THE PRACTICAL GUIDES
Recommendation system consultants in Australia: eligible products before ranking
A product-recommendation project has two jobs: determine which items may be shown, then choose a useful order for the eligible items. A model's preference score should not override stock, delivery or a withdrawn product. Write those rules into the supplier brief before judging a demonstration by how personalised it looks.
This guide helps Australian buyers commission a recommendation system using published scope from existing directory providers. It concerns selecting and ranking products, rather than the separate task of forecasting future demand. Browse AI implementation providers for service discovery.
Define the place where recommendations appear
Start with one placement: related products on a product page, alternatives when an item is unavailable, or a suggested next purchase. Those can use different inputs and answer different questions. A “similar item” list does not necessarily require personal history, while a returning customer's suggestions may.
Name the user, trigger and allowed action. For example: “After someone views one approved product, show up to three currently available alternatives deliverable to the selected region.” Showing a suggestion is different from placing an order, changing a price or contacting that person.
Agree the catalogue identifiers, current stock source, delivery rules, product removals and update frequency. If those records cannot be reconciled, use the data-readiness guide to scope the foundation before commissioning a ranking model.
Two advertised capabilities to discuss
Resolute AI lists recommendation systems among its services on its official site. It describes feasibility work, demonstrations, code and documentation. Ask what the feasibility phase will show for your catalogue and which implementation or continuing operation is separate.
Ippon Australia advertises personalisation and recommendations within its AI service page, alongside product design, engineering and evaluation. Ask which user interface, catalogue integration, evaluation and handover work belongs to the proposed engagement.
These are published capabilities, not measured results on your products. Resolute describes an hourly engagement with a ten-hour weekly minimum; obtain the current rate, currency, minimum duration and complete scope in a proposal. No equivalent fixed-price recommendation project was established across the two pages.
An original five-item catalogue check
Consider this fictional catalogue for a visitor in one defined delivery region. Assume only the stated conditions apply. The list is an editorial commissioning exercise, not customer data or a model trial.
| Item | Current catalogue condition | Expected eligibility under this brief |
|---|---|---|
| A | Active, in stock, deliverable to the selected region. | Eligible for ranking. |
| B | Active but out of stock. | Excluded from this available-alternatives placement. |
| C | In stock but unavailable in the selected delivery region. | Excluded for this visitor. |
| D | Active, in stock, deliverable to the selected region. | Eligible for ranking. |
| E | Withdrawn from sale. | Excluded. |
Only A and D are eligible. If the placement allows up to three suggestions, agree whether it should show two, use an approved fallback or show none. It must not invent another item or quietly relax the rules to fill a third slot.
A model might rank B highly because many people previously bought it. That historical preference does not establish that B is available now. Ask the supplier to demonstrate both the retrieved candidates and the final displayed items, with the exclusion reason available for review.
Google's recommendation-system guidance describes filtering and other constraints in re-ranking. The buyer still needs to specify the actual rules and where they are enforced; the technology label alone does not supply them.
Compare with a simple existing rule
Keep an agreed reference such as a manually curated related-products list or an eligible popularity list. Ask the supplier to explain whether the proposed system improves the decision you care about compared with that reference. An elaborate demo should not erase the simpler alternative from the proposal.
Choose observable outcomes with clear units: useful product selections, valid recommendations, add-to-basket actions or completed purchases. These are different events. Agree how repeated actions, unavailable products, returns and users seeing different placements are handled before reporting a result.
Historical testing and a future controlled experiment answer different questions. A good offline ranking score is not proof of additional sales. Any later live trial needs agreed audiences, exposure records, duration and operating limits. No experiment or test traffic was run for this guide.
Ask for the operating catalogue, not only a model
The handover should identify who updates stock and delivery rules, how fast removals reach the recommendations, and what users see when that update fails. Ask about new visitors, new products and items with little history. Agree a fallback and a way to pause the placement without breaking the original product pages.
Separate catalogue preparation, integration, model work, evaluation and interface delivery from continuing usage, monitoring, support and later changes. Specify the ownership and export of catalogue data, event definitions and evaluation records. Use the quote worksheet to compare the same scope.
Sources reviewed 26 October 2026, Melbourne time. No business system, store, customer record or model was accessed. Provider descriptions remain advertised claims. Browse all AI buying guides.
Editorial guidance with linked sources. Reviewed 2026-10-26.