Data Provenance and Due Diligence

Data provenance is the documented chain that connects a published figure to the raw observation behind it, including every transformation applied on the way. Due diligence is the procurement work that tests whether that chain exists, holds up, and will still be there in three years.

The two belong together because provenance is only useful if someone checks it, and buyers rarely do. A methodology page is read once during evaluation and treated as a credential. Provenance is what a buyer needs eighteen months later, when a figure is challenged and the question is not "how is your data made?" but "where did this number come from?"

Provenance is not methodology#

A methodology document describes the general process. Provenance is specific to a figure. The distinction is easy to state and consistently collapsed in practice, so it is worth being blunt about it:

A vendor can have a good methodology page and no provenance chain. That combination is common, and it is comfortable right up until the first serious challenge.

The six links in the chain#

Every published figure passes through some version of these stages. A provenance record should be able to say what happened at each.

1. Collection — which source, on what date, under what access arrangement. See Market Intelligence Data Sources for the source families and what each can support.
2. Cleaning — what was removed and why. Duplicate records, invalid entries, non-genuine activity. Removal rules are judgement calls and should be written down.
3. Classification — how a record was assigned to a brand, a category and a geography. This is where most silent error is introduced, because classification is largely automated and largely unreviewed.
4. Normalisation — how records from sources with different conventions were made comparable. Currency, unit, period boundary, variant handling.
5. Estimation — where a figure was inferred rather than observed, by what model, with what assumptions. The single most important link to have marked explicitly.
6. Aggregation — how individual records were combined, and what was excluded from the total.

The estimation link deserves emphasis. Most commercial market intelligence contains estimated figures, and estimation is legitimate. What is not legitimate is presenting an estimate in the same visual register as an observation, with no marker distinguishing the two. A buyer who cannot tell which figures on a dashboard are observed and which are modelled cannot calibrate how much weight to place on any of them.

What good provenance documentation contains#

• Per-source statements, not one aggregate statement. Sources differ, so their chains differ.
• Dated methodology versions. A methodology without a version history implies nothing has ever changed, which is never true of a dataset with multi-year history.
• A stated restatement policy. When methodology changes, is history restated to match, left as published, or forked into a parallel series? All three are defensible; silence is not. See Data Refresh Frequency, where the same policy governs revision after publication.
• Marked estimation boundaries. Which figures are observed, which are modelled.
• Known limitations, named. Every dataset has them. A limitations section that exists and is specific is the strongest single credibility signal in vendor documentation, because it is the section with no marketing incentive to write.

Due diligence beyond data quality#

Sample testing — covered in Sample Data Verification — establishes whether the data is good. Due diligence establishes whether the arrangement is sound. Four areas that no amount of sample testing reaches:

Sourcing basis. On what legal and commercial footing is each source obtained, and does the intended use fall inside it? The question that most often goes unasked: may the buyer redistribute figures to clients, publish them in marketing material, or include them in a filing? Those are three different permissions and a general subscription may grant none of them. Establish the answer in writing before the dataset is designed into anything client-facing.

Continuity. What happens to the buyer's history if a source relationship ends, a platform changes its access terms, or the vendor is acquired? A dataset is a dependency, and a multi-year analysis built on one inherits its fragility. Ask what the vendor's contingency is, and whether historical data already delivered remains usable if the relationship ends.

Change management. How are methodology changes versioned, dated and communicated to existing customers? A change communicated only in release notes nobody reads is, from the buyer's perspective, an undisclosed change.

Concentration. How much of the dataset rests on a single source or a single relationship? Concentration is where a sudden coverage gap would appear, and it is invisible in any sample.

A one-hour provenance test#

Run this during evaluation, on the sample dataset, before the commercial conversation gets serious.

1. Pick one figure you would actually cite in a decision.
2. Ask which source family produced it, and whether it is observed or estimated.
3. Ask to see the underlying records that compose it, and confirm the components sum to the aggregate.
4. Ask what the same figure was six months ago, and whether it has been revised since.
5. Ask what would have to change in the methodology for this figure to move by a material amount.

The fifth question is the most revealing and the least expected. A vendor that understands its own chain can answer it concretely — naming the classification rule or the estimation assumption the figure is most sensitive to. A vendor that cannot answer it has not decomposed its own numbers, which means nobody has, which means the chain is not there to follow when it is needed.

Where to look next#

For the source families that begin the chain, see Market Intelligence Data Sources. For the tests that establish data quality, see Sample Data Verification. For the full comparison framework, see Market Intelligence Platform Comparison Criteria and How to Evaluate Market Intelligence Providers.

Common questions#

What exactly is data provenance in a market intelligence context?#

Provenance is the recorded chain that connects a published figure back to the raw observation it came from, together with everything that was done to it along the way — collection, cleaning, classification, normalisation, estimation and aggregation. It is not the same as methodology. A methodology document describes the process in general; provenance is the specific answer to the question "where did this particular number come from?", for a particular number. A dataset with a published methodology but no provenance chain can tell a buyer how numbers are made in principle, and nothing at all about the one being disputed in a meeting.

Why does auditability matter if the vendor is reputable?#

Because the moment auditability is needed is the moment reputation has stopped settling the argument. Figures get challenged — by a competitor's agency, by a finance team reconciling against internal sales, by a regulator, by a board member who has read something different. At that point the useful asset is the ability to drill from the disputed aggregate down to the underlying records and show what it is composed of. A vendor's reputation is what gets a dataset bought; auditability is what keeps a conclusion defensible after it has been acted on. The two are not substitutes, and only one of them survives contact with a challenge.

What should due diligence cover beyond data quality?#

Four areas that quality testing does not reach. Sourcing basis — on what legal and commercial footing is each source obtained, and does the buyer's intended use, including redistribution to clients, fall inside the licence granted. Continuity — what happens to the buyer's history if a source relationship ends or a platform changes its terms. Change management — how methodology changes are versioned, dated and communicated. And concentration — how much of the dataset depends on a single source or a single relationship, since that is where a sudden gap would appear. None of these show up in a sample dataset, and all of them affect whether the dataset is still usable in three years.

How does a buyer test provenance during an evaluation?#

Pick one number from the sample dataset and follow it all the way down. Ask which source family produced it, what processing it went through, whether any part of it is estimated rather than observed, and to see the underlying records that compose it. Then ask what that same number was six months ago and whether it has been revised since. The exercise takes under an hour and is unusually diagnostic: a vendor with a real provenance chain answers from records, a vendor without one answers from general description and moves the conversation back to the methodology page.