Summarize with AI
To show an AI product on a website, put one real input next to the finished output it produced, instead of a picture of the empty interface. Pick a request a buyer would actually type, run it through the live product, and show the full answer with its sources on the first screen. A visitor should be able to judge the quality of the work without signing up.
This matters more for AI products than for any other software, because the interface is the cheapest part of what you built. A chat box looks the same whether eighteen months of evals and retrieval sit behind it or a single model call. You know which one yours is. A stranger on your homepage has nothing on screen that tells them.
The hero with a box that says Paste a URL
Picture an AI research tool for product marketers at B2B software companies. The headline reads Research at the speed of thought. Under it sits a browser frame holding the product's main screen: a wide text field with the placeholder Paste a competitor URL..., a small paper plane button on the right, and three grey chips underneath reading Track pricing, Compare plans and Draft a brief. Nothing has been typed. Nothing has been answered.
Behind that frame, the product does something genuinely hard. It pulls competitor pricing pages, archived copies of them, changelogs and public filings, works out what changed and when, and writes an answer with every claim cited. The team spent most of a year teaching it to say it could not confirm a date instead of making one up. None of that reaches the hero. The hero shows a text field, and every AI product in the category has a text field.
So the visitor guesses. Another chat tool? A wrapper with a nicer font? They have about four seconds of attention to spend on deciding what kind of company this is, and the image gave them nothing to spend it on.
Why the empty chat box is the least useful image you own
Most advice on AI landing pages says to show the real product. That is half right. The empty box is the real product, and the prompt box deserves real design work inside the app. On the landing page it is still the wrong picture, because it shows where the work starts and hides where it ends.
Here is the claim plainly: on an AI product's landing page, the interface is almost never the thing to show. The output is. An interface proves the product exists. An output proves it is good, and good is the only question the buyer has. A bare chat box also tends to read as a product whose interface question is still open, and the sharper visitors read it exactly that way.
The founder already knows this, judging by what he does on calls. When a prospect asks what the tool does, he does not describe the box. He shares his screen, types the long, specific request he has typed a hundred times, the one with a role, a source rule and a format, and lets the answer land. Those calls go well. Meanwhile the homepage shows the box as it looked before he typed anything.
We met the website side of this on Dualite, an AI builder competing with tools that looked alike to non-technical first-time visitors. The site had to make the difference obvious on sight, before anyone opened the product. That is the one job an empty interface cannot do, because empty interfaces look the same from company to company.
Forty words in, one finished page out
Now redesign the research tool's hero, still as an example. Keep the headline. Replace the browser frame with two panels side by side.
The left panel holds the input, set in the product's own font, exactly as a user typed it: Three of our competitors changed pricing this summer. For each one, tell me what changed, when, and which customer segment it targets. Use only their public pricing pages and changelogs. Return a table I can paste straight into Monday's board deck. That is forty words, written the way a product marketer writes on a Thursday afternoon with a deadline.
The right panel holds the answer. A table with four columns: Competitor, What changed, Date, Segment. Three rows. The first competitor dropped its free tier and moved its starter plan to per-seat billing on 3 June, aimed at teams under ten people. The second added an AI add-on sold per workspace on 11 July, aimed at mid-market. The third raised its annual discount, with the date marked approximate, aimed at renewals. Under the table sits one line in smaller type: Date for the third row estimated from two archived snapshots, the changelog does not say. Below that, three numbered sources, each with a page title and the date it was captured.
A visitor does not need to type anything to get this. They read forty words and recognise their own week. They scan three rows and can judge the quality in the time a hero gets. And the hedged line about the third row does more for trust than any sentence about accuracy, because it shows a product that knows where its evidence stops.
Run it for real, then leave the flaw in
The pair only works if it is genuine. Run the input through the live product, capture what came back, and change nothing except length. Do not tidy the answer by hand. People who use AI tools every day can spot a staged output quickly: every cell filled, every claim confident, every source from last week. A real run has texture, such as a date marked approximate or a source with an awkward title. That texture is the evidence.
Choose the input with care, because it does half the selling. It should be a request your best customers actually send, in their words, with their kind of constraint in it. If you can read your own logs with permission, start there, then change the details so no customer can be identified. A short prompt like Summarize this report tells the visitor that any tool could do it. A long one with a role, a rule about sources and an output format shows that this product handles the hard version of the job.
Build three pairs, one for each kind of buyer you sell to, and let a small toggle above the panels switch between them. Three shows range. Ten turns the hero into a gallery that nobody reads to the end.
Static, animated, or the steps as they run
Founders often want the hero to type itself out: letters appearing in the box, a spinner, then the answer. It looks lively in the design file. On the live page it makes a stranger wait for the one thing they came to see, and anyone who scrolls in the first few seconds never sees the output at all. If you animate, open with the finished pair already on screen and save the motion for a replay button.
Showing the steps beats showing the answer in one case: when the work itself is the argument. An agent that reads a long document, checks two systems and then drafts a reply is easier to trust when a visitor can watch the checklist tick through. For a research tool, a writing tool, a code reviewer or a meeting summarizer, the finished answer is enough.
Show it still, at a size where every word can be read on a phone. On mobile, stack the input above the output and cut the table down to its two most telling rows. And if your product really is a chat interface at heart, the in-product half of this problem, what the first reply should contain, is covered in our conversational UI design guide.
What being read as a wrapper costs
The empty box does not only lose clicks. It changes what kind of company a visitor thinks you are. A product whose only visible surface is a text field gets filed as a thin layer over someone else's model, and that reading travels: to the investor who opened the site before the meeting, and to the engineer deciding whether your problem is hard enough to be worth joining.
The money side is concrete. Seed-stage AI startups raise at valuations roughly 42 percent above non-AI peers, so they have to grow into a price that was set optimistically. The Series A bar for AI startups is now around 3.5 million in ARR (Carta, Q1 2026), up from roughly one million three years earlier. A company read as a wrapper has to clear that bar while also arguing it will still matter when a model provider ships something better. The homepage is often the first place that argument gets made, and an empty box makes it badly.
Every visitor was also paid for, in ads, in content, or in the founder's own hours posting on LinkedIn. That cost lands whether the hero explains anything or not. Sending more traffic to an empty-box hero buys more of the same confusion at the same ratio.
Swap the hero this week
- Write down the request your best customers send most often. Keep their words, their constraint and the format they ask for. Aim for thirty to sixty words.
- Run it through the live product today. Save the full output and its sources exactly as they came back.
- Trim the output to what fits in one panel at a readable size. Cut rows, but never rewrite the wording of what is left.
- Keep one honest edge in view: a hedged date, a low-confidence flag, or a cited source with an odd title.
- Place the input on the left and the output on the right, or the input above the output on a phone. Take the empty chat box out of the hero. It can sit further down the page if you want it.
- Write one plain label above the pair in the product's own voice, such as Found 3 pricing changes, 3 sources.
- Show the page to someone outside the company for ten seconds. Ask what the product does and whether it is good at it. If they can answer the first and not the second, the output is too small or too generic.
The sections below the hero, from proof to pricing to the AI trust question, are covered in order in our guide to AI SaaS landing page design. If you would rather hand the page to someone, this ranking of landing page design agencies for AI startups compares studios on public evidence.
Studio Maydit is a web and product design studio for AI founders in the US, UK and Europe. Every landing page we design starts from the question in this post: what does a stranger need to see to believe the product is good, not just real. You already know the answer, because you have watched your product earn it on a hundred calls. The work is getting that moment onto the first screen when you are not in the room, usually as a fixed scope of three to four weeks. If your hero still shows an empty box, book a 30 minute call and we will pick the input together.
