Multiple bots in one shop: personas that sell

Multiple bots in one webshop: when personas help instead of confuse

An AI product advisor does not need the same role everywhere in a webshop. A visitor looking for a gift asks different questions from a trade buyer comparing parts. Someone on a Dutch storefront expects different language and product terminology from a shopper in an English-speaking market. Fynd lets you create multiple bots for one shop, each with its own persona, welcome, suggested prompts, theme colours and, when needed, separate knowledge.

That does not mean every page needs a new bot. More bots only add value when they create a clearer job, better context or more reliable answers. This guide helps you make that decision in a practical way.

Start with a job, not a character

A persona is not a witty name or a marketing costume. It defines what a bot is trying to solve, how much detail is appropriate, and which information it may use. Before creating a bot, write one sentence beginning: “This bot helps visitors who …”

Useful examples include:

• choosing between consumer products;
• finding a gift within a budget;
• helping a professional compare compatibility, specifications and quantities;
• answering questions inside a separate brand line or collection;
• guiding visitors through the catalogue in another language.

“This bot is friendly” is not a job. Every bot can be friendly. “This bot helps an installer find a compatible part without recommending consumer accessories” is a job. Use that sentence as a filter: if a prompt, source or answer does not support it, it probably does not belong in that bot.

When to split bots

Split a bot when the question, required knowledge or desired conversation changes materially. Four situations appear frequently.

Different buying intent

A fashion retailer might offer a consumer style advisor alongside a corporate-gifting assistant. The first can help with fit, occasion and styling. The second should focus on quantities, personalisation, lead times and invoicing. A generic assistant will often recommend products too early, or ask too many questions of a shopper who simply wants a quick choice.

Make the distinction obvious in the welcome. “I can help you choose an outfit or gift” is different from “I help businesses with larger orders and customisation.” Visitors should not have to infer which route is right for them.

Separate brand worlds

A multi-brand shop may sell brands with distinct audiences, terminology and price points. A premium skincare line calls for ingredient and routine guidance with careful claims. A sports nutrition brand calls for usage moments, flavours and bundles. A shared bot can mix those brand voices and accidentally steer a visitor from one line into another.

A bot per brand line keeps guidance coherent. Use the brand’s theme colours in the widget, while keeping the basic interaction familiar: opening, closing and finding help should work the same way everywhere. Also provide a route back to the main shop when a visitor needs a wider selection.

Professional versus consumer data

For technical, B2B or parts catalogues, this is often the strongest reason to split. Professionals search by model number, standard, material or connection. Consumers search by outcome, size or “will this fit my device?” If one bot searches every technical document and every marketing page for both audiences, neither group gets the best answer.

A professional bot can offer a prompt such as “I have a part number” and ask for a SKU, series or dimension. A consumer bot can begin with “Which device do you have?” and explain options without specialist language. The catalogue may be shared; the conversation flow should not be.

Language is more than translation

Separate Dutch and English bots are useful when markets differ in substance, not just interface language. Dutch customers may see different delivery options, product names, measurements or legal information from visitors in the UK or US. Create NL and EN personas with local welcomes, prompts and sources in that case.

Do not simply machine-translate the same instruction. A size guide, units, returns expectation and shipping terminology can vary by market. Let the bot follow the storefront language by default, and make a language switch visible if visitors often change languages. A bot that understands an English question but links to a Dutch product page still feels unreliable.

Knowledge isolation: optional, sometimes essential

Fynd can let bots operate with separate knowledge. This is valuable when a bot must answer only from a brand collection, B2B range or country-specific service information. Isolation prevents it from pulling a product, policy or price context from another line into an answer.

Use isolation when:

• a brand has its own ingredient claims, manuals or compatibility rules;
• a B2B catalogue is not relevant or visible to consumers;
• countries have different delivery, returns or availability;
• internal guidance belongs to one sales route only;
• legal wording differs by brand or market.

Isolation is not automatically better. If every bot needs the same catalogue, stock logic and general service pages, total separation creates duplicate maintenance. Use a shared core for delivery, payment and returns, with bot-specific instructions for language and priorities. Still test that a specialist bot does not recommend outside its permitted range.

Define every persona in five decisions

Keep a brief internal definition for each bot. That stops differences from being merely cosmetic.

1. Audience and job. Who opens it, on which page, and for what purpose?
2. Tone and depth. Should it be concise and choice-led, or technical and systematic?
3. Permitted knowledge. Which collections, documents and service sources may it use?
4. Boundaries. Which claims, product lines, questions or actions should be handed off?
5. Next step. Should it compare products, open a PDP, suggest a bundle or offer contact?

A sleep-brand persona, for example, can be calm, practical and non-medical. It can compare mattress firmness and sizes, but should not diagnose conditions or make health promises. A business bot may ask for desired quantities, but must not invent a quote or stock commitment. Those boundaries matter more than an enthusiastic opening line.

Give each bot a recognisable start

The first few seconds determine whether a widget feels helpful or distracting. Match the welcome and two to four suggested prompts to the job. Avoid prompts that merely demonstrate what AI can do.

For a gift advisor, useful prompts include:

• “I need a gift under €50”
• “What suits someone who enjoys cooking?”
• “Which sets can ship directly?”

For a technical advisor, try:

• “Which filter fits my model?”
• “Compare these two connections”
• “I have a SKU or part number”

For an English storefront:

• “Help me choose the right size”
• “Compare these two products”
• “What can arrive this week?”

Never imply that the bot already knows the visitor. “How can I help?” is fine, but a concrete route lowers effort. Keep prompts short enough to read without truncation on mobile.

Place bots deliberately, not all at once

Multiple bots should not mean multiple floating buttons on the same screen. That creates visual noise and forces a decision without context. Show one primary Fynd widget per page, selected by storefront, collection, category, campaign or account state.

A practical setup could use:

• a general shop bot on the home page and broad search pages;
• a brand bot on that brand’s landing pages and PDPs;
• a B2B bot inside the trade portal;
• a local-language bot on a country storefront.

When someone moves from a brand page to a mixed category, do not abruptly change bots in the middle of a chat. Offer “Help with the whole shop,” or preserve the current bot until the session ends. Conversation continuity usually matters more than perfect segmentation.

Make visual differences functional

Theme colours help visitors understand which expert is assisting them, especially across clear brand worlds. Apply them to accents, avatars and the opening state, not to radically different controls. The button must remain visible, keyboard reachable and easy to use on mobile in every context.

Check contrast for text, icons and focus indicators. A campaign colour that looks attractive may have insufficient contrast in a closed widget. Give every bot a clear text name rather than relying only on a logo; screen reader users and new visitors need that context.

Test in the playground before going live

Put a new persona through the Fynd playground first. This is not a cosmetic final check: it reveals whether boundaries hold up in real questions. Create a small test set per bot with at least:

• three typical product questions;
• two questions that belong to another bot;
• one question outside its knowledge;
• one question about delivery, returns or price;
• one question in the other language for a multilingual shop.

Do not judge grammar alone. Confirm that suggested products belong to the right line, uncertainty is stated honestly, links go to the right pages and the voice fits the persona. Test follow-up questions too. Shared conversation context can accidentally carry a previous brand name or market condition forward.

Only then enable the live placement. Consider starting with one collection or limited storefront. Review conversations where visitors rephrase, abandon the chat or change topic. Those are signals that the persona, prompts or knowledge boundary is too broad.

Common mistakes

The first mistake is creating a bot for every internal team instead of every customer job. Marketing, support and sales ownership may make sense internally, but a shopper should not need to decide who is responsible. Hide that ownership behind a clear customer role.

The second is splitting the persona without splitting knowledge or instructions. A “premium bot” that uses the same sources and recommendations as a budget bot provides no distinct experience. The third is isolating too aggressively, so a brand bot cannot answer a simple shipping question that matters to every visitor.

Other problems include translated welcomes without local prompts, inaccessible brand accents and automatic bot switches without explanation. Finally, do not go live because a bot gave one good demo answer. Use the playground for variants, difficult follow-ups and wrong-route questions.

Reveal complexity only when it helps

The best multi-bot setup does not feel like a system with many bots. It feels like meeting the right product expert at the right time. Start with one general bot, then add a persona only when you see a clear question group, brand boundary or market difference. Document the job, knowledge, language and hand-off. Keep one widget per screen, keep interactions familiar and validate every configuration before launch.

That turns multiple bots from an extra layer of technology into focused product help that gets shoppers to a suitable choice faster.