Mike King’s case for cloaking LLMs is right. And a trap.

Mike King published a piece yesterday arguing that you should cloak content from ChatGPT, Perplexity, and the rest of the non-rendering AI crawlers. Not from Google. Not from users. From the LLM scrapers specifically — by putting your unique, hard-won content behind a JavaScript block they cannot execute.

He's technically correct. The mechanics work. The data he cites — Vercel and MERJ finding zero JavaScript execution across 500 million GPTBot fetches, 69% of major AI crawlers unable to render JS at all — is real. If you have a proprietary dataset or a framework that took six months to assemble, you can hide it from the scrapers while Googlebot and real users see everything.

And I still think you shouldn't do it. Or rather: I think most of the people who'll read King's piece and reach for this tactic are about to make a strategic error dressed up as a technical clever-clever.

The mechanic is sound. The premise is the problem.

King frames the case as defensive. Your competitors will use ChatGPT to extract and republish your unique content. Why make it easy? Cloak the differentiator content from the bots that don't render JS, and you've kept your moat.

The trouble is what "the moat" actually was in the first place. If your information gain is genuinely defensible — a real dataset, a real framework, real proprietary research — what's protecting it isn't obscurity. It's authority. It's the fact that you're the source everyone cites. AI systems disproportionately surface content from entities they recognise. They cite brands. The unique research is what gets you cited, which is what builds the brand, which is what gets you cited again.

Hide the research from the citation engines and you've broken that loop on purpose.

You don't build a moat by becoming invisible to the systems your buyers use.

What King is actually solving for, and who it serves

Read the piece carefully and it's solving a publisher problem, not a consultant problem. If you're a content site running on ad revenue and ChatGPT is cannibalising your clicks, then yes — the calculus changes. Every uncited extraction is a direct hit to the business model. Cloaking the unique stuff is at least a coherent response to a real economic harm.

But for almost every business reading this — the service business, the SaaS, the consultancy, the e-commerce site, the agency — the calculus is inverted. You don't make money from people reading your content. You make money from people knowing who you are when they have a problem. AI citation is not a leak. It's distribution.

This is the second time in a month the industry has reached for a clever tactical fix to what is fundamentally a positioning problem. The schema-for-citations pitch was the first. Ahrefs ran the test, and adding schema didn't move AI citations on any platform. Cloaking is the same shape of mistake: a technical lever being pulled because pulling levers feels productive.

The other 90% of the piece is the actual story

What's quietly more important in King's post is the part that isn't about cloaking. The 499 response code. The fact that AI crawlers abandon slow pages. The fact that Cloudflare's own HTML announcement weighed 16,180 tokens against 3,150 for the markdown version. The fact that pages with high failure rates to AI crawlers receive — per Profound's analysis of 700,000 pages — roughly 18 times fewer citation events than stable ones.

That's not a cloaking story. That's an eligibility story. And it matches what I keep saying about the protocol layer eating the SEO stack — the bit between your content and the systems consuming it is now where wins and losses actually happen, and almost nobody is auditing it.

Serve fast. Serve clean HTML or markdown. Don't time out. Don't return 499s to crawlers that have a one-shot retrieval budget. This is the real takeaway and it doesn't require any cloaking at all.

What happens when the bots start rendering

Here's the bit nobody wants to write about. The non-rendering crawler situation is temporary. It exists because rendering JavaScript at scale is expensive, and the AI companies have been optimising for training-data ingestion volume over fidelity. That math changes the moment retrieval quality becomes a competitive differentiator — which is happening right now, because every chatbot is bleeding into every other chatbot and the answer quality is the product.

OpenAI is not going to spend the next five years politely fetching raw HTML and giving up. The companies that build their differentiation on the premise that GPTBot can't execute JavaScript are building on a fault line. When the rendering arrives — and it will — every cloaked block lights up at once, and the people who deployed this tactic will need to explain why their site has a parallel content layer that was deliberately hidden from inference systems for two years.

That's not a great look to a buyer. It's not a great look to Google either, if it ever decides to update its definition of cloaking to include intent-to-deceive against AI surfaces. King's caveat — that this isn't against Google's guidelines because it isn't aimed at Google — is true today. It is exactly the kind of thing that becomes untrue with one policy update.

What I'd do instead

The honest version of this is much less fun to write. If you have genuinely unique content you're worried about being scraped and republished:

Publish it under your brand with enough authority signals — author, methodology, dataset citation, original-research disclosures — that downstream republishing reads as derivative. Get it cited by real outlets in the first 30 days, because that anchors the attribution graph. Mention the source page in your own subsequent writing. Get the dataset on a real domain people in your space know. Build the moat by being unambiguously the origin, not by hiding.

That's slower. It's less satisfying. It doesn't involve a clever `<div data-llm-protected>` and a robots-style block on a JS file. But it's the version that compounds.

The other version — cloaking — is a tactic that solves a real problem for a small number of publishers and a fake problem for everyone else who'll deploy it.

The wider pattern

I keep noticing the same loop in AI search commentary right now. A smart practitioner identifies a real technical asymmetry — bots can't render JS, schema isn't moving the needle, 499s gate eligibility — and the asymmetry gets repackaged as a tactic. Tactics travel faster than diagnoses do. Within a week, half the industry is implementing the tactic, and the underlying diagnosis — the one that actually matters — is the thing nobody talks about because it's harder to sell.

The diagnosis here is that AI search rewards entities with authority, distribution, and infrastructure that doesn't time out. Cloaking solves none of those. The 499 finding solves one of them. Brand-building solves the other two. That's the order to do the work in.

Mike King's piece is well-argued and the data is real. Read it. Then mostly do the boring thing.