- Confirm OAI-SearchBot can fetch your key pages, both in robots.txt and at your CDN or firewall, because a blocked site cannot appear in ChatGPT search answers.
- Open every section with a self-contained answer that names a specific number, date or entity, since retrieval lifts passages rather than whole pages.
- Track a fixed panel of 20 to 30 buyer prompts across repeated runs each month and log which domains get cited, rather than judging from one answer.
If you want to get cited by ChatGPT, start with the mechanism, because most of the advice circulating about it skips the mechanism entirely. There is no submission form and no ranking dashboard. A link appears in a ChatGPT answer when the model decides to search, retrieves a page from that search, reads a passage from the page, and judges the passage good enough to attribute. Each step can succeed or fail for a specific reason, and each one can be checked. Most tactics worth doing serve one of them.
That also explains why so much of the advice is useless. Many ChatGPT answers never search at all. The model answers from what it absorbed in training, which you cannot update on a schedule and cannot inspect. The citations you can actually influence come from the search path, so that is where this playbook spends its time. For the wider picture across answer engines, see our guide on getting cited by AI engines.
How ChatGPT decides which sources to cite
ChatGPT does not search for every prompt. When a question seems answerable from general knowledge, such as how a TCP handshake works, it usually responds without links. When a question depends on current or specific facts, such as today’s price for a named product, a recent regulation, or a comparison of tools released this year, it tends to search. That makes your first decision a targeting one. Prioritize the queries where the answer changes: prices, versions, rankings, statistics and comparisons. Those are the prompts that trigger retrieval, and the ones where a citation is possible at all.
When search does run, the path looks roughly like this.
Two details in that path surprise most people.
The query the model writes is not the query the user typed
The model rewrites the prompt before it searches. “What should a twelve-person agency use for client reporting?” might become three narrower queries: “client reporting tools for small agencies,” “agency reporting software pricing,” and “client dashboard tools comparison.” Each returns its own set of pages, and the model draws on the combined results. Google describes a similar pattern in its AI Mode as “query fan-out.”
The consequence is that optimizing for one head term is a losing bet. You need pages that answer the sub-questions a model is likely to generate from the prompts your buyers actually type. The practical method is to run those prompts yourself and note which sub-questions the cited sources seem to address. Your content should map onto those, not onto the phrase you wish people searched.
Retrieval works on passages, not pages
Retrieval systems generally split documents into chunks, score the chunks against the query, and pass the best few to the model. A 4,000-word page can get cited because of one 80-word section. That is good news for long pages and bad news for pages whose key answer sits three paragraphs deep, behind setup that only makes sense in context.
So the unit you are optimizing is the passage. Ask whether the best passage on your page still answers the question if someone pasted it into a stranger’s document with no surrounding text. If it doesn’t, rewrite it.
Check that the search crawler can reach you
Start with the unglamorous part, because it is a hard gate. OpenAI documents several crawlers, and they do different jobs. Confusing them is the most common way a site locks itself out of ChatGPT search without meaning to.
| Agent | What it’s for | What blocking it does |
|---|---|---|
| OAI-SearchBot | Surfaces websites in ChatGPT’s search features | Blocked sites stop appearing in ChatGPT search answers, though they can still show up as plain navigational links |
| GPTBot | Crawls content that may be used to train OpenAI’s foundation models | Signals that your content should not be used for training. It does not remove you from search |
| ChatGPT-User | Fetches a page when a user asks ChatGPT something that needs it | Robots.txt rules may not apply, because a person initiated the visit. It does not affect search inclusion |
If you want citations, OAI-SearchBot is the agent that matters. GPTBot is a separate decision about training, and you can block it without losing search visibility. Make that call deliberately rather than copying a blanket block from a template.
The failure that bites most sites is rarely in robots.txt. A team allows OAI-SearchBot in the robots file, then a bot-protection rule at the CDN or firewall returns a 403 to any crawler it doesn’t recognize. The robots file says yes, the front door says no, and nobody notices because the site still ranks in Google.
Check the logs first: filter your server or CDN logs for OAI-SearchBot over the last 30 days and group by status code. Mostly 200 responses means you're fine. Persistent 403s or 429s mean the crawler is being refused, and no content change will fix that.
Once access works, don’t assume the crawler will render your JavaScript for you. Put the core answer in the HTML the server sends. A page whose main content only appears after a client-side script runs is a page you’re gambling on.
Write passages a model can lift without rewriting
Models cite passages, so write passages that stand alone. Three habits do most of the work.
Answer first, then support
Put the answer in the first sentence or two under each heading, and use the rest of the section for support. Compare two versions of a claim about a hypothetical reporting tool called Ledgerline:
Weak: “Pricing is an important consideration when choosing a reporting tool, and plans vary depending on your needs.”
Strong: “Ledgerline’s Team plan costs $49 per seat per month on annual billing and includes unlimited client dashboards and a 14-day free trial.”
The strong version names an entity, gives a number, states the unit and the condition. A model can quote it without changing a word. The weak version gives the model nothing to use, so it will pick a competitor who said something concrete.
Specifics, with sources attached
“Faster” and “more affordable” are not claims a model can use. “Agencies in our 2025 customer survey cut weekly reporting time by roughly a third” is, but only if the survey exists and the number is right. Attach the source, the year and the method when you have them. Only publish numbers you can defend. An invented statistic is a liability twice over: readers will catch it, and a model choosing among several candidate sources has no reason to keep trusting one that contradicts the rest.
Name the subject every time
Pronouns and back-references fail when a passage is lifted out of its context. “This approach reduces the workload” tells the model nothing once the previous paragraph is gone. Put the subject back into the sentence that carries the claim. “Automated client reports cut weekly workload for account managers” survives extraction. “It cuts the workload” does not.
The same logic applies to comparisons. A page about your product should name the alternatives it is being compared against and say in which respect, instead of claiming to be the best. Comparison prompts need pages that name the things being compared, and a page that avoids naming competitors usually doesn’t get pulled into the comparison at all.
Earn the third-party mentions that comparisons lean on
For recommendation prompts (“best reporting tool for a small agency”), ChatGPT leans heavily on sources that have already done the comparing: review sites, roundup articles, trade publications and community threads. Your own site is a strong source for facts about you. It is rarely the only source a model uses to decide who belongs in a list. If your category has a dominant roundup, assume it is part of the answer whether or not you’re in it.
That makes off-site work non-optional for comparison prompts. Four moves do most of it.
- Get into the comparisons that already exist. Take your top twenty prompts, list the five pages cited most often, and work out how each one decides what to include. Some want a pitch to the editor, some want accurate product data, some only accept independent reviews. Each route is a different job, but the list is finite.
- Publish something other sites want to cite. A survey, a benchmark or a dataset gives journalists and roundup writers a reason to link to you. Restating a number that already exists gives them none. Unique figures get cited, and reworded opinions mostly don’t.
- Answer questions where buyers already ask them. Forum threads and community Q&A get retrieved, and they carry the exact phrasing buyers use. Disclose your affiliation, answer the question first, and mention your product only where it genuinely fits. A self-promotional reply gets removed, and the thread stays as the record.
- Keep the listings accurate. The pricing, category and feature summaries on review and directory sites are often what a model reads when it describes you. A listing that shows a $29 plan when you charge $49 produces a wrong answer that is hard to spot, because the model is faithfully repeating its source.
Make your brand impossible to confuse
Models resolve brand names the way search systems resolve entities, and the failure here is confusion. A company called Relay, Pulse or Atlas shares its name with dozens of products and open-source projects. When your name is generic, a model can attribute your features to someone else, or skip you because it can’t tell which entity the prompt means.
Four things reduce that risk.
- Use one canonical name everywhere. Same spelling, capitalization and suffix on your site, your structured data, your social profiles and your directory listings. Each variation splits the signal across several identities.
- Add Organization structured data. A JSON-LD block on your homepage stating your name, URL and logo, with
sameAslinks to profiles you control (a company LinkedIn page, a Crunchbase entry, a GitHub organization), is a clear, machine-readable identity statement. It won’t create authority by itself. It removes ambiguity about who you are. - Show real authors. Give articles a byline, a bio and a reason the author knows the subject. Anonymous articles are harder to attribute, and attribution is the thing a citation depends on.
- Pick one product vocabulary and keep it. If the homepage says “content engine,” the docs say “publishing platform” and the press release says “automation suite,” the model has three phrases to reconcile. It will often settle on whichever phrase appears most in third-party sources, which may not be the one you’d have chosen.
Measure citations with a prompt panel, not a single query
Typing your brand name into ChatGPT once tells you almost nothing. Answers vary between runs, phrasing changes which sources appear, and whether a search fires at all changes the result. You need a method that accounts for that variance.
- Build a fixed panel of 20 to 30 prompts. Include category questions (“best reporting tool for small agencies”), problem questions (“how do agencies automate client reports”) and comparison questions (“Ledgerline vs. the main competitor”). Keep the list fixed so you can compare month to month.
- Run each prompt several times. Three to five runs per prompt, across more than one session, will show you how much results shift. Record every cited domain, not just whether you appeared.
- Log the details. For each run, note the date, the prompt, the cited domains, which of your URLs were cited and which competitors appeared. A spreadsheet with one row per run is enough to start.
- Read the patterns. Which of your pages get cited? Which third-party sources dominate your category? Which prompts never mention you? The last question is the most useful, because it points at where content is missing.
- Change one thing at a time. Rewrite a page or fix a listing, then re-run the panel two to four weeks later. Changing five things at once leaves you unable to tell what worked.
Expect noise. Being cited in one run and absent in the next is normal. What you’re measuring is the share of runs in which you appear, and whether that share moves over time. The tooling side of this is covered in how to measure whether you show up in AI answers, but a spreadsheet is enough to begin.
Mistakes that waste the effort
Some tactics appear in every “GEO” guide, and most of them don’t hold up.
Treating llms.txt as the switch. Some guides tell you to publish an llms.txt file and expect citations to follow. OpenAI has not publicly said its search system uses llms.txt to decide what to cite, so treat the file as an optional convenience at most.
Stuffing keywords for machines. The most rigorous public work on this is an academic study on generative engine optimization by researchers from Princeton and several other institutions, presented at KDD 2024. It tested content edits against simulated answer engines. Adding citations, direct quotations and statistics to source text improved how prominently those sources appeared in generated answers. Plain keyword stuffing did little. The study used simulated engines rather than live ChatGPT, so read it as direction rather than a guarantee. The direction still matches the mechanism: models reward text that is specific and attributable.
Hidden text aimed at AI systems. Some sites place instructions in white-on-white text or metadata, telling models to cite them or describe them favorably.
Don't do it: hidden instructions aimed at AI models are manipulation, and they're exactly the kind of pattern that gets pages filtered. When they work at all, they work until someone notices, and the domain's reputation pays for it.
One page per prompt variant. It’s tempting to generate a thin page for every phrasing of a query. That is the programmatic SEO trap without the quality bar. Near-identical pages split your signal, and the model is less likely to find the one that matters.
A 90-day sequence to get cited by ChatGPT
- Week one: audit crawler access. Pull 30 days of logs for OAI-SearchBot, fix any 403 or 429 responses, and confirm your key pages return 200 with the answer in the HTML. Set the GPTBot policy separately, as a training decision.
- Weeks one and two: baseline the prompt panel. Record which domains get cited today for your 20 to 30 prompts. You can’t show progress without a starting point.
- Weeks two to four: rewrite five money pages. Make the first two sentences under every H2 a specific, self-contained answer, with dates, sourced numbers and named subjects. Leave the rest of each page alone for now.
- Month two: close the third-party gap. Identify the comparisons and roundups cited for your top prompts. Pursue accurate inclusion, fix your directory listings, and publish one piece of original data that others have reason to link.
- Month two: tighten entity signals. Canonical name, Organization markup, real author bylines and one product vocabulary across the site.
- Month three: re-run and decide. Compare the panel to the baseline, keep the changes that moved the cited share, and refresh any page whose cited passage has gone stale.
The honest wrap-up
You can’t guarantee a citation. Models choose among sources, and that choice shifts with phrasing, freshness and what else is available. What you can control is narrower and more useful: remove every technical blocker, make your key passages easy to lift, get your facts into the third-party sources comparisons draw on, and measure the result over time. It is less glamorous than the tactic lists suggest, and it’s the part that still works when the model changes.
AutoSEO Stack applies the same discipline to every article it publishes for a customer: research the questions buyers actually ask, write answer-first passages with specifics, and track whether the result shows up. If you want to see that loop run on your own domain, the AI visibility section explains how it works.