A shopper’s first name in a headline changes almost nothing about whether they buy. It is close to the least valuable thing an online store can personalise, and it is very often the first thing built, because it demonstrates well in a meeting and takes an afternoon. The versions that move revenue are duller. A returning customer who does not have to hunt for the size they bought last time. A catalogue of several thousand items that opens on the sixty a visitor could plausibly want. A restock notice that goes to the people waiting on one variant rather than to everyone who ever looked at the range.
Whether any of that repays the licence fee and the build time depends on two things a vendor demonstration rarely shows: how much behavioural data the store produces each month, and how much choice the catalogue actually presents. Where both are substantial, personalization is one of the more defensible investments available to an established retailer. Where either is thin, it produces attractive modules that no algorithm has enough evidence to fill, and a competent ecommerce personalization agency will often say so before quoting for the work.
What Counts as Personalization, and What Is Merchandising With a Condition Attached
BigCommerce describes the category in terms of software behaviour, noting that “Personalization platforms use AI and machine learning-driven algorithms to track onsite behavior and customer data points in real-time to deliver a unique experience to each site visitor”. That is a fair description of the upper end of the market. It is not a description of most of what stores actually run.
In practice the work divides into two families with very different economics. The first is rules-based. Somebody defines a condition, such as customers in a particular country or customers who have ordered three times, and attaches different content to it. Shopify’s own segmentation tooling sits here: customer segments let a merchant “group customers who have similar characteristics” using attributes such as order history, amount spent, subscription status and location. Adobe Commerce takes the same approach further into merchandising, where related product rules “give you the ability to target the selection of products that are presented to customers as related products, up-sells, and cross-sells”, and each rule can be tied to a customer segment.
The second family is model-based. Nobody writes the rule. A system infers what to show from patterns in behaviour across many sessions and many orders. This is the version that carries the higher price, the higher expectations and, critically, a minimum data requirement that rules-based work does not have.
The distinction matters commercially because the two fail in opposite ways. Rules-based personalization fails through neglect. Somebody writes twelve rules, leaves the company, and the store spends two years showing winter accessories to Australian visitors. Model-based personalization fails through insufficiency. The system is correctly installed, the module renders, and the suggestions are close to random because there was never enough evidence underneath them.
The Precondition Most Proposals Leave Out: How Much Data the Store Produces
Prediction has a floor, and the platforms that publish theirs are unusually candid about it. Google Analytics 4 will only generate predictive audiences, such as likely seven-day purchasers, once a property clears a stated volume test. The documentation requires that “In the last 28 days, over a seven-day period, at least 1,000 returning users must have triggered the relevant predictive condition (purchase or churn) and at least 1,000 returning users must not”. Eligibility is also conditional on the model holding up, and Google states that if quality falls below its threshold, Analytics stops updating the predictions and they may disappear from the interface.
Read that as a benchmark rather than a universal rule, because every vendor sets its own thresholds and most do not publish them. The useful part is the order of magnitude. A store with a few hundred returning purchasers per month is not in the same category as one with several thousand, and no amount of tooling closes that gap. It is one of the few constraints in ecommerce that cannot be bought around, because the input is your own trading volume.
The same limit shows up in what platforms give away for free. WooCommerce documents that “Related products are automatically generated by WooCommerce based on shared tags or categories with the currently viewed product”, that these blocks are “sorted randomly”, and that a merchant cannot select related products manually. That is not a recommendation engine. It is a category filter with a shuffle, and it will happily fill a product page with items nobody has ever bought together.
Shopify’s baseline is stronger but has a similar shape. Its theme documentation separates related products, described as “a mix of products that are similar to a product the customer is interacting with”, from complementary products, which are the add-on items shown in a “Pair it with” block. The important sentence for anyone budgeting this work is the next one: only the related set is auto-generated, while complementary recommendations need to be set up manually. The storefront API reflects the same split, exposing an intent parameter whose “accepted values are related and complementary” with a result limit that ranges from 1 to 10.
So the free tier of both major platforms produces either a rough approximation or a manual curation task. Neither is a reason to avoid personalization. Both are reasons to know which of the two you are buying, because the second one carries an ongoing labour cost that vanishes from most business cases. Curating complementary items across a catalogue of two thousand products is a merchandising job with no end date, and it needs an owner before it needs a budget.
Where Personalization Reliably Pays Off
The categories below share one property. In each, the personalised version removes work from the shopper rather than adding persuasion. That is the pattern worth looking for, and it is a more reliable filter than any list of tactics.
- Large catalogues where the visitor cannot see the range. Past a few hundred active products, the shopper’s problem changes from choosing to finding. Narrowing what is shown is then a navigation improvement, and it tends to survive contact with reality better than most conversion tactics.
- Replenishment and repeat-purchase ranges. Consumables, spares, filters, refills and trade supplies generate the repeat-order data that models need, and the shopper genuinely benefits from not re-specifying the same item. The signal quality here is usually the best a store has.
- Region, currency and delivery context. BigCommerce makes the point plainly, observing that a personalised homepage may see a lower bounce rate “because the content is curated according to the shopper’s IP address, ensuring the language, currency and shipping costs match their location”. This is often the cheapest form of the work and the least glamorous.
- Lifecycle and post-purchase email. Segmented messaging runs on order history rather than on live session data, so it clears the volume bar far earlier than onsite prediction does. For most mid-sized retailers it becomes measurable here before it does anywhere else.
- Returning-customer convenience. Remembered sizes, saved addresses, reorder shortcuts and back-in-stock alerts for a specific variant. None of it demonstrates impressively, and it rarely fails.
The catalogue case deserves more than a line, because it is the one most often underestimated. WD Market’s rebuild of Evelatus involved a catalogue the case study describes as 860,000 products across three markets, and the published account of that project puts the emphasis on Elasticsearch for search across that range and on Klaviyo for segmented email, rather than on onsite prediction. That ordering is deliberate. At extreme catalogue size, retrieval and segmentation carry most of the value, and the sophisticated layer sits on top of them rather than in place of them.
The email case is worth understanding for a different reason. It is the only one of the five where a store can usually act immediately, because the underlying data is transactional and already exists. Klaviyo’s own guidance is notably restrained about how far to take it, warning readers to “not create 30 separate campaigns because you have 30 data points” on the grounds that this level of personalization is not effective, and adding that data should not be collected unless it will change what gets sent. That is unusual advice from a vendor whose product sells segmentation, and it is the right advice.
Key takeaway: personalization that removes work from the shopper tends to hold its value, while personalization that adds persuasion tends to decay. Before approving a build, ask which of the two is being proposed. A reorder shortcut and a dynamic greeting cost roughly the same to develop and rarely produce comparable returns.
Where It Is Usually Overkill
Nothing in this section is an argument that personalization does not work. It is an argument about sequence. In each of these situations the money tends to return more if it is spent elsewhere first, and the personalization case usually improves once that other work is done.
- Small, curated catalogues. With a few dozen products the shopper can see the whole range in two scrolls. Filtering that range for them may remove options they would have chosen, and there is little to gain in exchange.
- Considered, infrequent purchases. Where customers buy once every few years, browsing history from a previous visit may be stale rather than informative, and prior purchase data is often a poor guide to the next decision.
- Stores losing orders at checkout or on delivery terms. A recommendation engine cannot recover an order abandoned over shipping cost. Personalising the path to a checkout that is itself the constraint tends to move the loss rather than reduce it.
- Cosmetic personalization. Greetings, name insertion and countdown widgets carry maintenance cost and privacy exposure while rarely changing the decision a shopper is making.
- Real-time individual targeting below the data floor. If the store cannot supply thousands of returning purchasers a month, one-to-one prediction may be producing confident-looking output from very thin evidence.
The third case is the expensive one, because it is the hardest to see from inside a personalization project. A store with a strong add-to-cart rate and a weak checkout completion rate has a problem the recommendation layer cannot reach. Work through where the loss actually sits, using something like a structured ecommerce CRO audit, before assuming the answer is relevance. Occasionally it is. More often the funnel data points somewhere less interesting.
There is also a stated-preference argument that gets quoted in most vendor decks, and it is worth handling carefully. BigCommerce cites Adobe research finding that 66% of customers say encountering content that is not personalised would discourage them from buying. That is a survey response about attitudes, not a measurement of revenue, and it should not be read as evidence that a specific personalization build will pay for itself in a specific store. What people report about relevance and what they do at checkout are different datasets.
Three Realistic Ways to Deliver This Work
Most retailers choose between platform-native features, an app plus an internal owner, or an ecommerce personalization agency. The table below sets out where each tends to fit. None of them is correct for every store, and the third is the one that most often gets bought too early.
| Approach | Best suited to | Internal ownership needed | Time to first measurable result | Main risk |
|---|---|---|---|---|
| Platform-native features | Stores testing whether relevance is the constraint at all | Low, but merchandising time for manual curation | Weeks | Ceiling is reached quickly on large catalogues |
| Personalization app plus an internal owner | Mid-sized stores with steady repeat purchase and a named owner | High and continuous | One to two quarters | Ownership lapses and rules quietly go stale |
| Ecommerce personalization agency or CRO partner | Stores with the data volume but no analytical capacity to direct it | Moderate, mainly decision-making and data access | One to two quarters, with diagnosis first | Cost is committed before incrementality is proven |
| Do nothing yet, fix search and segmentation | Stores below the data floor or losing orders further down the funnel | Low to moderate | Weeks | Competitors with better relevance may pull ahead in the interim |
Read the last row as a genuine option rather than as a placeholder. For a substantial share of established retailers it is the correct answer for a year or two, and the delay costs less than an underfed model does. The middle option is the one that most often disappoints, not because the software underperforms but because the ownership assumption is rarely tested. Rules-based personalization is a standing operational commitment, and a store that cannot name the person who will review those rules each quarter is buying a system that will drift.
When an Ecommerce Personalization Agency Is Worth the Cost
An ecommerce personalization agency earns its fee in a narrower band of situations than the market implies. Choose it when the store already produces meaningful behavioural volume, when several plausible explanations for a conversion problem exist and nobody internally has time to separate them, or when a previous personalization attempt produced modules that nobody can now evaluate. In each case the deliverable that matters first is diagnosis rather than implementation.
An internal hire may be the better choice when the work is continuous merchandising rather than analysis, because curating complementary products and reviewing segment rules is closer to a permanent role than to a project. A platform partner may be better when the constraint is technical, such as a catalogue whose attribute data is too inconsistent for any rule to be written reliably. Bringing in an ecommerce personalization agency before that groundwork exists usually converts a data problem into a more expensive data problem.
The questions worth asking a prospective partner are mostly about evidence rather than capability. Ask how they will establish a baseline before anything changes, how they intend to measure incremental revenue rather than attributed revenue, what they expect to be able to conclude if the first three months show nothing, and which parts of the build the internal team will have to maintain afterwards. A partner who answers the last question honestly is describing a cost the proposal may not contain. Internal expert input required: add WD Market’s own qualification criteria for personalization engagements, including the minimum monthly returning-purchaser volume below which the team recommends deferring the work.
A Sequence for Deciding What to Personalise
The order below is deliberate. Each step is capable of ending the project, which is the point of running them in sequence rather than in parallel.
- Count what the store actually produces. Monthly returning purchasers, repeat-order rate, and the number of products that receive enough traffic to generate any signal. This single exercise resolves the model-based question for most retailers in an afternoon.
- Locate the loss before choosing a remedy. Establish where sessions are ending relative to add-to-cart and checkout. If the largest drop is below the point personalization can influence, the sequence should stop here and resume later.
- Audit the product data. Rules and models both depend on attributes being consistent. Inconsistent categorisation, missing variant data and free-text attributes are the most common reason a technically correct implementation produces poor suggestions.
- Start with rules where the logic is already known. If merchandisers can articulate why two products belong together, encode that first. It is cheaper, it is inspectable, and it gives a baseline that any later model has to beat.
- Hold back a control group before launch. A share of traffic that never sees the personalised experience is the only reliable way to answer the question that follows six months later. Retrofitting this is rarely convincing.
- Price the running cost, not the build cost. Licence, curation hours, quarterly rule review and the engineering time to keep the integration working after theme and platform updates. This total, rather than the implementation quote, is what the revenue has to cover.
- Re-evaluate on margin, then decide whether to extend. Judge the result on gross margin per session against the control group. Extend only into the categories where that comparison held up, and be willing to remove the modules where it did not.
Mistakes That Make Personalization Look Like It Failed
- Counting attributed revenue as incremental revenue. A recommendation module that appears in the path of an order gets credited with it. Some of those customers were always going to buy the item.
- Personalising the homepage first. It is the most visible surface and often the least decisive one, since a large share of paid and organic traffic arrives on product and collection pages instead.
- Segmenting to a resolution the business cannot act on. Twelve segments that receive identical treatment are twelve maintenance obligations and no commercial difference.
- Leaving rules unreviewed after a range change. Seasonal ranges, discontinued lines and renamed categories break conditions written against them, usually silently.
- Treating consent as a compliance detail. In consented-tracking markets a meaningful share of visitors may never be eligible for behavioural targeting, which changes both the addressable audience and the measurement.
The first of these does more damage than the rest combined, because it produces a number that looks like success. A store reports that recommendations influenced a substantial share of revenue, the budget is renewed, and nobody establishes what would have happened without them. Adobe’s documentation hints at why this is easy to lose track of: customer segment membership “is constantly refreshed, customers can become associated and disassociated from a segment as they shop”. When the population being measured moves while the measurement is running, before-and-after comparisons become unreliable, and a stable holdout group is the only straightforward remedy.
The consent point is worth a moment for retailers selling into the UK and the EU. Behavioural personalization depends on data collected under conditions the visitor controls, so the effective reach of any onsite model may be materially smaller than total traffic. That does not invalidate the approach, but it does mean the business case should be built on the consented population rather than on sessions.
How to Tell Whether It Actually Paid
Three numbers settle the question, and only one of them is usually reported. Attach rate, meaning the share of orders containing a recommended item, describes engagement with the module. Incremental margin per session, measured against a holdout, describes whether the business is better off. Running cost, including the curation hours nobody logs, describes what that improvement has to exceed.
A store that reports only the first will almost always conclude that personalization worked, because the metric cannot produce any other answer. This is why the holdout decision belongs before launch rather than after the first quarterly review. It is also why the honest version of this analysis sometimes ends with modules being switched off, which is a legitimate outcome and considerably cheaper than maintaining them indefinitely on the strength of an attribution report.
For teams building the underlying discipline, the wider question of what customer data is worth acting on is covered in more depth in WD Market’s guide to turning customer insights into revenue, and the mechanics of recommendation systems specifically are set out in the guide to AI product recommendations. Stores that have already resolved the questions set out above and want the implementation detail will find it in the step-by-step guide to scaling personalization.
Deciding Where Personalization Earns Its Place
The decision is less about whether personalization works and more about whether this store, at its current size and with its current catalogue, can feed it. Large ranges, repeat purchasing and location context reward the effort consistently. Small catalogues, infrequent considered purchases and funnels that leak further down usually do not, at least not yet. Between those poles sits the practical test set out above: measure the returning-purchaser volume, find where the loss actually occurs, fix the product data, start with rules that can be inspected, and keep a holdout so the result can be believed. Judged on incremental margin rather than attributed revenue, personalization becomes an ordinary investment decision rather than a matter of conviction.
From a Personalization Shortlist to a Verified Revenue Case
A shortlist of personalization ideas is easy to produce and hard to rank. What settles the ranking is evidence the store already holds: the returning-purchaser count, the shape of the funnel, and the state of the product attributes. CRO and growth support at WD Market is built around assembling that evidence before anything is committed to a build, so the argument over which modules deserve funding runs against numbers instead of preference. A store that finishes the exercise with a shorter list than it started with has usually had the more valuable outcome, whether the work then runs in-house or through an ecommerce personalization agency.
Bring a catalogue size, a monthly order figure and an honest description of the current setup to the contact page, and the conversation can begin at the diagnosis rather than at a product demonstration. For anyone who would rather see how the reasoning runs first, working notes on ecommerce measurement go out through the company LinkedIn page.
Questions That Come Up When Scoping Personalization Work
At what size does predictive personalization start to work?
It depends on which type. Rules-based work, such as showing different content by country or by customer group, has no meaningful volume requirement and can be run by almost any store. Predictive, model-driven personalization is different, and published thresholds give a sense of scale: Google requires properties to have well over a thousand returning users on each side of a purchase or churn condition before it will build predictive audiences at all. Below that order of magnitude, expect rules and segmented email to outperform prediction.
Do the built-in recommendation features on Shopify and WooCommerce work well enough?
They are a reasonable starting point and a poor finishing point. WooCommerce assembles its related blocks from shared categories and tags and shuffles the result, which means the products shown may have no purchase relationship at all. Shopify generates its related set automatically but requires complementary items to be configured by hand, so the quality of that block is a function of merchandising effort rather than software. Both are worth switching on before paying for anything, precisely because they establish a baseline.
What should we ask an ecommerce personalization agency before signing?
Ask how the baseline will be captured before anything changes, how incremental revenue will be separated from attributed revenue, and what conclusion they would draw from a flat first quarter. Then ask which elements the internal team inherits and how many hours a month that is expected to take. Proposals tend to price the build accurately and the maintenance optimistically, and the difference between those two is where most disappointment with this work originates.
Is personalization worth it for a store with fewer than 100 products?
Rarely in the onsite sense, because a visitor can review that range without assistance and filtering it may hide items they would have picked. The exception is context rather than content: matching currency, language and delivery information to the visitor’s location helps at any catalogue size. Segmented post-purchase email is also usually worthwhile, since it depends on order history rather than on browsing volume and a small catalogue does not weaken it.
How long before personalization shows a result?
Context-level changes such as currency and language matching can show up within weeks. Recommendation and segmentation work generally needs a quarter or more, because the comparison has to cover complete purchase cycles rather than a favourable few weeks. Where the category has long gaps between orders, a fair read may take two quarters. Setting the review date and the evidence standard at the start prevents the more common outcome, which is an indefinite extension while the numbers are still being interpreted.
Can personalization damage conversion rather than improve it?
It can, in two ways that are worth watching for. Narrowing what a shopper sees may remove the product they intended to buy, which is most likely on small or highly varied ranges. Adding modules to pages that are already dense can also slow rendering and push key content further down, particularly on mobile. Both effects are detectable in a holdout comparison and largely invisible in an attribution report, which is another argument for running the control group.
Should personalization come before or after a checkout rebuild?
After, in most cases. Relevance work operates on the discovery and consideration stages, so it increases the number of shoppers arriving at a checkout that is already losing them. If the funnel shows healthy add-to-cart behaviour and weak completion, the sequence should address completion first. Where both stages are weak, the discovery work is usually cheaper to test and can proceed in parallel, provided the two changes are measured separately.