The request to add AI to your website usually arrives without a problem attached. A competitor launched a chat assistant, a vendor gave an impressive demo, or someone asked why the site doesn’t “do AI” yet. Understandable, but the wrong place to start.
Start instead with the job a visitor comes to do. Without one, AI becomes a widget that costs money every month, slows the page and answers from content nobody has checked. With one, the right feature can do that job faster than a page can.
Here are the five AI website features businesses are most often offered, with when each is worth it, what it needs and what goes wrong. For the wider picture, see how AI is changing websites.
The short answer
An AI feature earns its place on a business website when it passes three tests:
- It does a specific visitor job better than a well-organised page could. If a clearer page or a shorter form would solve it, do that first.
- It answers from your own current content, and shows where each answer came from.
- It keeps a plain route to a person for anything it can’t or shouldn’t handle.
AI site search and smarter enquiry intake are usually the easiest to justify, because the job is obvious and the result measurable. Chat, personalisation and summaries earn their place only in specific situations. Skip any feature whose main purpose is to show that you have AI.
Start with the visitor’s job, not the technology
Before comparing tools, find out where visitors struggle. Three sources usually tell you enough:
- Site search. GA4’s enhanced measurement can record what people type into your search box (as a
view_search_resultsevent), provided the term appears in the results page’s address, as WordPress’s?s=does. Searches that find nothing, or keep recurring, show where content fails. - Your inbox. The questions that arrive again and again, especially ones your website already answers somewhere.
- Your sales or support team. The questions they answer most, and the ones that decide whether a lead becomes a customer.
Then check whether a non-AI fix would do the job, often for less:
| Visitor job | Try this first | Consider AI when |
|---|---|---|
| Find a specific answer | Clearer headings, an FAQ, better search settings | Hundreds of pages, searched in visitors’ own words |
| Compare options | A comparison table or filters | The right choice depends on many combined needs |
| Check they’re a fit | Plain prices, criteria and examples | Questions vary too much to publish every answer |
| Get in touch | A shorter form with one clear next step | Sorting varied enquiries delays replies |
Make the non-AI fix either way: every feature below depends on the same well-structured content.
Five AI website features, reviewed
1. AI site search and answers over your own content
Traditional site search matches words. AI site search matches meaning, so “can I pay in instalments” finds your page on payment plans. Many tools go further, retrieving the most relevant passages and writing a short answer from them with source links. This is retrieval-augmented generation (RAG).
Worth it when you have lots of content (product ranges, documentation, policies, course catalogues) and visitors describe their needs in their own words, not yours.
What it needs. Current, consistent content, because the tool can’t tell an old price list from this year’s. Decide what is in scope, retire outdated PDFs, keep the index in step with what you publish and budget for usage costs. Preparing your content for an AI knowledge base covers this step by step.
What can go wrong. Confident answers from stale or contradictory pages. Show the source link with every answer, let the tool say it doesn’t know, and read the query log monthly.
2. A chat assistant
A chat assistant answers questions in conversation, usually from the same content. Decide first how far it should go: answering questions is lower-risk than booking or cancelling things. Chatbot, workflow or AI agent explains that choice, and AI customer service agents covers what to automate and what to keep human.
Worth it when visitors ask the same pre-sales questions outside working hours, or your team already spends hours a week answering chat.
What it needs. Clean content, a clear handover to a person, a line saying it’s AI, and someone who reads the conversations.
What can go wrong. Dead ends where the assistant loops instead of passing the visitor on, and a launcher that covers content on phones or pops up uninvited.
3. AI personalisation
Personalisation changes what a visitor sees based on who they are likely to be: a trade buyer or a homeowner, a first visit or a return. Much of it isn’t AI at all; simple rules do most of the work. AI adds value mainly in recommending related content or products, when you have plenty of both and enough traffic to learn from.
Worth it when you serve clearly different audiences with different next steps, and have content written for each.
What it needs. Variants written by people, a way to measure whether they help, and a consent plan. In many places, including the EU and the UK, personalising content from a visitor’s behaviour with cookies or similar identifiers generally needs consent (remembering a setting the visitor chose, such as their language, is usually treated differently). The site must work well for visitors who decline.
What can go wrong. Personalisation that feels intrusive, and slower pages when personalised content bypasses caching. Keep the core content the same for everyone, search engines included, and personalise the blocks around it. For languages, suggest rather than redirect: Google advises against automatic redirects based on the language you think a visitor speaks (see the multilingual SEO guide).
4. AI summaries for website content
Worth it when you publish long reports, specifications, policies or handbooks that most visitors only need the gist of.
What it needs. Less than you might think. The strongest option is usually an AI-drafted summary that a person checks and publishes as ordinary text at the top of the page, where search engines can read it and it costs nothing per view. A live “summarise this page” button adds usage costs and duplicates what several browsers and AI assistants already offer.
What can go wrong. Summaries that drift from the page after updates, or drop the conditions that matter, such as who is eligible. Add “check the summary” to your editing routine.
5. Smarter enquiry intake
Behind the scenes, AI can read a free-text enquiry, tell a sales lead from a support request or supplier pitch, pull out details such as budget and timing, and route it to the right person with a suggested reply.
Worth it when enquiries are varied and sorting them by hand delays replies to the good ones.
What it needs. Clear categories, a connection to your inbox or CRM, a person who reviews anything uncertain, and a check of your AI provider’s data terms, because every enquiry contains personal data. AI lead qualification goes deeper.
What can go wrong. A good lead misrouted, an automatic reply that promises too much, or a form turned into a slow conversation. Keep the visible form short and let AI work after the visitor presses send.
Risks that apply to every AI feature
Wrong answers and leaked data
Language models can produce fluent, plausible answers that are wrong, so limit answers to your own content and test with real questions from your inbox. Know what each feature sends to which provider, how long it’s kept and whether it’s used to train their models, and update your privacy notice before launch. AI agent risks covers hallucinations, prompt injection and data leaks.
Inaccessible widgets
Chat launchers and AI search panels often fail accessibility basics. Test them against WCAG 2.2 with a keyboard and a screen reader: every control must work without a mouse, a sticky launcher must not completely hide the element that has keyboard focus (success criterion 2.4.11), and buttons need a target of at least 24 by 24 CSS pixels or enough space around them (2.5.8). Answers that stream in word by word can flood a screen reader, so announce the finished answer instead. The web accessibility guide covers the rest.
Heavy scripts
Many AI widgets load a large script on every page, whether anyone opens them or not. That can slow how quickly the page responds to taps and clicks, measured as Interaction to Next Paint, which Google rates as good at 200 milliseconds or less. Load the widget only when someone clicks the launcher, and compare Core Web Vitals before and after. Third-party scripts and site speed shows how.
Ongoing upkeep and cost
Most AI services bill by usage, so costs rise with traffic, and without rate limits bots can run up the bill. Providers update their models, and answers can change with them. Set a spending cap with alerts, and name one person to review logs, costs and answers.
Checklist: before you add AI to your website
Tick every line before you commit budget:
The firm rule: always keep a plain route to a human
Whatever you add, visitors must be able to reach a person without going through the AI: a phone number, email, form or messaging link on the page, not hidden behind the chat. Inside any assistant, “talk to a person” should work at every step and pass the conversation along. Designing AI experiences people trust covers disclosure and handover.
For complaints, cancellations, payments and anything sensitive, send people straight to your team. An AI standing between a frustrated customer and a person only adds to the frustration.
Frequently asked questions
How much does it cost to add AI to a website?
Budget for the one-off work of building the feature and connecting it to your content, then running costs billed by usage, which grow with traffic. Ask providers for an estimate at your expected monthly volume. The hidden third cost is keeping that content current.
Will AI features on my website improve my SEO?
Not directly. Answers generated inside a widget aren’t part of your pages, so search engines don’t rank you for them. What helps is the clear, current content every AI feature depends on, and summaries published as text on the page.
Do AI features work in more than one language?
Most current models handle many languages, but quality varies, and answers are only as good as your content in each. Where a page exists in only one language, an assistant may translate it on the fly with nobody checking, so test in every language your customers use.
Should visitors be told they’re talking to AI?
Yes. It sets expectations and builds trust, and some regulations, such as the EU’s AI Act, include transparency obligations for AI systems that interact with people. A short line in the launcher or first message is usually enough.
What to do next
AI features sit on top of a website; they don’t replace one. A good AI-powered customer experience rests on unglamorous foundations: content with an owner, a clear structure, fast pages and forms that work.
Our website design and development work starts there: a sitemap planned around your visitors’ questions, content structured so people and search tools can find answers, and a WordPress site you own, checked for speed and accessibility before launch. Before adding any AI feature, make sure the website underneath is ready: talk to us about what your website needs.