AI Website Detector documentation
The AI Website Detector reads a live web page and weighs more than 80 signals to estimate whether the site was built and written by people or by AI. This guide explains every part of the report, how the score is calculated and how to get the most reliable result.
What it does
The AI Website Detector checks one public web page and estimates how likely it is that the site was produced by an AI website builder or written with AI. You get a single score from 0 (entirely human) to 100 (entirely AI), a plain verdict, a confidence rating and the full list of evidence behind it.
It looks at the same things an experienced web developer would: the platform the site runs on, the markup and code style, typography and design choices, imagery, metadata, business trust details and the writing itself. Findings that point toward people count just as much as findings that point toward AI, so a well-crafted site gets credit for its craft.
Use it to check a website before you pay for it, benchmark competitors, or review your own site and see which parts read as templated.
The score is an evidence-based estimate, not proof. A carefully edited AI-built site can read as human, and a template-heavy site can read as AI. That is why every report shows its confidence and its evidence.
Quick start
- Open the AI Website Detector.
- Type or paste a website address, such as
example.comorexample.com/about. You can leave outhttps://; the tool adds it for you. - Select Analyse site. A progress list walks through each stage while the page is fetched and checked.
- Read the verdict, score and confidence at the top of the report, then scroll down to see the evidence.
- Use Share this report to send the result to someone, or Scan another site to start again.
There is no sign-up. The address you scanned is added to the page link, so you can bookmark it.
Every option explained
The detector is deliberately simple. There is one input and no settings to tune, so every scan gets the same full analysis, including the deep checks.
| Control | What it does | Default |
|---|---|---|
| Website address | The public page to check. Accepts a bare domain, a full address or a specific inner page, up to 2,000 characters. | Empty |
| Analyse site | Starts the scan. It is disabled while a scan is running. | None |
?url= in the page link | When the tool is opened with an address in the link, for example /ai-website-detector/?url=example.com, the scan starts on its own. This is how shared reports work. | Not set |
| Show every signal | Each side of the evidence ledger shows its top 7 entries; this reveals the rest. Select any entry to read why it matters. | 7 shown |
| Context, not evidence | Lists findings that carry zero weight, such as the web server software in use. | Closed |
| Share this report | Share to LinkedIn, X, Facebook, WhatsApp or email, copy the link, or use your device's share menu where it is available. | None |
| Scan another site | Clears the report and takes you back to the address box. | None |
Which addresses work
Public http:// and https:// addresses on standard ports (80, 443, 8080 or 8443). Local and private network addresses, and addresses containing a username or password, are not accepted. Anything after a # is ignored.
Reading the results
The score and verdict
The score runs from 0 to 100 and always falls into one of six verdict bands:
| Score | Verdict | What the report says |
|---|---|---|
80 to 100 | AI-generated | Strong evidence the site was produced by an AI website builder or AI-written end to end. |
65 to 79 | Likely AI-generated | Multiple AI-generation markers across platform, structure and copy. |
52 to 64 | AI-assisted | Looks human-built but with AI-written copy or AI-assisted sections. |
42 to 51 | Inconclusive | Signals point both ways. Treat it as unclassified rather than either label. |
25 to 41 | Likely hand-built | Mostly human craftsmanship, possibly with light AI help on copy. |
0 to 24 | Hand-built | Consistent with a site designed, written and built by people. |
A zone bar and a gauge show where your score sits. A short sentence spells it out, for example "62 of 100 points of evidence pointed toward AI generation; 38 pointed toward human authorship."
Confidence
Confidence, shown from 20% to 92%, tells you how strongly the evidence agrees and how much of it there was. At 75% and above the verdict is reliable; from 55% it is a solid indication; from 40% the evidence is mixed; below 40% there was too little consistent evidence. A note under the stat tiles says which applies.
Two caps keep confidence honest. If fewer than 6 signals carried any weight, confidence cannot go above 45%. If the score lands between 42 and 58, it cannot go above 55%.
The stat tiles
- Platform: the detected platform and its type (AI website builder, Visual / template builder, Developer framework or CMS, or Unknown). It reads "Unidentified" when nothing matched.
- Signals weighed: all findings, split into human and AI.
- Words analysed: visible words read, plus the HTML size in KB.
The heaviest signals
Up to three cards show the findings that moved the score most: which way each points, what was found, why it matters and its share of its side's points. Strength tiers are Decisive (weight 25 or more), Strong (10 or more) and Supporting (under 10).
The evidence ledger
Two columns list every finding: "Evidence it was built by people" and "Evidence it was generated by AI". The AI column always adds up to the score and the people column to 100 minus the score.
Balance by evidence group
One bar per group shows its net pull, left toward people and right toward AI. The number is the raw total of that group's weights before caps, so read it for direction and size.
How it works
1. Fetching the page
The detector requests the page like a browser, identifying itself as RV-AI-Website-Detector. It reads the HTML the server sends but does not run JavaScript. It waits up to 15 seconds, follows up to 3 redirects and reads up to 1.5 MB.
2. Deep checks
Every scan also reads three extra sources on the same site, each with an 8 second time limit. Any that cannot be read are skipped.
- Stylesheets: up to two of the page's own stylesheets (up to 400 KB each, Google Fonts excluded), because fonts and theme settings usually live in CSS.
- robots.txt: a missing file is a small AI signal, and one that names an AI builder is a strong one. Rules for AI crawlers such as GPTBot, CCBot or ClaudeBot count toward people, because someone chose them on purpose.
- llms.txt: a real file is a small signal toward a thoughtful human operator.
3. Platform detection
The page, its generator tag and its server headers are compared against three fingerprint lists:
- 28 AI builders and AI writing tools, such as Durable, 10Web, B12, Lovable, Bolt.new, Vercel v0, Replit Agent and Base44, plus the hosting domains of 9 AI builders.
- 21 visual and template builders, such as Wix, Squarespace, Webflow, Shopify and Elementor. These add only a little AI weight.
- 18 developer stacks and CMSs, such as Next.js, Astro, Laravel, Hugo and WordPress. A build toolchain counts toward people; a CMS on its own is neutral.
When an AI builder is clearly identified, the score is raised to at least 82.
4. The 12 evidence groups
Every finding belongs to one group. Each group can add at most its cap to each side, so no single area decides the result alone.
| Group | Examples of what it checks | Cap |
|---|---|---|
| Platform & build stack | Generator tag, builder fingerprints, build toolchain | 45 |
| AI artefacts | Builder badges, leftover chatbot text, unreplaced placeholders, AI plugins | 45 |
| Writing style | Stock LLM phrases, sentence rhythm, vocabulary, concrete facts, personal voice | 42 |
| Page structure | Boilerplate headings, stock navigation, repeated calls to action | 26 |
| Typography | Default typefaces, monospace body text, licensed or self-hosted fonts | 24 |
| Visual design system | Untouched component kits, gradient habits, custom design tokens | 24 |
| Imagery | AI image hosts, stock photos, prompt-like alt text | 20 |
| Markup & code style | Utility-class density, BEM naming, developer comments | 16 |
| Metadata & SEO | Default titles, meta description, structured data, favicon | 14 |
| Business trust signals | Contact details, social profiles, dead links, testimonials | 12 |
| Hosting & infrastructure | Security headers, AI host headers, robots.txt, llms.txt | 8 |
| Accessibility care | ARIA labels and alt text coverage | 8 |
5. Calculating the score
Each finding carries a signed weight: positive toward AI, negative toward people. Hygiene that any CMS adds on its own, such as structured data or canonical tags, is scaled down. After the group caps, the score is the AI side's share of all the evidence, with 16 balancing points (8 per side) added in, so a page with little to go on lands near 50 instead of at an extreme.
Tips for the best results
- Scan the page with the most words. Writing checks need at least 60 words, and several need 150 or more. A homepage or about page usually gives the richest result.
- Read confidence before the verdict. A high score with low confidence is a hint, not a conclusion.
- Treat Inconclusive as a real answer. It means the evidence genuinely points both ways.
- Compare a few pages of the same site. A templated homepage and a hand-written blog post can score differently.
- Use the ledger as an editing checklist for your own site. Replace boilerplate headings, add real dates, names and figures, write meaningful alt text and publish working contact details.
- Pair it with a wider check. The Free SEO + AI Audit looks at how search engines and AI answers see the same site.
If you would like your website to read unmistakably as your own, my team at Thetechhacker can help. Let's talk.
Limits, privacy and fair use
| Limit | Value |
|---|---|
| Scans per visitor | 10 every 10 minutes |
| Result cache | 6 hours per exact address |
| HTML read per page | 1.5 MB (anything after that is not read) |
Caching
Scanning the same address again within 6 hours returns the saved report straight away, marked Cached result. example.com and example.com/about are cached separately. Cached results and shared links you open are free: only fresh scans count toward your limit.
Sharing
Share links carry only the site's hostname, without www. or the page path. The person who opens it gets a scan of that site's home address, so their report can differ if you scanned an inner page or the www version. When a report for that address is already cached, link previews on social networks show its score and verdict.
Privacy
- No sign-up and no account.
- The scan limit uses a one-way, salted fingerprint of your connection rather than your IP address itself.
- The detector does not save the addresses you check to a database. Reports live only in server memory for up to 6 hours.
- It only reads what any visitor can see on a public page.
Fair use
Please scan sites you have a genuine reason to look at. The limits keep the tool free for everyone.
Troubleshooting
- "Enter a website address first." The box was empty. Type or paste an address and try again.
- "That does not look like a valid website address." or "That hostname is not valid." Check for typos and make sure the address includes a domain ending, such as
.com. - "That domain could not be resolved." The domain does not exist or its DNS is not set up yet. Check the spelling.
- "Only http:// and https:// addresses can be analysed.", "That port is not allowed." or "Addresses with embedded credentials cannot be analysed." Use a plain web address without another prefix, a custom port or a username and password.
- "Internal or local addresses cannot be analysed." or "Private, loopback and reserved addresses cannot be analysed." Staging and local sites cannot be reached. Publish the page to a public address first.
- "That address is too long to analyse." Remove tracking parameters to bring it under 2,000 characters.
- "The website could not be reached (...)." The site is down, took longer than 15 seconds or has a certificate problem. Open it in your browser, then try again.
- "The website responded with HTTP status 403." (or another 4xx or 5xx code) The page is missing or the site blocks automated visitors. Check the address. If it is your site, allow the
RV-AI-Website-Detectoruser agent. - "The website redirected too many times." or "The website redirected somewhere that cannot be analysed." Open the site in your browser and scan the final address it lands on.
- "The website returned an empty page." Try the full address of a specific page.
- "You have reached the scan limit. Please try again in about N minutes." You have used 10 scans in 10 minutes. Wait the time shown.
- Very few words and low confidence. The site probably draws its content with JavaScript, so the server sends an almost empty page. Try an inner page with more text in its HTML.
- The result did not change after you edited your site. You are seeing the cached report. Check again after 6 hours.