What is a honeypot link?
Definition
A honeypot link is a link placed where no human visitor will see or follow it, so that any client requesting it reveals itself as automated.
The page that holds one loads normally, and nothing tells the crawler it has been caught. The verdict is recorded somewhere else and shows up later as blocks, challenges or quietly degraded content, which makes the cause hard to trace back to a single link.
A large deployment
Cloudflare announced AI Labyrinth in March 2025. When it detects unauthorised crawling, it inserts links into the customer’s existing pages that, in its words, “remain invisible to human visitors through carefully implemented attributes and styling”. They lead to pages generated in advance with an open-source model, factual but unrelated to the site, and those pages carry directives that keep them out of search indexes. Following the links is treated as evidence of automation and fed back into Cloudflare’s bot-detection models. A trap like this costs the site nothing in content and costs the crawler its bandwidth, its time and its classification.
How links are hidden, and what a crawler sees
We built a test page with seven links on 21 October 2026: a skip link, five hidden in different ways, and one normal link. Python’s html.parser, reading the markup, found all seven hrefs and had no way to tell them apart. A rendered check in Chrome 155, using the script below, separated them:
| Link | How it is hidden | Found by a markup parser | Rendered check in Chrome 155 |
|---|---|---|---|
| Skip to content | Off-screen until it receives focus | Yes | Follow: revealed on focus |
/trap-display |
display:none |
Yes | Skip |
/trap-visibility |
visibility:hidden |
Yes | Skip |
/trap-offscreen |
Positioned at left:-9999px, no focus rule |
Yes | Skip |
/trap-zero |
A 0 by 0 box with overflow hidden | Yes | Skip |
/trap-ghost |
Text coloured like the background | Yes | Skip |
/real |
Not hidden | Yes | Follow |
function shown(a) {
const r = a.getBoundingClientRect();
const s = getComputedStyle(a);
return a.checkVisibility({ checkOpacity: true, checkVisibilityCSS: true })
&& r.width > 0 && r.height > 0 && r.right > 0 && r.bottom > 0
&& s.color !== getComputedStyle(document.body).backgroundColor;
}
function focusRule(a) { // a rule such as ".skip:focus" that restyles this link
for (const sheet of document.styleSheets) for (const rule of sheet.cssRules) {
const sel = rule.selectorText || "";
if (!/:focus/.test(sel)) continue;
try { if (a.matches(sel.replace(/:focus(-visible|-within)?/g, ""))) return true; } catch (e) {}
}
return false;
}
const verdicts = [...document.querySelectorAll("a[href]")].map(a => ({
href: a.getAttribute("href"),
verdict: shown(a) ? "follow" : focusRule(a) ? "follow (revealed on focus)" : "skip",
}));
One detail cost us a run. The first version called a.focus() and measured the link again, but in headless Chrome launched with --dump-dom the window never had focus, so :focus never matched and the skip link was wrongly marked as a trap. Reading the stylesheet for a focus rule avoids depending on window focus. The colour test is also crude: it catches only an exact match with the body background, not a link placed behind another element.
Hidden is not the same as trapped
Plenty of legitimate links are invisible by default. W3C technique G1 describes the skip link at the top of a page and accepts one that becomes visible only when it receives keyboard focus, which is exactly the first row above. Treat every off-screen link as a trap and you will skip real navigation on accessible sites. The useful question is whether any person could reach the link, not whether it is visible in the default state.
The robots.txt version
The simplest trap needs no styling. The site disallows a path in robots.txt and links to it from its pages. A crawler that reads and honours the file under RFC 9309 never requests it; one that does not identifies itself in a single fetch. Honouring the file removes this whole class of trap and only costs you pages you were not entitled to. Read the policy before you read the page covers doing that at scale, and our survey of AI crawler policies shows how sites now write rules for specific agents.
Bounding the crawl helps with the traps no check catches. A maze of generated pages is endless by design, so caps on depth, pages per domain and bytes, as set out in bound the walk, stop it costing more than a few requests. A fresh address after a trap may not help: the classification can attach to your browser fingerprint or session as easily as to the IP.