Big Data: decisions that start with a good question

Big Data can expand what an organization observes, but gathering information does not resolve a decision by itself. Before choosing tools or building a dashboard, specify which question matters, which data can answer it and what limitations interpretation will have.

Big Data does not replace defining the problem

A specific question distinguishes useful data from available data. If a team wants to improve a campaign, clarify which part of the journey is insufficient: awareness, participation, inquiries or commercial continuity. Mixing objectives produces large reports and vague decisions. The first analysis can be small if it identifies missing information and what is unnecessary for the actual problem.

Also decide which action might change because of the result. If no answer would change the work, the question may not yet be well formulated. This does not make exploration useless, but distinguishes it from evaluation. Exploratory analysis can reveal patterns; turning them into a recommendation requires explaining context, reviewing alternatives and recognizing which additional evidence would make the decision more reliable.

Understand the architecture before accumulating sources

NIST’s Big Data work develops an interoperability framework with components and interfaces independent of a particular provider. Its existence does not demonstrate that a company needs a large-scale platform. The useful idea is understanding information flow, who produces it and which transformations happen before it reaches someone who must interpret it.

One source may record posts, another inquiries and another sales. Before combining them, describe their units, timing and constraints. Without a suitable shared identifier, a join may create mistaken associations. If an update arrives late, the dashboard may present an incomplete period as finished. Drawing the journey reveals these problems before trying to resolve them through more charts or processing capacity.

Define comparable units and periods

A figure makes sense only alongside its definition. A connection, session, account, person and purchase are distinct units. A report should explain what it counts and what cannot be inferred. It also needs a period, time zone and criterion for recording late events. Otherwise, sources may seem contradictory while actually observing different phenomena.

Campaign comparisons must retain relevant conditions. A multi-day activity is not equivalent to a brief interaction, nor an invited audience to passersby. The article on measuring brand experiences offers questions for beginning this work. The aim is not to make comparison impossible, but document limitations so a numerical difference does not automatically become a commercial conclusion.

Review quality before interpreting a pattern

Missing data does not mean zero, and a repeated record does not necessarily represent another person. Identify duplicates, empty fields, format changes and coverage differences. Then decide how each situation will be handled. That decision is part of analysis and should accompany the result, especially when cleaning changes case counts or excludes a group unevenly.

A useful check compares a sample against its source and reviews examples the process classifies incorrectly. Opening the file and seeing apparently consistent fields is insufficient. If a provider changes a definition, the team needs to know when and which periods remain comparable. Keeping that history prevents a technical variation being presented as changed audience behavior.

Separate observation, association and explanation

Two variables changing together does not prove one causes the other. A campaign may coincide with more inquiries while season, price or product availability also changes. List possible explanations and seek evidence distinguishing them. Sometimes descriptive interpretation suffices; other questions require a more demanding evaluation design and specialist knowledge.

Communicating analysis must respect that evidence level. Phrases such as “was observed,” “coincided” or “requires checking” can be more precise than total-attribution promises. A team does not lose value by explaining uncertainty. It does when an overly confident conclusion drives investment, stops an action or changes an offer without reviewing potentially influential external factors.

Design a view that supports decisions

A dashboard should answer questions, not exhibit every database column. Begin with a period summary, source status and a few objective-linked metrics. Then allow access to detail needed to understand differences. Incomplete snapshots and definition changes need visible signals: hiding them in distant notes encourages incorrect interpretations.

Presentation also belongs to visual content work. A scale, label or color can change what readers understand. Check whether someone outside the analysis recognizes the unit, period and main limitation. If a lengthy explanation is needed to avoid a basic mistake, the view still needs adjustment.

Work with the minimum necessary information

Data’s usefulness does not remove obligations to handle it carefully. Analysis must distinguish public, aggregated, confidential and personal information. Before gathering sources, identify purpose, authorized people and retention appropriate to context. This article does not establish a legal basis or replace privacy review; it proposes including these decisions in design rather than postponing them until after building the system.

For a prototype, a fictional dataset preserving the relationships to be tested may suffice. This practice allows calculations, filters and presentations to be reviewed without copying real records into a new environment. A demonstration working with that sample does not establish permissions or operation with customer data. Separating testing and implementation allows progress without turning a technical test into implicit authorization.

An exercise to discover what is missing

Imagine a fictional campaign with many views and few inquiries. The first reaction should not be blaming the content. The team can check whether the link works, the offer is understandable and inquiries are recorded in the same period. It can then compare pieces and journeys. This exercise shows that an aggregate signal must connect with its producing process before justifying intervention.

The conclusion might be that information is missing to distinguish interest from difficulty. The next task is then improving records or testing a bounded variation, not manufacturing an explanation. Document what is known, assumed and which data would change the decision. That small map makes analysis more useful than a general recommendation based only on one number being below another.

Close analysis with a reviewable decision

The result should conclude with an action, owner and review condition. It may mean maintaining a campaign, adjusting a page, correcting a source or stopping an invalid comparison. Not every conclusion requires further investment. Sometimes the most valuable decision is to stop using an indicator that does not answer the question or acknowledge an experiment still lacks sufficient coverage.

If you want to organize a data and marketing project, describe it to GX with its question, sources and known constraints. The link implies neither automatic system connections nor verified operational receipt. A useful conversation starts by identifying which decision needs support; then information, visualization or testing appropriate to supporting it can be evaluated.

A check to bring to the review. Before distributing a report, ask someone to reconstruct a figure from its authorized source. They should identify the filter, unit and period without needing a new explanation from its author. If the number cannot be reproduced, record the difficulty before deciding with it. This review need not reveal individual data: a synthetic sample and calculation documentation can be used. The point is checking that interpretation does not depend on untransferable personal memory.

Explore more insights in our X Files Blog

Measuring brand experiences with clear criteria · The X Files Blog · GX: Useful measurement begins with the question you need to answer, not the biggest figure.

Visual content: bringing clarity to the brand message · The X Files Blog · GX: Choose images and formats for what they help explain, not just their visual impact.

Sources

NIST: Big Data Interoperability Framework. Questions, the fictional exercise and review sequence are editorial work; they do not establish GX implementation or results.