How to get cited by ChatGPT.

What ChatGPT does differently, measured

We measured this across 50 buying questions put to four assistants in July 2026, with the methodology, prompts and a run-to-run variance control published alongside the dataset. ChatGPT behaved unlike the other three in a way that changes where the work is.

Per answer, averaged over 50 questions in five industries. Treat the directory figures loosely: they move depending on where you draw the line between a directory and a publisher.

ChatGPT named the most businesses from the fewest sources. It cited 4.9 sources per answer and named 8.9 companies, roughly the inverse of Perplexity, which read 7.8 sources to name 4.5. It also leaned on directories harder than any other assistant, at about a quarter of everything it cited.

Both facts point the same way. With so few sources behind each answer, the competition for a slot is tighter than on an assistant that reads widely, and a large share of the slots that exist are held by listings rather than by businesses' own sites. Step 6 below is not an afterthought for that reason. The full study lists every source each assistant used.

How does ChatGPT decide what to cite?

ChatGPT does not rank pages the way Google does. When it answers a question with live sources, it retrieves a small set of candidate pages, extracts facts it can verify, and composes an answer that attributes those facts. Getting cited means winning at each of those three stages, retrieval, extraction, and attribution.

This is why sites that rank well on Google are often invisible in ChatGPT. Google rewards links and relevance. Retrieval-and-quote systems reward machine-readability and quotable structure. They are different problems.

One scoping note: Google's own AI features work differently. Google's official AI optimization guide says AI Overviews and AI Mode use the same ranking systems as regular Search, so your Google rank matters there, and Google explicitly notes that structured data isn't required for its generative features. Everything in this guide targets the assistants that don'tuse Google's rankings: ChatGPT, Perplexity, and Claude.

The three OpenAI agents, and which one matters

Most robots.txt confusion in this area comes from treating these as one thing. They are three, they do different jobs, and blocking one does not block the others.

• GPTBotis the crawler that collects pages for OpenAI's model training. Blocking it opts you out of training and leaves you citable.
• OAI-SearchBot is the crawler behind ChatGPT search results and citations. This is the one that decides whether you can appear as a source, and blocking it removes you from answers.
• ChatGPT-User is the fetcher that retrieves a page live when somebody pastes your URL into ChatGPT and asks about it. Block it and that person gets nothing back.

All three are documented by OpenAI in its overview of OpenAI crawlers, which is the page to check rather than any third-party summary, this one included, because the agent list changes.

Step 1: Make sure ChatGPT can actually reach you

This is the step almost everyone skips, and it invalidates all the others. OpenAI runs three separate crawlers, and they do different jobs:

Plenty of sites blocked "the AI bots" in 2024 to protect their content from training, then wondered why they never appear in ChatGPT answers. If you want citations, allow OAI-SearchBot:

That last block is the nuance worth understanding: you can stay out of training data and still be cited in answers.

Step 2: Add JSON-LD structured data

Structured data tells a machine what your page is instead of making it infer meaning from your layout. At minimum, publish Organization sitewide and Article on every content page:

This page carries exactly that markup. You can view source and check. A guide telling you to add schema should have schema.

Step 3: Write answers that can be quoted

An AI composing an answer needs a self-contained sentence it can lift without losing meaning. Content built as narrative prose rarely offers one. Content built as question → immediate answer offers one in every section.

Practical rule: after every H2, answer the heading in the first two sentences. Put the conclusion first and the reasoning after. If a sentence only makes sense with the paragraph above it, it cannot be cited.

Step 4: Make freshness machine-readable

"Updated June 2026" typed into your page as plain text is invisible to a parser. Freshness has to exist in markup, dateModified in your JSON-LD and a <time dateTime="…"> element in your HTML. And only update the date when you genuinely change the content; stamping today's date on unchanged pages is a pattern search systems learn to discount.

Step 5: Build topical depth, not one page

Retrieval systems favour sources that cover a subject repeatedly and consistently. One strong page is weaker than six connected pages that agree with each other and link together. Pick your core topic, write the cluster around it, and interlink them with descriptive anchor text.

Step 6: Get onto the sources AI already reads

AI answers lean heavily on a handful of source types: review sites, community threads like Reddit, YouTube, and editorial roundups. If your category's answers consistently cite a comparison site you are absent from, that absence is the gap, not your own page. Being named on the sources AI already trusts is often faster than trying to make it trust you directly.

Which sites, though. That is the part usually left as an exercise for the reader, so here is what ChatGPT actually cited when we asked it fifty buyer questions across five industries in July 2026.

Counts are how many of that industry's ten answers cited the domain. One run per question, July 2026.

The directory route is not universal, and that is the finding. For an agency, being absent from Clutch is close to disqualifying: it was cited in seven of ten answers. For an interior designer the same advice is useless, because ChatGPT cited no directory at all and linked five separate practices' own websites instead. Two industries, opposite instructions, same assistant.

The B2B SaaS row is the one worth sitting with. ChatGPT named 63 software products across those ten questions and cited nothing whatsoever, while the other three assistants cited normally on the same questions. It answered from what it already knew rather than from anything it read at the time. If that holds, there is no citation to win in that category, only the slower work of being known well enough to be recalled: mentions on sites and communities that outlast a single crawl.

Ten questions per industry, one run each, so treat these as a direction to check rather than a constant. The point is not the exact counts. It is that you should look up what your own category gets cited, because the answer differs enough to reverse the advice.

How do I check whether ChatGPT cites me?

Three checks, cheapest first:

1. Ask it. Put your real buyer questions to ChatGPT with search enabled and see whether your domain appears in the sources. Not your brand name, the question a customer would actually type.
2. Check your logs. Search your server access logs for OAI-SearchBot. No hits means you are not in the retrieval pool at all.
3. Measure it repeatedly. A single check is an anecdote. Track the same questions over time to see whether changes move the needle.

Frequently asked questions

Sources

The technical claims above rest on these. Each was read on 2026-09-01, and all three change without notice, so check the page rather than trusting a date.

• Overview of OpenAI crawlers, OpenAI. Names GPTBot, OAI-SearchBot and ChatGPT-User and what each is for.
• Introduction to structured data markup, Google Search Central.
• Optimizing your website for generative AI features, Google Search Central. Worth reading in full: it states that no special markup or AI text file is needed for Google.

The citation figures are our own, counted from the study, whose dataset is published under CC BY so you can recount them.