E-commerce is one of the sectors where AI adoption has moved fastest — partly because the problems are well-defined (get people to buy things), partly because the data is rich (every click, view, and purchase is logged), and partly because the competitive pressure is intense enough that any edge matters.
But the conversation about AI in e-commerce has become noisy. Marketing copy about "personalised shopping experiences" and "AI-powered product discovery" obscures what's actually working at the store level. This is an attempt to cut through that and give a clear picture of where AI is genuinely moving the needle for e-commerce businesses today — and where it still falls short.
Product Content at Scale
The most widely adopted AI use case in e-commerce is product description generation, and it's one of the more defensible ones — not because the output is remarkable, but because the problem it solves is genuinely painful.
A catalogue of a few hundred products is manageable to write by hand. A catalogue of five thousand products, updated seasonally, with variants for size, colour, and material, is not. That's the situation most mid-sized e-commerce businesses face, and AI makes it tractable.
What works in practice
The workflow that produces usable output: feed the AI a structured data sheet (product name, category, key specs, target audience, brand tone guidelines), and ask for a description of a specific length for a specific placement — PDP headline, long-form description, meta description, and social caption are all different formats that need different treatments. Generate in bulk, do a human review pass on anything going to high-traffic pages, and automate the rest.
The trap to avoid: asking AI to generate from minimal input and accepting the output without review. Generic product copy on a high-intent page is a conversion killer. The AI needs enough specificity in its input to produce specificity in its output. "Blue women's jacket, size 8-16" will produce generic output. "Recycled-shell technical jacket, rated to -10°C, women's sizing 8-16, target customer is urban commuters who cycle year-round" will produce something usable.
The quality check: Read the generated description and ask whether a competitor could publish it without changing a word. If yes, the input brief wasn't specific enough. Rewrite the brief, not the output.
Search and Discovery
On-site search is where AI is delivering some of the most significant conversion improvements in e-commerce, and it's underreported relative to the results being seen. Traditional keyword search requires shoppers to use the exact terms your catalogue uses. AI-powered semantic search understands intent.
The difference in practice: a customer searching for "something to wear to a garden party that isn't too formal" gets served relevant results with AI-powered search. With keyword search, they get nothing or get served results matching "garden" or "party" as literal strings. That gap — between what shoppers actually type and what keyword systems can handle — is where AI semantic search earns its cost.
Tools worth knowing: Searchanise and Boost Commerce for Shopify offer semantic search as a drop-in addition. For larger operations running custom infrastructure, Algolia's NeuralSearch or Elasticsearch with vector embeddings give you more control. The implementation cost has come down substantially in the last year, to the point where it's viable for stores doing above roughly £500k annual revenue.
Personalisation That Converts
Personalisation in e-commerce exists on a spectrum from trivial to transformative. Most implementations sit closer to the trivial end. "Customers also bought" recommendation engines have been around since Amazon built the first one — they're table stakes, not competitive advantage.
Where AI personalisation is creating real lift today:
- Homepage hero personalisation. Tools like Nosto and Dynamic Yield can serve different hero content based on a visitor's session history, traffic source, or inferred intent. A returning visitor who previously browsed running shoes sees running content on the homepage. A first-time visitor from a cycling blog sees cycling content. The lift on click-through to product pages is consistently measurable.
- Email product recommendations. Klaviyo's predictive analytics engine uses purchase history, browsing behaviour, and cohort data to determine what a given customer is likely to buy next. Campaigns built around these recommendations consistently outperform manually curated product emails on revenue-per-recipient.
- Post-purchase sequences. AI can identify the optimal timing and product recommendation for a replenishment email — sending a reminder for consumables before the customer has run out, based on average consumption rates for their purchase quantity. Sounds simple; produces meaningful repeat purchase rates when implemented correctly.
Personalisation earns trust when it's genuinely useful and creeps people out when it's merely demonstrating how much data you have. The goal is to serve the right product, not to prove you've been watching.
Visual Search and AI-Generated Imagery
Two distinct but related areas where AI is changing the visual layer of e-commerce.
Visual search
The ability to search by uploading an image rather than typing a query is genuinely useful for fashion, home furnishings, and any category where aesthetics drive decisions. Pinterest pioneered this at scale. Platforms like ASOS and Zalando have rolled out their own implementations. For independent retailers, tools like Syte offer visual search as a Shopify integration without requiring a custom ML build. Adoption by shoppers is still low compared to text search, but the conversion rate for visual search sessions is typically higher — people using it are committed to finding something specific.
AI product photography
This is the area with the most active development right now, and the results are genuinely impressive for certain use cases. Tools like Caspa and Pebblely allow you to generate lifestyle context around existing product photographs — placing a white-background product shot into a kitchen scene, or on a model in a garden. For brands with large catalogues and limited photography budgets, this is a meaningful cost reduction. The quality is good enough for supporting imagery; for hero images on high-traffic PDPs, a human photographer still produces more trustworthy results.
Pricing and Inventory Intelligence
This is the area with the highest potential and, currently, the widest gap between enterprise capability and what's accessible to mid-market brands.
Dynamic pricing — adjusting prices in real time based on demand signals, competitor pricing, and inventory levels — is standard practice for airlines and hotels. In e-commerce, it's used aggressively by Amazon (which updates prices millions of times per day) and by larger fashion retailers for end-of-season clearance. The tools to do this at mid-market scale are maturing. Prisync and Wiser offer competitor price monitoring with alerting; automated repricing based on those signals is the next step most brands haven't taken yet.
Demand forecasting is arguably more immediately valuable for most businesses. Accurate demand forecasting reduces overstock (which erodes margin) and stockouts (which lose sales). Tools like Inventory Planner, which integrates with Shopify, use historical sales data and seasonality to surface reorder recommendations. This isn't glamorous AI, but it's the kind that directly protects margin.
Where E-commerce AI Still Falls Short
Not everything pitched as AI-powered e-commerce improvement lives up to the claim. Three areas where the gap between pitch and reality is still significant:
- AI styling advice chatbots. The idea is compelling — a conversational interface that helps shoppers find the right product based on their needs and preferences. The execution is consistently disappointing. The conversation logic is too shallow to handle the nuance of real fashion or fit questions, and shoppers quickly lose patience with a bot that can't meaningfully understand what they're looking for. Better UX filtering is usually a more effective investment.
- Fully automated ad creative. Meta's Advantage+ and Google's Performance Max can optimise delivery, but the creative quality from fully AI-generated ad sets is still noticeably below what a thoughtful human creative produces. Use AI to generate copy variants for testing; keep a human in the loop on which creative concepts actually get produced.
- One-size-fits-all personalisation engines. Smaller stores don't have enough data for sophisticated personalisation models to work well. A recommendation engine trained on 200 purchases will make poor recommendations. The data requirements for meaningful AI personalisation are higher than most vendors disclose. Under roughly 10,000 monthly orders, rules-based personalisation often outperforms AI-based approaches.
The e-commerce businesses getting the most from AI right now are solving specific, data-rich problems — catalogue content at scale, semantic search, email recommendations for customers with purchase history — rather than trying to AI-enable their entire operation at once. That selective approach produces measurable results. The ones trying to deploy AI everywhere simultaneously tend to spend a lot on tools and see diffuse, hard-to-attribute improvements. Pick the highest-leverage problem first, solve it properly, then move to the next one.