Summarize with AI
Yes, let users try before signing up whenever one run of your product shows what it is for. Put the real product on the homepage, let a stranger get a result with no account, and ask for the account at the moment they want to keep that result.
This matters more for an AI product than for most software. You know what a good output looks like, because you have seen hundreds. A stranger has seen none. Every screen you put before the first result asks them to trust a headline, and the result is the only proof that the headline is true.
Drop a photo of an empty room
Picture an AI virtual staging tool for real estate agents, used here as an example. An agent uploads a photo of an empty room. The product furnishes it so the listing looks lived in. The team spent a year on the hard parts: matching the window light, sizing a sofa to the real floor, keeping the walls straight.
Version one of the homepage says Stage any listing in one click, with a Start free trial button. The button opens Create your account: Full name, Email, Password, Brokerage, Phone number. Then Check your inbox. Then an empty upload box. Then Choose a style. The first furnished room appears on screen five.
Version two keeps the headline and puts the product under it. A drop zone reads Drop a photo of an empty room, with JPG or PNG, up to 20 MB in small grey type. Under it sits a link: Or try ours: 2-bed condo, living room. The agent drops a photo of a bare room with a bay window.
Three short lines tick past. Finding the walls and window light. Measuring the floor: about 14 by 18 feet. Placing furniture. About twenty-five seconds later, three rooms sit side by side, labelled Modern, Scandinavian and Coastal. Each has a small Preview mark in the corner. Under them are two buttons, Download full size and Stage another room.
Clicking Download opens a small panel on top of the rooms. Save these 3 rooms to your account. Continue with Google. Email me a link instead. The rooms stay visible behind the panel the whole time.
The agent decided before you asked
In version one, the agent decides on the headline. In version two, she decides on her own living room. By the time the sign-in panel appears, she is no longer asking whether the product works. She is asking where to keep the result. That is a much easier question to say yes to.
The founder of this example company already knows this, judging by his calls. He keeps a folder of before and after pairs on his laptop and opens it in the first minute of every demo. Those calls go well. His cofounder has suggested twice that the model run on the homepage. Both times the answer was the GPU bill.
Meanwhile a competitor with flatter lighting and smaller sofas has a drop zone on its homepage. Agents use it, because they saw it work before anyone asked for a phone number. They did not pick the better tool. They picked the one they understood first.
What an anonymous room actually costs
The GPU bill is a fair worry, so price it. OpenAI's image generation guide lists GPT Image 2 at about 4 cents for a medium quality landscape image and about 17 cents at high quality. Three medium previews cost roughly 12 cents of image output per visitor. A thousand anonymous visitors who each stage one room cost about 120 dollars.
Now look at the other side. You already paid to bring those thousand people to the homepage, in ads, content or your own hours on LinkedIn. The ones who leave at a signup form cost the same as the ones who stay. Product-led benchmarks, measured on SaaS companies, put activation at 20 to 40 percent for most products. On the same benchmarks, a ten point gain in activation typically drives a 15 to 25 percent increase in free-to-paid conversion. A visitor who never sees a result cannot activate at all.
Here is the claim, plainly. For most AI products, the cheapest growth available is spending a few cents of inference on people who have not signed up. Most advice treats inference cost as the reason to put up a wall. It is really the reason to cap anonymous runs. AI product builders do run at around 52 percent gross margin against 70 to 80 percent for traditional software, so the caps matter. A wall in front of the first result is the more expensive choice.
There is a common middle ground: take the photo first, ask for sign-in, then show the rooms. It makes sense when one run costs dollars, such as a long agent run or a minute of video. For a 12 cent run it is the worst place for the wall. The visitor has done the work and still has no proof.
Put the wall at keeping, not at seeing
The rule is simple. Let people see a result for free. Ask for an account when they want to keep it, share it, or come back for more. For the staging tool, that moment is Download full size. For a writing tool, it might be Copy or Export. For a research tool, it is Save this report.
Duolingo is the best known version. An onboarding teardown on Appcues describes it as onboarding that begins with the product and ends with optional account creation. Signup becomes more tempting as people want to save their progress. The account is the answer to a need the product has already created.
We saw the same shape on the website we designed for Chariot, a speech AI lab. The page has a Try Chariot TTS button and shows typed text next to the speech the model makes from it. After our work, Chariot-1 TTS moved from private enterprise access to a public research preview, with the demo on the site as the way in.
Some products should keep the wall early. If the first result needs the customer's own data, like a connected CRM, a sample run on fake data proves little. If the buyer is a general counsel who never browses, a sandbox will not move her, which is when an interactive demo or a pilot does more.
Caps that stop abuse without stopping agents
Anonymous runs need limits. The trick is to limit what a single person can take, not whether anyone can try. These are the caps we would start with on the staging tool.
- Three rooms per browser per day. Store a count in a cookie and back it with a per-network limit.
- Anonymous runs at medium quality, signed-in runs at high. The preview is honest and the full size is the reason to sign in.
- Cache the sample room. Its three results cost nothing after the first run, and they cover everyone without a photo at hand.
- Add a bot check only on the second run, never the first.
- Set a daily budget. When it runs out, the drop zone quietly shows the sample results instead of an error.
- When someone hits the cap, say what an account gives them: Sign in to stage 10 more rooms today. Never just say You have used your free tries.
Carry the three rooms into the account
The claim step is where most try-first flows break. The agent clicks Continue with Google and lands on a dashboard that says No listings yet. Her three rooms are gone. She has just learned that work done in this product can vanish.
Instead, tie every anonymous result to a session ID on the server. When she signs in, attach those results to her new account and land her on them. The screen should read Untitled listing, 3 rooms, with the download she asked for already starting. Her first screen inside the product is her own work, not an empty state.
Then count it. Of everyone who got an anonymous result, how many claimed it? That share tells you whether the wall sits in the right place. If almost nobody claims, the result is not good enough yet, or the keep moment is wrong. For the questions you ask after sign-in, our guide to signup flow best practices for AI products covers which fields to drop.
A try-first homepage in five working days
You can test this without a rebuild. Here is the order we would do it in.
- Day one: open your homepage in a private window and count every screen between it and the first output. Write the number down.
- Day one: price one run in tokens or images. Multiply by the visitors you expect in a month.
- Day two: pick your keep moment, the one action people take when they like the result.
- Day three: put one input and one output on the homepage, plus a sample input for people with nothing to paste.
- Day four: add the per-browser cap and the daily budget.
- Day five: make sign-in carry the result into the account, then ship it.
- For two weeks, track two numbers: the share of homepage visitors who see a result, and the share of those who claim it.
This turns the homepage into the product's first screen, which changes what the page has to do. If that page needs a designer, compare the studios in our ranking of landing page design agencies for AI proptech startups, or see how we approach website design for AI products.
Studio Maydit designs the first run of AI products, the screens where a stranger decides whether to stay, and that now often starts on the homepage. A fixed-scope project takes three to four weeks and ends with a diagnosis of what is leaking in the product, so you know which wall to move first. If your best demos happen on calls and your homepage still asks for a password first, book a 30-minute call with Studio Maydit.
