"Just add AI": Introducing Conversational Search for your website

Webinar Q&A

Skip ahead:

• Setup and content
• Implementation and customization
• Security and other capabilities

How do you know if your content is ready for AI, or what needs to be fixed first?

At Squiz, we help you assess and prepare your content for AI in a few ways. We use an automated content auditing tool that analyzes content fragments across your indexed URLs. These tools color-code each fragment using a traffic light system – green for content that's ready for AI use, yellow for content needing minor optimization, and red for content requiring significant improvement. This gives you an instant visual overview of which content fragments are AI-ready and which need work.

We also do semantic testing, where we run common questions through the AI system to test the quality of answers. This testing quickly reveals where content needs improvement, where content gaps exist, and how well your existing content performs with AI.

Even if you're not planning to implement AI search, this process acts as an excellent content audit. When an AI can't provide good answers to common questions, it's usually because the content doesn't exist on your site or isn't easily accessible – indicating areas for improvement regardless of AI implementation.

What kind of content can we include in the knowledge base? And any limitations around content that wasn’t built on Squiz?

Because Conversational Search is powered by Funnelback Search, any content that works with Funnelback can be included in your knowledge base. This includes web collections, text files, and PDFs, with a focus on text-based content rather than images. You can also index content that wasn't built on Squiz – the system can pull from various sources and platforms to create a comprehensive knowledge base for your Conversational Search.

Can the Squiz conversational search draw on content not contained within our Squiz CMS? E.g. Can we point it to our knowledge base?

Yes, absolutely. Conversational Search can use any content that the search engine can access through crawling, as well as any content you push into the search index. The same mechanisms we currently use to include external content in traditional keyword search indexes are fully available for Conversational Search. This means you can incorporate your own knowledge bases, external documentation, or any other content sources that the search engine can access.

How granular can we get with specifying which content to include or exclude within the knowledge base?

You can get very granular. Similar to how Funnelback Search works, you can include or exclude content at the individual URL level, or scale up to entire site sections and subsections. The level of granularity is entirely up to you.

Seeing you can go as granular as you want - do you then have to set up multiple rules to cover different sections etc.?

You don’t need to set up different guardrail rules for different content sections, but you could if it adds value to your implementation. The system works perfectly well with a single set of guardrail prompts that apply across different sections.

However, if you have content areas that would benefit from different approaches – for example, different tone of voice for HR content versus student services, or specific restrictions for certain departments – you have the flexibility to create separate content slices with their own tailored guardrail settings. It’s about what makes sense for your organization, not a technical requirement.

Is Conversational Search a "Funnelback killer," or do we still need Funnelback if we implement this tool?

Conversational Search doesn't replace Funnelback – it's actually powered by Funnelback. Think of Conversational Search as an additional layer that sits on top of your existing Funnelback search infrastructure, adding AI-powered conversational capabilities to the robust search engine you already have.

Your traditional Funnelback search continues to work exactly as it always has, handling keyword searches and delivering search results. Conversational Search uses that same Funnelback foundation but adds the ability for users to ask questions in natural language and receive conversational responses. This means you get the best of both worlds: the reliable search functionality you're familiar with, plus the new AI-powered conversational experience.

How does this work on our actual site? Is it a separate page, or can it sit as a hovering chat on the bottom right? If we go with the hybrid system, can they still refine their search by asking further questions?

We offer several standard implementation options to get you started quickly. You can implement it as a separate, self-contained page (like the conversational search on the Squiz website), or as a hybrid version embedded within your existing search results page. However, customization possibilities are extensive.

Whether you choose a standard option or go fully custom, users can always ask follow-up questions to refine their search – including in the hybrid approach.

Do you have to have different searches as you start to do more than one slice? How do you tell users that you are searching only over certain content?

We recommend building your slices incrementally, and the hybrid model makes this particularly seamless. Similar to how Google sometimes shows AI summaries and sometimes doesn't, your users will only see conversational search responses when their question falls within your configured content slices – otherwise, they get the standard search experience they're already familiar with.

For example, if you start with just housing or accommodation content, students asking about those topics will get AI responses, while other queries won't trigger the conversational search at all. This creates a graceful fallback that feels natural to users, so they don't need to be explicitly told what content is or isn't included – the system handles it transparently.

Are all of the guardrails in the back-end easily set up by non-technical users? It looks like they are natural language prompts, so we’d be able to control that without having to have technical knowledge, is that right?

Yes. As you can see in the demo, the setup interface uses natural language prompts rather than complex code, so no technical expertise is required. We also provide pre-filled templates as a starting point, which you can customize and extend based on your specific needs.

Can we configure the conversational search to follow our tone of voice and guidelines?

Absolutely, and we strongly encourage you to do so. While we provide a solid foundation of default instructions, we recommend supplementing these with your own tone of voice guidelines. Your conversational search should reflect how you uniquely communicate with your audience – it should sound authentically like your brand and speak to users in the same way you do across other touchpoints.

Will there be a change in your analytics reporting given the likely increase in complexity in search term data?

Yes, we've built enhanced monitoring capabilities specifically for Conversational Search. Our analytics now capture the full conversational history, along with tagging to help you identify recurring themes and user satisfaction metrics.

We're also introducing new analytics around the number of conversations, questions per conversation, trending topics, and unanswered questions. This adds a new layer of insight into how users are engaging with your content through natural conversation, providing additional depth to your existing analytics capabilities.

Assuming you use Funnelback to power this, does it require you to have your Funnelback collections in great order or can it work with fairly vanilla setups? For example, we've not been great at adding in best bets, etc!

You don't need extensive configuration within Funnelback to power Conversational Search – all you need is a search package and some data sources. You might even  choose to create a specific search package for Conversational Search, if you want to limit the content for your Conversational Search to a subset of what you’re using in your full search.

Do you have any live use cases you can point us to - locally or internationally?

Yes! You can check out our own implementation of Conversational Search on squiz.net — you can ask it anything about the Conversational Search feature. Check it out here.

Is it 100% hallucination-proof?

We've built robust safeguards specifically to prevent hallucinations and ensure reliable responses. Our system includes multiple protection layers: we validate that questions are legitimate to prevent prompt injection, then generate answers using our RAG system that only draws from your specific content.

Most importantly, we use a guardian agent that validates every response for accuracy and relevance. If the system detects any issues with the generated answer, it defaults to a safe response rather than providing potentially incorrect information.

This multi-layered approach ensures users consistently get reliable information sourced directly from your content, with strong controls in place to catch and prevent any issues before they reach your users.

Does using conversational search train the AI model, and would our content then be accessible to other Squiz customers?

No, there's no training on customer data, and there's no cross-pollination between clients. Conversational Search uses a RAG (Retrieval Augmented Generation) system, which retrieves the most relevant content from your specific sources every time a question is asked – it doesn't remember or train on your content. Your content remains completely separate from other customers' content.

The AI only has access to the content you've specifically configured for it and doesn't look beyond your designated sources to the wider internet. Your content and conversations stay private to your organization.

Can your search also develop drafts and checklists apart from answers just like other AI platforms?

The Conversational Search may include checklists when they're part of the natural answer to a user's question, but it's specifically designed for answering questions based on your existing content rather than being a general-purpose AI tool.

Would you say content jobs are dead now? Is it just typing info into the LLM?

Not at all – in the age of AI search, content on your website is more important than ever. Content teams now play a crucial role in auditing and optimizing content for high quality, not just from an SEO perspective, but also for AEO (Answer Engine Optimization). This ensures your content can be discovered and effectively utilized wherever users are searching – whether through traditional search engines or AI-powered platforms. The role is evolving, but it's becoming more strategic and impactful, not obsolete.