On-site AI
AI shopping assistants
When a store should add an AI assistant, the risks, and how to test it.
AI and conversion
AI ecommerce personalization: what works, what to skip and how to test it
AI ecommerce personalization uses a store's data to change what each shopper sees: product recommendations, search results, sorting and messages. Done well, it helps people find the right product faster. Done badly, it adds cost, slows pages and takes credit for sales you would have made anyway. Here is how to tell the difference.
Short answer
AI ecommerce personalization uses machine learning to show each shopper different product recommendations, search results, product order and messages, based on what they and similar shoppers have viewed, searched and bought. It suits busy stores with large catalogs and clean product data, and it should be proven against a random holdout group, not judged by attributed revenue.
- Rules are easy to control and test; AI handles large catalogs and many visitor types but needs more data.
- Fix product pages, tracking and product data before you personalize anything.
- Start with one busy placement and a simple non-AI baseline that the AI version has to beat.
- Measure against a random holdout group on revenue per visitor, not revenue attributed to recommendations.
- Respect consent choices, and give visitors who decline tracking a good default experience.
What is AI ecommerce personalization?
AI ecommerce personalization is the use of machine learning to tailor an online store to each shopper, based on what they and similar shoppers have viewed, searched and bought. It decides which products to recommend, how to rank search results and category pages, and which message or offer to show, instead of showing everyone the same page.
Personalization is not new. Stores have shown "customers also bought" lists for a long time. What AI changes is scale: a model can make a different choice for each visitor, across thousands of products, and adjust as behavior changes. Whether that is worth paying for depends on your traffic, your catalog and how good your product data is, more than on which tool you choose.
Rule-based vs AI personalization
Rule-based personalization follows instructions a person writes, such as "show the free shipping threshold to visitors from the US". AI personalization learns its own choices from data and keeps updating them. Rules are easy to control, explain and test. AI handles large catalogs and many visitor types, but needs more data and is harder to check.
| Question | Rule-based | AI-driven |
|---|---|---|
| Who decides what a shopper sees? | A person writes the rule | A model predicts from data |
| Data needed | Little: location, device, source, cart contents | A lot: browsing, search and order history across many visitors |
| Best for | Small and mid-sized stores, a few clear segments, messages and offers | Large catalogs, high traffic, recommendations and search ranking |
| Easy to explain? | Yes, you can read the rule | Often not, which makes mistakes harder to spot |
| Main risk | Rules go stale and nobody updates them | Confident choices built on thin or messy data |
Most stores that do personalization well use both: rules for the things they want to control, such as which promotion runs where, and AI for the things that are too large to manage by hand, such as which of several thousand products to show in a recommendation slot.
What can AI personalize on an ecommerce site?
AI can personalize product recommendations, site search results, the order of products on category pages, "recently viewed" and "pick up where you left off" modules, on-site messages and offers, emails, and product content. Each needs different data and carries different risks, so it pays to choose one or two places to start rather than switching everything on.
| Placement | What it does | What it needs | What to watch |
|---|---|---|---|
| Product recommendations | Suggests related, complementary or "you may also like" products | Enough orders to see what sells together; clean product categories | Recommending items that are out of stock, or what the shopper just bought |
| Site search | Ranks results by likely relevance to this shopper; handles typos and synonyms | Good product titles, attributes and search logs | Searches with no results, which often reveal missing product data |
| Category sorting | Changes the order of products on collection pages | Traffic and sales data per product | Burying new products that have no sales history yet |
| Recently viewed | Brings back what the shopper looked at last time | A returning visitor and their consent to be remembered | Little risk; one of the simplest and most useful modules |
| On-site messages | Shows a banner, pop-up or offer to some visitors and not others | Clear segments and a reason each one needs a different message | Pop-ups that interrupt, and discounts given to people who would pay full price |
| Email and cart recovery | Chooses products, timing and content for each subscriber | Purchase history and email engagement | Sending more email rather than better email |
| Generated content | Writes or adapts product descriptions and copy | Accurate product facts to start from | Errors and invented features published at scale |
| Personalized pricing | Changes prices or discounts for different shoppers | Pricing data and legal review | Customer trust, legal risk, and prices that disagree with your product data |
When AI personalization does not pay off
AI personalization rarely pays off on stores with low traffic, small catalogs, poor product data, or product pages that are not doing their basic job yet. A model needs volume to learn from, and it can only rearrange what is already there. If shoppers leave because of unclear pages or surprise costs, personalization will not fix it.
- Small catalog. With a few dozen products, a person can choose good related items by hand, and shoppers can see most of the range anyway.
- Low traffic. Models need many visits and orders to find patterns, and you need enough traffic to measure whether the result is better than doing nothing.
- Mostly first-time visitors. If most shoppers arrive once from an ad and never return, there is little history to personalize on.
- Messy product data. Missing attributes, inconsistent categories and vague titles lead to poor recommendations and poor search.
- Broken basics. Slow pages, weak product photos, unclear shipping costs and a hard checkout cost more than any recommendation slot can earn back.
For most stores, the order is: fix the product page and checkout, clean the product data, then personalize. A CRO audit tells you which of those is costing you most.
How to start with AI personalization in 7 steps
To start with AI personalization, fix your product pages and tracking first, clean your product data, pick one busy placement, set a simple non-AI version as the baseline, test the AI version against it with a holdout group, check privacy and consent, and only then expand to other placements or stop.
-
Fix the basics and the tracking
Make sure conversions in your analytics match your real orders, and that product pages, shipping costs and checkout work well on a phone. Personalization measured on broken tracking will look like a success or a failure for the wrong reasons.
-
Clean your product data
Complete titles, attributes, categories, variants and stock status. Every personalization system works from this data. So do site search, product feeds and the AI agents that read your pages.
-
Choose one placement with real traffic
Pick one place many shoppers see, such as recommendations on product pages, the cart, or site search. One placement is easier to build, measure and fix than ten.
-
Set a simple baseline
Before AI, decide what the slot shows without it: bestsellers in the same category, or related items chosen by hand. The AI version has to beat this baseline, not an empty space.
-
Test against a holdout group
Show the AI version to a random share of visitors and the baseline to the rest. Compare revenue per visitor and conversion rate over full weeks, and plan the sample size first with the A/B test sample size calculator.
-
Check privacy and consent
Confirm what consent your tracking and profiling need where you sell, that the tool respects it, and what the vendor may do with your customer data. Visitors who decline tracking should still get a good default experience.
-
Expand, adjust or stop
If the holdout test shows a clear gain, keep it and move to the next placement. If it does not, stop or change the approach. A personalization feature that cannot beat a simple baseline is a cost, not an asset.
How to measure personalization honestly
Measure personalization by comparing visitors who get it with a random holdout group who do not, on revenue per visitor and conversion rate across the whole site. Do not rely on revenue "influenced by" or "attributed to" recommendations, because those figures count sales that would have happened without the recommendation.
Attributed revenue is the most common way personalization looks better than it is. A shopper who came to buy a specific product, clicked a recommendation on the way and bought it anyway shows up as recommendation revenue. A holdout group removes that problem, because both groups contain the same kinds of shoppers and only one gets the feature.
- Use whole-site metrics. A recommendation slot can steal clicks from other parts of the page. Measure the effect on the visitor, not on the widget.
- Run full weeks. Shopping behavior changes between weekdays and weekends, and around paydays and holidays.
- Watch order value and returns. Recommendations can raise basket size and also raise returns if they push items that do not fit.
- Keep a small holdout running. Some teams keep a small share of visitors on the baseline permanently, so they can see if the gain fades over time.
- Read the result properly. Use a statistical significance calculator, not the vendor's dashboard alone.
Privacy, consent and trust
Personalization runs on personal data, so privacy law applies wherever you sell, such as the GDPR in the EU and state privacy laws in the US. Beyond the law, shoppers notice when a store seems to know too much. Collect only what you use, respect consent choices, and make personalization feel helpful rather than watchful.
- Consent changes your data. Visitors who decline tracking cannot be profiled the same way. Plan for a good experience without personalization, not only with it.
- Know where your data goes. Ask the vendor whether your customer data trains models used for other stores, and whether you can export or delete it.
- Avoid sensitive inferences. Recommendations that reveal health, pregnancy or financial situations can upset customers even when they are accurate.
- Explain the choice. Labels such as "Because you viewed linen bedding" make recommendations feel useful instead of strange.
- Be careful with personalized prices. Different prices for different people carry legal and trust risks. Take legal advice before you try it.
- Get advice for your case. This page is not legal advice. The rules differ by country, state and product type.
Agent readiness
Personalization when the shopper is an AI agent
Some AI assistants now research and compare products on a person's behalf. They usually arrive without your cookies and without the shopper's history on your site, so they see your default page.
PERSON VIEW (RETURNING CUSTOMER)
Product page
Linen duvet cover
$89
Welcome back: free delivery on your next order.
Goes well with: linen pillowcases, waffle throw.
Illustrative example
AGENT VIEW (DEFAULT PAGE)
- product.name
- Linen duvet cover
- PASS
- offers.price
- 89.00
- PASS
- recommendations
- null
- LOADED BY SCRIPT
- returning_customer_offer
- null
- NOT SHOWN
Illustrative example: a returning customer sees a loyalty offer and recommendations based on past orders. An agent reading the same page sees the default product data, with no recommendations and no offer.
That has three practical consequences for a store that personalizes:
- Your default page carries more weight. An agent judges your product on the unpersonalized version, so it needs complete details, clear prices and plain-text policies in the page HTML.
- Keep prices consistent. If a personalized price or discount differs from the price in your page and structured data, a person and an agent may see different numbers. Make sure offers are clearly labeled as offers and never contradict your listed price.
- Do not hide product information inside widgets. Details that only load in a script-driven module, such as sizing help or bundle contents, may never reach an agent.
This is a new area and nobody has years of data on it yet, including Convertica. Start with the guide to agentic commerce, then work through the agent readiness checklist.
Questions to ask a personalization vendor
Before you buy a personalization tool, ask how it proves its own value, what it does with no data, how it affects page speed, what happens to your data, and how you leave. A vendor that cannot support a holdout test, or only reports attributed revenue, is asking you to take its results on trust.
- Can I run a holdout group and see revenue per visitor for both groups?
- What does the tool show a first-time visitor with no history?
- Does it load on the server or with a script in the browser, and what does that do to page speed?
- How does it handle visitors who decline tracking?
- Is my customer data used to train models for other clients?
- Can I export my data and turn the tool off without breaking my pages?
- How do I override or pin what it shows, for example during a promotion?
- How does it stop recommending out-of-stock or discontinued items?
- What exactly is included in the price, and what costs extra as my traffic grows?
Audit and advisory
Where Convertica fits
Convertica has worked on ecommerce conversion since 2017, with more than 2,000 clients, and its CRO case studies publish each result with its metric, timeframe and source. The free CRO audit, an app you can run now, checks your page for people and for AI agents, scores it out of 100 and gives you three fixes you can make now. CRO advisory, led personally by founder Kurt Philip with Convertica's CRO team, helps you choose what to personalize, plan the holdout test and read the result. In advisory your own team builds the changes; with full implementation, Convertica's team builds the fixes from your audit. Both are priced after your audit.
In this series
More on AI and conversion
Testing
AI A/B testing
What AI can and cannot do for your tests, and a 7-step workflow.
Checklist
Agent readiness checklist
Check what AI agents can read on your product pages.
Questions about AI ecommerce personalization
What is AI ecommerce personalization?
AI ecommerce personalization is the use of machine learning to tailor an online store to each shopper, based on what they and similar shoppers have viewed, searched and bought. The most common forms are product recommendations, personalized site search, product sorting and targeted on-site messages.
How is AI used for personalization in ecommerce?
AI models look for patterns in browsing and purchase data, then predict which products, search results or messages a shopper is most likely to want. A store uses those predictions to fill recommendation slots, re-rank search results and category pages, choose emails and decide which offer to show.
What is the difference between AI personalization and rule-based personalization?
Rule-based personalization follows rules a person writes, such as showing a free shipping message to visitors from one country. AI personalization learns its own rules from data and updates them as behavior changes. Rules are easier to control and explain. AI scales to large catalogs and many visitor types.
Does AI personalization work for anonymous visitors?
Partly. With no history, a model can use what the visitor does in the current session, the page they landed on, their device and location, and what similar sessions did. That is useful, but weaker than personalization for returning customers. Your default, unpersonalized experience still has to work well.
How do you measure whether personalization works?
Hold personalization back from a random share of visitors and compare revenue per visitor and conversion rate between that holdout group and everyone else over full weeks. Revenue attributed to recommendation clicks is not proof, because it includes sales that would have happened without the recommendation.
Is AI personalization worth it for a small store?
Usually not as a first step. Models need enough visitors, orders and products to learn from, and a small store rarely has them. Simple rules, such as showing bestsellers or related items chosen by hand, and fixing the basics of the product page usually come first.
Is AI personalization legal under privacy law?
It can be, but it uses personal data, so privacy law applies, such as the GDPR in the EU and state privacy laws in the US. Check what consent you need for tracking and profiling where you sell, and what your vendor does with your data. Take legal advice for your own case.
How does personalization affect AI shopping agents?
An AI agent shopping for someone usually arrives without your cookies or the shopper's history on your site, so it sees your default page. Make sure that default page has complete product details, prices and policies in the HTML, and that personalized prices or offers do not contradict them.