Order value may rise because prices or the customer mix changed. Review purchase count, contribution, returns and why an item was added. A useful recommendation solves the customer’s next task rather than merely expanding a basket.
Compare the same categories, periods and conditions. Separate an additional offer’s effect from discounts or seasonal demand changes.
AI proposes; a person checks compatibility
Start with an approved catalogue: the task a product solves, compatibility, exclusions and limitations. AI drafts bundles with a reason for each recommendation.
A reviewer removes unavailable, incompatible or irrelevant items. For RU/EN offers, check that names, specifications and terms are equally clear in both languages.
Example recommendation card
Illustration. Customer task → core product → related need → suitable item → reason → limitation → customer decision. Record a declined offer and subsequent returns separately.
Stop the use case when specifications are invented, pressure increases or returns and complaints worsen. Continue based on usefulness and contribution after costs, not order value alone.
Common questions
Do we need a recommendation engine on day one?
No. First test a few rules and AI-drafted recommendations with a team member. Automate after validating demand and recommendation quality.
How does purchase frequency differ from order value?
Order value describes one purchase; frequency counts purchases during a defined period. Test them separately with the natural buying cycle and comparable groups in mind.
AI helps identify patterns and generate hypotheses. Verify outputs against source evidence, use de-identified data and keep a person accountable for decisions.
Choose a segment and period; define order value, purchase frequency and contribution margin.
Baseline metrics, data source and comparison boundaries.
Step deliverable: Baseline metrics, data source and comparison boundaries.
2. Find a useful pairing
Use de-identified orders to ask AI for pairings and value explanations. Review customer need and avoid confusing correlation with causation.
One pairing, customer value, uncertainty and recommendation guardrail.
Step deliverable: One pairing, customer value, uncertainty and recommendation guardrail.
3. Test the offer
Present the offer to a small group at an appropriate moment. Compare order value, frequency, margin and declines.
Channel, audience, message, duration, metric and risk limit.
Step deliverable: Channel, audience, message, duration, metric and risk limit.
4. Decide
Record actual results and side effects. Revise the hypothesis if trust or margin worsens.
Result, limitations and next test or decision to stop.
Step deliverable: Result, limitations and next test or decision to stop.
How to judge whether it helped
You should leave with a document that supports a decision and a next action. Completing steps does not guarantee a commercial result. Measure the chosen change separately over an agreed period. If evidence is missing, state that and plan how to collect it.
A baseline, actual result and data limitations are recorded.
A person made the decision after reviewing quality and side effects.
How this differs from S01–S13
This is a user route for a specific problem. The GRI AI research stages S01–S13 and the 13-week programme remain in the library as separate methodological work. This route is not a validated GRI diagnostic scale.
Your route0 / 4 · 0%
Completed 0 · remaining 4
Next action: Record baselineThe percentage measures completed steps, not proven business growth.
PRACTICE / 1
Record baseline
Choose a segment and period; define order value, purchase frequency and contribution margin.
Step deliverable: Baseline metrics, data source and comparison boundaries.
No names, contacts, client secrets or sensitive details. Saving happens only when you press the button.
Discuss opens WhatsApp. Your answers are not sent automatically.
Data & privacy
Matching works without saving your query. Practice answers remain in this tab until you enable saving. Do not enter names, contacts, client secrets or sensitive information.
Server saving: answers and steps are linked to a random key in a secure browser cookie for up to 90 days. This is pseudonymous, not anonymous data. It does not sync across devices.
Server saving is not enabled yet: the data controller and hosting configuration are being confirmed. You can work in this tab and download your results.
