How to get your website cited by ChatGPT, Google AI Mode, Perplexity and Copilot, without tricks that backfire

Take a question someone types into ChatGPT or Google's AI Mode. The model does not pick the pages it links to from memory. It runs a search first, often several searches at once, which Google calls query fan-out, then it reads the pages that came back and writes an answer with links to some of them. So a citation is the last step of a search, and if your page never comes back from that search, nothing you write on it matters.

Each engine searches its own index. Google's AI Overviews and AI Mode use Google Search's index. ChatGPT search uses its own crawler plus outside search providers, and Microsoft's Web IQ documentation from 23 October 2026 says its service, built on Microsoft's own web index, powers both Copilot and ChatGPT. Perplexity has its own crawler, and Claude has one for search too. A University of Toronto study from 10 October 2026 looked at hundreds of thousands of AI citations. When the researchers replayed questions through ChatGPT, every page it cited had first come back from ChatGPT's own search.

So the job has two halves. First, be findable by each engine's search, which is mostly settings. Second, be the page worth quoting once you are found, which is mostly writing.

Before anything else, make sure you are not blocking the search bots by accident. An old security plugin, a "block AI bots" button at your CDN or a robots.txt copied from someone else can shut out a search crawler along with the training ones. Open yoursite.com/robots.txt right now and read it, then check your host or CDN's bot settings too, because a firewall rule that blocks a crawler beats any line in robots.txt.

Now the part people mix up. The companies that train models keep search and training apart. OpenAI has OAI-SearchBot for search and GPTBot for training. Anthropic has Claude-SearchBot for search and ClaudeBot for training. Google has no separate training crawler, and uses a robots.txt token called Google-Extended for training and the Gemini apps, which never decides whether you show up in Search. Perplexity says PerplexityBot is not used for training at all. So you can say yes to being cited and no to being trained on, and blocking GPTBot does not remove you from ChatGPT search.

The builder below writes a robots.txt with both choices made clearly. If your current file already has rules under User-agent: *, copy those lines into each named group too, because a bot with its own group ignores the * group completely. We ran its output through Python's robots.txt parser and checked each crawler against it, and the search bots were allowed while the training bots were blocked exactly as chosen. Bots that fetch a page because a user asked, like ChatGPT-User and Perplexity-User, are documented as generally ignoring robots.txt, so do not count on it to stop those.

Google added a switch in 2026 that many site owners have never seen. In Search Console there is now a setting called Search generative AI, and Google rolled it out to every site in the world on 31 August 2026. It decides whether your site can be linked and used in AI Overviews, AI Mode and the AI features in Discover. Every site is included by default, and a smaller Search Console property, like one set up for your blog folder, follows the setting of the property above it. So you only need to check that nobody on your team excluded the domain or one of those smaller properties. Google says excluding a site does not affect your normal rankings, but an excluded site gets no AI traffic or impressions at all.

Then check two meta tags, because they quietly do the same job. A nosnippet tag on a page stops Google using it as a direct input for AI Overviews and AI Mode, and max-snippet limits how much of the text they can use. Some themes add these for unrelated reasons, so search your page source for the word snippet.

While you are in the webmaster tools, open Bing Webmaster Tools too, add your site and submit your sitemap. Bing's index feeds Copilot and, per Microsoft, ChatGPT, and IndexNow lets you tell Bing about a changed page the moment you publish it. Google does not use IndexNow, so Google still comes from your sitemap and normal crawling.

Now we come to the writing half of the job. Google's own guide for generative AI search came out in spring 2026, and it says the thing that works is non-commodity content. That means first-hand experience and a point of view of your own, and not a remix of what is already online. Its example of commodity content is a list like seven tips for first-time homebuyers, which a thousand sites already have. Bing says the same thing in practical words: state facts directly, put the key information near the top and keep one topic per page, and its AI Performance launch post adds evidence like examples, data and cited sources.

So take any page you want cited and check it like this. Does the first paragraph answer the question in a sentence someone could quote word for word? Does it use the words people actually type, since the Toronto study found that matching the question was the one page feature that went with more citations within the same site? Does it show a real number, a date, a test you ran or a source you link to? Does it say when it was last updated? That is the whole formula, and it is the same thing that makes a page good for people.

The 2024 study that coined the term GEO found that adding real quotations and statistics raised visibility in AI answers by up to about 40 percent. The biggest gains went to pages ranked lower down in the search results. Read that as a reason to add your own data and named sources, not as a reason to invent numbers.

There is a whole industry selling AI visibility tricks right now, and most of it either does nothing or gets you punished. Google says plainly that Search does not use llms.txt files and that you do not need special schema for AI features. It also says you do not need to chop your pages into chunks, and that no outside tool can see inside its AI systems. An llms.txt file will not hurt you, and Google's John Mueller has said it can be useful to an agent that is already on your site, but do not expect a citation from it.

Then there are the tricks that are actually dangerous, and October 2026 gave us two clear examples. In a deliberate experiment, a marketer built about 50,000 pages of made-up statistics in one day for roughly 50 dollars. ChatGPT's bots fetched those pages tens of thousands of times and the fake numbers turned up in published research, until he wrote up the experiment himself on 7 October and marked the data as false. The same month, the data-deletion company Redact found hidden instructions for chatbots inside WhatsApp's help pages, including one telling the bot not to point people to data protection regulators. WhatsApp called them notes for its support bot shown by a bug, and they were gone by 18 October, the day before Redact published.

Both are against the rules now. Google's spam policies now cover attempts to manipulate generative AI responses, on top of the old rules against hidden text and mass-made pages. Bing's guidelines have a section on prompt injection, and breaking it can get a site demoted or delisted. Fake data and hidden prompts get found, and the trust you lose is the exact thing that made you citable.

You cannot improve what you cannot see, and in 2026 you can finally see some of it. In Google Search Console, the Generative AI performance report shows how many times links to your site appeared in AI Overviews and AI Mode. It counts impressions only, with no clicks, and it cannot split AI Overviews from AI Mode. Clicks from those features are still counted, but they are mixed into the normal Performance report under the Web search type.

Bing Webmaster Tools has the most useful report of all right now. Its AI Performance section shows how often Copilot and Bing's AI answers cited your pages and which pages they were. It also shows a sample of the search phrases the AI used when it cited you, and since June your citation share compared with other sites.

For ChatGPT, Perplexity and Claude there is no dashboard, so test by hand once a month. Ask each one the five questions your best pages answer, write down whether you were cited and who was cited instead, then read those pages and work out what they say more clearly than you do. Keep the list in a sheet so you can see the trend.