"Best carry-on luggage under $300", "a gift for a dad who cycles", "non-toxic cot mattress Australia": questions that used to open twenty browser tabs now go straight to an assistant, which replies with a handful of specific products and reasons.
For e-commerce brands, that reply is a new shelf. Small, brutally curated, and stocked entirely from what the engines can read and verify about your products.
Getting onto that shelf is an evidence contest, not a paid placement; there's nothing to buy yet. The assistants recommend products whose facts are explicit, whose claims are corroborated by third parties, and whose data is clean enough to be retrieved confidently.
That's a playbook, and most Australian brands haven't started it.
Make every product fact explicit
Assistants can't recommend what they can't pin down. Every product page should state, in crawlable text, the facts a recommendation needs: current price in Australian dollars, concrete specifications (dimensions, materials, weight, compatibility), availability, shipping to Australia, and warranty or returns terms.
Vague lifestyle copy ("engineered for the modern adventurer") gives an engine nothing to compare. "1.8 kg, 55 × 35 × 23 cm, fits most Australian domestic carry-on limits" gives it a reason to shortlist you for a specific question.
Freshness matters too: a price or stock status that contradicts reality teaches engines your data can't be trusted, and stale sources get quietly dropped.
Reviews and retailer listings are your proof layer
An engine recommending a product is staking its credibility on it, so it leans on evidence beyond your own site. Aggregated reviews (on your product pages via review platforms, on Google, on marketplaces) are the strongest signal, especially when reviewers mention the specific use cases buyers ask about.
Retailer and stockist listings corroborate too: a product that appears consistently at multiple reputable retailers, with matching names and specs, reads as an established product rather than a claim. If you're stocked anywhere beyond your own store, make sure those listings carry correct, consistent data: they're testifying about you whether you curate them or not.
Feed and schema hygiene: the unglamorous edge
Structured data is how you remove ambiguity at scale. Product schema with price, availability, ratings and identifiers (brand, GTIN or MPN where they exist) lets engines extract facts without guessing, and a well-maintained merchant feed keeps those facts synchronised as your catalogue changes.
The hygiene test is boring and decisive: no contradictions between page copy, schema and feed; no dead variants lingering in the index; one canonical product name used everywhere. The same citability principles that govern content apply to catalogues: we've written up the general version in how to write content LLMs cite.
What the evidence says about conversion
Does AI recommendation actually sell product? The most honest data point comes from randomised retail experimentation, and it's a range rather than a headline.
0–16.3%
sales lift measured in a randomised retail experiment when AI surfaced products: a range, not a promise; outcomes varied by category and context
Read both ends. The zero says AI recommendation is not magic: categories, price points and buying contexts differ, and some saw no lift at all.
The 16.3% says that in the right conditions the channel moves real revenue. The practical conclusion is to instrument your own funnel rather than believe either number: track assistant-referred sessions, watch which products get recommended for which questions, and measure through to orders.
Treat published figures as a reason to run the experiment, not a substitute for it.
A 30-day starting sequence
- Week 1: audit. Ask the major assistants the buying questions your category wins on, and record which products and brands get named.
- Week 2: facts. Rewrite your top 20 product pages so price, specs, availability and shipping are explicit and current.
- Week 3: proof. Fix schema and feed inconsistencies, then tighten your review-collection flow so every delivered order generates a request.
- Week 4: measure. Set up tracking for assistant-referred traffic and revisit the audit questions to establish your baseline.
From there it compounds monthly, the same way search always has, except the shelf is smaller and mostly empty. The broader recommendation playbook is in how to get recommended by ChatGPT, and our e-commerce programme is on the industries page.
Both: the shelf is built from evidence, not brand size. A small brand with explicit facts, clean schema and genuine reviews frequently outranks a bigger name with a vague catalogue, particularly for specific Australian-context questions like local shipping, compliance or sizing.
They work together. Your pages are the canonical fact source; marketplaces and stockists are corroboration. Contradictions between them are the real killer: an engine that sees two prices or two spec sheets for one product recommends the competitor whose story is consistent.
Start with referral sources and landing-page patterns in your analytics, and add a "how did you hear about us?" field at checkout: assistant mentions show up there before they're visible anywhere else. Perfect attribution doesn't exist yet; directional evidence is enough to justify the work.
Sponsored formats may well emerge, but recommendation engines stake their usefulness on trustworthy answers, so verified facts and third-party proof are likely to stay decisive. The evidence trail you build now is the asset either way.