An AI builder ships a website in an evening, and Lighthouse hands it an average SEO score of 96 out of 100. Sounds great. But when we measured 93 of these sites, the median mobile load was 6.6 seconds, two thirds had no main heading, and three quarters had no structured data at all. That 96 out of 100 can mislead you. Here are the numbers.
As building websites with AI builders becomes the norm, you hear the same line more and more: "You can make a website in one click now, so why pay extra...". Is that true or not? So we decided to run our own study (spoiler: sometimes). First, I want to thank Zhadyra Askarkyzy. I am Tair, I run DevNova in Bratislava, and it was that 96 out of 100 that pushed me to write this.
Let me get it out of the way: at DevNova we build sites in custom code, so AI builders are, to a degree, our competition. That is exactly why I will be straight here. No bashing, no "hire us". Just the raw numbers from the test, and an honest look at when an AI builder is genuinely ENOUGH and not worth paying more for.
1. A website in one click: what AI builders promise
Open LinkedIn and within a week you will hit the same post: someone built a site over a weekend, "no developer, no agency", with a hundred likes underneath. The ads run the same way. You describe what you want and the AI generates the site, copy and images included. And I will give them one thing. Visually, it often looks decent.
The problem is what a pretty screen does not show. How the site looks in the editor on your laptop is not what a visitor feels on a four-year-old phone on mobile data. Between "looks finished" and "loads fast and Google likes it" there is a gap. And the builder usually does not fill it for you.
"Looks finished" and "works" are two different things. An AI builder gives you the first by evening. Not the second.
2. Reality: what we measured across 93 sites
So this would not be one more opinion off the internet, we ran our own measurement. We took 93 real, live sites built on AI builders: Wix (37 sites), Lovable (31), Gamma (20), and five more on various AI and no-code tools including Mobirise. We ran each through Google PageSpeed Insights on mobile and checked the basic page structure. Here is what came out:
| What we measure (mobile, n=93) | Result | Verdict |
|---|---|---|
| SEO score (Lighthouse) | average 96 / 100 | "looks SEO-ready" |
| LCP, main content load | median 6.6 s · 70% poor | slow |
| No main heading (h1) | 66% | two in three sites |
| Structured data (schema) | missing on 75% | Google/AI cannot read it |
| No canonical | 44% | duplicate-content risk |
| CLS, visual stability | 93% good | this they get right |
Method: we collected 93 publicly available, live sites regardless of speed (we did not hand-pick slow ones); the only condition was that the given AI tool built the page. Measured via Google PageSpeed Insights on mobile in lab (Lighthouse) mode; real users on better connections may see lower numbers, but the ratio between sites does not change. The dataset is free to cite (CC-BY) on the AI-Builder Audit page.
Notice the contradiction. For SEO those sites averaged 96 points out of 100, a score that suggests they are ready to rank. And yet half are missing the basics that decide whether Google moves you up at all. How does that add up?
3. Why a 96/100 SEO score misleads
One thing has to be said plainly: the Lighthouse SEO score does not measure whether you will rank. It measures whether you have a few technical basics on the page, a title, a meta description, readable text, a language attribute. Those are things a builder fills in automatically. That is why the score jumps to 96. It is like having clean windows and a full tank. Nice, but it tells you nothing about whether the engine actually pulls.
Real ranking in 2026 stands on something else. On speed, which Google measures through Core Web Vitals, where an LCP over 2.5 seconds already counts as a weak signal. On semantic structure, where the main heading tells the engine what the page is about. And on structured data, which lets Google and language models like ChatGPT or Perplexity understand that you are a business, a review, or a product. The Lighthouse SEO score barely touches any of those three.
A 96/100 score does not measure whether Google moves you up. It measures whether your fields are filled in.
4. What exactly is missing on these sites
Let us take it in order, because each of those three numbers has its own concrete consequence.
Speed (LCP median 6.6 s). Seven in ten sites loaded in the band web.dev calls poor. Why, when it flies in the editor? Because the builder loads layers onto the page that nobody turns off for you: a full CSS framework, animation libraries, a set of plugins, and full-size images with no compression. You do not see them. The visitor on mobile downloads every last one.
Missing main heading (66%). The main heading (the h1 in code) is the first thing by which a search engine grasps the topic of a page. Two thirds of the sites had none at all: either the builder replaced it with plain large text, or it scattered five headings and none was the main one. To Google that is a chapter with no title.
No structured data (75%). Schema (schema.org) is the invisible description that tells a search engine "this is a local business in Bratislava, these are the hours, these are reviews". Three quarters of the AI sites had none. Without it you lose star ratings and rich results, and what matters more in 2026, language models read you worse and cite you less. You can check it in a minute with the Rich Results Test.
And to be fair, one thing those sites do well. Visual stability (CLS) was fine on 93% of them. The page does not jump while loading. Builders handle that for you decently. But speed, heading, and schema are three of the four legs of the table. When they are missing, a pretty design will not hold you up.
5. A real example: a site we rebuilt
So these are not only aggregate numbers. One of those 93 sites is fb-n.com, built on the no-code tool Mobirise. In our audit it scored Performance 56/100 and the main content loaded in 11.4 seconds. And this is the very site a B2B company from Kazakhstan later asked us to rebuild (they found us through profiles on Clutch and TechBehemoths). The pattern was exactly like the table above: a good-looking site, disastrous speed. We rebuilt it from scratch in custom code (web development): 1.5 seconds and 99/100, and the SEO score rose from 91 to 100. The full before and after is on the project page and in detail in an AI builder site vs a custom site.
I will admit I hesitated at first whether it was even worth a rebuild. The client had a site that worked. It was the measurement that showed how much that slowness was costing them in visitors. This is the moment where it gets decided: a cosmetic tweak in the builder will not move those numbers. What moves them is what sits under the hood.
A site that works and a site that loads in 1.5 seconds are often a full rebuild apart.
6. What to do about it: how to check your own site
Here is what I would do in your place, free and with no strings, before paying anyone anything:
- Measure reality. Open PageSpeed Insights on mobile and enter your page. If LCP is over 2.5 s, you have weak Core Web Vitals and it is holding you back.
- Check the heading and the schema. The Rich Results Test shows whether you have structured data. In the browser (right click, View Source, search for h1) you confirm the main heading. If they are missing, you are in those two thirds.
- Look inside Search Console. Google Search Console shows you directly which pages Google is not indexing and why. It is free and it is the truth firsthand.
Once you find the problem is real, the question becomes fix or rebuild. And here is the value a builder cannot give you by design, and it is not about paying more for the same tool. It is about depth: knowing exactly why LCP is six seconds and solving it at the code level, designing a semantic structure Google understands, and adding structured data so language models read you too. That is the difference between custom web development and a generator. If you run an e-shop it counts double, see custom e-commerce. Ballpark prices are in the pricing, and a direct comparison is on DevNova vs Wix.
7. Who an AI site is genuinely fine for
Now the part a site builder is supposed to keep quiet about. For a chunk of people an AI builder is exactly the right call, and paying for more would be foolish. Let us split it honestly.
A builder is genuinely enough if: you are making a personal portfolio or a freelancer card; you need a one-off page for an event, a wedding, or a conference; you are launching a temporary coming-soon page; you are testing an idea or an MVP and only want to learn whether it has legs; you run a small blog with no traffic ambitions; you have an information page for a nonprofit; or you just want a digital menu or flyer for a restaurant with no ordering. In these cases the site is fast and cheap, and its limits will never hurt you.
A builder hits a wall if you are: an e-shop; a B2B distributor or manufacturer; a real estate agency; a clinic, dentist, or doctor; a lawyer or accountant; a gym or wellness with membership and bookings; a restaurant that takes orders and reservations online; an auto service; a construction firm; or any local service that depends on whether Google shows it on the map and in search. Here the site is not a card, it is meant to bring in inquiries. And that is exactly where slowness, a missing heading, and no schema show up on the bottom line.
The question is not AI or not. It is: should your site bring customers from Google? If yes, a builder hits a ceiling.
A simple test I give everyone: if the site could vanish tomorrow and nothing would happen to your business, a builder is ENOUGH. If your inquiries, calls, or orders would vanish with it, you need a site that is built properly.
Frequently asked questions
So are AI sites bad? No. They are fast and cheap to launch, and for a portfolio, an event, or an idea test they are ideal. The problem starts when you expect traffic and inquiries from Google, that is when the limits show.
Why did a site score 96/100 for SEO and still be slow? Because the Lighthouse SEO score measures a few technical basics (title, description, language), not speed or structure. Those are scored separately through Core Web Vitals. So a high SEO score and a slow site go together just fine.
Can an AI site be sped up without a rebuild? Partly. Compressing images and turning off unused plugins helps by a few tenths of a second. The big numbers (from 6 s to 1.5 s) only change with a rebuild in custom code, because the bloat is built into the builder itself.
What is LCP and why 2.5 seconds? LCP is the time until the main content loads for a visitor. Google treats up to 2.5 s as good and over that as poor. In our measurement the median was 6.6 s.
Does Google index sites from builders? Yes, it does. But indexing does not mean ranking high. A missing heading, no schema, and slow loading mean it places you below a better-built competitor.
How many of the 93 sites had structured data? Only 25%. Three quarters had none, which costs them rich results and clean readability for language models.
Which builder did worst? We deliberately do not rank by individual tool. The samples for Wix, Lovable, and Gamma are not large enough for a fair comparison and we will not claim more than the data carries. The numbers here are an aggregate across the whole sample. The pattern, high SEO score, weak speed, missing structure, was shared across all of them.
Where is the data if I want to cite it? The full dataset is freely available (CC-BY) on the AI-Builder Audit 2026 page, including a CSV download.
About the author
Tair Khamitov, founder of DevNova, a web studio in Bratislava that builds sites and automations for small and mid-sized companies. This piece grew out of measuring 93 real websites, because that 96 out of 100 demanded an explanation.
Have a builder site and suspect it is slow? Send a brief and I will measure it and tell you straight whether a tweak is enough or a rebuild is worth it. No strings.
External sources
Curious about the speed of the Slovak web in general? See our separate dataset, State of SK Web Speed 2026.