AI for ecommerce: search, recommendations, support and catalogue at scale
Semantic search, personalised recommendations and assistants that know your catalogue
There are four places AI reliably earns its cost in a store. Search that understands "warm jacket for hiking" rather than demanding the exact words. Recommendations tuned to margin as well as clicks. A support assistant that can actually see an order. And catalogue enrichment at a scale no one is going to do by hand. Everything else is worth waiting on.
AI where it moves ecommerce numbers
On-site search is usually the fastest win and the most neglected. On most stores a quarter of buyers use the search box, and they convert at several times the rate of browsers, yet the search returns nothing for anything but an exact match. Semantic search fixes that, with typo tolerance and your merchandising rules still on top.
Recommendations need an objective before they need an algorithm. Optimise purely for click-through and you will be shown your own bestsellers on every page. We set the goal with you, usually a blend of margin and discovery, and measure it against a holdout group.
A support assistant that can verify a customer and look up an order removes the bulk of your tickets, since most of them are where is it and can I return it. That is built as an AI assistant connected to your commerce platform.
Catalogue enrichment is the unglamorous one. Attributes, descriptions and translations drafted across thousands of SKUs and pushed through a review queue, because nothing publishes unreviewed. Better attributes then improve filtering, search and ecommerce SEO at the same time.
One at a time, with a holdout
Each capability is piloted on its own and measured against a group that does not get it. Ship four at once and you will never know which one worked, or which one quietly cost you money.
What we build for stores
Semantic search
Search that understands intent and tolerates typos, with your merchandising and stock rules layered on top rather than overridden.
Recommendations
Cross-sell and upsell weighted for margin and discovery, not just click-through, and always measured against a holdout group.
Support assistant
Order status, returns, sizing and product questions answered from your own data, with the customer verified before anything is disclosed.
Explore →Catalogue enrichment
Attributes, descriptions and translations drafted across thousands of SKUs and pushed through a review queue before anything goes live.
Pricing & stock signals
Alerts from your sales and inventory data: lines about to sell out, products dead for ninety days, margins drifting the wrong way.
Measurement
Every capability A/B tested or held out, with the result reported even when it is that the feature did nothing.
One capability, start to proven
Pick the capabilityweek 1
From your analytics: search exit rate, ticket categories, catalogue gaps. Whichever of the four has the most money sitting behind it goes first.
Define the winweek 1
The metric, the holdout size and the threshold that would make it worth keeping, all agreed before anybody builds anything.
Build & connectweeks 2–5
Integration with Shopify, WooCommerce, Magento or your headless stack, plus whatever your existing search or helpdesk tooling supports.
Pilot against a holdoutweeks 5–8
Live for a share of traffic while the rest carries on as before, until the difference is large enough to be believed.
Keep, kill or tuneweek 8
The numbers decide. If the lift is not there we say so and either tune it or turn it off rather than declaring victory.
From $6,000/project
From $6,000 per capability, pilot first.
Ecommerce AI, answered
Which ecommerce platforms do you work with?
Shopify, WooCommerce, Magento and headless stores. Where you already pay for a search or helpdesk tool, we would rather configure that properly than sell you a replacement, and we will say when the existing tool is genuinely the limit.
Will recommendations just push the same bestsellers?
That is what happens when the only objective is click-through, and it is why so many recommendation widgets feel useless. We set the objective with you, weighting margin and discovery alongside clicks, and check the result against a holdout.
Can the assistant see customer orders?
Yes, once the customer is verified through an email or order-number check. That is what turns it from a FAQ widget into something that resolves tickets, since most support contact is a question about a specific order.
How is generated catalogue content reviewed?
Everything goes into a review queue and nothing publishes unreviewed. For a large catalogue that means a person approving in batches, which is still an order of magnitude faster than writing two thousand descriptions from nothing.
How much does AI for ecommerce cost?
From $6,000 per capability, with a measured pilot first. We would rather prove one thing on your own numbers than sell a bundle of four, and the pilot result decides whether the next one is worth starting.
How long does each capability take?
Four to eight weeks including the pilot period, which is the part that cannot be compressed because it needs enough traffic to be conclusive. A store with modest volume may need longer before the holdout comparison means anything.
Prove one thing first.
Send us your search exit rate and your top support ticket categories. We will tell you which capability is worth piloting.