Is your website readable by AI? A field guide to getting cited by ChatGPT, Perplexity and Google AI Overviews

What does “readable by AI” actually mean?

It means a machine can fetch your page, find your words as text, and reuse them with attribution. Answer engines like ChatGPT, Perplexity, Google AI Overviews, Gemini and Claude work by retrieval-augmented generation: they first retrieve documents relevant to a question, then write an answer that cites a small handful of the sources they retrieved. Increasingly they shop this way too — our guide on agentic commerce covers what “discover in AI, buy on your own site” means for a store. And once AI browsers can act on your pages, being readable is only half the job — is your website ready for AI agents? covers whether an agent can actually complete a task once it arrives. If a retrieval crawler cannot reach your page, or reaches it but finds an empty shell that only fills with content after JavaScript executes, your page is not in the running — it is invisible at the exact moment a buyer is asking your question.

The stakes moved quickly. Google AI Overviews now appear in roughly half of all searches, and by some counts up to 60% (Geoptie, 2026). ChatGPT passed 800 million weekly users, and Perplexity handles around 780 million queries a month (Geoptie, 2026). Meanwhile the analyst benchmark everyone cites — Gartner’s February 2024 projection — expects traditional search volume to fall about 25% by 2026 as people move questions to chatbots (via Mersel, 2026). For a decade, “position zero” meant a featured snippet; today it means a citation inside an answer, which is harder to earn and worth more to hold (Leapd, 2026).

This is why we describe an AI assistant as your most important visitor. It does not browse, scroll or admire a hero animation. It reads. So the question is not whether your site looks impressive to a person — it is whether the same page is legible to a reader that only consumes structured text.

How do answer engines choose what to cite?

They retrieve, then attribute the sources that are clearest, most specific, and most credible — and each engine weighs those things differently. An analysis of 680 million citations found that only 11% of domains were cited by both ChatGPT and Perplexity; the platforms run on largely separate logic (Leapd, 2026). A separate study of 34,234 AI responses measured a 46-fold gap in how often platforms cite brands at all — ChatGPT around 0.59%, Perplexity 13.05%, Grok 27% (Leapd, 2026). Optimizing for “AI search” as one thing is like running one campaign across two unrelated networks.

The differences are practical. ChatGPT favours authoritative long-form content and historical tier-one sources, and sources heavily from Bing’s top results. Perplexity favours fresh, well-cited articles and structured community content. Google AI Overviews favour pages already ranking in the organic top ten, with schema and domain authority (Frase, 2026). That top-ten tie has loosened sharply in 2026, though — our guide on how to appear in Google AI Overviews has the current picture. And one source type dominates across all of them: community platforms — Reddit, Quora, forums — captured 52.5% of citations in one analysis of a million AI citations, with Reddit the single most-cited domain (Geoptie, 2026).

Underneath the platform differences, the peer-reviewed research that named this field measured which content changes actually lift visibility. Across 10,000 queries in 25 domains, the largest gains came from adding quotations (+41%), statistics (+32%), citations (+30%) and clearer, more fluent writing (+28%) (via Mersel, 2026). None of those are tricks. They are what good, trustworthy writing already does.

Why can’t many modern sites be read by AI?

Because their content is delivered as something a crawler cannot read: JavaScript that never runs, or pixels on a canvas. A single-page application sends an almost empty HTML file and then paints the content with JavaScript in the browser. A person sees a finished page; a retrieval crawler that does not execute that JavaScript sees a blank one. The same is true of sites built inside a WebGL canvas — visually rich, and, to a machine, a sealed box. Some AI website builders ship exactly this pattern by default, which is the quiet irony of the moment: tools sold as the future of the web often produce pages the AI web cannot read. We take one of them apart in detail in is Lovable good for SEO? — including why an AI-built site can rank on Google and still be invisible to ChatGPT. Translation tools hit the same gap from another angle: is Weglot good for SEO? shows why JavaScript-injected translations vanish from AI search — and why you may be renting them.

This is the gap we build against. A site assembled from a template or generated by a builder tends to look acceptable and read poorly to a machine; the craft that makes a page legible — semantic structure, real text, answer-first order — is precisely what those shortcuts skip. Rendering your content as static HTML, where every word is present in the page source before any script runs, is the single most important technical decision for AI readability. It is also the decision that makes a site fast and accessible, which is why we treat it as one rule, not three.

What makes a page citable? The checklist that matters

A citable page answers the question in its first lines, backs claims with specific numbers, is structured so a machine can parse it, and is served as crawlable text. Researchers who reverse-engineered AEO signals found that answer engines weight FAQ schema quality at about 20%, answer-first formatting at 19%, and statistical density at 16% — ahead of the backlink and keyword signals that dominated classic SEO (AuthorityTech, 2026). In practice, that means a short, useful checklist:

• Answer first. Open each section with the answer in one or two sentences, then explain. Engines extract the lead. Our guide on how to structure content so AI will cite it turns this into a full method.
• Be specific and cite. Include a concrete statistic with a source roughly every 150–200 words; verifiable data is what an engine borrows to sound credible (Frase, 2026).
• Structure for parsing. Use real headings as questions, lists where lists fit, and a visible FAQ. Ship FAQPage, Article and BreadcrumbList schema as JSON-LD so a machine knows what each block is.
• Render statically. Make sure your words are in the raw HTML, not injected later by JavaScript. If you can read it in view-source, so can a crawler.
• Stay crawlable. Do not block the bots that cite you. A page that cannot be fetched cannot be quoted (Geoptie, 2026). Our guide on whether you should block AI crawlers covers which bots to allow and which to block, and its companion on whether you should charge AI crawlers covers the newer pay-per-crawl option — and why most businesses shouldn’t use it.
• Show who wrote it. First-hand experience and clear authorship feed the credibility signals engines lean on.

Read that list again and notice something: there is nothing in it that a careful, well-built website does not already do. The reason most sites are not cited is not that AEO is hard — it is that most sites were not built with this care in the first place. For how AEO relates to SEO — an extension of it, not a replacement — and how to measure it, see our guide on AEO vs SEO.

Is GEO different from SEO, or the same thing with a new name?

It is different in target and overlapping in foundation. SEO optimizes for a ranked position in a list of blue links; GEO and AEO optimize for a citation inside a written answer. The strongest performers in AI search almost always have strong SEO foundations, because both depend on crawlable, fast, well-structured pages — but the execution differs, and chasing one does not automatically win the other.

	Classic SEO	GEO / AEO
Goal	Rank in a list of links	Be cited inside an answer
Heaviest signals	Backlinks, keywords, domain authority	Answer-first writing, citable stats, FAQ schema
What the visitor is	A person who clicks	A machine that reads and quotes
Where it shows	The ten blue links	The generated answer above them
Shared foundation	Crawlable, fast, semantic, accessible HTML	Crawlable, fast, semantic, accessible HTML

The bottom row is the point. You do not choose between being found by Google and being read by an assistant; one well-built page earns both. This is the connective view we build to: search legibility, performance, accessibility and AI readability are not four projects — they are one page, made well.

What is llms.txt, and do you need one?

It is a short, plain-text file at your domain root that gives AI systems a clean map of your most important pages and facts — and it is worth publishing. Think of it as a sibling of robots.txt: where robots.txt tells crawlers where they may go, llms.txt tells AI systems what matters and how to summarize you. It is an emerging convention, not a guaranteed ranking factor, but it costs minutes and signals a site built to be understood. A minimal version looks like this:

# Your Company
> One-paragraph summary of who you are and what you do, in plain language.

## Key pages
- Pillar guide: /guides/your-best-guide/
- Pricing: /pricing/

## Facts
- Founded: 2001
- What you sell: ...

Publish it, keep it short, keep it true, and update it when your key pages change. It is the cheapest legibility win available, and almost no one your competitors hire will have done it.

How do you prove a page is actually readable by AI?

You test it the way a crawler would, and you let the result be public. Open your page’s source and confirm your words are there before any script runs. Fetch it the way a bot does and confirm it returns content, not an empty shell. Validate your structured data. Check that the pages you care about are not blocked. Structural changes made this way tend to surface fast — within two to seven days on Perplexity and roughly one to three weeks on ChatGPT, with Google AI Overviews and Claude slower (AuthorityTech, 2026). Most sites see measurable movement within four to eight weeks of getting the structure right (Mersel, 2026).

This is the discipline we hold for our own site, in public — the home page audits itself in front of you, and every figure links to the report that produced it. A studio that tells you to be legible to machines should be the most legible site you read that day.

Is beautiful, animated design bad for AI readability?

No — beautiful design does not hurt AI readability; hiding your content inside the design does. The problem is never art direction in the abstract, it is where the meaning lives. If your headline, your argument and your evidence are real text in the HTML, an answer engine reads them whether or not the page also carries a cinematic hero. If those same words are painted on a WebGL canvas, baked into an image, or injected by a script a crawler never runs, the machine reads nothing — however striking the result.

This is the quiet cost at the award-winning end of the web. The most spectacular sites often live inside a canvas, which to a person is a world and to a retrieval crawler is a sealed box. The research that ranks citation signals is blunt about the fix: structure and substance — answer-first text, citable data, clean markup — outweigh decoration, with FAQ schema, answer-first formatting and statistical density carrying the most measured weight (AuthorityTech, 2026). Decoration a machine cannot parse is not a tie-breaker; it is simply absent.

We build to keep both. The spectacle goes in CSS and SVG — motion the browser composes from text and vector, not from an opaque scene — and the substance stays in semantic HTML. The result reads as design to a person and as a quotable source to an assistant, with no second, machine-only version to maintain. You should never have to choose between a site that moves people and a site machines can read; that trade-off is a symptom of how a site was built, not a law of the web.

Does AI readability stop at your own pages?

No — a large share of what answer engines cite never lives on your domain at all, so authority has to be built off the page too. In one analysis of a million AI citations, community platforms — Reddit, Quora, niche forums — accounted for 52.5% of citations across ChatGPT, Perplexity and Google AI Overviews combined, with Reddit the single most-cited domain (Geoptie, 2026). When an engine answers a question in your field, it often reaches for a discussion thread or a first-hand account before a brand’s marketing page. Which sources an engine trusts enough to cite in the first place is decided by E-E-A-T — the subject of our guide on E-E-A-T for AI.

That has two consequences. First, the credibility signals engines lean on — genuine expertise, first-hand experience, consistent mentions across independent sources — are earned in public, not declared on an About page. Showing up usefully where your buyers already ask questions feeds the same retrieval index that decides whether you are cited. Second, the facts about you need to be consistent everywhere an engine might read them: your founding date, your services and your claims should match across your site, your profiles and third-party listings, because contradiction reads as uncertainty, and uncertainty does not get cited.

This is the part most “AI optimization” advice skips, because it cannot be bought as a plugin. It is slower and it is durable: a body of genuinely useful, consistent, first-hand material — on your site and in the places that get cited — compounds into the kind of authority an answer engine treats as a default source.

A ten-minute audit you can run yourself

You can gauge your own site’s AI readability in about ten minutes, with tools you already have and no budget. Work through six checks:

• Read the source. Open an important page, view its HTML source, and search for a sentence you can see on screen. If it is in the source, a crawler can read it; if it appears only after scripts run, an answer engine may see an empty page.
• Turn off JavaScript. Disable JS and reload. What remains is roughly what a non-rendering crawler gets. If the page is blank or skeletal, that is your most urgent fix.
• Validate your structured data. Run the page through a Rich Results or schema validator and confirm your Article, FAQPage and BreadcrumbList blocks parse cleanly. Invalid schema is schema an engine ignores. Our guide on schema markup and structured data covers which types move the needle and the one catch most sites miss.
• Check what you allow. Open robots.txt and confirm you are not blocking the crawlers that feed AI answers. A page that cannot be fetched cannot be cited (Geoptie, 2026).
• Publish and read your llms.txt. Confirm the file exists at your root and that its summary is current and true.
• Ask the engines. Put a question your page should answer to ChatGPT, Perplexity and Google, and see whether you are cited — and how you are described.

Then fix the highest-impact gaps first and re-check. Structural changes tend to surface on Perplexity within days and on ChatGPT within a few weeks (AuthorityTech, 2026), so this is a loop you run monthly, not a project you finish once. The sites that win AI visibility are the ones that measure it on a schedule — not the ones that guessed right one time. How to measure it — across the broken analytics of AI search — is its own discipline, covered in our guide on how to measure AI traffic.

Going deeper: the supporting guides

Three companion guides expand the pieces of this one. On the architecture that makes a page readable in the first place — static, zero-JavaScript HTML a crawler receives instantly — see is Astro good for SEO and AI search. On the file everyone debates, with the honest evidence on whether AI crawlers actually use it, see what is llms.txt, and does your site need one. And on where AI answers actually come from — and why one community platform dominates them — see why Reddit dominates AI citations. Together they cover how to be readable, what genuinely moves citations, and where those citations are sourced.

---

One last framing, because it is the whole idea in miniature. The mark in our name is an infinity, and AI readability is one loop within it: the same static, semantic, fast, accessible page that an assistant can quote is the page that loads instantly, opens for everyone, and ranks. You do not bolt “AI optimization” onto a finished site. You build the site so that being read — by a person, a search engine, or an assistant — was never in question.