Marketing Analytics
Customer and marketing analytics you can defend
Segmentation, customer value, retention, look-alike, promotion and campaign measurement models for UK financial firms — where choosing who is offered credit, an account or a price is a decision a firm has to evidence as fair. Built the same way as everything else we do: the method written down, every segment re-derivable from the data, and every campaign judged against a group that did not receive it.
Segmentation and campaign work delivered for international retail, consumer-credit and membership brands.
Targeting Record
Re-derivableCustomer data
Segment assigned
Offer targeted
Outcome compared
Fairness checked
Every customer’s segment traces back to the data and the method that placed them there — and every offer made on it can be shown to have been fair.
The shift
Targeting became a conduct question
Under Consumer Duty, who is shown which product, at which price, is an outcome a firm has to evidence as fair — including for customers in vulnerable circumstances. A segment is no longer only a marketing idea. It is a decision about people.
The gap
The segments are real, the evidence is not
Most segmentations live in a deck. The method that put a customer in a segment was never written down, so nobody can show which customers received which treatment, or why — and a campaign is judged on a before-and-after that cannot tell the campaign from the season.
The standard
A segment you can re-derive
Every customer’s segment reproducible from the data and a written method. Every campaign read against a group that did not receive it. Every input that could stand in for a protected characteristic tested for doing so.
Know the customer
Who the customers are, what they are worth, who a contact saves, and what to send them
The models a customer strategy rests on, each reported in margin rather than revenue. Each is built so the firm can re-run it without us.
Customer segmentation
Groups customers on how they buy — frequency, margin, category, channel, sensitivity to promotion, returns. The number of groups is chosen by how reliably they come back on resampled data, not by eye. Each market is compared by who it has and how they behave, separately, and a rule is handed over that places a new customer in a segment.
Customer value
The margin each customer is expected to bring next year, after discounts and returns — not the revenue. Checked against a year the model did not see. A heavy spender who buys on discount and returns much of it tops a revenue list and brings little margin; this is the model that shows it.
Retention
Who a contact actually saves, learned from a randomised offer test — not who is most likely to leave, because many of those leave whatever is sent and some are pushed away by being contacted. The saving is measured on customers the model never saw, with an interval.
Next-best-action
One action per customer — a message, an offer, a discount, or nothing — judged against the policies a team would otherwise run before anything is sent. Where it cannot show it beats a simpler policy, it says so.
Find the customer
Where the customers a firm wants are, and how many of them
Acquisition models built from the firm’s own best customers outward. These are the ones most likely to lean on area and credit data, which is why they carry the heaviest testing below.
Geographic opportunity and look-alike
Where more of the customers a firm wants live, and how much margin is waiting there — checked on regions the model did not see, with areas too far to serve kept off the list. The same profile scores individual prospects, and whether the opportunity tracks the age or deprivation of an area is measured, not assumed.
Prove the spend
What a campaign, a promotion or a channel actually caused
Measurement that separates the campaign from the season, the stock position and the spend it was never going to need. Every result carries an interval, and one too wide to act on is reported as not determined rather than as nothing.
Test design and readout
Every test sized before it runs and read after it with one of four answers: lift, harm, nothing material, or not determined. A test too small to see an effect worth acting on says so, and how many customers it needed — never “no effect”. The other models are judged by it.
Promotion and price
Each promotion read as margin after the discount on sales that would have happened anyway, the dip the week after, and the sales taken from its sister product. Weeks where the product sold out are not read as low demand. Markdowns are planned from the same fit.
Media mix
What each channel adds, with carry-over and the point past which more spend stops paying back. Paid search is calibrated to a regional test, because its spend follows demand. Each channel is called grow, hold or cut in margin — or not determined where the evidence is thin.
Retail media measurement
What each campaign added, by a holdout that never saw it, read by one method for every campaign — set beside the attributed figure usually reported, which counts what targeted customers would have bought anyway.
What makes it defensible
Six properties decide whether a segment will stand up
The same test we apply to a fraud rule or a credit model. Where a property cannot be evidenced we say so, and never record it as met — an untested proxy is not a clean one.
01
Method, written
How a customer is assigned to a segment, in a form somebody else can follow — not a notebook only its author can run.
02
Re-derivable
The same data and the same method land every customer in the same segment months later, or the difference is explained.
03
Outcomes compared
What each segment was offered and what it received, read side by side, including for customers in vulnerable circumstances.
04
Proxies tested
Postcode and area data can stand in for age or ethnicity. Whether they do is tested and recorded, not assumed away.
05
Data basis stated
The lawful basis for each source, and the licence terms for census and credit-reference data used in marketing.
06
Effect measured
A claimed uplift is read against a group that did not receive the campaign. A before-and-after is a description, not a result.
Where this has been done
Four engagements, names withheld
Delivered for retail, consumer-credit and membership brands. The methods carry straight across to a bank, a lender or an insurer; what a regulated firm adds is the evidence each segment was fair.
A global fashion retailer
Over a million customers across four markets segmented into distinct groups, each with a data-driven persona, product preferences and an activation plan for email and paid social — alongside a review of where the firm’s own customer data could not yet support the strategy.
A UK credit retailer
The customers the firm most wanted to acquire defined from its own book, then located: census, credit-profile and spending-profile data combined to show where more of them live and how much value was waiting there.
A sportswear brand
A national campaign read across media spend, impressions, sales and stock. Impressions stopped growing with spend well before the budget did — and the fall in sales traced to products selling out, not to the media.
A weight-management membership
Members segmented by how they engaged in their first month, and followed over their first six months. Early engagement separated the groups; most members still lapsed, which reframed the retention problem.
How engagements are shaped
Start with the segmentation you already use
Most firms already target on some segmentation, often one nobody can re-derive. The first piece of work reads it and says where it stands, on documentation alone.
Start here
Segmentation review
Fixed fee, one segmentation
- The segments and targeting rules already in use, read cold
- Scored against the six properties above
- Proxy and fair-outcome exposure named
- Documentation only: no customer data, no system access
Then
Build
Scoped on the question, not the technique
- Any of the nine models above, scoped to the firm’s question
- Method written so the firm can re-run it
- Fair-outcome and proxy tests recorded as results
- The firm owns the model, the code and the record
Ongoing
Campaign measurement
On the campaign calendar
- Test and control groups designed before launch
- Media mix and promotions read in margin
- Segments re-derived as the customer base moves
- Independent review stays with the firm or a third party
Where this stops
We build the model and the evidence. We do not run the campaign.
Buying media, writing the creative and operating the CRM stay with the firm and its agencies. What we hand over is the segmentation, the targeting logic and the measurement design, with the record that shows each was fair and each can be re-run.
The independence line that governs our other work governs this too: a firm that commissions a targeting model from us has not also commissioned its independent review. Where the same customers are being scored for credit, that model is credit risk work, and the pipeline that runs either every day is AI & automation.
Starting
Bring the segmentation you would least like to explain
The one a campaign was targeted on last quarter, with the method nobody wrote down. That conversation takes about twenty minutes and needs no customer data. Every enquiry is treated in confidence.