How to master personalization with AI

Picture ecommerce in its early days: businesses were just beginning to discover the power of personalized marketing. They’d divide their customers into segments based on categories such as their age demographic, location, and past shopping habits, then adjust their marketing strategies to target the various segments, such as by sending out personalized email.

For example, if someone fit a “luxury shopper” segment, the ecommerce platform would recommend high-end products. If they fit a “discount seeker” profile, the site would recommend discounted items.

It was relatively simple and it worked fairly efficiently, to a degree. But there were glitches with this segment-based approach, which was akin to trying to make a one-size-fits-all shirt actually fit everyone.

Here’s an example: maybe in those days, you shared a Netflix account with a family member or friend. You might have been a fan of crime documentaries while they preferred romantic comedies. But because you were lumped in with each other in the same viewer segment, the recommendation engine targeted your combined customer behavior. So unfortunately, your Netflix home page was suggesting “Bridgerton” when you were in the mood for “Mindhunter”.

So much for reliably relevant content personalization.

personalization AI banner

The era of true personalization

Let’s fast forward to a world of AI-driven advancements with a high personalization experience bar that’s been set by Amazon and other ecommerce giants. With the help of machine-learning algorithms, businesses are creating infinitely better-quality omnichannel shopping experiences that are unique to each customer — the equivalent of one-on-one personalization.

With AI algorithms, for an online shopper, it’s like walking into a store where everything on the shelves has been picked out and displayed with their specific interests in mind — pretty revolutionizing. It’s like being with a personal shopper who, after following you around on social media sites such as LinkedIn and Facebook, knows your tastes to a tee.

This modern form of personalization is reshaping how businesses interact with first-time ecommerce site visitors, returning visitors, and repeat customers alike.

AI: the secret sauce

The key to this sharper personalization functionality is, of course, artificial intelligence. Deep learning can expertly comprehend user intent and act on its “understanding” by providing a finely tailored experience for site visitors and app users.

This personalization requires investment in technology and skills to analyze data, of course, but the prospective payoff is huge.

First and foremost, you need the ability to collect the right type of information — user behavior data, events, clickstreams — at various shopper touchpoints. When you have that material, you can build a user profile that can be leveraged to personalize user experiences across your channels.

Your personalization must also:

  • Respect data privacy
  • Be applied in real time
  • Support user identity consistency across sessions, user states, and devices
  • Be able to predict user intent
  • Be composable (API first)
  • Use explainable AI technology and transparent business metrics
  • Be used and applied ethically

Let’s consider the who, what, when, where, and why of orchestrating a modern personalization experience.

Here are the four types of website visitors and how personalization levels can be applied to them.

The user types

1. New visitors

For first-time visitors, you can collect location and device data.

If a session is a bit longer, or it’s rich in terms of interactions  like clicks, views, and conversions, you can apply real-time in-session personalization.

New visitors vs. returning ones

What’s your new vs. returning visitor rate? Do you have many more new shoppers than repeat buyers?

This is an interesting thing to know because different types of visitors mandate different personalization approaches.

In the ecommerce industry, there’s an expectation that in terms of marketing campaigns, new users should get the same level of personalized customer experiences as returning shoppers enjoy. However, that’s not possible, as new visitors haven’t interacted enough with your platform for you to collect their preferences and respond with relevant personalized content.

New visitors may come to your site in greater numbers, but the returnees are the ones more likely to be responsive to your personalization efforts. Plus, their conversions and average order values (AOV) will tend to be better.

That’s why it pays to not worry about first-time visitors so much, and instead focus on devising a strong personalization strategy that targets your returning users.

Returning visitors

The returning-user types include:

2. Non-authenticated visitors

These visitors are the first ones for which you can gather valuable data that you can then feed to your personalization system.

A visitor can interact with your website or application repeatedly over a period of hours, days, weeks, or months. You can track their time spent on the website, number of sessions, number of pageviews, and bounce rates.

At some point, non-authenticated users can of course be authenticated, such as when they create an account. If your website is structured in a certain way, they might convert without taking that formal step.

If they don’t create an account, it could signify that they’re lukewarm on your business, but they could also just have some mitigating concern, such as not wanting to share their personal information.

3. Authenticated visitors

When a visitor converts or creates an account, indicating trust and an intention to return and perhaps place an order, they become a bona fide authenticated user.

With these people, you now have more data points — for instance, their name, mobile number, email address, birth date, gender, and occupation — that you can utilize to orchestrate a meaningful customer journey for your newly acquired user.

This brings us to the question of how personal data should be treated in the context of user privacy laws. We won’t go into that now, but if you’re interested, here are some reflections on user privacy.

4. Buyers

Finally, buying users are, of course, those who’ve purchased, whether during an initial session or on a subsequent visit. These are the prospects most rich in potential in terms of future interaction and embracing your brand. They’ve indicated preferences in regard to your products. Based on their orders —  especially if they’ve returned to buy more items — you can know whether they like a certain type of item or brand, their sizes, their favorite colors and patterns, and more.

Here’s a summary of the information you might gather when new and returning visitors come to your website:

The personalization maturity model

What’s the best personalization strategy for each of these segments?

The personalization maturity pyramid (below) consists of using content-based strategies (non-personalization), followed by orchestrating experiences for user segments (weak personalization), and applying AI-powered personalization, also known as 1:1 and hyper personalization; the pinnacle of the customer experience.

Content-based strategies

This baseline form of personalization factors in item popularity and relies on manual creation of personalized content. This approach is suited mainly for scenarios in which you don’t have enough individual user data. It’s not a personalization strategy per se, but you’re still looking for ways to tailor the experience for prospective new-customer needs.

We can divide content-based mechanisms into two categories: those that need the product catalog (e.g., images, titles, descriptions, attributes) and those that require additional aggregated user data (e.g., clicks, views, purchases).

A catalog-based recommendation system produces images, articles, and products similar to given images, articles, and products. No user interactions such as clicks, views, or conversions are taken into account.

With item-based collaborative filtering, the set of items most similar to the target item is determined based on aggregated user interactions. The similarity between each pair of items is calculated according to how many clicks, views, and purchases shoppers generate while interacting with the two items.

There’s also a scenario in which both content and aggregated user features are assessed to generate a list of similar images, articles, or products. 

Here’s more-comprehensive information about recommender systems.

For new users

Unless new visitors have interacted meaningfully with your website in terms of clicks, views, and purchases, you can’t give them true real-time in-session personalization.

In the purely content-based scenarios applied for new visitors, a real-time approach doesn’t change or improve the quality of the experience due to the fact that the content is mostly popularity based or manually curated.

Here are some content-based methods of engaging new users.

When no user data is available (see code sample #1 above):

On product detail pages (PDPs)

Related items and products

This involves providing alternatives to the current item a user is viewing or showing interest in, whether they’re similar in functionality, belong to the same product category, or are comparable.

For instance, if a shopper is engaged in decision making about buying a smartphone, they might be shown other models with similar features or models from the same brand in the same price range. 

Image-based similarity

This method, often used in the fashion and home-decor industries, leverages visual similarities in items in order to suggest other items. The technology uses convolutional neural networks (CNNs) or similar deep-learning techniques to identify what’s appearing in images that are associated with items, and can thereby compare items.

Complementary items

The goal is to increase the order value by encouraging the purchase of complementary items. The system suggests items that complement an item a shopper is viewing or that’s in their purchase history. For example, if someone is looking at buying a laptop, they might be shown a laptop bag, a wireless mouse, or antivirus software before they check out.

When aggregated user data is included (see code sample #2 above):


Item-to-item collaborative filtering

This method works by creating an item-to-item matrix to determine relationships between items, and then recommending items closely related to those that someone rated positively. For example, if people who bought item A also bought item B, then, when another shopper buys item A, the system suggests item B.

Frequently bought together

This recommendation system encourages additional purchases by leveraging data from past transactions to identify items that are often bought together. For example, if a grocery retailer’s customers frequently buy bread, butter, and jam on the same online shopping trip, then if someone selects only bread and butter, the system helpfully suggests jam as well.

Trending items

Popular items can be identified by a sudden increase in sales views or another indicator of customer engagement. By showing shoppers what other people find appealing, you may be able to spark even more interest and ride the trend wave to higher sales. 

On search results pages

Query categorization and suggestion

This method classifies people’s search queries in categories and provides suggestions based on the categories. This helps new shoppers refine their searches and make their way to the most relevant results.

For instance, a query for “shoes” might be categorized in footwear, and the system would then suggest subcategories such as running shoes, sandals, and boots. This can not only improve search relevance but  help shoppers discover items they might not have even thought to search for.

Dynamic reranking

This approach modifies the ordering of search results based on user interaction data such as clicks, views, and conversions. For example, if a significant number of visitors search for “summer dresses” and then click on “floral summer dresses”, the system might dynamically re-rank the search results to prioritize items that match both the queries “summer dresses” and “floral.”

Dynamic synonym suggestion

This method automates the process of providing search-term alternatives. If your record contains the keyword “pants” but someone is searching for “trousers”, adding “trousers ⇔ pants” leads people using the slightly different terms to the relevant items. 

For returning users

Content-based strategies can be applied for returning visitors, whether or not they’re authenticated. However, because you can know more about these people at the individual level, it makes more sense to explore segment- and AI-based personalization strategies, which can produce better results.

As a rule of thumb, applying only content-based strategies to returning users is a suboptimal strategy because it does not utilize the explicit or implicit information that the visitors have provided.

Consider a bookstore owner who’s familiar with her regular customers’ reading habits. A content-based strategy would entail recommending books solely based on bestsellers, regardless of whether a customer is a regular or a newcomer. The owner might recommend a certain fantasy novel because the author is a trending writer that week. This strategy would apply to all her customers because it’s not reliant on knowing anything about them.

Personal knowledge pays off

However, for regular customers, the bookstore owner has more information: their tastes, preferred authors and genres, how often they buy new books.

Using a segment- or AI-based personalization strategy would be akin to applying this additional information to tailor her recommendations.

For example, she might know a regular customer enjoys mystery novels but also has a passion for historical non-fiction. The result? She could recommend a historical mystery novel that a purely content-based strategy might miss.

Relying solely on content-based recommendation for these repeat customers would obviously be less than ideal. It’s like ignoring the additional knowledge about regular customers and treating everyone as a stranger. While she might still recommend a book someone enjoys, she’d be missing out on the chance to give a more personalized recommendation that the customer might really appreciate.

User segmentation

With access to individual customer data (clicks, views, purchases), we can begin thinking about offering people a more personalized experience.

In the first level of personalization, the segmentation-based approach, associating a segment with a new user implies that there’s a predefined segment (or more) in the system. When the required user attributes have been correctly identified, the user can be linked to one or more segments and different actions can be triggered based on that.

Let’s take user123 as an example. As they navigate and interact with products, they generate clicks, views, and possibly purchases:

Let’s say you’ve defined these segments:

Given the above segment1 and segment2 definitions, let’s say that the following actions are set to be triggered on the website:

Action #1: An invitation to sign up is shown and a promotional discount is displayed only for new users who are part of segment1

Action #2: Free shipping is offered for returning buyers who are part of segment1

Action #3: For users who are part of segment2, based on price preferences, product listings are re-ranked

For new users

The user profile is empty the first time someone lands on your website. When there’s more activity, such as clicks and views, the visitor can be associated with a segment and an action can be triggered.

In their first session, user123 has been clicking on a few products priced between $500 and $1,000, which are associated with segment1. Based on the actions defined above, the user is invited to sign up and receive a discount for their first purchase. As a result, user123 buys an item.

Visitors have seen this prompt only if they’ve satisfied the condition of being a new user as part of segment1. At the same time, because user123 is not yet a returning user, the second action above, free shipping, is not triggered.

However, upon user123’s revisiting the Books category, because the third action above is triggered, books priced $10–25 are boosted to the top.

For returning users

A few days later, user123 returns to the website, and this time, because they’re part of segment1, they’re greeted with a free shipping banner.

Additionally, when revisiting the Books section, user123 is again presented with a list of books, this time boosted based on their pricing preferences. Notice that segment2 is not dependent on the type of user (new or returning), so this action is triggered for both scenarios.

You might ask whether you can embed the user type in the definition of your segments. Short answer: yes. The only nuance is that you’d end up with three segments instead of two, and of course the third action would be tied to segment3.

1:1 Personalization

AI tools and machine learning make hyper personalization possible. In turn, 1:1 personalization makes human intervention optional. As opposed to manual segmentation, AI-driven personalization doesn’t require a merchandiser to define the conditions under which personalization kicks in.

In the context of 1:1 personalization, when user123 conducts a search, the results would immediately be re-ranked based on their affinities. The same would be true for the rest of the visitors with valid user profiles. So even if they search using the exact same keywords, they would each get slightly different results. Or, to be more precise, the results would be re-ranked based on the customer preferences.

That’s true personalization.

As with the other strategies, there are nuances when it comes to applying personalization for new vs. returning users.

For new users

1:1 personalization for new users implies real-time personalization. But there’s that caveat: there’s not much you could know about the user in the first second they land on your website. You might have access to location and device details, but that doesn’t say anything about their individual preferences.

Would a 60-second session make any difference in the richness of the user profile?

Maybe, if the user interacted with the website. Otherwise, it would be back to content-based strategies instead of 1:1 personalization.

Then the question would be how many interactions are enough to build a decent user profile? Can you start personalizing after the first click?

Not really; you’d run into over-personalization — the problem of boosted content being overly narrow and lacking diversity, which can lead to several issues:

  • Shoppers getting stuck in a loop of similar content, limiting their exposure to new products and categories, which can translate to a poor user experience and low customer satisfaction
  • Shoppers having fewer opportunities to discover new interests. Ecommerce thrives on variety and novelty, and people often enjoy browsing and discovering, so this is problematic
  • False assumptions being made based on limited data. For instance, if someone is buying a gift, their session data might not accurately reflect their own interests, leading to inaccurate future recommendations

Do most people want a personalized experience? One way to find out if your shoppers do is to include the option in the search bar, like Instagram does:

Based on our experience here at Algolia, the best user profiles in the context of real-time personalization balance session duration with volume and depth of interactions. That’s a personalization-ready session capable of generating decent user profiles that, in turn, can represent the basis of a 1:1 personalization strategy.

Here’s a user profile generated based on a visitor’s product clicks and views.

Note: this is a simple representation; a real-world scenario would require more-sophisticated techniques.

The resulting user profile can be then passed on to Algolia for generating personalized search results in real time:

1:1 personalization shines for non-authenticated, authenticated, and repeat buyers because these types of users tend to convert better than new users.

In the previous section, you saw the example of personalizing search results. Now let’s look at how you might personalize product recommendations for returning shoppers.

The sky’s the limit

1:1 personalization applies mostly to search, browse, and recommendations, but you can also make great use of it for sending intelligent push notifications and doing email marketing. You can even utilize it in conversational interfaces like AI-powered chatbots; explore some use cases in our DevCon ’23 live-coding session on a generative AI ecommerce framework to assist with “long-tail” merchandising.)

Master personalization with AI

Want to enjoy the benefits that effective personalization can afford your ecommerce business?

Overhaul and streamline your search and discovery with the Algolia API, which combines vector-based natural language processing (NLP) and keyword matching. We power 1.5 trillion search-provider requests a year, enabling more than 17,000 customers to build blazing-fast, high-quality experiences for their visitors.

Request a demo on how you can optimize your search and discovery for higher conversion rates and better customer retention, or simply contact us and let’s chat.

About the authorCiprian Borodescu

Ciprian Borodescu

AI Product Manager | On a mission to help people succeed through the use of AI

Recommended Articles

Powered by Algolia AI Recommendations

Search personalization 101: how to capitalize on personalized search and discovery

Search personalization 101: how to capitalize on personalized search and discovery

Catherine Dee

Catherine Dee

Search and Discovery writer
E-commerce personalization: pitfalls, tradeoffs, and solutions

E-commerce personalization: pitfalls, tradeoffs, and solutions

Eunice Lee

Eunice Lee

Matthew Foyle

Matthew Foyle

Solutions Engineer @ Algolia
Introducing AI Personalization (𝛽)

Introducing AI Personalization (𝛽)

Ciprian Borodescu

Ciprian Borodescu

AI Product Manager | On a mission to help people succeed through the use of AI