer experience.

For a board, the useful question is broader than the particular matching method. What business action does a confident data label authorise? Who decides whether the distinction it removes is commercially or ethically material? Can the organisation explain and revisit a decision when new evidence arrives? Directors need not choose individual catalogue fields to ask those questions. They can examine whether management has matched the strength of an automated action to the evidence and judgement available behind it.

I would now begin a merge proposal by asking what must remain true for the customer after the change. That does not produce a universal formula, but it gives the teams a shared purpose. A catalogue should make related books easier to find while preserving the edition the reader actually intends to buy. The best relationship model is therefore not necessarily the one with the fewest records. It is the one that makes a useful connection without quietly taking away an important choice.