Summarize with AI
Add an AI chatbot to your website only if it answers a stranger's real question better than the page does, and only if it never covers the button you want them to press. If it pops up to greet everyone and then repeats your tagline, remove it. For an AI startup the widget is the first piece of your AI a visitor ever talks to, so a weak one argues against your product before anyone signs up.
This matters more for you than for a shoe shop. A visitor to an AI company's site is already asking whether your model is any good. A chat widget hands them a free test of a model, on your domain, under your logo. Nobody on your team designed that test. Usually it went in one afternoon with the vendor's default settings, and it has run unwatched since.
Hi! How can I help you today? on top of Start free
Picture an AI tool that checks an architecture firm's drawings against the local building code before the permit goes in, used here as an example. The hero reads Catch code issues before the plan checker does. Under it sit a green Start free button and a grey link, Watch the 2 minute demo. On a laptop the page is clear.
Now open it on a phone. Three seconds after the page loads, a round purple bubble in the bottom right corner bounces once and slides up a card. The card says Hi! How can I help you today? Under the greeting are three chips: Pricing, Book a demo and Talk to a human. On a screen 390 pixels wide, the card is about 340 wide and 420 tall. It covers Start free completely. The X that closes it is a grey mark 16 pixels across in the top corner.
Say the visitor plays along and types what does this actually do. A typing animation runs, then the reply: Transform your design workflow with intelligent, AI-powered compliance. That is last year's tagline, lifted from the meta description. Next question: does it read Revit files? The bot says it is not sure and asks whether the visitor would like to book a demo. The visitor wanted a yes or a no. The product does read Revit files. That fact lives on the file formats page, which the bot was never given.
Visitors read the widget as your model
Most advice about website chatbots comes from support and lead capture. It counts conversations started and emails collected, and on those numbers almost any widget looks like it works. That frame is wrong for an AI startup. Your visitor is not checking whether you offer chat. They are checking whether your AI understands things, and the widget is the one place on the site where they can find out without signing up.
So the bot in the example does three kinds of damage at once. It hides the main button. It shows that the company's own AI does not know the company. And it meets a plain question with a sales move, which tells the visitor the bot exists to catch leads, not to help. Someone who just watched your AI fumble a question about itself will not trust it with a permit set.
The starting bar is low, too. In moderated usability tests with nine participants, Nielsen Norman Group found that people will not experiment to work out what a site chatbot is for. They reach for search, filters and navigation first, because those are faster. Your widget rarely gets a second try.
What the founder sees from the desk
The founder almost never sees the phone version. He opens the site on a large monitor, where the card floats in an empty corner far from the hero. He has closed the greeting so often that the vendor's cookie keeps it shut for him. When he tests the bot, he asks questions he already knows the answers to, phrased the way his team talks, so the bot finds the right page every time.
The widget went in after a prospect asked on a call whether the site had AI chat, since the company sells AI. It took a script tag and an afternoon. Since then the vendor dashboard has reported conversations started, a number that climbs every month and goes into the investor update. Nobody has read the transcripts. Most are one line long, because opening the card and closing it again counts as a conversation.
What a covered button costs before the raise
Every visitor on that page was paid for, in ads, in content or in the founder's own hours on LinkedIn. A card over the main button on mobile taxes all of that traffic at the same rate. Sending more traffic makes the loss bigger, not smaller.
Google states the general problem in its guidance on intrusive interstitials: page elements that block the content can frustrate visitors, erode their trust in the site and make the page harder for search engines to understand. A chat card that opens itself over the hero on a phone is that element, whatever the vendor calls it.
The loss also carries downstream. Measured on SaaS and product-led companies, activation runs 20 to 40 percent for most products, and a ten point gain typically drives a 15 to 25 percent increase in free-to-paid conversion. Those rates begin at the signup. A visitor who never found Start free never enters the funnel, so your onboarding never gets its chance. With the Series A bar for AI startups at around 3.5 million in ARR (Carta, Q1 2026), growth that leaks on the first screen has to be bought twice.
Three questions that decide keep or remove
Before you tune the widget, decide whether it should exist at all. Answer these from your real transcripts. If there are none worth reading, use the last twenty questions buyers sent by email.
- Do strangers ask things the page cannot answer in one scroll? Which file types it reads, which cities it covers, where uploaded drawings are stored. If every question is about price or demos, the page should answer them, not a bot.
- Can the bot answer from your own material? Docs, the supported formats list, the security page, pricing. A bot that can only read the homepage will repeat the homepage.
- Will someone on the team read fifty transcripts a month? A bot nobody reviews falls behind every launch, and the visitor notices first.
Three yes answers: keep it and rebuild it as a demo. Any no: remove it this week and move the missing answers onto the page. The NN/g study points the same way. It advises giving a site chatbot one area where search and navigation fall short, and not making it copy them.
Removal is a good outcome, not a failure. A clear hero with one main path does more for a stranger than any greeting, which is the argument of the hero section guide. On the SATHI launch site, the hero names the job in one line, See the True Impact of your Affiliate Program, and puts Start a free Trial and Book a demo right under it. That is the work a chat widget is often installed to do, done by the page itself.
If it stays, make it show what the product can do
A site bot for an AI company should behave like a short, honest demo, which means somebody designs it like any other screen. The answer rules in this guide to in-product chat still apply. The marketing widget needs a few rules of its own:
- No auto-open. The bubble stays shut until someone taps it. On phones, keep it off the first screen, or shrink it to a small labelled pill that sits below the fold.
- Name what it knows. Swap How can I help you today? for a line with a range, such as Ask about file formats, the cities we cover, or how a drawing gets checked. That label is your capability statement.
- Seed three real questions. Chips like Does it read Revit files? and Is the Austin code covered? show depth. Pricing and Book a demo are not questions, so they do not belong in the chips.
- Answer from the docs, with the link. Every answer ends with the page it came from, so the visitor can check it. That one habit makes the bot read as grounded instead of merely plausible.
- Admit the gap, then route. When there is no answer, the bot says so and offers docs search or a person, with the question already filled in.
- Hand off with context. When a visitor books a demo after three questions, those questions go to whoever takes the call, so the call starts at question four.
A one-afternoon test to run this week
Open your homepage on your own phone, on cellular data, in a private tab. Note whether anything slides over the main button, and when. Then ask the bot the three questions buyers raised on your last three sales calls, word for word. Write each answer next to what your docs actually say.
If two of the three come back wrong, vague or as a demo pitch, switch the widget off today. Turning it off costs nothing, and it can come back once it passes. If what you want is a way for strangers to try the product before signing up, an interactive demo usually does that job better than a chat box.
The widget is one small part of a larger question: can a stranger on a phone tell what your AI does from the first screen? That question sits at the centre of website design for an AI startup. Studio Maydit designs and builds those sites for AI founders in the US, UK and Europe, in Framer, Webflow or custom code, as a fixed-scope project of three to four weeks or on a monthly retainer. If your widget failed the phone test, book a call and bring three transcripts.
