Adding an AI Chatbot to Your Website: Every Question Business Owners Ask Us, Answered

By Lyle Heartman, ELRQ Agency

Every time we pitch a chatbot build, the same questions come back. Not the ones in the marketing brochures. The real ones.

Will this break my site? Can someone hack it? Will Google punish me for it? Where does my customer data actually go?

They're fair questions, and most vendors dodge them. So here are straight answers, including the parts that don't flatter the technology.

Part 1: Installation and Risk

1. Can I add a chatbot without disturbing my existing website?

Yes. In almost every case we don't touch your existing site code at all.

A modern chatbot is a single script tag that loads in an isolated container. Your pages, your theme, your plugins, your checkout flow, all of it stays exactly as it is. If we pulled the script out tomorrow, your site would be byte for byte what it was before.

There are three ways we deploy it:

Embedded widget. One line of JavaScript in your footer. Works on WordPress, Shopify, Webflow, Squarespace, custom builds, anything. Ten minutes of work.

Isolated iframe or shadow DOM. The chat UI lives in its own sandbox, so its CSS can never leak into your styles and your styles can never break its layout. This is our default, because it kills off the entire "the chat box broke my navbar" category of problem.

Native integration. For clients on SvelteKit, Next.js, or Astro, we build the chat as a real component inside your app. More work, but it's faster, it matches your design system properly, and it can read your app state.

We also stage it first. The bot goes live on a staging copy of your site, you click around, you try to break it on purpose, and only then do we ship to production.

2. Can a chatbot be used to attack my website? How do we prevent that?

Almost nobody asks this one, and it's the most important question on the list. The honest answer is yes. A badly built chatbot is a genuine attack surface.

Research accepted to the IEEE Symposium on Security and Privacy in 2026 looked at 17 third-party chatbot plugins running on more than 10,000 public websites. The findings were ugly. Eight of those plugins, in use on roughly 8,000 sites, didn't verify the integrity of the conversation history moving between the visitor's browser and the chatbot backend. In practice that means an attacker can forge the conversation, including inserting fake system messages, and push the bot into behaviour it was never meant to have. The researchers measured this as making attacks three to eight times more effective.

The second finding was worse. Fifteen of the plugins offered to scrape your website content to give the bot context, but none of them separated your trusted content, like product descriptions, from untrusted content, like customer reviews or user comments. So an attacker leaves a review containing hidden instructions. Your bot reads that review as though it were your own copy. Now it's following someone else's script. About 13% of the e-commerce sites in the study had already exposed their bots this way without knowing.

There's a third risk, and it's more traditional. If the chat window renders the bot's replies as raw HTML or unsanitised markdown, an attacker can get JavaScript onto the page through the chat. That's textbook cross-site scripting, and it isn't hypothetical: a self-hosted chatbot project had exactly this patched in early 2026, where an iframe payload in a chat message could reach cookies and localStorage tokens and lead to account takeover through the admin inbox.

Here's how we prevent all of it:

Server-side prompt only. Your system prompt, your instructions, and your business rules never touch the browser. The browser sends a message. The server decides everything else.
Conversation state on the server. The client can't forge history because the client doesn't hold history. It holds a session ID.
Trusted and untrusted content kept apart. Your product data goes in as trusted context. Reviews, comments, scraped third-party pages, and user uploads go in fenced and clearly labelled as untrusted, with explicit instructions that anything inside those fences is data and never an instruction.
Output sanitisation. Every bot response is sanitised before it renders. No raw HTML, no javascript URIs, no inline event handlers, no iframes. Links are validated and given rel attributes.
Strict Content Security Policy. The widget can only talk to the endpoints we allow. Nothing else.
Rate limiting and abuse detection. Per IP and per session. This stops attack automation, and it stops the thing people forget about, which is someone burning through your API budget for entertainment.
Least privilege on tools. If the bot can look up an order, it can only look up an order for a session that's already authenticated. It can't read arbitrary customer records, and it can't issue refunds, change addresses, or delete anything.
Full logging. Every conversation, every tool call, every failure. When something goes wrong you can see exactly what happened.

If you already have a chatbot on your site from another vendor, we'll audit it against this list for free. It takes an afternoon, and it's usually revealing.

3. Will the chatbot make my site slower?

It can. Most of them do. Ours doesn't, and the whole difference is in how the script gets loaded.

The measured reality on common widgets: the heaviest ones, Zendesk and Tawk.to among them, ship between 500 KB and 750 KB of JavaScript just to draw a chat button. Lighter options like Crisp and Zoho stay under about 155 KB. On average a chat widget adds 300 to 600 milliseconds of main thread blocking time, which is enough on its own to push an otherwise healthy Interaction to Next Paint score into the failing range.

That matters because Core Web Vitals feed into rankings. A widget competing with your hero image for bandwidth hurts Largest Contentful Paint. A widget that pops into the corner and shoves content sideways hurts Cumulative Layout Shift. Google doesn't penalise the widget. It penalises the page the widget slowed down.

What we do about it:

The chat script doesn't load during page load at all. We wait for the browser load event, then wait again for requestIdleCallback, so it only downloads once the browser has genuinely finished the work that matters.
We inject the script through JavaScript rather than a static script tag, which stops the browser's preload scanner grabbing it early and racing your real content.
We use a facade. What loads first is a tiny button, a few kilobytes. The actual chat application only downloads when someone clicks. Most visitors never click, so most visitors never pay for it.
The launcher gets a reserved position and fixed dimensions, so it can't cause layout shift.

The standard we hold ourselves to is simple. Run PageSpeed Insights before install and after install. The scores should be identical. If they aren't, it isn't finished.

Part 2: Does It Actually Work

4. What do the numbers say about chatbots and user experience?

The data has moved a long way since the frustrating decision-tree bots of a few years ago.

On waiting and speed:

82% of people said they'd talk to a chatbot rather than wait for a human, and only 18% are willing to wait 15 minutes for a live agent (Tidio, 2026).
62% of consumers now prefer a digital assistant over sitting in a queue (Master of Code, 2026).
72% expect a website to offer instant messaging support at all.

On satisfaction:

87.2% of consumers rate their chatbot interactions as neutral or positive, with only 13% calling them ineffective (Master of Code, 2026).
Roughly 80% of customers who've used AI-powered support reported a positive experience (Zendesk, 2026).
About 90% of queries get resolved in fewer than 11 messages (Tidio, 2026).

On revenue:

AI chatbots deliver conversion improvements of 20% or more, and proactive chat that opens at the right moment has been measured driving lifts up to 40% (Which-50).
Salesforce reports 30% of service cases are now resolved by AI, and projects 50% by 2027.
43% of consumers say a bad customer service experience stops them buying from that business again. Support quality is a revenue line, not a cost line.

On adoption:

91% of businesses with more than 50 employees already use chatbots somewhere in the customer journey.
Gartner projects that by 2027, 25% of organisations will use chatbots as their primary customer service channel.
64% of small businesses planned to adopt one during 2026, which means your competitors probably already have.

Two numbers matter more than everything above, and they shape how we build.

86% of customers want a clear path to reach a human. A bot that traps people in a loop with no exit generates more anger than having no bot at all. Every bot we ship has a visible, one-click handoff to a real person.

69% want more human-like conversation, and 41% say tone and personality significantly affect their experience. That's why we don't ship template bots. Voice and tone are part of the build, not a finishing touch.

5. Will it actually increase sales, or just deflect support tickets?

Both. The sales side is usually the bigger number, and it's the one people ask about last.

Support deflection is easy to measure and easy to sell. Where's my order, what are your hours, do you deliver to my area, what's your return policy. That traffic leaves your inbox on day one and you notice immediately.

The sales side is quieter. A visitor at 11pm who can't work out whether you carry a size 42 doesn't email you. They just leave, and you never find out. The bot answers, and you keep a sale you'd never have known you lost. That's where most of the conversion lift comes from, which is exactly why we instrument the bot to report on it specifically.

Part 3: What Can It Actually Do

6. Can it add items to the cart on my customer's behalf?

Yes, and this is the point where a chatbot stops being an FAQ machine and starts being useful.

We connect the bot to your store's API so it can search your live catalogue, check real stock levels, add items to the cart, apply a valid discount code, and hand the customer a ready checkout link. Someone can say "I need the blue one in medium, two of them" and land on a filled cart.

What we deliberately don't let it do:

Complete a payment. The bot builds the cart and passes the customer to your normal checkout. Card details never go near the conversation.
Act without confirmation. Anything that changes state gets shown first. "I'm adding 2 x Blue Shirt, size M, for 4,500. Confirm?" No silent actions, ever.
Touch money after the fact. No refunds, no cancellations, no price overrides, no address changes on paid orders. Those go to a human every single time.

Beyond the cart, the same mechanism lets it book appointments into your real calendar, check order status against your fulfilment system, capture and qualify leads into your CRM, and open a support ticket with the full conversation attached.

7. What happens when the bot doesn't know the answer?

It says it doesn't know, and it hands over. That's a designed behaviour, not a failure.

We ground the bot in your actual content, so it answers from your documents, your product data, and your policies rather than from general knowledge. When a question falls outside that, it does three things: says plainly that it doesn't have the answer, offers the handoff, and logs the question.

That log is one of the most valuable things you'll get out of this. After a month you have a ranked list of exactly what your customers want to know and can't find on your site. Most clients end up rewriting a chunk of their website because of it.

8. Will it make things up about my products or prices?

This is the risk that keeps serious people away from AI support, and it's a real one. Our answer is architectural rather than hopeful.

The bot doesn't answer from memory. Every factual response is retrieved from your live data at the moment of asking. Prices come from your database, not from something the model absorbed during setup. Stock levels come from your inventory system. Policies come from your policy pages.

On top of that, we hard-fence the categories where a wrong answer costs money or gets you into legal trouble. Prices, stock, delivery dates, refund eligibility, and anything regulated are answered from source or not at all. If the data isn't available, the bot says so instead of guessing.

Before launch we run an adversarial pass where we actively try to get the bot to say something false, offensive, or off-brand, and we tighten it until we can't.

9. Does it work in more than one language?

Yes, and it's essentially free. The same bot handles English, Sinhala, Tamil, Arabic, Mandarin, or whatever your customers bring, and it replies in the language it was addressed in.

For businesses in multilingual markets this is often the single highest-value feature, because it removes the need for separate language support staff.

10. Can it work on WhatsApp and Instagram too?

Yes. The knowledge base and the logic live in one place, and we connect channels to it. Website widget, WhatsApp Business, Instagram DMs, Facebook Messenger, all answering with the same brain and the same tone.

Website chat is still the largest channel, but messaging apps are growing roughly three times faster. If most of your customers already message you on WhatsApp, that's where the bot should go first.

Part 4: Data and Privacy

11. Is my customers' personal information saved inside the chatbot?

Partly, and you should know exactly what that means. "The vendor handles it" is not an answer that will protect you if something goes wrong.

What normally gets stored: the conversation text, a session identifier, timestamps, the page the visitor was on, and anything the visitor voluntarily typed, which can include a name, an email, a phone number, or an order number.

What we make sure doesn't get stored: card numbers, passwords, and government ID numbers. A redaction filter strips these patterns out before anything is written to disk, so even if a customer pastes their card number into the chat, it doesn't survive to the database.

Where it lives: on infrastructure we control on your behalf, in a region you choose, rather than scattered across a vendor platform you can't inspect.

The model provider question: when the bot generates a reply, the conversation goes to a model API. The major providers offer zero data retention and no-training terms for API traffic, and we configure those settings explicitly. Your customer conversations don't become training data. We'll show you the exact configuration.

Retention: you set it. Thirty days, ninety days, a year. It deletes automatically after that. If a customer invokes a deletion right under GDPR or a similar regime, we can find and purge their conversations by identifier.

There's a business argument here as well as a compliance one. Companies with clearly stated AI privacy policies score about 23% higher on customer trust than those without. Worth being explicit about publicly, not just internally.

12. Who owns the conversation data?

You do. Your data, your export, any time, in a standard format. If you stop working with us, you leave with everything.

We don't hold customer data hostage as a retention strategy, and you should be suspicious of any vendor whose contract lets them.

13. Do I need to update my privacy policy?

Yes, and it's a short update. It needs to say that you use an AI assistant, what it collects, how long you keep it, who processes it, and how a customer requests deletion. We draft the section for you. Have your lawyer check it if you're in a regulated sector or handling EU or UK customers.

We also add a short disclosure in the chat window itself. Customers should know they're talking to an AI. Hiding it is bad practice, and in a growing number of jurisdictions it's illegal.

Part 5: Running It

14. How long does it take to set up?

For a standard business site with an FAQ, a product catalogue, and a contact flow, about one to two weeks from kickoff to live.

Most of that isn't engineering. It's gathering your content, agreeing the tone, and defining what the bot is and isn't allowed to say. Add another one to two weeks if you want live cart actions, calendar booking, or a CRM connection, because that means working against your existing systems.

15. What does it cost to run each month?

Two components.

Hosting. Small. A widget and a lightweight backend, usually a few dollars a month unless your traffic is very high.

Model usage. Priced per conversation, depending on volume and how long conversations run. A small business site handling a few hundred conversations a month typically lands in single or low double-digit dollars. We cache aggressively, keep prompts tight, and route simple questions to cheaper models, which cuts this a lot.

We also set a hard monthly spend cap. If something goes wrong, a traffic spike or someone abusing the endpoint, the bill stops at a number you chose. Nobody gets a surprise invoice.

16. What if my products or prices change?

Nothing breaks. Live data, meaning prices, stock, and order status, is pulled at the moment of the question, so it's always current.

Static content like policies and FAQs is re-indexed on a schedule, or immediately when you edit it if we hook into your CMS. There's an admin panel where you can update the bot's knowledge yourself without calling us.

17. What happens if the AI provider has an outage?

The bot degrades, it doesn't disappear. On a model API failure it falls back to a secondary provider. If that fails too, it switches to a simple contact capture form so the visitor can still leave a message. It never shows a broken widget or a raw error.

Your site is completely unaffected either way, because the chat is isolated from it.

18. Does this replace my support staff?

No, and any agency telling you otherwise is selling something they can't deliver.

What it does is absorb the repetitive volume, which is usually the majority of incoming messages, and hand your team the ones that need judgement. Around 74% of businesses using chatbots say the benefit is scaling operations without adding headcount, which is a different claim from replacing the headcount they already have.

The failure mode to avoid is using it as a wall. Customers spot that immediately, and it costs more goodwill than the bot saves in labour.

19. Does a chatbot help or hurt my SEO?

Neutral to positive, if it's built properly.

The negative risk is entirely about performance. A heavy widget that damages your Core Web Vitals damages your rankings. That's solved by the loading approach described earlier.

The positive side is behavioural. Visitors who get their question answered stay longer and bounce less. And the log of unanswered questions is free keyword research, straight from people who are already on your site trying to buy something.

One thing to avoid: don't stuff your chatbot content into the page HTML hoping to rank for it. Google doesn't reward that, and it slows the page down.

20. How do I know it's working?

You get a dashboard, and we agree on the numbers before we build, so there's no arguing about it afterwards.

Here's what actually goes on it:

Conversations started, and how many ended with the customer's question resolved.
Deflection rate, meaning how many conversations never needed a human.
Handoff rate, and the reason behind each one. A 30% handoff rate isn't a failure. Thirty percent of handoffs because the bot couldn't find your opening hours is.
Unanswered questions, ranked by frequency. This list is usually the most useful thing we hand you. It tells you what your website should have said in the first place.
Assisted conversions, meaning sessions where someone chatted and then bought or booked.
Average messages to resolution. If that number is climbing, something is broken.
Customer rating, a thumbs up or down at the end. Blunt, but people are honest with a thumb.

Give it 60 days. If the numbers aren't moving, we fix it, or we tell you it's the wrong tool for your business. We'd rather have that conversation than keep invoicing for something that isn't earning its keep.

Where to Start

You don't need to commit to a full build to find out whether this is worth it.

Write down the 20 questions you get asked most. The real ones, including the boring ones. If most can be answered from information that already exists somewhere on your site or in your head, a chatbot will pay for itself quickly. If most need genuine judgement, negotiation, or someone the customer actually trusts, it won't, and we'll tell you that.

We'll run that assessment on a call at no charge, and give you a straight answer either way.

About the author

Lyle Heartman is the founder of ELRQ Agency, which builds web platforms, AI automation, and security-hardened integrations for businesses in Sri Lanka, the Gulf, and North America. Get in touch for a free chatbot readiness assessment.