A director asks for last quarter’s brand awareness and gets two answers: 75% from marketing and 27% from insights. Both teams pulled from the same tracker, and neither made an arithmetic error. This is the problem a single source of truth solves for research teams. It is one of the main payoffs of data harmonization for research and insights, and this guide covers what it contains, how it differs from a warehouse or a repository, and how to build one without stopping everything else.
Key takeaways
- A single source of truth means one agreed place and one agreed definition for each metric, not one database for all data.
- For research, it has four parts: a governed data layer, a metric dictionary, lineage back to study, wave and base, and access rules.
- Most conflicting numbers come from definitions (aided or unaided, weighted or not, which base), not from errors.
- Start with the 10 metrics leaders ask about most, and add more only once those are trusted.
What is a single source of truth, and what is it not?
A single source of truth is an agreement before it is a system. For every metric that matters, the business agrees on where the number lives and exactly how it is calculated. Anyone who asks gets the same answer, and anyone can trace that answer back to its source.
It is not one database holding everything. Survey data, retail scan data and CRM records can live in different places. What has to be single is the definition and the approved route to the number. A metric with two definitions is two metrics, and it needs two names.
Why do insights teams rarely have one?
Research data grows study by study, and each study tends to end as a deck. After a few years, the numbers live in slides, inboxes and shared drives, and the person who knows how a figure was based may have left.
Three patterns make it worse. Trackers run by different research agencies use different wording and scales. Teams define the same metric differently, so “consideration” means one thing to brand and another to sales. And each team keeps its own copy of the data, which drifts as people re-cut it. If your trackers already differ by market or agency, harmonizing brand trackers is usually the first job.
What does a single source of truth for research contain?
A governed data layer
This is where approved data sits: survey and tracker files, CX surveys, and outside sources such as retail scan, consumer panel and industry studies. “Governed” means data enters through a known process and is checked before anyone reports from it. Many teams import or connect survey, provider, customer and market data into one governed layer so every metric is calculated from the same files.
A metric dictionary
The dictionary is the heart of it. Each entry states the name, question wording, scale, base, weighting and calculation.
| Metric | Definition | Base | Weighting |
|---|---|---|---|
| Net Promoter Score | % scoring 9–10 minus % scoring 0–6 on a 0–10 likelihood-to-recommend scale | Customers who bought in the past 12 months | Weighted to customer mix by region |
| Consideration | % choosing the brand at “Which of these brands would you consider next time you buy?” | Adults 18–64 aware of the brand (aided) | Weighted to population by age and gender |
Lineage back to study, wave and base
Every number should carry its origin: which study, which wave, which question, which base and how many respondents. When someone challenges a figure, lineage lets you answer in minutes instead of hunting through old decks.
Access rules
Not everyone should see everything. Some studies are confidential, and some data comes with license terms that limit who can use it. Write down who can see which studies and metrics, and apply it in the system rather than by email.
How is it different from a data warehouse or a research repository?
The three are related but solve different problems.
| Single source of truth | Data warehouse | Research repository | |
|---|---|---|---|
| Main job | One agreed answer per metric | Store and query large volumes of company data | Help people find past studies and findings |
| Holds | Governed data plus metric definitions and lineage | Transactional and operational tables | Reports, findings, verbatims, tags |
| Unit | The metric | The table | The study or insight |
| Fixes | Conflicting numbers | Scattered raw data | Forgotten research |
A warehouse can hold the data without settling what “awareness” means. A research repository makes past work findable, but two reports in it can still disagree. You often want all three, with the dictionary connecting them.
Why do two teams report different brand awareness for the same quarter?
Here is the opening example worked through. The numbers are illustrative. A quarterly tracker interviews 1,000 adults aged 18–64, split 500 and 500 between two age groups. The population is 40% aged 18–34 and 60% aged 35–64, so the weights are 0.4 ÷ 0.5 = 0.8 for the younger group and 0.6 ÷ 0.5 = 1.2 for the older group.
| Age group | Respondents | Weight | Aided aware | Unaided aware | Category buyers | Buyers aided aware |
|---|---|---|---|---|---|---|
| 18–34 | 500 | 0.8 | 340 | 160 | 350 | 280 |
| 35–64 | 500 | 1.2 | 245 | 120 | 250 | 170 |
| Total | 1,000 | 585 | 280 | 600 | 450 |
Marketing reported aided awareness among category buyers, unweighted. Insights reported unaided awareness among all adults, weighted. Walk from one to the other one change at a time:
| Step | Definition | Arithmetic | Result |
|---|---|---|---|
| Marketing’s number | Aided, category buyers, unweighted | 450 ÷ 600 | 75.0% |
| Apply weights | Aided, category buyers, weighted | (280 × 0.8 + 170 × 1.2) ÷ (350 × 0.8 + 250 × 1.2) = 428 ÷ 580 | 73.8% |
| Widen the base | Aided, all adults, weighted | (340 × 0.8 + 245 × 1.2) ÷ 1,000 = 566 ÷ 1,000 | 56.6% |
| Insights’ number | Unaided, all adults, weighted | (160 × 0.8 + 120 × 1.2) ÷ 1,000 = 272 ÷ 1,000 | 27.2% |
Weighting moved the number 1.2 points, the base 17.2 points and the question type 29.4 points, which adds up to the 47.8-point gap. Nobody was wrong. One label covered three different metrics.
The dictionary resolves it by naming each one. “Brand awareness” becomes aided awareness among all adults 18–64, weighted: 56.6%. “Unaided awareness” (27.2%) and “aided awareness among category buyers” (73.8%) get their own entries. When the base or weighting is set in the definition, tools like crosstabs and banner tables with weighting and significance testing apply it the same way every time.
How do you build a single source of truth incrementally?
Trying to define every metric at once stalls. Build in this order:
- List the 10 metrics leaders ask about most. Look at the last few quarterly reviews and board packs. Awareness, consideration, NPS, satisfaction and market share usually make the list.
- Write a dictionary entry for each. Agree wording, base, weighting and calculation with every team that reports it. Where teams genuinely need different versions, give each its own name.
- Map each source to the entries. Note which trackers, CX surveys, retail scan feeds or CRM fields feed each metric, and where definitions differ by source. Joining survey and customer records is covered in customer data integration.
- Record lineage and set access rules. Tag every published number with study, wave, base and sample size, and decide who sees what.
- Retire the old copies. Point people to the new source and stop updating the spreadsheets it replaces.
- Add metrics in batches. Once the first 10 are trusted, add the next tier, such as brand attributes or outside sources like syndicated data.
What are the common mistakes?
- Building the database before the dictionary. Moving data into one place without agreeing definitions just puts the conflicting numbers side by side.
- Forcing one definition on everyone. Sales may genuinely need awareness among buyers. Name both metrics rather than picking a winner.
- Dropping the base. A percentage with no base and no sample size cannot be checked. We see this most often in figures copied from decks.
- Overwriting definitions when a questionnaire changes. Version them, or your trend line silently breaks.
- Treating it as a one-off project. New studies arrive every quarter. Someone has to own the dictionary and approve changes.
What should you do first?
Pick the metric that caused the most recent argument, write its dictionary entry with the teams involved and recalculate last quarter’s number under the agreed definition. That single fix shows people what the wider effort buys. As more studies join, connecting findings across studies and waves lets teams build on what is already agreed instead of starting over.