How to Build an AI Knowledge Base for a Website

01

Define which sources may support an answer

List the public pages, documentation, policies, FAQs, and connected store data that contain customer-facing facts. Remove obsolete or contradictory material before importing it. Private or login-only information needs a separate approved workflow.

02

Inspect the knowledge before customers use it

Check names, numbers, dates, prices, policy conditions, product attributes, and destination links. Add approved owner notes where the public source is incomplete, and assign a business owner to facts that require judgment.

03

Use real questions and rescan after changes

Test short questions, follow-ups, misspellings, competing sources, and cases that should go to a person. When the website, documentation, catalog, or policy changes, rescan the affected source and rerun the relevant question set.

04

Treat unresolved questions as source gaps

If the sources do not support a reliable answer, the assistant should ask for clarification, point to an approved contact route, or notify an operator. Measure unresolved intents and time to a usable answer instead of promising universal accuracy or a fixed ticket reduction.

05

Map every answer to a source hierarchy

Start with the source that owns each fact. Product and service pages should own offer details; policy pages should own delivery, returns, privacy, and contractual conditions; approved owner notes should supplement public sources only where the website is intentionally incomplete.

When two pages disagree, correct the source or record an explicit owner decision before launch. A larger knowledge base is not automatically a better one: clear ownership is more useful than repeated, conflicting text.

06

Control scan scope and language duplicates

Use the sitemap and crawlable navigation as discovery inputs, then keep only canonical customer-facing pages. Exclude private areas, internal search, filter combinations, expired campaigns, and translated duplicates that are not part of the chosen language branch.

For multilingual websites, build each assistant from the intended primary language source instead of mixing copies of the same facts. The assistant can answer in the visitor's language without storing every translated URL as duplicate knowledge.

07

Build a practical evaluation set

Create test questions from real visitor intents: exact factual questions, paraphrases, short follow-ups, misspellings, comparisons, requests for a link, and questions whose answer is absent. Include sensitive or account-specific cases that should move to an approved human route.

For each question, record the expected fact, source, destination, and fallback behavior. Test the same set after source or configuration changes so a polished demo does not hide regressions in ordinary conversations.

08

Treat updates as an operating workflow

Rescan the affected pages when products, services, prices, policies, forms, or navigation change. Review what was added, changed, or removed before treating the refreshed knowledge as ready for visitors.

Use unresolved questions as a maintenance queue. If several visitors ask for the same missing fact, improve the authoritative website page or add an approved owner note, then repeat the failed question and its natural variations.

09

What a website scan becomes in AiRep24

AiRep24 does not turn every sentence into an unreviewed answer. The owner can inspect the detected pages and structured blocks, correct the text, add a site-only note, and test the assistant before the widget is published.

Website input	What the owner reviews	Launch check
Home, service, and product pages	Grouped knowledge records with their source URL and detected type.	Ask for the exact offer, qualification, and destination.
Delivery, return, privacy, and other policies	The policy text and the conditions the public page actually states.	Test a normal case, an exception, and a missing condition.
Contact links, forms, and primary calls to action	Detected website actions stored separately from descriptive copy.	Open every suggested destination and submit only a safe test form.
Owner-approved facts absent from the public site	A manual note that belongs only to the selected website knowledge base.	Ask a conflicting question and confirm the approved note wins.
Questions the assistant could not resolve	Knowledge gaps visible for review rather than silently converted into claims.	Update the source or note, then repeat the exact failed question.

A repeatable release check

Open Knowledge → Content and inspect the imported groups.
Review Quality for missing, thin, or conflicting source material.
Use Sync settings to choose how the website stays current.
Test exact questions, paraphrases, follow-ups, missing facts, and destination links.
Publish the widget only after the expected answers and escalation path are clear.

Real implementation

Verified implementation examples

See how maintained sources support real visitor questions without inventing performance claims.

Real implementation

How SoloCruz makes a large cruise catalogue easier to navigate

SoloCruz helps solo travellers find a cabin mate around a specific sailing before booking. The information problem spans destinations, dates, ships, cabin expectations, listing rules, and public guidance rather than one short FAQ.

Example visitor request: “I want a Mediterranean cruise from Barcelona in October. Can I find a cabin mate before booking?”

Open customer story

Real implementation

How MyUGC Studio guides visitors to the right AI creative workflow

MyUGC Studio turns product images into ecommerce photos and UGC videos using different models, locations, formats, and publishing paths. Visitors may understand the desired output without knowing which workflow or source material to choose.

Example visitor request: “I have a flat-lay shoe photo and need a vertical social ad. Should I start with product placement, an AI model, or video generation?”

Open customer story

Next practical step

Build the knowledge base from a real website

Add the public website you already maintain, inspect the pages AiRep24 discovers, and correct or supplement the resulting knowledge before publishing the widget.

Build your website knowledge base Review automatic knowledge setup