Your next customer may never see a list of blue links. They'll ask ChatGPT which agency to hire, which software to buy, or which organization deserves their money — and the answer will either name you or it won't.
This practice has collected a pile of names: AEO (answer engine optimization), GEO (generative engine optimization), LLM SEO. They all mean the same thing — making sure AI systems can find your content, read it, and cite it.
The traffic is still small, but it behaves differently. AI referrals are only about 1% of web traffic, growing roughly a percentage point a month. Semrush found AI visitors convert about 4.4x better than organic search visitors, and Ahrefs saw AI search drive 12% of signups from 0.5% of visits. Someone who arrives from an AI answer has already been briefed on you. They show up ready to act.
Meanwhile, the old channel is shrinking. AI Overviews now appear on roughly a quarter of Google searches and cut clicks to the top result by a third or more.
Here's what actually moves the needle, and what you can skip.
How AI search actually reads the web
Four facts explain most of the playbook:
- AI crawlers don't run JavaScript. Vercel analyzed the major AI crawlers and found that none of them — GPTBot, ClaudeBot, PerplexityBot — execute JavaScript. If your content only appears after scripts run, it doesn't exist to them.
- ChatGPT search leans on Bing. Seer Interactive found 87% of ChatGPT search citations match Bing's top organic results. Bing, long ignored, suddenly matters.
- AI answers are built from passages, not pages. Google's AI Mode splits one question into many sub-queries and retrieves the best individual passages. Each section of your page competes on its own.
- Citations follow consensus, not links. Ahrefs found branded web mentions correlate with AI visibility three times more strongly than backlinks. Models recommend what many sources agree on.
1. Make sure you're not blocking the bots
Each AI company runs separate crawlers for training, search indexing, and live user requests. OpenAI has GPTBot (training), OAI-SearchBot (search), and ChatGPT-User (live fetches). Anthropic mirrors this with ClaudeBot, Claude-SearchBot, and Claude-User. You can block training while staying visible in search:
User-agent: OAI-SearchBot
Allow: /
User-agent: ChatGPT-User
Allow: /
User-agent: GPTBot
Disallow: /
The catch: robots.txt isn't the only gate. Since July 2025, Cloudflare blocks AI crawlers by default on new domains, and many firewalls and bot-protection tools do the same. Plenty of sites are invisible to AI without anyone having decided that. Check your CDN and security settings, not just your robots.txt.
We checked ours while writing this. loopdash.com allows every crawler. Now you know why.
2. Put your content in the HTML
Because AI crawlers don't execute JavaScript, anything rendered client-side — single-page apps, content loaded after the fact — is invisible to them.
The test takes ten seconds: view your page's source (not the inspector — the raw source) and search for a sentence from the page. If it's not there, AI can't read it. Server-side rendering or static generation fixes this. Most WordPress sites pass by default. Many modern JavaScript builds don't.
3. Show up in Bing
If ChatGPT's citations track Bing's index, you need to be in it. Set up Bing Webmaster Tools, submit your sitemap, and fix whatever it flags. It's an afternoon of work most organizations have never done, because for twenty years Bing didn't matter.
4. Write passages an AI wants to lift
The original GEO research from Princeton tested nine tactics across 10,000 queries. Three won consistently: citing sources, adding statistics, and including quotations — each boosted visibility in AI answers by 30–40%. Keyword stuffing performed worse than doing nothing.
In practice:
- Answer first. Put the direct answer in the first sentence under a descriptive heading, then elaborate.
- Make each section self-contained. Retrieval grabs chunks, not pages. A section that only makes sense in context won't get cited.
- Use real numbers and name your sources. Claims with evidence attached are the ones models repeat.
- Keep it current. Ahrefs found ChatGPT-cited pages had been updated within 30 days 76% of the time. Refreshing your best pages beats writing new ones.
One more finding worth knowing if you're a small business or organization: in the Princeton study, adding citations lifted visibility for sites ranked fifth by 115% — while the top-ranked site lost ground. Generative engines flatten the winner-take-all dynamic of Google. Underdogs get a real shot.
5. Be mentioned where the models read
Models cite what the web agrees on. ChatGPT's most-cited domain is Wikipedia. Perplexity and Google's AI Overviews lean heavily on Reddit. YouTube mentions correlate with AI visibility more strongly than almost anything else Ahrefs measured. And ranked lists dominate: studies of hundreds of millions of citations found listicles account for 40–65% of the most-cited URLs.
So the work looks less like link building and more like reputation building: get included in the "best of" roundups for your niche, keep your Wikipedia presence accurate if you have one, show up in the Reddit and forum threads where your community actually talks, and publish the comparison content in your space before someone else defines it.
Semrush's study of 50,000 brands found AI visibility is won topic by topic, not by domain authority — only 15% of categories had a clear owner. Pick the topics you can genuinely own and go deep.
What this looks like in practice
We recently ran this playbook for a client whose site had strong data but no AI visibility. Three moves did most of the work.
We started with questions, not keywords. Using DataForSEO, we pulled the actual queries people ask around the client's topic — long, specific, conversational — and looked at which ones AI engines were answering without a good source to cite. Those gaps became the content plan. AI prompts run far longer than searches, so the opportunity lives in the long tail traditional keyword research skips.
We built calculators. A calculator answers a question with a concrete, personalized number — exactly what a model wants to hand a user. It's the same logic behind the Princeton finding that statistics win citations, packaged as a tool: instead of one statistic, you're the source for every variation of the question. One caveat from earlier applies — the calculator page needs its purpose, inputs, and methodology described in plain HTML, or the crawlers can't tell what the JavaScript does.
We gave the charts a markdown twin. The client's best pages were chart-heavy — and charts render with JavaScript, which means AI crawlers see none of the data. So each data-heavy page now links to a plain markdown file containing the tables behind the charts. Humans get the visuals; crawlers get a clean, token-cheap version of the numbers they can actually quote. The site went from invisible to cited.
What you can skip
llms.txt. The proposed standard — a markdown index of your site for AI — has real enthusiasm behind it and almost no usage. Ahrefs checked 137,000 sites and found 97% of llms.txt files received zero requests. Google's John Mueller compared it to the keywords meta tag. It's harmless and takes five minutes, but don't expect it to do anything yet.
A schema markup project. Structured data helps the search indexes that feed AI answers (especially Bing), and basic Organization and FAQ markup is worth keeping. But the evidence that schema itself drives AI citations is correlational at best. Don't buy a five-figure schema engagement on AI grounds.
A "GEO agency" retainer. The uncomfortable truth practitioners keep landing on after months of testing: most of this is SEO fundamentals with a new name. Clear writing, fast pages, real authority, technical hygiene. If someone's pitch depends on AI search being magic and different, be skeptical.
Get ready for agents, not just answers
The next step past AI search is AI that does things — agents that fill out forms, complete signups, and finish purchases on a user's behalf. OpenAI already runs agent-completed checkout inside ChatGPT.
What makes a site agent-friendly turns out to be exactly what makes it accessible: semantic HTML with real buttons, labeled form fields, no forced account creation, no CAPTCHA walls in critical flows, layouts that don't shift as they load. If your checkout, signup, or donation form works for a screen reader, it'll work for an agent. Accessibility work and agent-readiness are the same work — one more reason to do it.
Final thoughts
None of this requires abandoning SEO. It requires noticing that a second audience is now reading your site — one that doesn't run JavaScript, pulls from Bing, cites passages instead of pages, and rewards evidence over keywords.
The organizations that win AI citations early will be hard to displace, because models keep recommending what's already part of the consensus. The window where this is easy is now.
At Loopdash, we build AI-ready websites and handle the SEO fundamentals underneath all of this. If you want to know whether AI can actually see your site, ask us — that audit is quick, and the answer is often surprising.
