For about twenty years, writing for the web meant satisfying two readers who happened to want the same thing. A human, who wanted something worth reading. And Googlebot, which fetched your page to rank a link that the human would click. Serve the first well and structure it for the second, and the click was the payoff you were both working toward.

A third reader changed the deal. The answer engine — the thing behind an AI chat’s “here’s what I found” — reads your page, synthesizes an answer, and hands it to the person who asked, who now never arrives. Zero clicks. Your page stopped being a destination and became an ingredient. And an ingredient gets used whether or not anyone ever learns your name.

I got curious whether my own site was even legible to that reader, so I built a scanner to look at it the way a model would. The headline was humbling: my static field-notes blog scored 93 out of 100 for classic Google SEO — “excellent” — and 79 for AI visibility — “moderate.” An A-student for search engines and a C for answer engines, on the exact same pages. Two readers, two very different report cards.

What “AI-legible” actually means

It comes apart into three layers, and only the last one looks like anything SEO ever taught you.

Can the bot reach the text at all? There are dozens of AI crawlers now, and they do different jobs — some fetch pages to train models (GPTBot, ClaudeBot, Google-Extended), some fetch to cite in live answers (PerplexityBot, OAI-SearchBot). Blocking the training crawlers while allowing the citing ones is a perfectly legitimate, deliberate choice — but you have to actually know which is which, and most robots.txt files are cargo-culted. Worse, robots.txt can say “allowed” while your CDN quietly challenges the same bot at the edge — a friendly front door with a bouncer nobody told you about. My scanner’s single loudest finding was that my own robots.txt was blocking my sitemap. An own-goal I’d never have caught by eye.

Can it parse what it reaches? Content that only appears after client-side JavaScript, or that’s tucked inside tabs and accordions, may never be extracted at all. The rule is blunt: the content that matters has to be in the initial HTML. This is the one place a plain static site quietly wins — the text is just there, nothing to execute, nothing to expand.

Is it shaped to be quoted? This is the genuinely new axis. Answer engines lift chunks, so self-contained sections of roughly 100–500 words travel well and a wall of text doesn’t. Leading a section with the direct answer, using headings phrased the way people actually ask the question, offering the occasional list or table or specific number, keeping a visible byline so a model can attribute you — these are the things that decide whether you get quoted or paraphrased into anonymity. SEO’s unit of value was the page and its keyword rank. The answer engine’s unit is the entity and the quotable chunk.

Classic SEO checks — title length, meta descriptions, canonicals, Core Web Vitals — still matter, but they’re now one axis of two. There’s also a frontier of “AI discovery” files (llms.txt, agents.json) that the hype cycle wants you to add today, and I’d hold those very loosely: llms.txt is a proposed standard that no major platform actually consumes yet. Building tightly to this month’s file conventions is optimizing for something that’ll be renamed by next quarter.

The tension I didn’t expect

Here’s what the scanner taught me that I didn’t want to learn: optimizing for the machine can degrade the human read. On the same run, my site scored 47 for readability — “difficult, college level.” And the fixes that would raise the AI score — shorter self-contained chunks, question-shaped headings, lead-with-the-answer, sprinkle in extractable stats — are, taken far enough, a recipe for writing like the quote machine. The two readers do not always want the same thing, and if you serve only the model, you sand off the voice that made the writing worth quoting in the first place.

There’s a deeper doubt underneath, too. Being citable is not being visited, and it is definitely not being paid. A perfectly AI-legible page can feed the zero-click answer flawlessly and get nothing back — you’ve optimized yourself into a free input. The value that survives is in the things a model can’t lift from your public text: proprietary data, a tool that actually does the work, a live workflow, someone accountable for the answer. Answer-engine optimization is a visibility play, not a moat — worth doing so you’re not invisible to the reader that increasingly matters, but not to be confused with being safe.

So I made the cheap changes and left the rest: content in the HTML (already true — static site), a real byline, question-shaped headings where they didn’t fight the prose, the sitemap un-blocked. The number turned out to be the least useful thing the scanner gave me. What mattered was seeing the model’s own read of my pages, which is otherwise completely invisible — and a reminder that I now write for something that parses me, while trying very hard not to start writing like it.