DEV Community: Tair Khamitov

Manipulating AI answers is officially spam, and it can get you deindexed

Tair Khamitov — Mon, 14 Sep 2026 10:14:01 +0000

Since 15 May 2026, Google officially treats attempts to manipulate AI answers in Search as spam, with the same sentence it uses for ordinary ranking spam. The penalty is the same framework: a site can rank lower, or not appear in results at all. Here is exactly what changed and what it means for you.

The same pitch keeps landing in my inbox: "for a monthly fee, ChatGPT and Perplexity will start recommending you". I'm Tair, and at DevNova I build websites and automations, we add schema and structured data on every project, so whether AI cites us is something I watch first-hand. And that exact pitch turned from a "growth hack" into a risk on 15 May 2026. I'll walk you through what Google changed, what is still fine, and where the line actually sits.

1. The LinkedIn promise: write it so AI recommends you
2. What Google changed on 15 May 2026
3. It targets AI Overviews and AI Mode
4. Three tactics that are now spam
5. The penalty: same framework, up to deindexation
6. May 2026: the end of schema and AI-answer tricks
7. What safe citability looks like (how we do it)
Frequently asked questions

1. The LinkedIn promise: write it so AI recommends you

Behind that LinkedIn message is always the same kind of seller, an agency or a self-styled "GEO expert" who, for a monthly retainer, promises "guaranteed recommendations from AI". They call it "GEO" or "AEO", optimization for generative answers, and the recipe they send next sounds simple. Write your pages so the models love them. Seed mentions of yourself into forums and review sites where AI pulls its citations. Add phrases like "the best in town" so the model recommends you. And suddenly, they claim, ChatGPT, Perplexity and Gemini all cite you.

It sounds tempting because part of it is true. Generative answers really are changing how people search, and being cited in them has value. The problem is where exactly the line runs between "make genuinely good content so models enjoy reading you" and "trick those models into recommending you". The first is legitimate work. The second is what Google named in May 2026 and threw into the same bucket as spam.

Optimizing for AI answers is not banned. Manipulating them is. There is a whole canyon between them.

2. What Google changed on 15 May 2026

On 15 May 2026 Google rewrote the opening of its spam policies (the document itself is stamped "Last updated 2026-05-15"). The definition of spam used to talk about attempts to "manipulate Search systems into ranking content highly". Now the sentence continues with words that were not there before: "...or attempting to manipulate generative AI responses in Google Search". For the first time, manipulating an AI answer is named directly as spam.

And right after it comes the consequence, in black and white: "Sites that violate our policies may rank lower in results or not appear in results at all." This is not a new penalty for AI. It is the same sentence Google has used for years to cover ordinary spam. The change is in scope, not in punishment.

Google did not write a new penalty. It wrote that the old one now covers AI answers too.

3. It targets AI Overviews and AI Mode

An important detail: the wording says "generative AI responses in Google Search", not just the classic blue links. Industry read-outs of the change (Search Engine Land among them) confirm it: the rule covers AI Overviews (the summaries above the results) and AI Mode (the conversational search mode). Trying to game your way into them is treated exactly like trying to game the ranking.

In practice that means one thing. If you have been treating AI Overviews as a "grey zone" where the rules do not apply yet and you can bend things, that zone just closed. Google explicitly said the same yardstick that governs results governs the AI summaries.

4. Three tactics that are now spam

So what concretely falls under the new sentence? Google did not publish a closed list, but its policies and the industry coverage point to three recurring patterns:

Biased "best of" rankings and listicles. A page that poses as an independent comparison ("the 5 best studios in town") but is engineered so the model lifts one specific name out of it, yours. The goal is not to inform the reader but to feed the answer.
Sabotage through third-party forums and review sites. Deliberately seeding mentions and "reviews" where models pull their citations, so the AI remembers a recommendation that never came from real experience. Industry and security research coined a name for this class of attack, recommendation poisoning. Microsoft described it as manipulating the "memory" of AI systems. It is an industry term, not Google's wording, Google simply files it under spam.
Prompt injection into your own pages. Hidden instructions in the text or code that a human does not read but a model does, designed to force it to recommend you. Invisible to the visitor, a script for the AI.

The common denominator is clear: the goal is not to serve the reader but to deceive the machine sitting between you and the reader. That is precisely what Google calls manipulation.

5. The penalty: same framework, up to deindexation

Let us be precise here, because there is a lot of needless fear around this. Google did not introduce some new "AI death penalty". It used its existing spam framework, and that framework is a scale. At one end is a demotion: the site ranks lower. At the far end is what the document literally says, the page "may not appear in results at all", that is deindexation. Deindexation is the ceiling, not the standard rate.

And enforcement is not just theory in a document. On 24 June 2026 Google rolled out a separate June spam update, which finished on 26 June and runs on SpamBrain. These are two distinct events: 15 May was the policy text change, 24 June was the algorithmic wave that enforces it. By Google's own comments the June wave did not target link spam, which I read as a sign that other forms of manipulation are where the enforcement lands.

Deindexation is the ceiling, not the standard rate. But it is a ceiling you do not want to test.

Why is this serious even without panic? Because manual and algorithmic penalties are hard to appeal and even harder to recover from in time. If your growth in AI answers rests on a trick that falls under enforcement tomorrow, that growth is really risk with deferred maturity.

6. May 2026: the end of schema and AI-answer tricks

May 2026 was not only about AI answers. Eight days before the policy change, on 7 May 2026, Google finally stopped showing FAQ rich results, those expandable questions under a link, for every site including the government and health sites that had kept them. This was not sudden: back in August 2023 Google already restricted FAQ rich results to well-known authoritative pages, a response to schema being abused by everyone. May 2026 just finished a slow burial.

I know this one from my own work. For years FAQ schema was sold as "easy stars in search". We kept it on sites for a different reason, because of how language models read it, and told clients outright not to expect a rich result from it anymore. Anyone still ordering FAQ schema "for the rich results" in 2026 is buying last year's snow. To be precise: FAQ rich results are gone, but local-business results and star reviews still work, let us not blur the two.

In 2026 schema will not earn you FAQ stars. It earns you readability for the models. That is a different game.

See the pattern? Google is systematically closing shortcuts. First schema tricks (rich results for FAQ), now AI-answer tricks. Whoever builds on shortcuts keeps hitting the same wall. Whoever builds on depth walks around it.

7. What safe citability looks like (how we do it)

The good news: the way into AI answers without risk is boring, and that is exactly why it works. Models cite for the same reasons Google moves you up: real depth, a credible author, correct technical basics. ZERO tricks. Here is how we approach it on projects:

Real content with depth and a date. Text that actually answers the question, with a named author (Person in structured data) and a modified date. That is what models and Google read as a trust signal, not as a trick.
Correct schema, not schema to deceive. On every project we ship the same structured-data baseline — Article, Person and LocalBusiness JSON-LD, plus FAQPage or Dataset where it fits. We mark via schema.org what the page genuinely is, a business, a product, an article, an author, so models and the search engine understand you and can cite you correctly. Not to pretend you are something you are not. That difference is the whole difference between legitimate work and spam.
Speed and clean structure. A model and a search engine pull text, headings and data, not pretty graphics. That is why we build on custom code, where we know exactly what a page contains and how fast it loads.

What it looks like in practice: when we do SEO and content, we do not start from "how do we trick the model" but from "what question should this client be the best answer to on the internet". Then we write and tag the page so that it is true, and verify it through the Rich Results Test and Search Console. How exactly this differs from a generator-built site that "looks SEO-ready" I lay out in whether AI-built sites are good for SEO. Ballpark prices for this work are always public in the pricing.

And honestly about the line too: if you came to me wanting to "feed" forums with mentions or hide instructions for a model inside a page, I would say NO. Not out of caution, but because that is exactly what now falls under enforcement, and I will not sell you risk wrapped as growth.

If you want to know where you actually stand in AI answers, I will run a GEO/SEO check: are you growing in AI answers safely, or one manual action away from a penalty? Write via contact and I will tell you straight what is fine and what I would change. No strings attached.

Frequently asked questions

Do FAQ or How-To schema still earn a rich result in 2026? No. Google stopped showing FAQ rich results on 7 May 2026 for all sites (they were already restricted from August 2023). Schema still matters, but as a signal for language models and page understanding, not as a path to stars.

Is optimizing for AI Overviews (GEO) against Google's rules? No, optimization itself is not. What is banned is manipulation, tricks that deceive the model into recommending you. Making genuinely deep content that AI is happy to cite is legitimate and encouraged.

What exactly did Google change on 15 May 2026? It added a line to the definition of spam saying spam includes "attempting to manipulate generative AI responses in Google Search". For the first time, manipulating an AI answer is named directly as spam.

Can a site really be deindexed for trying to influence an AI answer? In the extreme case yes, the document says a page "may not appear in results at all". But that is the ceiling of the framework, not the standard. The more common consequence is a demotion. It is not a new AI death penalty, it is the existing spam penalty.

Is "recommendation poisoning" Google's official term? No. It is an industry and security label (Microsoft described it too) for seeding manipulative mentions where AI pulls its citations. Google does not use the phrase, it simply files that kind of attempt under spam.

What is the difference between the 15 May policy change and the June 2026 spam update? They are two things. 15 May was the policy text change (what counts as spam). 24 June was the algorithmic wave (the June spam update, completed 26 June) that actually enforces the rules through SpamBrain.

How do I get cited by AI safely, without risking a penalty? Through honest depth: real content, a named author, a modified date, correct schema and a fast page. No hidden instructions for models, no fake reviews. Boring, but risk-free, and long term it is the only thing that works.

About the author

Tair Khamitov runs DevNova in Bratislava, a web studio that builds websites, e-shops and AI automation for small and mid-sized businesses across Central Europe. He translates regulatory and search changes into the language of "what do I actually do with my page". This is not legal advice, it is an orientation overview from practice. Contact: b2b@devnova.eu.

External sources

The EU AI Act for a small business: what actually applies from August 2026 (and what is just panic)

Tair Khamitov — Sat, 04 Jul 2026 09:33:21 +0000

If your business uses an AI chatbot or generates content with AI, one rule from the EU AI Act applies from 2 August 2026: you must clearly tell visitors they are talking to AI, not a human. For most small businesses that is the whole core, not the 35-million-euro fines from the LinkedIn headlines.

I read the full Regulation (EU) 2024/1689, including the omnibus changes approved in June 2026, so you do not have to. I am Tair Khamitov, I build websites and AI automation for small businesses in Bratislava. Most articles about the AI Act are written either for corporate lawyers or for clickbait headlines. This is the version for a cafe, e-shop or clinic owner: what actually applies to you, what you can safely ignore, and what you can settle in one afternoon.

1. The panic you keep hearing about the AI Act

Open LinkedIn and you find the same loop: "the AI Act changes everything", "fines up to 35 million euro", "your business must act now". Most of it is written by consultants who want to sell you an audit, or by law firms targeting corporations. For a sole trader with an e-shop, or a cafe with a booking bot, it is needless fear.

The reality is duller and cheaper. The AI Act is tiered by risk, and almost everything a normal small business does with AI (a chatbot, generating text, automating email) sits in the two lowest tiers, where the only real duty is transparency. Not an audit, not a certificate, not a lawyer on retainer.

The AI Act is not about banning AI. It is about people knowing when they are talking to a machine.

2. The AI Act sorts everything into four risk levels

The entire Regulation (EU) 2024/1689 rests on one idea: the higher the risk to people, the stricter the rules. Four levels:

Unacceptable risk, banned outright (social scoring of citizens, manipulative subliminal techniques). A normal business is not here.
High risk, AI-driven hiring, credit scoring, medical diagnostics. Strict documentation. Most small businesses are not affected.
Limited risk, chatbots and AI-generated content. This is where the vast majority of small businesses sit. The duty: transparency under Article 50.
Minimal risk, spam filters, AI in games. No obligations.

The question a small business should ask is not "do I have to comply with the AI Act?", but "which tier does what I do fall into?". For a chatbot and marketing text, the answer is almost always: limited risk.

3. Timeline: what applies when (after the omnibus)

The AI Act phases in over time. In 2026 the EU institutions agreed on the so-called digital omnibus, which pushed back the deadlines for high-risk systems. That matters, because the original dates still circulate online as if they were current. The state of play:

Date	What starts applying	Small business?
2 Feb 2025	Banned practices + AI literacy (Art. 4)	Yes, literacy
2 Aug 2025	General-purpose AI (GPAI) + governance	Indirectly
2 Aug 2026	Transparency (Art. 50) + enforcement	YES, the main date
2 Dec 2027	High risk, Annex III (deferred by omnibus)	Rarely
2 Aug 2028	High risk in products, Annex I (deferred)	Rarely

Source: Regulation (EU) 2024/1689 plus the omnibus amendment the EU institutions agreed in 2026. The high-risk dates moved; the general 2 August 2026 date stays in force (marking of AI-generated content has a transition until 2 December 2026 for systems placed on the market before then).

4. What a small business must do (and what it need not)

If you run an AI chatbot on your site, Article 50 applies from 2 August 2026: at the first interaction you must state clearly and distinguishably that the visitor is talking to AI, not a human, unless that is obvious. That is one sentence in the welcome message. That is all.

The second duty, in force since February 2025, is AI literacy. It sounds official; in practice it means the people in your business who use AI understand it at a basic level, what it can do, where it makes things up, when to double-check its output. For a small business that is a short internal briefing, not a certificate.

And what you need NOT do: no formal audit, no registering your chatbot, no lawyer on retainer, and no paying for an "AI Act certification" that someone is selling you. Those are either obligations for high-risk systems, or inventions to sell a service.

If you generate visuals or video with AI (deepfake-type content), those must be visibly marked and machine-readable. Ordinary marketing text that a human has read and approved does not fall under the marking duty, human editorial review is an exception written into Article 50 itself.

5. Fines: why 35 million euro is not about you

Yes, Article 99 allows fines up to 35 million euro or 7% of worldwide turnover. But that is the top band for banned practices, social scoring, manipulative systems. For breaching other obligations the ceiling is 15 million euro or 3%.

The key point for a small business: the regulation expressly requires the interests of SMEs and startups to be taken into account. For small and mid-sized enterprises the lower of the two values (the percentage or the amount) applies, not the higher. Fines must be proportionate and must not wipe out the firm's viability. The real risk for a cafe that forgets to label its chatbot? A request to fix it, not a business-ending fine.

Those headline millions are for firms doing banned things. You just have to be transparent.

6. Who supervises this in Slovakia

Member states had to designate national supervisory authorities. In Slovakia the framework is being built by MIRRI through a national law moving through 2026. For the part where AI processes personal data, the Office for Personal Data Protection is competent; product conformity is covered by other bodies (such as the Slovak Trade Inspection).

For an ordinary business owner this means one thing: the exact split of powers is still being finalised, but the obligations from the regulation apply directly regardless of which office ends up enforcing them. Do not wait for the national law, you can handle chatbot transparency either way.

7. Your compliance in one afternoon

A concrete path for a small business that uses AI. Three steps, no lawyer:

Label the chatbot. One sentence in the first message: "You are chatting with an AI assistant; we will connect you to a human if needed." Article 50, done.
Brief the team. 30 minutes: what AI can do, where it makes things up, when to verify output. That is your AI literacy, write down the date of the session.
Mark AI visuals. Visibly mark AI-generated images and video. Text a human approves needs no marking.

When we build an AI chatbot or automation, the Article 50 disclosure is set by default, it is part of DevNova AI Automation, not an extra charge. How it fits the wider digitalisation path is covered in our guide to digitalising a small business, and a concrete bot deployment on an e-shop in our article on AI chatbots. Prices are public on our pricing page.

Frequently asked questions

Does my small business have to do anything right now? If you use an AI chatbot or generate content with AI, yes: from 2 August 2026 you must clearly label that it is AI. Most other obligations do not touch small businesses.

Am I at risk of 35-million-euro fines? In practice, no. The top rates are for banned practices that a normal small business does not do. For SMEs and startups the lower of the two values applies, and the fine must be proportionate.

Is the AI chatbot on my site a high-risk system? No. A customer-support chatbot falls under transparency (Article 50), not high risk. The only real duty: tell the visitor they are talking to AI.

What is the AI literacy that applies from February 2025? That the people using AI in your business understand it at a basic level, what it can do, where it errs, when not to rely on it. For a small business that is a short internal briefing, not a certificate.

Did the 2026 omnibus postpone any deadlines? Yes. The obligations for high-risk systems (Annex III) moved to 2 December 2027 and for product-embedded ones (Annex I) to 2 August 2028. The 2 August 2026 date for transparency stays in force.

Do I have to label AI-generated text on my blog? If the text informs the public on matters of public interest and has not had human editorial review, yes. Ordinary marketing text that you have read and approved does not fall under this duty.

Who enforces this in Slovakia? The supervisory framework is being built by MIRRI; for AI processing personal data the Office for Personal Data Protection is competent. The exact split of powers is being finalised by a national law during 2026.

About the author

Tair Khamitov is the founder of DevNova in Bratislava. He builds websites, e-shops and AI automation for small and mid-sized businesses across Central Europe, and reads regulatory texts to extract the one percent that actually touches an ordinary business owner. He is not a lawyer; this article is an orientation overview, not legal advice. For a specific case, consult a lawyer. Contact: b2b@devnova.eu.

This article was originally published on devnova.eu.

External sources

AI-built websites score 96/100 for SEO. Google still cannot see them. We tested 93 sites.

Tair Khamitov — Fri, 05 Jun 2026 12:10:12 +0000

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.

DEV Community: Tair Khamitov

Manipulating AI answers is officially spam, and it can get you deindexed

Contents

1. The LinkedIn promise: write it so AI recommends you

2. What Google changed on 15 May 2026

3. It targets AI Overviews and AI Mode

4. Three tactics that are now spam

5. The penalty: same framework, up to deindexation

6. May 2026: the end of schema and AI-answer tricks

7. What safe citability looks like (how we do it)

Frequently asked questions

About the author

External sources

The EU AI Act for a small business: what actually applies from August 2026 (and what is just panic)

1. The panic you keep hearing about the AI Act

2. The AI Act sorts everything into four risk levels

3. Timeline: what applies when (after the omnibus)

4. What a small business must do (and what it need not)

5. Fines: why 35 million euro is not about you

6. Who supervises this in Slovakia

7. Your compliance in one afternoon

Frequently asked questions

About the author

External sources

AI-built websites score 96/100 for SEO. Google still cannot see them. We tested 93 sites.

1. A website in one click: what AI builders promise

2. Reality: what we measured across 93 sites

3. Why a 96/100 SEO score misleads

4. What exactly is missing on these sites

5. A real example: a site we rebuilt

6. What to do about it: how to check your own site

7. Who an AI site is genuinely fine for

Frequently asked questions

About the author

External sources