It is hard to picture a successful online business today that does not use segmentation and personalization. Companies plug in an external technology vendor or build their own, hire outside specialists or assemble an in-house team. The way personalization work is organised differs from business to business — the mistakes, however, are remarkably consistent. In this article we go through the most common mistakes in personalization and what you can do to fix them.
Mistake #1. Stopping at standard segments
When a company first gets to know personalization tools, its team usually wants to work with the segments that already exist in its CRM strategy. Out of habit, marketers segment customers by gender, loyalty tier or purchase activity. Familiar segmentation tends to produce shallow ideas along the lines of “let’s show skirts to women and trousers to men”, which at best deliver a small uplift in average order value. Many people assume this is what personalization is, celebrate a 1–2% uplift and stop there. That is a mistake.
In practice (and in the campaign data on key metrics), men and women often like the same products. And within the segment of women there can be very different sub-segments that differ in behaviour, interests and preferences.
Example: a well-known fashion and beauty retailer. To fill a recommendation block on the product page, the client was set on a cross-sell strategy that showed women’s products to women and men’s products to men. We suggested testing that strategy against a control group (the recommendation block was hidden for them) and ended up with just +1.5% in average order value for the test variation versus control. Unfortunately, that is well below market benchmarks for cases of this kind.
How to fix it: look at the test results broken down by sub-segment
In this particular case we found a sub-segment of women who buy both women’s and men’s products (likely partners shopping for the household, or shoppers who prefer unisex items). That segment showed a statistically significant drop in average order value and was dragging the campaign average down.
With that in hand, we relaunched the test with new logic: the “shopping for two” sub-segment started seeing a mixed selection (women’s and men’s products), while everyone else kept seeing products matched to their gender. After the relaunch we reached +15% (!) in average order value for that sub-segment and +11% for the rest.
Mistake #2. Copying your competitors’ cases
A classic beginner’s assumption in personalization is that you can simply copy a peer’s successful campaign and get exactly the same effect.
Example: a client heard about a competitor’s success story in which a popular-products block worked brilliantly for new users and a personalized recommendation block worked for returning ones, and decided to repeat it. Two campaigns were launched on product pages with the matching targeting, and neither produced a statistically significant uplift.
How to fix it: compare the uplift from scaling the single best variation across the whole audience against scaling the best variation for each sub-segment.
Gravity Field, for example, has a dedicated engine built in — predictive targeting, which identifies the best “variation to segment” combinations and even forecasts the uplift you would get by relaunching campaigns in those combinations. Using that engine we established that both campaigns were being pulled down by the search-traffic segment. That produced a logical hypothesis: users in this segment arrive looking for a specific product, so showing them a block built on the “Similar products” algorithm should improve their visit-to-order conversion.
The relaunch confirmed the hypothesis: +6% conversion into orders in the search-traffic segment and +4% on average across the other segments.
Mistake #3. Personalizing recommendation blocks only
This mistake is most common among eCommerce and retail companies. Their sites and apps are usually well equipped with product recommendation carousels on all the main page types. The products in those blocks are picked by whichever recommendation strategy won the A/B tests, and the blocks themselves sit in the positions those same A/B tests found optimal.
Example: one of our clients preferred to run the personalization platform themselves, and they were very good at launching tests with product recommendations. Yet the cumulative effect of using the platform stayed at +2–3% revenue per user, while the global benchmark for their industry was +8–10% from personalization on average.
How to fix it: use a low-code solution to personalize other elements of the site.
Gravity Field, for instance, lets you build custom campaigns to improve the performance of literally any interface element. We helped this client move beyond recommendations and built several fundamentally new campaigns for them: personalized tags, personalized filters and a personalized category tile grid.
The first two campaigns targeted category-listing conversion. The personalized tag block showed the top attributes the user leans towards: in the “Milk” category listing, for example, you might see the tags “skimmed”, “lactose-free”, “Parmalat”. Clicking a tag took the user into the matching sub-category. In the personalized filter campaign we moved the filters the user interacts with most to the first positions, which made navigating the page easier. And we placed the personalized category tile grid on the home page, so that users could jump into their favourite sections in one click right after landing on the site.
Within a few weeks of launching the new experiments we reached +11% revenue per user compared with the global control group, which saw the site without any personalization at all.
Mistake #4. Ignoring product margin
Most companies treat product margin as sensitive data and prefer not to share it with vendors. As a result, running a personalization platform carries a risk: profit can fall even as revenue grows.
Example: an eGrocery client noticed that the popular-products blocks were frequently filled with the lowest-margin items in the assortment, and started seriously questioning whether the platform was worth it.
How to fix it: pass a margin level in the product feed and use the matching filters in the settings.
To raise the margin of the recommendation blocks we asked the client to add a rough margin marker for every product to the feed: high, medium, low. We then restricted the recommendation strategies to the first two groups only. We were somewhat concerned that this restriction would cut the revenue the platform generates, so we decided to check with an A/B test whether it actually would. The test group saw a site-wide recommendation block with medium- and high-margin products; the control group saw one with all the popular products from the feed.
Over a month of the experiment we found no statistically significant difference in conversion, average order value or revenue per user between the control and test groups — while the gain in margin was dramatic. So we kept the margin filter switched on for all users with a clear conscience.
Mistake #5. Skipping personalization because the assortment is narrow
When people in digital hear that Gravity Field works with telecom operators, banks and fast-food leaders, they are usually surprised. The question I hear most often is: “What is there to personalize when you have so few products?”
Example: a large fast-food client. A few dozen items on the menu. Tests with classic product recommendations produce no statistically significant uplift on key metrics. Is personalization of any use here at all?
How to fix it: use personalization to build the product mix inside bundles.
We suggested that the client assemble ready-made sets of several items using the “Bought together” algorithm. That gave restaurant guests the option to put an order together in literally one click and get a discount on it. Even with the discount, the mechanic delivered a statistically significant lift in average order value versus the control group (which saw no bundles at all) and versus the test group that saw manually curated bundles.
And one more interesting take on bundle personalization: the company defines the offer mechanic (say, “burger + side + drink = X% off”), while the affinity algorithm fills the offer with specific items based on each user’s preferences. Some people prefer beef, some chicken, some fish; some drink coffee, some never touch it — so bundles can end up in hundreds of different combinations. Which combination is best for each signed-in user is calculated by the personalization engine. The result is that those users order from what is effectively an individual menu — with a personal selection of dishes and personal prices.
These are only some of the typical mistakes made by marketers, product managers and their leaders as they start out — and continue — in personalization. And essentially all of them come from a fear of experimenting. Yet effective personalization is impossible without experimentation: constant work with a large number of hypotheses and analysis of test results across different segments. Only that approach shows you which logic works for each segment and lets you build personalization on data rather than on “prior knowledge” or shallow assumptions. Five more mistakes — funnel stage, context, traffic source, the cart and prior knowledge — are covered in part two.
