The 5 things that actually matter

Most AI chatbot buyer guides focus on features. Features barely matter. They all do roughly the same thing — answer customer questions from your content. What separates a good purchase from a regretted one is five things that almost nobody puts on a comparison page.

1. Data residency

Where are the conversations between your customers and the chatbot stored? Where is your business content stored? Where does the AI processing happen? These three questions have legal weight. If you serve EU customers, you have GDPR obligations about where personal data lives. If you serve healthcare patients, location of records matters under various health-data regimes. If you serve government or enterprise clients, they will ask you these questions during procurement, and "I don't know" is an instant disqualifier. The honest answer from a vendor should sound like: "Conversations and content are stored in [specific region], processed in [specific region], with [specific] safeguards on cross-border transfers." Vague answers — "we use cloud infrastructure," "stored securely," "we comply with all regulations" — mean the vendor either doesn't know or doesn't want you to know. Both are bad signals. For European-served businesses, EU data residency is the modern baseline. For Israeli-served businesses, EU residency also satisfies most domestic compliance needs.

2. Language support quality

Most platforms claim "30 languages." Most are running an English-trained model with a translation wrapper, which fails badly on Hebrew, Russian, Arabic, and other non-Latin or RTL languages. The right test isn't the marketing page; it's the live demo. Run the multilingual test protocol in your target languages before you sign. Specifically check: code-switching mid-sentence, idiom handling, RTL layout, locale-aware date and currency formatting. If you only serve English-speaking customers, this matters less. If you serve a multilingual customer base, this is one of the biggest differentiators between platforms — and the most often misrepresented in marketing.

3. Integration paths

A chatbot widget on your website is the minimum. A modern platform should also support: a standalone shareable chat link (URL you can post to social or email), QR codes that open into the bot, a Telegram bot tied to the same knowledge, and webhook integrations for sending leads to your CRM. Why this matters: customer behavior isn't single-channel. A customer might first scan a QR at a clinic, then ask a follow-up via Telegram, then click a link from your email campaign. If the chatbot is tied only to one channel, you lose that continuity. If it spans channels, you don't. Ask vendors specifically: "What channels does the bot run on, and is it the same conversation context across them?" If the answer is "widget only" — and your business has any meaningful offline-to-online flow — keep looking.

4. Knowledge update mechanism

Your business content changes. Prices update, hours shift, services get added, policies are rewritten. The chatbot has to keep up. Two questions to ask:

  • How often does the system reindex my content? Best answer: automatically, when changes are detected on your website, plus a daily full pass. Acceptable: scheduled reindexing every 24 hours. Bad: manual refresh only, or weekly.
  • What's the latency between a content change and the bot reflecting it? Best: minutes. Acceptable: hours. Bad: days. If you change pricing on Monday morning and the bot is still quoting old prices Tuesday afternoon, the platform isn't built for businesses that actually change content. Adjacent question: can the bot cite the source page for each answer it gives? If yes, you can audit and correct quickly. If no, you're flying blind on quality.

5. Exit cost

This is the question vendors don't want you to ask. What does it cost you — in time, money, and business continuity — to leave? Three sub-questions:

  • Data portability. Can you export your conversation logs and uploaded knowledge in a usable format? CSV, JSON, plain export — something. If the only export is a screenshot, you're locked in.
  • Contract terms. What's the notice period for cancellation? Are there auto-renewals? Is the contract month-to-month or annual? A vendor who insists on a 24-month contract for SMB is signaling lock-in expectations.
  • Workflow lock-in. If you've built booking flows, payment links, and CRM integrations on this specific platform, switching means rebuilding. Estimate the rebuild cost before signing — not after. A good vendor lets you leave easily. A vendor that makes leaving expensive is telling you they expect to retain customers through friction, not value.

Red flags to watch for

Five patterns that, taken together, mean keep looking. Vague data residency answers. "We use enterprise cloud infrastructure" is not an answer. The vendor should be able to name regions, processing locations, and cross-border transfer safeguards within 5 minutes of being asked. If they can't, they've never thought about it. No DPA template. A Data Processing Agreement is the standard contract any serious B2B vendor signs with EU customers. If a vendor doesn't have a template ready or asks "what's a DPA?" — they don't serve EU markets seriously, regardless of what their marketing says. Blocked or restricted free trial. If you can't get hands-on with the platform without a sales call, you're being filtered for budget instead of fit. Serious vendors offer real trials with real data, not 5-minute scripted demos. Vague pricing. "Contact sales" pricing for SMB-tier products usually means the vendor is going to price-discriminate based on perceived budget. Real SMB pricing is published. If the vendor will quote different prices to different customers for the same plan, ask why. Lock-in signals. Proprietary embedding format, no data export, mandatory annual contracts, "professional services" required for setup — any one of these alone might be fine. Together, they signal a vendor whose retention strategy is making it expensive to leave.

The 10 questions to ask any vendor before signing

These are the questions that separate a good purchase from a regretted one. Bring them to every demo. 1. Where is conversation data stored, and where is it processed? Good: "Conversations stored in EU region X, processed in EU region Y, with [specific] safeguards on transfers."
Bad: "We use secure cloud infrastructure." 2. Do you have a DPA template I can review? Good: "Yes, here's the link" — sent within 1 business day.
Bad: "We can put one together" or no response within 3 days. 3. Can I test the bot in my actual target languages with my actual content? Good: "Yes, the trial uses your real content and runs in any language you want."
Bad: "We have a demo dataset you can try." 4. How long between a content change and the bot reflecting it? Good: "Minutes — auto-detected when your site updates, plus a daily full pass."
Bad: "Whenever you click refresh in the dashboard." 5. Can the bot cite the source page for each answer? Good: "Yes, every answer links to the source page in your content."
Bad: "It uses your content but doesn't show the source." 6. What channels does the bot run on, and is conversation context shared? Good: "Widget, standalone link, QR, Telegram, all sharing context per session."
Bad: "Just the widget for now; other channels on the roadmap." 7. How do I export my conversation logs and uploaded content? Good: "CSV or JSON export, available anytime in the dashboard."
Bad: "Contact support for export options." 8. What's your pricing for my expected volume of conversations per month? Good: A specific number from a published price list, with overage costs disclosed.
Bad: "Custom pricing based on your needs." 9. What does cancellation look like? Notice period, refunds, data retention? Good: "Month-to-month, cancel anytime, data exportable for 30 days post-cancellation."
Bad: Auto-renewing annual contract with 90-day notice. 10. Show me a customer reference in my industry or my country. Good: A real customer name with permission to contact them.
Bad: "We have many customers but can't share names due to confidentiality." If a vendor stumbles on more than two of these ten, look elsewhere. If they answer all ten cleanly, you've found a serious one.

How to test for 30 days before committing

A concrete protocol for de-risking the purchase decision. Week 1: Setup and content load. Connect your sitemap, upload your documents, configure the widget, set up tracking. Time how long this takes — if it's more than 4 hours of your time, the platform is harder to maintain than the marketing suggests. Week 2: Internal testing. Have 3-5 people from your team run real customer conversations against the bot. Test in every language you serve. Run the multilingual test protocol if applicable. Note every wrong, weak, or weird answer. Week 3: Soft launch. Deploy to a low-traffic page or a single channel. Watch the conversation logs daily. Fix content gaps as they emerge — a bot's quality is a function of your content quality, and you'll learn fast where the content is thin. Week 4: Full deployment, with review. Roll out to your full traffic. End the trial period with a structured review: read the last 100 conversations, count clean answers vs. wrong vs. weak, calculate first-pass ROI math using actual numbers. If at the end of week 4 the bot is handling 60%+ of conversations cleanly and your gut says "this is helping," sign. If not, walk. Most vendors will let you extend a trial by another 2 weeks if you ask honestly.

Pricing models compared

Three pricing patterns dominate, and they reward different usage shapes. Per-message pricing. You pay for each conversation or each message. Predictable for low volume, expensive at scale. Works well for businesses with seasonal or unpredictable traffic — you only pay for what you use. Watch for: usage caps that throttle the bot during peak periods. Per-tier pricing. Flat monthly fee for a tier (Starter, Growth, Business), with a quota of conversations included and overage rates beyond. Most common SMB model. Works well for businesses with predictable traffic. Watch for: tier upgrade pressure when your traffic grows. Per-seat pricing. Charged based on number of admin users, not customer conversations. Common for enterprise platforms that bundle live chat. Works poorly for SMB, because it scales cost with your team size rather than with the business value the bot creates. For most SMBs, per-tier with transparent overage rates is the cleanest model. Per-message can win at very low volumes. Per-seat is usually the wrong fit unless you specifically need enterprise live-chat features alongside the bot. A note on price as a buying signal: cheapest is not best, but most expensive is rarely best either. The right price for an SMB chatbot is in the $30-500/month range depending on volume and verticals served. Anything dramatically below that range is usually missing one of the five things that matter; anything dramatically above is usually paying for enterprise features you don't need. The point of all this isn't to scare you off chatbots — they're useful tools when bought well. The point is that the buying decision is procurement, not marketing. Treat it that way and the rest follows.