Best practices

These are operational starting points derived from product behavior and common failure modes-not a claim about an unpublished group of highest-converting stores. Adapt them to your risk, data, traffic, and team.

Setup week

Choose a monitored launch window. Make sure someone can check answers, handoffs and integration errors after installation.
Start with the knowledge base. Add the questions, policies, and exact facts your team repeatedly uses, then test retrieval.
Set a brand voice. Pick a tone preset and add clear instructions and exceptions. Test the wording against real store policies.
Review rescue before enabling it. Check eligibility, consent, offer limits and purchase suppression with a test shopper.
Review the AI cost cap. Choose a cap from expected traffic and your acceptable interruption risk. A low cap can stop AI calls before the conversation allowance is used. See AI budget.

System prompt patterns

Things to include and exclude:

Do include: brand POV, hero product callouts, do-not-say list, escalation triggers, tone in 1 sentence.
Don’t include: long product taxonomies (the AI retrieves them anyway), exhaustive pricing rules (use the discount band setting), case studies, marketing copy.

Knowledge base patterns

Use clear question-shaped titles where helpful, and test retrieval with real shopper phrasings.
Keep approved policy details and exclusions accurate. Test whether the answer cites the right source instead of assuming it will quote verbatim.
Add one section per “first principle” - what makes us special, what we don’t do, our point of view on the category.
Refresh knowledge when products or policies change. Wait for indexing to finish before testing the new answer.

The common-question test

Start with a reviewed sample of the questions your support team actually answers. Add approved sources, rerun the sample, and record which answers improved, remained unsupported, or should still hand off.

Widget patterns

Launcher on every page is the right default.
Add an inline widget on landing pages where the AI is the hero - gift-finder, quiz, product comparison.
Test popup timing and email recovery separately. Neither channel is universally better; check shopper experience, consent, opt-outs and orders.
For mobile-first stores, test the launcher panel and inline placement. There is no separate fullscreen embed mode. See supported placements.

Recovery patterns

Use a verified sender and a monitored reply-to address. Test delivery and replies; no reply-rate improvement is guaranteed.
Don’t discount on the first email. Save the offer for the last step of the sequence.
Review the timing and suppression controls on the workflow you use. Check consent and overlapping campaigns in connected tools before sending.
Test a supported delay with a controlled audience and inspect the actual scheduled send. Browse-rescue scanning and journey-step delays are different mechanisms; see the recovery and journey guides.

Training patterns

Triage the gaps inbox once a day for the first two weeks. Weekly after that.
Reject unsupported or speculative drafts. Edit from an authoritative source and approve only accurate answers, then verify retrieval.
Use Locked Facts for approved shipping cutoffs and warranty terms, then test exact questions, paraphrases and exceptions. Locked Facts guide replies but do not guarantee error-free answers.
Skim the AI Feedback page after big changes to see real impact on answer quality.

Analytics patterns

The metric that matters: assisted revenue / billable conversation. Watch the ratio, not the raw numbers.
Review delivery, clicks, purchases and opt-outs together. Opens alone cannot diagnose a copy problem or rule out a delivery problem.
Use the weekly review alongside underlying orders and conversations. Follow up on examples before changing policies or offers.

Operations patterns

Configure the store-level SLA settings your team can meet. They do not provide separate per-channel SLA policies. See SLA settings.
Route VIP escalations to a single owner - the founder during early stage, the CX lead at scale.
Re-train quarterly. Set a calendar reminder; product copy and brand voice drift.
Review handoff reasons regularly. A growing low-confidence share is a reason to inspect examples, sources and integration state, not proof of one particular cause.

Anti-patterns to avoid

Putting unsupported claims or catalog dumps in store instructions. Use the structured AI settings and knowledge sources.
Changing many widget settings at once without recording what changed.
Discounting on the first recovery email - sets a bad expectation.
Ignoring the weekly review.
Mixing staging and production stores in one workspace.