AI & Content
Improve store AI answers
When customer chat AI is weak or inconsistent — add better data, keep sources aligned, and only then adjust models.
When merchants ask this
Typical questions:
- “My customers say the AI does not answer well enough.”
- “Replies are vague, invent facts, or contradict each other.”
- “How do I make WhatsApp / web chat AI more consistent?”
Model choice alone rarely fixes this. Quality comes from what the AI can read (catalog, Knowledge Base, instructions, procedures) and from those sources not contradicting each other.
What the store AI actually uses
| Source | Path | Role |
|---|---|---|
| Live catalog & bookings | Catalog → Products / Sales → Bookings | Prices, stock, variants, staff, availability — truth for sellable items |
| Knowledge Base | Assistant → Knowledge Base (or Assistant → Knowledge) | Policies, FAQs, how-to text — reference facts for answers |
| AI Instructions | Assistant → AI Instructions (or Assistant → Instructions) | Brand voice, hard constraints, what not to say |
| Support procedures | Assistant → Procedures | Return, delay, cancel, live-support playbooks |
| AI Models | Assistant → AI Models | Capability / cost — not a substitute for missing facts |
Booking slots, live prices, and product options should come from your live catalog and bookings, not from a stale paragraph in the Knowledge Base. Put policy and FAQ in KB; put sellable truth in Products / Bookings.
Symptom → what to fix first
| Customer complaint | Likely cause | Fix first |
|---|---|---|
| Invents products or prices | Thin catalog or stale price text in KB | Publish accurate Products; remove price lists from KB |
| Vague / “I don’t know” on policies | Missing or unindexed FAQ | Add clear KB docs; wait until status is ready |
| Two different answers on different days | Two sources with different numbers | Keep one source of truth; delete obsolete KB docs |
| Booking: asks date again / wrong slots | Hours or services incomplete; old KB hours | Fix Bookings + store hours; do not mirror slots in KB |
| Ignores brand tone | Empty or essay-length instructions | Short hard rules in AI Instructions |
| Never escalates / never cancels correctly | Procedures off or conflicting with Instructions | Enable Procedures; align refund/cancel rules |
| AI silent on WhatsApp | Inbox not in AI mode / WhatsApp not connected | Inbox AI mode + WhatsApp Business connection |
| Answers sometimes English, sometimes store language | Mix of EN/TR docs without clear language rule | Prefer one language per FAQ doc; set reply language in Instructions |
What belongs where (do not duplicate)
| Put here | Examples | Do not put here |
|---|---|---|
| Products / Bookings | Name, price, duration, staff, variants, stock, publish status | Long legal essays |
| Knowledge Base | Shipping policy, returns, hours, prep rules, “how do I…” | Live price tables that change weekly |
| AI Instructions | “Keep answers short”, “only catalog products”, “escalate when unsure” | Full product catalogs or PDF dumps |
| Procedures | Return window days, cancel rules, live-support task | Marketing slogans |
One policy fact → one place. If shipping is free over 500 in KB and over 300 in a campaign, the AI will look “random.”
Step 1 — Add more high-quality data
Prefer clear, current, customer-facing text over dumping the whole website.
- Open Assistant → Knowledge Base (
/en/settings/knowledge-base). - Add (or re-sync) documents that answer the questions customers actually ask:
- Shipping, returns, opening hours, payment methods
- Service descriptions and preparation rules (for bookings)
- Common “how do I…” flows
- Prefer short FAQ / policy pages over megabyte spreadsheets or noisy full-site crawls.
- Wait until indexing status is ready, then test the same question on WhatsApp or Inbox.
- If answers are still thin, add the missing URLs or paste the missing FAQ — do not only switch models.
Catalog hygiene (Products / Bookings)
- Publish only services/products customers can actually buy
- Clear titles and short descriptions (the AI quotes these)
- Correct variants, add-ons, and staff prices when they differ
- Inactive or draft items should not stay “almost published” if customers still ask about them
See Knowledge Base for upload, URL, crawl, and size limits. Bookings: Bookings.
Step 2 — Make sure data does not contradict
Inconsistent sources produce inconsistent replies. Before blaming the model, check alignment:
| If this says… | …but this says the opposite | Customers see |
|---|---|---|
| Knowledge Base: free shipping over 500 | Product / campaign: free shipping over 300 | Random or conflicting answers |
| Instructions: “never promise refunds” | Procedures: auto-guide refunds in 14 days | Soft refusals vs refund talk |
| KB: “open until 22:00” | Store setup / booking hours: until 18:00 | Wrong hours |
| Old crawled page still in KB | New policy uploaded as a second doc | Whichever document the AI happens to pick |
Rules of thumb:
- One policy fact should live in one place (preferably KB or Procedures — not both with different numbers).
- When a policy changes, update or remove the old KB doc (re-sync URL sources; delete obsolete files).
- Do not paste full price lists into KB if Products already hold live prices — the AI will mix stale KB with live catalog.
- AI Instructions should reinforce catalog/KB truth (“only sell listed products”, “use live availability”), not invent a second catalog.
Step 3 — Tighten instructions (after data is solid)
Assistant → AI Instructions (/en/settings/ai-instructions):
- Brand voice and languages
- Hard constraints (“short answers”, “do not invent products”, “escalate when unsure”)
- Channel notes if WhatsApp should differ from web chat
Keep rules short and enforceable. Long essays are ignored more often than short constraints. See AI Instructions.
Step 4 — Enable support procedures
For returns, delays, cancellations, and “talk to a human”, configure Assistant → Procedures so the AI follows your playbook with your live order and booking data instead of improvising. See Assistant & Support Procedures.
Step 5 — Models (optional polish)
Only after data and alignment look good, review Assistant → AI Models (/en/settings/ai-models):
- Stronger models can help on complex multi-step support
- Faster / cheaper models suit high volume when facts are already clear
- Changing models does not fix missing or contradictory Knowledge Base content
See AI Models.
How to test (same question, same channel)
- Pick one real customer question that failed.
- Confirm Products / Bookings / KB / Instructions / Procedures are updated and KB is ready.
- Ask again in the same channel (WhatsApp Inbox or web widget) with AI mode on.
- If still wrong, ask: which source still has the bad number? Fix that source — do not add a third copy.
- Metered chat uses AI credits (see Billing).
Quick checklist
- WhatsApp connected; Inbox AI mode on for the thread you are testing
- Products / bookings accurate and published
- Knowledge Base covers real customer FAQs — indexed and ready
- No conflicting numbers across KB, instructions, procedures, and catalog
- Instructions state voice + hard limits (short rules)
- Procedures enabled for return / cancel / live support
- Re-test the failing question on the same channel
- Then consider a stronger chat model if needed
