Ecommerce personalisation due diligence begins by defining what the system actually decides. It can describe a first name in a subject line, a product carousel based on what other people bought, a segment assembled by a marketer, or a system that independently recalculates what every customer is most likely to buy next. All four may be sold under the same heading. They do not solve the same problem.
That ambiguity matters because the attractive part of a demonstration is seldom the difficult part of the system. A supplier can show a polished email, a plausible recommendation and a dashboard full of attributed revenue. None proves that the software considered the whole catalogue, made a genuinely individual decision, operated without continuous campaign construction or caused sales that would not otherwise have occurred.
The buyer’s task is therefore not to decide whether personalisation sounds useful. It is to establish exactly what is being personalised, who or what makes the decision, how much of the catalogue participates, what the retailer must continue doing, and whether the commercial result survives proper measurement.
The following twelve tests provide a practical basis for that due diligence.
1. Ask for the unit of personalisation
Begin with a deceptively simple question: what, precisely, becomes individual?
If the answer is a segment, persona or journey, the system is personalising groups. If it chooses from a small manually approved collection, it is personalising within a campaign. If it rearranges products on a page only after somebody returns to the website, it is personalising an onsite experience.
Individualisation is different. The unit of decision is one identifiable customer. The system should maintain a separate, changing view of that person’s affinity for relevant products and use it to decide what to present next. Two customers who browsed the same category should not be assumed to want the same product, at the same price, for the same reason or at the same time.
Ask the supplier to describe the decision in a single sentence without using the words personalised experience. If it cannot, the proposition is not yet clear enough to buy.
2. Establish how much of the catalogue is genuinely eligible
Many systems appear intelligent because the demonstration starts after most of the difficult choices have already been removed. A marketer selects a campaign, category, collection or handful of products; the software then chooses among them.
That can be useful, but it should not be confused with perpetual catalogue-wide product selection. Ask:
- How many SKUs can be evaluated for each customer?
- Are long-tail products included or only popular inventory?
- Who creates the candidate set?
- How are new, low-data and seasonal products treated?
- Are unavailable, unsuitable or suppressed products removed automatically?
Established recommender-system literature separates candidate generation, scoring and re-ranking. That distinction is useful during procurement: a convincing ranking means little if the candidate pool was too narrow in the first place. The peer-reviewed Recommender Systems Handbook examines recommendation methods, evaluation, value, impact, fairness, novelty and diversity. In retail, stock, margin, suitability and product governance belong in that final decision too.
3. Separate similarity from next-purchase probability
“Customers who bought this also bought that” is not the same as deciding what one person is most likely to buy next.
Collaborative filtering identifies relationships between users and items. Content-based approaches identify similarity between item attributes and observed preferences. Both are legitimate techniques, but a due-diligence process should establish what the supplier’s score actually represents.
Is the model predicting similarity, a click, an open, an order, expected revenue or the next most probable purchase? These objectives are not interchangeable. A model optimised for clicks can become very good at generating curiosity while doing little for profitable sales. The evaluation literature in the Recommender Systems Handbook shows why recommendation quality must be tested against the intended task and commercial outcome: the objective chosen becomes the behaviour obtained.
For SwiftERM, the commercial question is not merely whether a product resembles something the customer once viewed. It is whether changing behavioural and transactional evidence makes that product a stronger candidate for their next purchase now.
4. Examine how the system interprets behaviour over time
A product view is not a declaration of lasting intent. It may represent interest, comparison, hesitation, a gift, an accidental click or an idea already abandoned.
Ask how the system treats:
- recency and repetition;
- the order in which behaviours occur;
- purchases that invalidate earlier browsing;
- changing brand, price, colour, size or category affinity;
- replenishment intervals;
- inactivity and decaying signals;
- contradictory evidence.
A credible answer should describe continual revision, not the permanent accumulation of labels. The same customer is not the same commercial proposition on every visit. Their evidence changes, and the product ranking should change with it.
This is where autonomous individualisation departs from a static profile. It is not a more elaborate segment. It is a decision that remains provisional because the person remains capable of surprising the model.
5. Test whether timing is individual or merely scheduled
The correct product can still fail when it appears at the wrong moment. Ask whether send-time optimisation is calculated separately for each recipient or selected as a campaign-level hour.
A supplier should explain:
- which behavioural evidence informs timing;
- whether timing updates continuously;
- how sparse or new-customer data is handled;
- how frequency, consent and suppression rules constrain delivery;
- whether the retailer can distinguish individual timing from a platform’s general “best time” feature.
The point is not to send more frequently. It is to avoid wasting a good decision at a moment when the customer is unlikely to notice or act upon it.
6. Follow the work back to the retailer
Every demonstration should be accompanied by a list of the work the retailer must still perform.
Who chooses the campaign? Who supplies the products? Who writes the rules? Who builds the journey? Who approves the audience? Who checks inventory? Who decides when the campaign ends? Who creates the next campaign when buying behaviour changes?
If those tasks remain with the retailer or its agency, the software may improve campaign execution, but it has not removed campaign dependency. This matters commercially because a system’s true cost includes internal labour, agency retainers, delay, maintenance and the opportunity cost of decisions that never get made.
The distinction is set out more fully in autonomous hyper-personalisation in email: automation repeats a defined process; autonomy makes and revises the decision within governed limits.
7. Demand an explicit account of data boundaries
“AI-powered” is not a data-governance answer. Establish what data enters the system, where it is processed, how it is secured, who can access it, how long it is retained and what happens when the relationship ends.
The buyer should map:
- behavioural events;
- product and inventory data;
- transaction history;
- customer identifiers;
- consent and suppression status;
- any inferred attributes;
- exports to other processors or model providers.
The UK Information Commissioner’s direct marketing guidance should form part of that review, particularly where profiling supports electronic marketing. Legal advice remains the retailer’s responsibility, but a competent supplier should be able to explain its role, safeguards and processing boundaries without evasive language.
Data minimisation is not in conflict with useful personalisation. The question is whether each input is necessary for a defined decision—not whether the supplier can ingest everything available.
8. Inspect governance when the model is wrong
Every model will sometimes make an unhelpful prediction. Due diligence should examine what happens next.
Can products be prohibited globally or for defined circumstances? Are stock and eligibility enforced at decision time? Can sensitive inferences be excluded? Are anomalous outputs detected? Is there an audit trail? Can a deployment be paused without corrupting customer or learning state?
The relevant test is operational rather than bureaucratic. Governance is not a document completed before launch; it continues while the system changes. Buyers should therefore require evidence of live safeguards, exception handling and accountability rather than accept a supplier’s reference to a borrowed framework as proof of competence.
9. Require evidence that survives a holdout
Attributed revenue is not necessarily incremental revenue. A customer may have purchased after receiving a personalised email and still have purchased without it.
The supplier should be able to support a proper comparison between eligible customers exposed to the intervention and a contemporaneous group withheld from it. Eligibility rules, duration, contamination, sample size and the chosen commercial outcome should be agreed before the result is known.
This need not turn every trial into an academic experiment. It does require intellectual honesty. The practical method is explained in how to measure incremental ecommerce revenue, including revenue per eligible customer, conversion lift and incremental return on spend.
Ask to see both figures: attributed revenue and defensible incremental revenue. A supplier confident in its contribution should not object to the distinction.
10. Test implementation against the unglamorous day
An integration is not proved by its launch demonstration. Ask what happens when:
- a product disappears or changes identifier;
- a feed is late;
- an order is refunded;
- consent changes;
- the email platform is unavailable;
- an event arrives twice or out of order;
- the retailer changes theme, catalogue structure or platform configuration.
The answer should include monitoring, reconciliation, failure handling and ownership. Reliable operation is often less photogenic than model design and more important to the retailer.
11. Calculate the whole cost, not merely the licence
Compare like with like. A low software fee can conceal continuing campaign labour; a larger fee can include work otherwise distributed across employees, agencies and multiple tools.
Include:
- implementation and integration;
- internal management time;
- creative and campaign production;
- agency cost;
- data engineering;
- monitoring and exception handling;
- experimentation and reporting;
- opportunity cost before deployment;
- the cost of keeping the system current.
Then calculate how many additional gross-profit pounds—not merely attributed sales—are required to cover the total. A trial should establish that threshold with the retailer’s own economics.
12. Make the supplier state what the system does not do
This is often the most revealing test.
A serious supplier will identify boundaries: what remains with the ecommerce platform, the existing email provider, the retailer’s consent management, its merchandising policy and its own people. It will distinguish current capability from roadmap, prediction from certainty, and measurement from marketing language.
SwiftERM does not replace the retailer’s ecommerce operation or general email platform. It analyses first-party behavioural, product and transaction evidence to calculate what each identifiable customer is most likely to buy next and when; it then delivers that individual selection autonomously through the established environment. The existing systems continue doing the jobs for which they were selected.
Ecommerce personalisation due diligence worth documenting
The completed due-diligence record should fit on a small number of pages. It should name the individual decision being made, the eligible catalogue, the required data, the retailer’s residual workload, the measurement design, governance controls, implementation dependencies, whole cost and evidence standard.
The most impressive demonstration is not necessarily the best system. A better supplier is the one prepared to let the buyer inspect the ordinary machinery behind the impressive moment: how products entered consideration, why one outranked another, what prevented an unsuitable selection, what work remained for the retailer and whether the resulting sale was genuinely additional.
Ecommerce personalisation due diligence is the difference between buying a persuasive use of the word personalisation and buying a capability that can be held commercially accountable.
References and accreditations
- McKinsey & Company — Nidhi Arora, Daniel Ensslen, Lars Fiedler, Wei Wei Liu, Kelsey Robinson, Eli Stein and Gustavo Schüler, The value of getting personalization right—or wrong—is multiplying — consumer expectations, repurchase behaviour and the revenue value of personalisation.
- Forrester — Brendan Witcher and colleagues, Evolve Now To Personalization 2.0: Individualization — the strategic progression from segmentation to individualisation.
- Boston Consulting Group — Mark Abraham, Mitch Colgan, Shane Fisher, Florian Kogler, Chris Lynes and Rachael Aiken, Retail Spotlight: Personalization in Action — first-party data, product recommendations and measured retail growth.
- Bain & Company, Making it personal: Rules for success in product customization — personalisation, customer loyalty and commercial differentiation.
- Peter Fader, Frances and Pei-Yuan Chia Professor of Marketing at Wharton, Customer Centricity and Why It Matters — individual customer heterogeneity, customer lifetime value and accountable targeting.
- Sheena Iyengar, Columbia Business School professor and TED speaker, How to make choosing easier — original behavioural research on choice, relevance and the commercial consequences of excessive options.
