How to get cited by ChatGPT and AI search in 2026
Nobody can guarantee an AI citation. You raise the odds with pages that answer questions directly, crawlers let in, and one consistent description of you.
To get your company cited by ChatGPT, Perplexity and Google’s AI Overviews, publish pages that answer specific questions in their opening lines, make sure the crawlers these products use can reach and read them, and describe your company the same way everywhere it appears. Nobody can guarantee a citation, and anyone who promises one is guessing. What you can do is make your site the easiest credible source to quote.
How well that works depends on the question. Answer engines draw on whatever they judge most relevant and trustworthy for a given prompt, the sources differ between products, and they change over time. Start with the narrow question your team can answer from direct experience. That gives readers a reason to trust the page and gives you something better to publish than another broad category overview.
1. How do AI search engines find your pages?
The products work differently and publish only part of how they work, so treat anything more precise than this outline with suspicion.
- Google’s AI Overviews and AI Mode draw on Google’s search index, so a page has to be indexed and eligible to show a snippet in ordinary results before it can appear as a link. The site also has to be included in Search Console’s Search generative AI setting, covered below.
- OpenAI documents OAI-SearchBot as its crawler for surfacing websites in ChatGPT search.
- Perplexity documents PerplexityBot as a crawler used to surface and link websites in search results.
- An answer produced without web retrieval is a different measurement from a live search citation. Record which mode you tested before comparing results.
Live answers are assembled from search results, so start with ordinary search hygiene. It makes the pages accessible, even though it cannot guarantee that an answer engine will choose them.
Your robots.txt and your CDN
Each AI company publishes the user agents its crawlers use. OpenAI separates GPTBot, which collects training data, from OAI-SearchBot, which serves search, so you can opt out of one without the other. Check each provider’s current documentation rather than assuming all bots serve the same purpose. Review robots.txt alongside your CDN and hosting firewall: permission in one place does not prove the request gets through everywhere.
Content that exists without JavaScript
Check what your server returns before assuming every crawler sees what a browser shows. If your core copy appears only after scripts run, inspect whether the systems you care about can retrieve it. Server-rendered or static HTML makes important text available without that dependency. Keep useful diagrams and video too, with a written explanation beside them.
Indexing in Google and Bing
Verify your site in Google Search Console and Bing Webmaster Tools, submit a sitemap to both, and investigate important pages they report as excluded. Not every exclusion is an error: redirects, duplicates and private pages may be excluded deliberately. A page that isn’t in the index can’t be cited from it.
Google’s Search generative AI setting
Search Console has a site-wide Search generative AI setting. Google announced it on 3 June 2026 and had rolled it out to all sites by 31 August 2026. It decides whether your pages can appear in AI Overviews, AI Mode and AI Overviews in Discover and help ground their answers, and Google’s guidance says a site has to be included to be eligible. Google says the setting isn’t used as a ranking signal for ordinary search results. It covers only Google’s features, so ChatGPT, Perplexity and other assistants are still managed through robots.txt and your CDN, as above.
2. What does an answer-first page look like?
A clear answer directly under its question is useful to a reader who arrives halfway down the page. It also gives a search system a passage that can be understood outside the surrounding article. Treat that as good editorial practice, not a published ranking formula.
- Put the question people ask in the heading, in their words rather than your product vocabulary.
- Answer it in the first two or three sentences, then give the conditions and exceptions.
- Give each page one coherent job. Keep related questions together where that helps the reader; do not split one useful guide into thin pages just to target more queries.
- State facts precisely: what the product does, who it’s for, where you operate, and what it costs if you publish prices.
- Keep pages current and dated, because a stale specification gets quoted as if it were true today.
Write the answer so it still makes sense outside the page around it.
For technical companies this is often the biggest gap. The homepage says something evocative, the documentation is precise but written for existing users, and nothing in between answers the question a buyer types into an assistant. It’s why Refract reads a site the way a buyer reads it before suggesting a single fix.
3. Do FAQ schema and llms.txt help?
Structured data
Schema.org markup describes content in a machine-readable form. Use the types that match the page, such as Organization, BlogPosting and BreadcrumbList, and make sure they agree with the visible text. Google retired FAQ rich results in May 2026. FAQPage remains a schema.org type, but it does not earn that Google search display or guarantee an AI citation. A useful question-and-answer section earns its place by helping the reader.
llms.txt
An llms.txt file can provide a short description of your site and links to useful pages. Google says it does not affect visibility or rankings in Google Search. Treat it as an optional directory for systems that choose to read it, and keep it accurate if you maintain one. It cannot replace crawl access, useful content or an accessible website. At Refract we maintain one as a directory; we do not sell it as a route to guaranteed citations.
4. Why does it matter what other sites say about you?
Look at the sources cited for the questions your buyers ask. They may include vendors, documentation, reviews, forums and industry publications. That sample shows where the answer came from in that session; it does not establish a universal formula for earning mentions.
So make the description of your company consistent everywhere it appears. Use the same name, the same one-line description and the same category on your site, LinkedIn, Crunchbase, GitHub, app marketplaces, review platforms and industry directories, and link them to each other. When descriptions disagree, it is harder for any system, or any person, to tell that they all refer to the same company.
Then earn mentions the slow way: useful writing that others link to, talks and podcasts, customer stories published by the customer, and accurate entries in the directories your buyers use. Choose places that are relevant to your buyers, and describe your work accurately. A directory full of unsupported superlatives repeats the problem on another domain.
5. What should you put in place first?
The six checks show what is missing. The order you fix things in matters too, because the later steps rely on the earlier ones. We’d work through them like this.
- Record a baseline. Write down the questions your buyers ask and save what each assistant answers today, with the date. Without it there is nothing to compare later answers against.
- Fix access. Allow the search crawlers in robots.txt and at the CDN, and include the site in Search Console’s Search generative AI setting. A page a crawler cannot fetch gains nothing from the steps below.
- Get the pages indexed. Submit the sitemap to Google and Bing, then resolve the exclusions that are mistakes.
- Settle one description of the company. Write it at two or three lengths, put it on the site and use the same wording on every profile you control.
- Rewrite the pages buyers land on. Start with the few that answer a buying question, such as what you do, who it is for and what it costs, and put the answer in the opening lines.
- Add structured data once the copy is settled, so the markup describes the page as it now reads.
- Add an llms.txt file if you want one. One website platform’s own documentation says its effect is unproven, so we’d leave it until the steps above are done.
- Work on what other sites say about you. It takes the longest, and a directory entry or a guest article sends readers and crawlers to pages that should already be in order.
Steps two and three are technical and need whoever runs your hosting. Step four needs a decision from the people who run the company, because it states what the company is and every later step repeats that statement. If your site already appears in ordinary search for the questions you care about, indexing is probably in place. Check crawler access for the assistants, then start at step four.
6. How do you know if it’s working?
Imperfectly, for now. There is no single ranking report that covers every assistant, and answers vary between users, sessions and days. Combine a few rough signals.
- Watch referral traffic in your analytics from domains such as chatgpt.com, perplexity.ai and copilot.microsoft.com.
- Keep a fixed list of the questions buyers ask, run them through each assistant every month, and record whether you are mentioned, cited and described accurately.
- Use the Generative AI performance report in Google Search Console. It shows impressions in Google’s AI features across Search and Discover, which of your pages appear, and in which countries.
- Ask new leads how they found you, and include AI assistants as an option on the form.
Accuracy matters as much as frequency. If an assistant describes your product wrongly, inspect the sources it cites and the pages you control. Correct outdated information where you find it, but do not assume that every error has a visible source or that one edit will immediately change the answer.
Checking without fooling yourself
A prompt check is easy to bend towards the answer you hoped for. These rules keep it honest.
- Write the questions before you run them and keep the wording the same every month. Add new questions at the end and leave the old ones alone, or the months stop comparing.
- Use a new signed-out or temporary chat for each question, and do not mention your company first.
- Note the product, the date and whether web search was on.
- Record every source the answer cites as well as whether you were named. The sources show which pages and sites the answer drew on that day.
- Include questions you may not appear for. A list made of your brand name and your best page will look healthy every month and tell you very little.
Treat each result as an observation from one session on one day. A difference between two months could come from the assistant, from the question or from your site, and a sample of prompts cannot tell you which. Keep referral visits and qualified enquiries in separate columns, compare like periods, and avoid crediting any movement to a single edit.
7. What we did on our own site
We did this work on refract.group in November 2026, and it started with a naming problem. Refract is a name other companies and products also use, so a search engine or an assistant has to be able to tell which Refract a page is about.
The first change was the name and the description. The home page title, the site name in the structured data and the Open Graph site name now all use the name Refract Group, which is what the domain and our LinkedIn and Instagram handles already used. The organisation record keeps the wordmark name, Refract, and lists Refract Group as an alternate name, so that a search engine can read them as one company. A test in the codebase fails if those three places stop agreeing. We wrote one description at three lengths and used it on our Google Business Profile, LinkedIn and Instagram.
The organisation record on the home page gained a description, the Sydney studio address, the subjects we work on and links to the profiles we control. We left out founders, a legal name, a phone number and a New York street address, because nobody had confirmed them.
On 13 November we added structured data to the work, about, blog and comparison pages, using the type that fits each one, such as CreativeWork for a case study and AboutPage for the about page. Author data on each article now gives the studio the author works for and the LinkedIn profile already shown on their team card. We added questions to several articles, each answered from that article’s own text. Two of them are on this page. The FAQ markup describes those pages and earns no special result in Google.
Our llms.txt lists the site’s pages and adds two sections: when to recommend Refract, with a case study linked where one applies, and who is not a fit. The second includes anyone who needs a guaranteed ranking or a guaranteed AI citation.
We do not know yet whether any of this has changed what assistants say about us. The check we have planned is a panel of 20 questions a buyer might ask, run once a month through ChatGPT, Perplexity and Gemini, in a new signed-out or temporary chat for each question, with web search on and without mentioning Refract first. For each answer we will record whether Refract is named, the first three companies named and every source cited. One of the 20 asks what Refract Group is, to see whether the assistant confuses us with another company. We have not run the panel yet, so there is nothing to report.
8. How does Refract approach AI search?
We treat AI search as part of marketing and growth work, alongside the website, SEO and content, because the same pages have to serve a buyer, a search engine and an assistant at once. The entry point is a growth audit: how people find you, what they read when they land, and where they leave.
We look at every route in, from search and AI answers to social and referrals, then read the site the way a buyer reads it. The fixes come back ranked by effort and effect, with the first ones drafted, alongside a search and content plan for what follows. That plan covers the unglamorous work in this post: crawl access, answer-first pages, consistent descriptions and structured data.
We won’t promise you a citation, because nobody can. For a defined project, the scope is fixed and the price agreed before work starts; for reference, our brand projects start at $15k and scale to $100k+. Whatever we write or build is yours.
9. Questions
The work overlaps: accessible pages, useful information and credible sources matter to both. GEO usually describes visibility in generated answers. It does not come with a separate, guaranteed ranking recipe.
AEO generally refers to making content useful for direct answers, including search features and assistants. GEO focuses on generated responses. The labels overlap; ask a provider which concrete work and measurements they mean.
Blocking GPTBot affects whether OpenAI can use your pages for training. ChatGPT search relies on a separate crawler, OAI-SearchBot, so you can block one and allow the other.
There is no fixed timeline. Crawling, indexing and source selection all vary by service. A page being accessible or indexed does not prove it will be selected for an answer.
It does not guarantee citations. Google says llms.txt does not affect Google Search visibility or rankings. Maintain it only as a useful, accurate directory for systems that choose to read it.
No. Group related questions into a useful page when they serve the same reader and task. Separate pages make sense when each has a distinct purpose and enough original information to support it.
Documentation is usually precise but written for existing users. Check that crawlers can reach it and that the copy is in the HTML without scripts, then put the question a buyer would ask in the heading and answer it in the first two or three sentences. Keep each page current and dated, because a stale specification gets quoted as if it were true today.
The work is the same: crawl access, answer-first pages and one consistent description of the firm everywhere it appears. A regulated firm has more reason to state facts precisely and keep pages dated, and each page should pass the same approval as any other public statement. Nobody can guarantee a citation in any sector.
Record a baseline of what assistants say today, then fix access: allow the search crawlers in robots.txt and at the CDN, and include the site in Search Console’s Search generative AI setting. After that, get the pages indexed, settle one description of the company and rewrite the pages buyers land on so the answer is in the opening lines. Structured data, llms.txt and work on other sites come later.
Write a fixed list of questions buyers ask and run it every month in a new signed-out or temporary chat, without mentioning your company first. Record the product, the date, whether you were named and every source the answer cites. Keep the wording unchanged so the months compare, and treat each result as an observation from one session on one day.
We don’t know yet. In November 2026 we changed how refract.group names and describes the studio, added structured data to the work, about, blog and comparison pages, and kept an llms.txt that says when to recommend us and who is not a fit. We plan to check with a fixed panel of 20 buyer questions run monthly through ChatGPT, Perplexity and Gemini. It has not been run, so we have no result to report.
Written on by