AI visibility: whether your site shows up in answers

A growing share of questions no longer ends in a list of results but in a finished answer — in ChatGPT, Perplexity, Copilot or Google's AI Overviews. These systems build their answer from passages they found on the web and name their sources. Whether your site is one of those sources depends on four things: whether the crawlers are let in, whether the content is in the HTML the server delivers, whether it is written in passages that can be quoted, and whether it is clear who stands behind it.

Isn't this just SEO?

Only partly. Ahrefs put 15,000 questions to Google and to AI assistants at the same time: only 12 percent of the addresses the assistants cited ranked in Google's top ten for the same question. For ChatGPT, Gemini and Copilot, around 80 percent of the cited addresses did not rank in Google for that question at all. Perplexity sits closer to Google — 28.6 percent of its citations were in the top ten.

Google's own AI Overviews are tied more closely to search: Google writes that a page has to be indexed and eligible to be shown with a snippet, and that there are no further requirements. How close the tie is keeps moving. Ahrefs found in July 2025 that 76 percent of the cited pages ranked in the top ten; in the 2026 repeat it was 38 percent — with the note that the measurement has improved and the two numbers are not directly comparable.

Two channels with a related but not identical logic. Good search engine optimisation is the foundation. It does not cover AI answers by itself.

Prerequisite 1: the content has to be there without JavaScript

This is the most important technical point, and the one most often missed. Vercel analysed the requests of the large AI crawlers across its network: none of them executes JavaScript — not OpenAI's crawlers, not Anthropic's, not Perplexity's, not Meta's. Some download script files, but none runs them. The exceptions are Gemini, which uses Google's crawling infrastructure, and Apple's crawler.

A page whose text only appears once a script has run is therefore nearly empty for these crawlers — however good it looks in the browser. That also applies to:

• text in accordions and tabs that is loaded on click. Content that is in the HTML and only visually collapsed is fine.
• content loaded while scrolling.
• reviews, prices and availability pulled in from a separate service.

The test takes thirty seconds: switch JavaScript off in your browser and reload your most important page. What is still there is roughly what an AI crawler sees. Or open the page source with Ctrl+U and search for a sentence from your main text with Ctrl+F. If it is not there, it is not in the HTML the server delivers.

Prerequisite 2: the right crawlers have to be let in

The large vendors now separate their crawlers cleanly by purpose — and this is where most settings go wrong:

Vendor	Training	Search	Fetch on a user's request
OpenAI	GPTBot	OAI-SearchBot	ChatGPT-User
Anthropic	ClaudeBot	Claude-SearchBot	Claude-User
Perplexity	—	PerplexityBot	Perplexity-User

To appear in ChatGPT's answers you need OAI-SearchBot, not GPTBot. OpenAI states explicitly that sites blocking OAI-SearchBot will not appear in ChatGPT's search answers. GPTBot, on the other hand, collects material for training the models. Anthropic works the same way: ClaudeBot trains, Claude-SearchBot builds the search index. So you can block training and allow search — every crawler is addressed by its own name in robots.txt.

Google has no separate door. AI Overviews depend on the regular Googlebot. Blocking Googlebot removes a site from search and from AI Overviews at once. Google-Extended only controls whether content is used for training and answers in Google's other systems.

The most common trap is not in robots.txt but in front of it. Since July 2025, Cloudflare — which by its own account serves around a fifth of the web — blocks AI crawlers by default for newly added domains. robots.txt can allow everything while the service in front turns the request away. Reading only robots.txt, you never see it.

A decision that is yours: whether to let AI crawlers in is not purely a technical question. If you do not want your content in other people's answers, you can block them. The price is not appearing in this channel. For most businesses whose content is public anyway and who want enquiries, allowing the search crawlers is the obvious choice — but it should be made deliberately.

Writing passages that can be quoted

Once the technology is right, the text decides. A passage gets quoted when it can stand on its own: you can lift it out of the page and it still makes sense — without the paragraph before it, without "as described above".

What demonstrably helps was studied by a research group around Princeton University and presented at the KDD conference in 2024. They changed texts deliberately and measured how visible they were afterwards in generated answers. Three things worked best: citing sources, adding quotations and adding numbers — visibility rose by up to around 40 percent. Stringing keywords together did nothing.

What does not count: length. Ahrefs analysed 174,048 pages cited in AI Overviews. More than half of them — 53.4 percent — have fewer than 1,000 words, and there is practically no relation between length and being cited. In this field, density beats length.

The structure that works for both channels

The good news: the structure AI systems need is the same one that helps readers in a hurry. You do not build two versions.

• The main heading names the question or the topic the way someone would phrase it.
• The answer comes first. The first paragraphs deliver what the heading promises — without a run-up. That reverses what many are used to.
• Every subheading is a sub-question, answered first in a short, direct passage, then explained.
• Lists and tables form clearly bounded units that are easy to lift out.

This article is built that way itself.

Who stands behind it

Google's guidance on helpful, reliable content asks explicitly: who wrote this, how was it made, and why? For readers and for systems building answers from other people's texts alike, a source is easier to place when the page answers those questions. In practice:

• an author's name — a person, not "the editors",
• publication and update dates, visible on the page,
• a real imprint with a postal address,
• evidence for facts — where there is a number, its source is next to it,
• a recognisable body of work — several connected articles on a subject rather than a single one.

For a sole trader this is an advantage over anonymous advice sites: a name, a face, details anyone can check.

How to measure whether it works

To be honest: there is no Search Console for AI answers. There are two workable substitutes.

The question list. Write down 20 to 30 questions your customers would actually ask — full questions, not keywords. Put them to ChatGPT and Perplexity once a month and note who gets named. It is manual work, but it is the only measurement that means something for your business. On the side you see which competitors are named, and for what.

Visitors from AI services in your analytics. In Google Analytics, referrals from chatgpt.com, perplexity.ai and similar addresses can be grouped into a channel of their own. Do not expect large numbers: studies currently measure between about 0.3 percent (Ahrefs, an ongoing analysis of around 112,000 websites) and about 1 percent (Conductor, over 13,000 domains) of all visits. The channel is growing — and whoever arrives from it brings a question that has already been answered.

What to do

1. Run the JavaScript test. If your most important page is empty without scripts, nothing else matters. On a cleanly built WordPress site with content rendered on the server, that is the normal case — with heavily script-driven themes and site builders it is not.
2. Check robots.txt and the service in front of it. Both. And keep training and search apart.
3. Rebuild your most important pages: the question as the heading, the answer first, subheadings as sub-questions.
4. Add evidence: numbers, sources, quotations. That is the lever with the measured effect.
5. Show who writes — name, date, imprint.
6. Start the question list and run it once a month.

And one expectation to set straight: AI visibility does not replace being found in search engines, and for most businesses it is currently the smaller channel. But the effort is small, because almost all of it is good practice anyway. A text with a clear structure, sourced statements and a recognisable author is better for people too.

Sources

• Ahrefs, Only 12% of AI Cited URLs Rank in Google's Top 10 for the Original Prompt — 15,000 questions, overlap per assistant: ahrefs.com
• Google Search Central, AI features and your website — no special requirements, indexing and snippets, Googlebot and Google-Extended: developers.google.com
• Ahrefs, July 2025 study of rankings and citations in AI Overviews (76 percent): ahrefs.com
• Ahrefs, 2026 repeat (38 percent, 863,000 keywords): ahrefs.com
• Vercel and MERJ, The rise of the AI crawler — no JavaScript execution: vercel.com
• OpenAI, Overview of OpenAI Crawlers — OAI-SearchBot, GPTBot, ChatGPT-User: developers.openai.com
• Anthropic, Help Center on ClaudeBot, Claude-SearchBot and Claude-User: support.claude.com
• Perplexity, Perplexity Crawlers — PerplexityBot and Perplexity-User: docs.perplexity.ai
• Cloudflare, press release of 1 July 2025 — AI crawlers blocked by default: cloudflare.com
• Google Search Central, Creating helpful, reliable, people-first content — Who, How and Why: developers.google.com
• Aggarwal et al., GEO: Generative Engine Optimization, KDD 2024 — sources, quotations and numbers as the strongest methods: arxiv.org
• Ahrefs, Short vs. Long Content in AI Overviews — 174,048 pages, 53.4 percent under 1,000 words: ahrefs.com
• Ahrefs, ongoing analysis of referrals from AI services and Google: chatgpt-vs-google.com
• SE Ranking, study of the share of AI referrals in website traffic: seranking.com
• Digiday on the Conductor analysis — about 1 percent of visits from AI referrals: digiday.com