AI product recommendations explained for AU retailers

AI product recommendations are systems that decide, for one shopper at one moment, which products to show and in what order, using what that shopper and others like them have done before. Next-best-action personalisation extends the same idea from products to prompts: whether to suggest a wish list, a reorder, a delivery slot or nothing at all.

How it works

Most production systems have two stages. A retrieval stage narrows a catalogue of tens of thousands of items to a few hundred candidates, using signals such as what the shopper viewed or bought, what shoppers with similar histories went on to buy, and which items are frequently bought together. A ranking stage then scores those candidates for the specific context: the page they are on, the season, stock availability, margin rules and whatever the retailer has chosen to optimise, which might be click-through, add-to-cart, or revenue. Both stages are trained on historical behaviour and retrained as new data arrives.

The newer variation is the agentic or conversational assistant. Instead of a carousel, the shopper types or speaks a goal, and a language model translates it into searches and recommendations, often assembling a whole basket. Under the surface the same retrieval and ranking machinery does the work; the assistant is a different front door. Seasonal personalisation, the back-to-school or Christmas case, adds a time dimension: the model is told that the shopper is in a known buying mission and re-weights toward that mission's categories.

Who has deployed it in Australia

Woolworths Group's F26 results announcement (26 August 2026) describes a "transformed digital shopping assistant, Olive, leveraging agentic AI" and a "digital Smart Basket feature which builds an online basket with a single click". The same document reports group eCommerce penetration of 15.9 per cent and 31.5 million average weekly visits to group digital platforms, which is the scale of data these systems learn from.

Wesfarmers' 2026 full-year results (27 August 2026) state that "Bunnings, Kmart and Officeworks launched their agentic commerce shopping assistants 'Buddy', 'Joy' and 'Ollie'", and that the group's shared data asset holds approximately 12 million customer records, with OnePass members spending more than three times non-members.

Coles Group's 2026 Annual Report describes analytics for "store-specific ranging" and a Flybuys base of 10.3 million active members, but does not describe a customer-facing recommendation product in the filing.

The most specific public result this season is from the United States. Retail Dive reported on 3 September 2026 that Target's senior vice president of technology, Brad Thompson, said AI-generated wish-list recommendations and a "next best action" widget were used for back-to-school, that shoppers who built wish lists generated about 45 per cent higher category demand, and that Target plans per-shopper dynamic page layouts. That figure is Target's own, given to a trade publication, and has not been independently verified.

What it costs

Costs fall into three buckets, and vendors price them differently enough that a single figure would mislead. Licence models are usually one of: a percentage of the revenue attributed to recommendations; a fee per monthly tracked visitor or per API call; or a platform subscription tiered by catalogue size and traffic. Implementation effort is dominated not by the model but by data plumbing: product feeds with clean attributes, an identity layer that recognises the same shopper across web, app and store, and event tracking that records what was shown as well as what was clicked. Retailers that already have those assets can be live in weeks; those that do not should expect a data programme first. Ongoing load is a small team to manage rules, exclusions, experiments and retraining, plus the analytics work to prove incrementality. No public Australian retailer has disclosed the cost of its recommendation programme in a filing, so this guide gives structure rather than figures.

Where it breaks

Cold start is the first failure. A new shopper or a new product has no history, and a system that leans on history will either show popular items to everyone or show nothing useful. Range-heavy Australian retailers that launch hundreds of seasonal lines each quarter feel this most, because the items they most want to sell are the ones the model knows least about, and the usual workaround, seeding new products with attribute-based similarity, only works if the product data is clean. Popularity bias is the second: without deliberate constraints the model recommends what already sells, which is good for conversion this week and bad for range discovery over a year. Seasonal drift is the third, and the seasonal use case makes it worse: a model trained on winter behaviour will keep recommending winter into spring unless someone tells it the mission has changed, which is exactly the problem Target's next-best-action layer is trying to solve.

The measurement failures matter more than the model failures. Click-through on a recommendation carousel is easy to lift and almost meaningless; the shopper may have bought the item anyway. The only honest measure is incremental sales against a holdout group that did not see the recommendations, run for long enough to include a full purchase cycle. Retailers that skip the holdout end up with a vendor report saying the tool drove 30 per cent of revenue and no way to know if any of it was new.

Finally, recommendations are the wrong tool when the catalogue is small enough to browse, when the purchase is infrequent and considered, or when the retailer's advantage is curation. A 200-product specialist does not need a ranking model; it needs a good buyer.

Questions to ask a vendor

Ask how the system handles a shopper with no history and a product with no sales. Ask what it optimises for by default and whether you can change it. Ask to see a holdout-based incrementality result from a retailer of your size, not a click-through chart. Ask what data leaves your environment and where models are trained. Ask how seasonal missions are configured and who does it. And ask what happens to recommendations when an item goes out of stock in one state but not another, because in Australia that is not an edge case.

Last reviewed 9 September 2026.