Chief Data Officer / India
Keep the distinction the customer needs to make
- Book retail estate
- 34 shops
- Online catalogue and 620 company employees
- Data and research function
- 26 colleagues
- Four leads;₹6.4 crore function expense and separate₹1.9 crore catalogue programme
- Catalogue and evidence work
- 24 years
- From July 2002 analysis to CDO in January 2023
Decisions & consequences
01Linking related editions without merging the reader's choice
In 2023 a catalogue improvement proposed merging records that shared strong title and author similarities. Duplicate-looking results frustrated customers and increased the work of maintaining descriptions. The buying team wanted a simpler offer, and I initially supported the rule because the visible repetition seemed unnecessary. Before release, catalogue colleagues found examples in which the apparent duplicate was a translation, a revised edition or an abridgement. A cleaner result could have hidden a meaningful difference in what the customer would receive.
We reviewed 1,204 records within 217 proposed merge groups at a cost of ₹32 lakh. The review classified 840 records as duplicate listings suitable for controlled consolidation and retained 364 as distinct publications. Those are counts of source records, not the number of final products or customers affected. We kept an explicit record of the evidence supporting each decision. Similarity remained a way to propose relationships for review; it ceased to be sufficient authority for collapsing them.
Omar Qalzeri argued for a linked presentation in which a reader could move between editions without treating them as interchangeable. The commercial alternative was to keep the original merge rule and rely on customer-service corrections for unusual cases. That would have reduced maintenance work sooner, but it placed the discovery of an important distinction after the customer had made a choice. I recommended the linked approach, with edition and translation details retained and unresolved cases held for review. Technology colleagues implemented the display and controlled migration.
The outcome was a revised catalogue structure, not evidence of a measured rise in sales or a complete end to wrong-edition orders. The ₹32 lakh review remained a cost, and keeping separate records created continuing stewardship work. I also recorded my own mistake: I had treated matching confidence as though it settled the business meaning of a match. The more useful question was whether substituting one record for another would preserve the choice the reader believed they were making. That question now precedes a recommendation to merge, rather than appearing only after a problematic example has been found.
Keeping transliterated queries visible in a search decision
In 2024 we compared the existing catalogue search with a proposed candidate using an 800-query evaluation set. We deliberately included four equal groups of 200: English-title queries, original-script queries, transliterated queries and mixed descriptive queries. This was a diagnostic design, not a claim that each group represented one quarter of live traffic. Editorial reviewers agreed what would count as an acceptable result before comparing the systems, including when more than one publication could reasonably answer the request.
The candidate produced acceptable results for 186, 172, 116 and 150 queries respectively, or 624 of 800 overall. The existing system produced 168, 154, 144 and 138, or 604 overall. The aggregate favoured the candidate, but the transliterated group fell from 144 to 116. A customer writing a familiar title in a different script could be worse served by the proposed improvement. I had initially expected the overall score to provide a straightforward release decision; it did not describe the distribution of the trade-off.
Product colleagues reasonably wanted to release the stronger aggregate performer and improve weak cases later. Omar argued that the catalogue should not quietly make a common route to regional-language books less useful. We spent ₹18 lakh on the evaluation, including review and disagreement resolution. I recommended retaining the existing route while the candidate's transliteration handling was revised, with the four group results carried into the next review. The accountable product and technology owners accepted that recommendation.
We did not claim the balanced set measured live conversion, revenue or the experience of every language community. Nor did the decision establish that the current system was generally superior: it had fewer acceptable results overall on this set. The result was an explicit release trade-off and a defined weakness to address. The review cost was incurred even though the candidate was held. In a future evaluation I would agree the treatment of meaningful subgroup regressions before colleagues see an aggregate winner, so the decision rule does not appear to be invented to protect a preferred conclusion.
Leaving nonrespondents outside a subscription forecast
A 2025 reader study invited 900 adult customers from a defined recent-purchaser list to discuss a possible curated-book subscription. Four hundred and fifty completed the survey. Of those respondents, 279 preferred the proposed monthly package, or 62%. A commercial paper proposed applying that percentage to the wider customer base when modelling an ₹84 lakh launch programme. The study cost ₹21 lakh. It had generated a clear opinion among respondents, but we had not designed it to establish a representative adoption forecast.
The invitation frame already excluded people who had not purchased recently. Within it, half the invited customers had not responded. Commercial colleagues argued that those willing to answer were the most plausible early subscribers, so their enthusiasm should guide a pilot. That was a useful hypothesis about a possible audience. It did not establish how many such people existed across all shoppers, whether they would pay at the proposed price or whether silence meant disinterest. I opposed relabelling the response rate as a measure of demand.
We separated the 450 respondent accounts from the 450 nonresponses and retained the 279 preferences as a descriptive finding. Available purchase-frequency information helped us examine who had answered, but weighting on that one characteristic could not establish that unobserved reading habits or enthusiasm were balanced. I recommended removing the population adoption calculation and asking a narrower commercial question through a separately approved follow-up. Management held the ₹84 lakh broad launch proposal while considering that next step.
The outcome was not a finding that subscription demand was low, and it was not ₹84 lakh saved. The launch remained a proposal that might later be revised; the ₹21 lakh research expense had already been incurred. The interviews supplied reasons people liked the offer and questions to test, which were useful without being a representative sales forecast. My own responsibility was to specify the intended inference when the research was commissioned. I should not have allowed a percentage with a clear respondent denominator to arrive in a commercial model with a much larger, unexamined population behind it.
Separating publication sequence from suggested reading order
In early 2026 we reviewed 180 book series whose catalogue records contained a single order field. The field was being used for several purposes: publication sequence, an explicit publisher sequence and an editor's preferred starting point. Some companion volumes fitted more than one path. A reader could reasonably interpret a number as an authoritative order even when it reflected a bookseller's suggestion. The project cost ₹26 lakh and was led jointly with editorial colleagues.
The review found 74 series with a sequence explicitly supported by publisher material, 58 with an editorially recommended path and 48 for which no reliable order had yet been established. These were mutually exclusive classifications for the reviewed set, not a statement that publication dates were unavailable or that every series required one fixed path. Omar wanted the source and purpose visible. The merchandising alternative was a single simple order that would keep the browsing interface consistent and avoid explanatory text.
I recommended separate relationships for the different uses. A publisher-supported sequence carried its source; an editorial suggestion was labelled as such and retained its review owner. The 48 unresolved series remained browsable without an invented sequence number. Publication dates continued as dates rather than becoming a universal instruction about where to start. Product colleagues adapted the presentation so the added distinction did not require a reader to understand the internal data model.
The immediate result was clearer provenance and responsibility for the reviewed series. It did not prove that every reader followed the recommended route, or that one interpretation of a complex series was permanently correct. New volumes and revised editorial views require maintenance. We also accepted that the interface would be less mechanically uniform than a single numbered list. I would make the field's intended meaning explicit earlier in any new catalogue feature: a convenient place to store a number is not enough reason to make different kinds of authority look the same to a customer.