E-commerce & AI for Online Stores

AI in e-commerce – what it actually does in an online store

AI in e-commerce is a set of tools that take over work on your store's data: writing product content, choosing recommendations, handling orders and tickets, and finding the places where the store loses customers. It is not one system but several separate projects, each with its own inputs and its own measure of success.

Before you pick a first project, look at what each one needs as input. Most failed projects I have seen did not fail on the model. They failed on missing data.

The service page explains how I work in each of these areas. Here I focus on the decisions you make as the owner: what to do first, how much to pay and how to tell whether it works.

AI for an online store – when it makes sense and when it doesn't yet

AI for an online store makes sense when you have repetitive work or data at a scale your team can't handle by hand. If the real problem is a lack of traffic or an unclear offer, a language model won't fix it.

Signs your store is ready

• The catalogue grows faster than the team can write copy, or descriptions come from the wholesaler and match every competitor.
• You have order history that shows what customers buy together.
• Someone on the team retypes stock levels, order statuses or answers to the same order questions every day.
• Your platform (Shopify, WooCommerce, PrestaShop, BaseLinker) has an API you can connect to.
• Analytics works correctly, so you can measure before and after a change.

Signs it is better to wait

• Processes aren't settled and everyone does them differently. Automation will lock in the chaos.
• Product data has gaps that suppliers can't fill.
• The store has too little traffic for A/B tests, and the main problem is acquiring customers.

E-commerce audit – where to start with AI in your store

The safest start is an online store audit and a baseline measurement, followed by one project with the largest estimated impact. Without a baseline, any result after launch is just an opinion.

In an audit I price two areas first because they usually hide the most money: checkout including shipping costs, and page speed. Then I check product data, because descriptions, recommendations and AI visibility all depend on it. The full scope is on the e-commerce audit and strategy page.

1. Baseline. I check that GA4 records purchase events correctly and record the starting values of the chosen metrics.
2. Checkout and speed. I look for obstacles on the purchase path and Core Web Vitals problems.
3. Product data. I assess attribute completeness, because it feeds descriptions, recommendations and product feeds.
4. Pilot. One process, one category or 50–100 descriptions for approval before the whole catalogue.
5. Test. An A/B test where traffic allows it, otherwise a before-and-after comparison.
6. Scale. I extend what worked, with alerts and documentation for your team.

AI product descriptions – how to generate them without SEO risk

AI product descriptions are safe for SEO when they are built from real product data and add something for the buyer. Google's 2023 guidance judges content quality regardless of how it was produced, but its spam policies explicitly list using generative AI tools to produce many pages without adding value for users as scaled content abuse.

In practice the line runs between copy built from data and copy built from a product name. A model that only gets the name invents features the product doesn't have. A model that gets attributes, materials, dimensions, use cases and customer questions writes copy that answers real pre-purchase doubts.

What I check before generating the whole catalogue

• Every category has its own template, not one prompt for the whole store.
• Rules block forbidden claims, such as health claims, and features not present in the data.
• Variants (colour, size) get short, distinct paragraphs instead of copies of the same text.
• Marketplace copy differs from the copy on your own store.

On my own GiftFinder project an agent system generated more than 11,000 pages, so I know where AI produces repetitive text. I described a description and parameter pipeline connected to Base (formerly BaseLinker) in the post on BaseLinker AI automation, and the project scope on the AI product descriptions page.

AI product recommendations – personalization you can measure

AI product recommendations show each customer similar, complementary or history-based products. You only know their value from an A/B test against your current setup, because a "similar products" block on its own proves nothing.

Customer expectations are high. McKinsey's 2021 report found that 71% of consumers expect companies to deliver personalized interactions and 76% get frustrated when that doesn't happen. That doesn't mean every store needs its own engine. With a small range and little order history, an off-the-shelf plugin is often the sensible choice.

Three decisions that decide the outcome

• A method that fits the data. With short history I start with attribute similarity (content-based) and add collaborative filtering as orders accumulate.
• Business rules. The model must not recommend out-of-stock products or items the store loses margin on.
• Placement and number of blocks. Product page, cart and post-purchase email behave differently, and too many blocks slow the page down.

Data requirements and the rollout are covered on the AI product recommendations page.

Online store automation – which processes to hand over to AI

I start online store automation with the most frequent, repetitive tasks, not the exceptions. Most steps are plain logic (stock sync, statuses, reports), and a language model only appears where something has to be read and understood, such as a return request.

Typical candidates for automation

• Syncing stock, prices and orders between the store, marketplaces and the warehouse.
• Recognising ticket type (return, complaint, status question) and collecting missing details.
• Draft replies to order questions based on system data, not guesswork.
• Alerts: stock below a threshold, a stalled sync, a delayed shipment.

I walk through the full cycle from cart to complaint in the post on AI order management automation, and the service scope is on the e-commerce AI automation page.

Store conversion optimization – cart, checkout and page speed

Store conversion optimization means removing the obstacles that make customers give up and proving each change with a test. The biggest gains usually sit in checkout and page speed, not in button colours.

Baymard Institute puts the average documented cart abandonment rate at 70.22%, based on 50 studies (updated 2025). In its survey of US shoppers, the most common reason for abandoning an order was extra costs being too high (40%), followed by slow delivery, not trusting the site and being forced to create an account.

Speed is money too. In Deloitte's "Milliseconds Make Millions" study (2020), a 0.1-second improvement in mobile site speed was linked to an 8.4% increase in retail conversions. As a technical target I use the Core Web Vitals thresholds: LCP up to 2.5 s, INP up to 200 ms and CLS up to 0.1 for 75% of visits.

I run A/B tests following Google's guidance: no cloaking, rel="canonical" on variant URLs, 302 rather than 301 redirects, and removing test code as soon as there is a result. The testing programme is described on the conversion rate optimization (CRO) page.

Product content for AI shopping answers – Google, ChatGPT and feeds

For products to appear in AI answers, you need complete, consistent data in three places: the product page, structured data and the product feed. Assistants compare products on price, availability and features, so a missing attribute means you're missing from the comparison.

• Google. Google's documentation says there are no special optimizations needed for AI Overviews and AI Mode, but it recommends keeping Merchant Center up to date and making structured data match visible content. Merchant listings require a name, an image and an offer with price and currency.
• ChatGPT. According to OpenAI's help center, the merchant list is built from product metadata supplied by third-party providers or directly by merchants, and ranking considers factors such as availability, price and quality. Shopify stores are connected through Shopify Catalog; others can apply for a direct feed.
• Feed. OpenAI's product feed spec requires, among other fields, an ID, title, description, page URL, brand, image, availability and price. That is the same data a good product description needs.

I cover store visibility in Google and AI shopping in more depth in the post on e-commerce SEO and GEO. Product structured data is part of my Schema.org implementation service.

How much AI in e-commerce costs and when you see results

My projects start at €1,200 for a store audit, and the cost depends mainly on the number of integrations, catalogue size and number of sales channels. The first measurable results usually come after a few weeks, because that's how long it takes to collect test data.

Add running costs to the build cost: AI model calls, hosting for automations and your team's time for quality checks. For product descriptions the maths is linear (setup plus a per-SKU rate), so you can work it out before deciding.

How to measure AI results in a store and choose a provider

You measure the effect against a baseline: an A/B test where traffic allows it, and a before-and-after comparison elsewhere. Set the metric before the project starts, or after launch someone will find one that happens to look good.

What to measure

• Descriptions: share of pages with unique copy, organic traffic and conversion of pages with new copy.
• Recommendations: average order value and revenue from recommendations in the test and control groups.
• Automations: hours saved, stock errors and response time on tickets.
• Conversion: conversion at each funnel step and revenue per session.
• AI visibility: traffic from AI Overviews (included in the "Web" search type in Search Console) and product presence in assistant answers.

How to choose a provider

• Ask how they will measure the result and what they'll do if a test shows no difference.
• Check whether they ship changes in code and integrations themselves or only hand over recommendations.
• Ask for a pilot on a small part of the catalogue or one process before the full engagement.
• Make sure you get documentation and access to the automations, not a black box.

Common mistakes

• Generating copy from the product name alone and publishing without quality control.
• Recommendations and checkout changes without a test, so nobody knows if they helped.
• Automations without alerts that break silently.
• Testing details while shipping costs or a slow site drive customers away.
• Skipping the feed and structured data, so products never make it into AI comparisons.