How a zero-authority domain got cited by ChatGPT in 6 weeks

The premise: buyers now ask an LLM what tool to use before they ask Google. If you don't exist in ChatGPT's answer, you don't exist. So the goal isn't just ranking — it's getting the answer engines to name you.

AEO sits on top of SEO. Answer Engine Optimization (getting cited by ChatGPT, Claude, Perplexity) isn't separate from Search Engine Optimization — it rides on it. LLMs run live web searches for anything current, then read the top results. And they only pull a handful of sources into an answer. You're in the top few, or you're invisible. That's brutal for a new domain: incumbents with years of authority start way ahead, and you cannot outspend them. So don't try. You grind fresh, specific, honest content that LLMs prefer — and you make your site trivial for their crawlers to read.

Here's what we actually did, in order.

1. Target the questions LLMs actually get asked

Not vanity keywords — buying questions. The highest-intent, most-cited queries in a tool category are comparisons and shortlists: "X vs Y", "best X for Z", "X alternatives", "top orchestrators 2026." When someone asks an LLM "what should I use," it reaches for exactly those pages. Map the real questions in your category and write the page that answers each one.

2. Write honest head-to-head comparison pages

We published 11 "vs" pages — defract vs each real alternative — and the rule was ruthless honesty: say where the competitor is better, name the cases where you're not the pick. Fair comparisons get cited; puff pieces get ignored (and an LLM can smell the difference as well as a developer can). Accuracy is the moat here, not word count.

3. Publish fresh, specific guides — consistently

We shipped 22 blog posts in six weeks: ~60% practical guides and shortlists ("spec-driven development, how to", "best Claude Code orchestrators"), ~40% opinion pieces that stake a position. LLMs favor content that is fresh, specific, and structured — dated, well-headed, answering one question cleanly. Volume matters only because it compounds; a thin post helps nobody.

4. Write it with AI — but keep a human loop that guards the truth

This is how you produce comparison pages and guides at real volume without shipping slop. The AI does the drafting; a human does the judgment. Skip the loop and you get generic prose the exact readers you want can smell in a sentence.

Source from facts, never memory

Ground every piece in primary material — your own product, real docs, competitor pages you've actually read, your own analytics. The model drafts from the facts you give it, not from what it half-remembers. A hallucinated spec or an unfair competitor claim doesn't just embarrass you; it gets you cited wrong, or not at all.

Make it sound like a person

Generic AI prose is instantly recognizable, and it repels technical readers. Feed the model samples of your real writing and give it hard voice rules — ours: brand always lowercase, verb-first, concise, no hype, no emoji, write like an engineer who ships. Anything that reads like a brand account gets cut.

Revise like an editor, not a spectator

Treat the first draft as raw material. A human reads every line, verifies every claim, removes the padding, and sharpens the argument. The model gives you a running start on volume; you supply the taste and the accountability.

Nothing ships without human approval

Every fact checkable, every comparison fair, every number real — signed off by a person before it goes live. One inaccurate post costs more trust than ten good ones earn, and trust is the whole moat.

5. Make your site machine-readable for AI crawlers

This is the part almost nobody does — and it's the difference between being crawled and being understood.

• llms.txt + AGENTS.md — a plain-language map of the site written for machines, so an AI crawler knows what you are and what each page covers.
• Markdown mirrors — a clean .md version of key pages (homepage, sitemap, glossary) with no layout noise for the model to trip over.
• Content negotiation — serve markdown to clients that ask for Accept: text/markdown (the crawlers), HTML to browsers, with Vary: Accept at the edge so caches don't cross the wires.

We ran our pages through an AI-readability audit and moved the score from 63 to 100. The payoff: when an LLM does cite you, it describes you accurately, because it could actually parse you.

6. Nail the boring technical foundation

None of the above matters if crawlers can't render or trust the page.

• Pre-render your SPA. Our marketing site was client-rendered React — which crawlers index unreliably. Pre-rendering it to static HTML was the single highest-leverage fix.
• Lighthouse SEO 100 / 100, CLS 0 (zero layout shift), first paint 1.4s. Speed and stability are ranking inputs and reduce bounce.
• Shrink everything. We cut page weight ~90% (megabytes to kilobytes) — WebP images, woff2 fonts, no render-blocking scripts.
• Structured data (JSON-LD), clean canonicals, a real sitemap, robots, and security headers. Table stakes, but skip them and you leak authority.

7. Optimize for Bing — it's the back door to ChatGPT

The one almost nobody tells you: ChatGPT Search doesn't read Google — it reads Bing. Analyses of ChatGPT citations find ~87% track Bing's top results. Your Google rank has no direct effect on whether ChatGPT cites you; your Bing rank is what decides it.

So everyone sets up Google Search Console and stops there. Do the Bing side too — it's where the AI answers actually come from:

• Bing Webmaster Tools. Verify your site (there's a one-click import from Google Search Console) and submit your sitemap. It's the only place you can see your Bing rankings — which, now you know where ChatGPT reads, is the closest thing to a ChatGPT-visibility dashboard that exists.
• IndexNow. A Microsoft-backed protocol that pings Bing the instant you publish, so a new page lands in the index in seconds instead of waiting days for a crawl — and therefore reaches ChatGPT's answers that much sooner. Drop a key file at your site root and ping the endpoint on every deploy. Google doesn't participate, so it costs your Google setup nothing.
• Bing rewards exact matches. It weighs literal keywords in your title, headings, and URL more heavily than Google's semantic approach — so the keyword-bearing URLs and titles from step 1 pay off even more here.

8. Measure what actually moves — including the AI referrals

You can't tune what you can't see. Two signals told us it was working:

• AI referrals. ChatGPT auto-tags links it surfaces with utm_source=chatgpt.com; other engines pass a referrer. That's how we know ChatGPT and Claude now send readers to our posts — first-party proof the content entered the citation set.
• Non-branded impressions. At the start, ~all search impressions were people typing our name. Now non-branded impressions outnumber branded — people find us for the topic, not just the brand. That's the flywheel starting to turn.

9. Be honest — it's the whole moat

Every claim on every page has to survive a skeptic. LLMs and developers both punish marketing fluff, and inaccurate content gets you cited wrong or not at all. Fair comparisons, real numbers, admitted trade-offs — that's what earns the citation and the click. Credibility compounds; hype decays.

The honest part

It feels slow. For weeks the graph is a flat line and you wonder if any of it lands. Then the compounding shows up — a post starts ranking, an LLM starts citing it, a stranger installs from a blog page they found by asking a question. If you're a founder staring at zero: it is possible. Don't rush to paid ads or growth hacks. Grind the best practices into an honest organic strategy and let time do the compounding.

We built this while building defract — a desktop app that runs your whole development process on parallel Claude Code agents (story → design → review → ship). The playbook above is free; the tool is too.