How to Tell if a Website Is AI-Generated: 12 Signs to Check

Updated October 15, 2026. 8 min read

There is no single tell that proves a website was made by AI. There are a dozen small ones, and they are much more convincing together than alone. This guide walks through the signs we actually use when we check a site, from the strong ones you can see in the page source to the weak ones people over-trust.

We built Bot or Built to run these checks automatically, and we tested it on 100 real websites: AI content farms named in press investigations, sites made with AI website builders, and long-running human blogs and newsrooms. Some of what we learned surprised us, so the last section covers the signs that look convincing but don't hold up.

Strong signs: the site's own code gives it away

The most reliable evidence is technical. AI website builders and AI coding tools leave fingerprints in the HTML, and they rarely get cleaned up.

1. Builder fingerprints in the page source

Open the site, press Ctrl+U (Cmd+Option+U on a Mac) and search the source for a few telltale strings:

• Lovable: image paths containing lovable-uploads, attributes named data-lov-id, or hosting on a .lovable.app address.
• Durable: images loaded from cdn.durable.com.
• AI blogging plugins on WordPress: paths like /wp-content/plugins/aiomatic or /plugins/ai-engine. These plugins write or auto-publish posts.

One caution: count a match only when it appears in the site's code, such as a script, a stylesheet or an image path. A page that simply links to Lovable or mentions Durable in an article is not built with them. We learned this the hard way when an early version of our checker flagged stripe.com, which shows Lovable as a customer logo.

2. An empty page that only JavaScript fills in

Many sites generated by AI app builders ship a nearly blank HTML file with a single <div id="root"> and a script bundle named something like /assets/index-a1b2c3.js. All the text appears only after the script runs. For a long time this was the default output of tools like Lovable and Bolt, although Lovable projects created since May 2026 render on the server instead (more on spotting Lovable sites). It is a hint, not proof, because plenty of developers build React apps the same way by hand.

3. The default look of AI coding tools

AI coding tools tend to reach for the same component kit: Tailwind CSS with shadcn/ui and Lucide icons. In the source you will see class names such as bg-background, text-muted-foreground and lucide lucide-arrow-right. Again, humans use this stack too, so treat it as supporting evidence.

4. Template text nobody replaced

Lorem ipsum in a footer, "Your Company Name" in a contact block, an email ending in @example.com. Generated sites often go live before anyone reads them end to end. One placeholder is a sloppy human. Several on a brand-new site is a pattern.

Medium signs: how the site was published

5. Too many pages, too fast

Look at the site's sitemap, usually at /sitemap.xml. A domain registered six months ago with 1,500 articles is publishing around 250 pages a month. Very few human teams write at that pace. One site in our test set had close to 1,500 pages on a domain about eight months old.

6. Hundreds of pages dated the same day

Bulk-generated sites often push content in batches, so their sitemaps show hundreds of URLs with identical dates. Be careful here: some blogging platforms rewrite every date whenever the site is rebuilt. In our test set, three long-running human blogs had the same pattern for exactly that reason.

7. A fresh domain, or an old one with a new owner

You can look up a domain's registration date with any WHOIS or RDAP lookup tool. A very new domain is a weak signal on its own. The more interesting case is the opposite one: AI content farms increasingly buy expired domains that used to belong to real newspapers, radio stations or businesses, because old domains already have search-engine trust. Nieman Lab documented this pattern in 2025. To spot it, open the domain in the Wayback Machine and see whether the site looked completely different a few years ago.

Content signs: read like an editor

8. Authors who don't exist anywhere else

Click on the author names. Real writers usually have a footprint: other bylines, a LinkedIn profile, a social account with a history. Content farms often invent authors, sometimes with AI-generated headshots, and the same "writer" publishes dozens of articles a day on unrelated topics.

9. Every article has the same skeleton

Open five posts. If each one has the same introduction length, the same "Key takeaways" box, the same FAQ block at the end and the same closing paragraph that starts with "In conclusion", you are probably looking at a template being filled by a model.

10. Leftover chatbot messages

This one sounds too obvious to happen, but it happens. NewsGuard found published articles that still contained chatbot error and refusal messages, the kind of text a model writes when it declines a request. Search the site for phrases like "as an AI language model" or "I cannot fulfill this request".

11. Confident mistakes and stale "news"

Language models invent details and present old information as current. Look for a story that reports a years-old event as breaking news, quotes that you cannot find anywhere else, or biographies with wrong dates. One mistake proves nothing. A site full of them, with nobody correcting them, says a lot about how it is produced.

12. AI-generated images

Check image file names and URLs. Downloaded generator images often keep names like ChatGPT Image Apr 2 2025, DALL·E 2024-… or include the word "midjourney". Visually, watch for hands, text inside images and backgrounds that melt into patterns. Keep in mind that news sites legitimately use AI images when they write about AI, so this sign only matters in context.

Signs that don't hold up

A lot of advice online focuses on writing style. We measured the most popular style tells across our 100 test sites, and most of them did not separate AI sites from human ones.

• "AI words" like delve, tapestry and seamless. AI content farms active in 2025 used these words no more often than human writers did, roughly 0 to 2 times per 1,000 words in both groups. Newer models and their editors have learned to avoid the famous ones.
• Em dashes. Several human-written sites in our set used more than 8 em dashes per 1,000 words. One well-known essay blog used more than 15. Punctuation is a writer's habit, not a model's signature.
• Very even sentence length. This is a real tendency of language models, but the ranges overlap so much that it only helps alongside other evidence.

The practical lesson: style alone should never be the reason you call a site AI-generated. Technical fingerprints and publishing patterns are far more trustworthy.

Putting it together

Think of the signs as weights, not verdicts. A builder fingerprint in the code is close to proof that the site itself was generated. Fast publishing on a repurposed domain with invented authors is a strong pattern even without fingerprints. A few em dashes and a stock photo are nothing.

That is exactly how Bot or Built scores a site: each clue adds or subtracts points, the result is a probability, and every clue is listed so you can judge it yourself. Checking by hand still matters, because no automated tool can confirm who wrote a page. It can only tell you where to look.