Summarize with AI
Yes, you can use ChatGPT to review a landing page, but only if you stop asking it for feedback. Ask for feedback and it returns the average of all landing page advice: stronger proof, a clearer call to action, more urgency. Give it a stranger with a task instead, make it answer from your page alone, and it will tell you the one thing you cannot see yourself, which is what the page never says.
That matters because you are the worst reader your homepage has. You know what the product does, so every vague line gets finished in your head before you notice it is vague. A model can stand in for the stranger you cannot be, for free, in ten minutes. It just will not do it by default.
A property tool that passed two AI reviews
Picture an AI tool for small property management firms, the kind that run 200 to 2,000 rental units. Tenants text the office line about a leak or a broken heater. The tool reads the text, asks the tenant for a photo, sorts the job by how urgent it is, and books the right plumber or electrician from the firm's own vendor list. It sends the property manager one line when the job is done.
Now the homepage, as an example. The headline reads Run your portfolio on autopilot. The subhead reads AI-powered operations for modern property teams. Below it sit three cards: Faster, Smarter, Always on. Two buttons of the same size, Book a Demo and See How It Works. The words maintenance request first appear in the fourth section. The word tenant appears once, in the footer of a testimonial.
The founder copies the page into ChatGPT and types: review my landing page and tell me how to improve it. The answer opens with Your value proposition is strong and clear. Then a list. Add social proof such as logos and testimonials. Make the primary CTA more prominent. Add a sense of urgency. Consider an FAQ section. He pastes the same page into a second assistant and gets the same list in a different order.
So he does what the list says. He adds four logos, a testimonial slider and an orange button. He rewrites the headline for the third time this quarter, to Save 10 hours a week on property operations. Demo bookings stay flat, and two of the calls that do come in are from people who thought it collected rent. Nothing on the page ever said it handles the leak at 11 at night.
Why every review comes back as the same list
A language model answers with the most likely response to the question it was asked. Review my landing page is one of the most common requests on the internet, and the most likely answer to it is the checklist that thousands of blog posts already repeat. That list is not wrong. It is just true of every page, so it tells you nothing about yours.
The deeper problem is that a model judges the words that are there. It praised the value proposition because Run your portfolio on autopilot is a grammatically clean benefit line. It has no way to notice that the noun is missing unless you give it a reason to look for one. Social proof, a CTA and urgency are things it can point to on any page. The absence of the word maintenance is not.
This is the part most advice about AI reviews gets wrong. It tells you to write a better prompt with more context about your business. That makes the review worse.
Do not tell it what your product does
The instinct is to brief the model the way you would brief a designer. Here is our product, here is our ICP, here is what we do, now review the page. The moment you do that, the model knows what you know. It reads Run your portfolio on autopilot and fills in maintenance requests and vendor booking, exactly as you do. You have turned the one reader who could be a stranger into a second copy of yourself.
So give it less. Paste the page text and nothing else. No about section, no pitch, no notes on who the buyer is beyond the role and the task you write into the prompt. If the model cannot work out what the product does from the page, neither can the person who clicked through from a search result, and that is the finding you came for.
You can disagree with this. Plenty of people will tell you context always helps an AI. For writing, it does. For reviewing a page whose whole job is to explain itself to someone with no context, every line of background you add is a line the page no longer has to earn.
The prompt: one buyer, one task, and the words Not on the page
Here is the version that works, as a set of steps you can run this week without hiring anyone.
- Copy the page text from top to bottom, in order, including button labels, nav items and image alt text. Mark where the first screen ends with a line that says FOLD.
- Write one buyer and one task from a real sales call or a real signup reason. For the property tool: You run a firm with 400 rental units. Tenants text your office about repairs after hours and you miss some. Your partner asked you to find something that fixes this. You have 20 seconds on this page.
- Ask four questions. What does this product do, in one sentence? What would it replace in your week? Which button would you press first, and why? What would you need to know before you press it?
- Add the rule that does the work: for every answer, quote the exact line on the page that told you. If no line told you, write NOT ON THE PAGE. Do not guess and do not suggest improvements.
- Run it three times with three different buyers, in a fresh chat each time so the earlier answers do not leak into the next.
- Fix only the NOT ON THE PAGE answers that sit above FOLD. Leave the rest of the list alone for now.
On the property page, the first answer comes back as: it seems to be operations software for property teams, quoting the subhead. The second comes back NOT ON THE PAGE. The fourth comes back NOT ON THE PAGE too: does it work with the property software I already use, and what happens if the AI books the wrong plumber? That one run says more than both feedback lists combined. The headline needs the job in it, something close to Every tenant repair request, sorted and booked, even at night.
A simple decision rule follows. If the answer to what would it replace comes back NOT ON THE PAGE, no button colour, logo strip or slider will help, because the visitor still does not know what they are buying.
What a model catches well
Used this way, an assistant is good at a narrow set of things, and they happen to be the things founders miss most. It catches the missing noun: the category, the job, the thing the product replaces. It catches jargon a buyer would not use, like operations layer or workflow orchestration on a page for property managers. It notices two buttons with equal weight and asks which one is the path. It notices when the hero promises one thing and the pricing table sells another.
It is also fast and it does not get tired of your page. You can run it after every headline change in a few minutes, which is more than any friend will put up with.
What it cannot see: the image, the stature, the first four seconds
A pasted page is words. The model never sees the abstract gradient that sits where a screenshot should be, or that the only picture above the fold is a stock photo of a city skyline. Even when you upload a screenshot, it describes what is there rather than judging whether it looks like a company a 400-unit firm would trust with its tenants. That judgement is stature, and on a website it is a large part of the job.
A stranger decides what kind of company this is in about four seconds, from the surface, before reading a full sentence. A model reads every word with equal attention and never skims. It cannot tell you that the fourth section, where the real explanation lives, is a section nobody reaches. Look at how the Sawdust website handles a dry subject: it shows who owns each layer of a music license as a visual structure instead of a wall of text. No text review would have asked for that, and it is what made one scroll enough.
If you are weighing a draft from a designer rather than your own page, the checks in how to judge design work without a designer cover the visual side a model skips.
Then put the page in front of five people
Treat the model as the cheap first pass, not the verdict. Once the NOT ON THE PAGE answers are fixed, show the first screen to five people outside the company for five seconds each and ask what it is, who it is for and what they would click. The five second test post walks through how to score it. Where the model and the five people agree, you have a fix. Where they disagree, believe the people.
The cost of skipping this is easy to miss because it is spread out. Every visitor was paid for, in ads, in content or in your own hours on LinkedIn. A page that never names the job turns that spend into demo calls with the wrong buyers, and the filtering lands on your calendar instead of on the page. For a company working toward a Series A, where the bar for AI startups is now around 3.5 million in ARR (Carta, Q1 2026), those wasted calls are runway. If the fixes need more than a headline and you have nobody in house, the shortlist of landing page agencies for startups without designers compares studios on public work.
An assistant can find the missing noun. Deciding how the page should look so a stranger trusts it is the design work Studio Maydit does, for AI and SaaS teams like Wave, PixelFlow and Mi-VAD, built in Framer, Webflow or custom code. If your prompts keep coming back NOT ON THE PAGE and you are not sure what to change first, book a 30 minute call and bring the page with the transcript of your best prompt.
