Short answer: to get ChatGPT to cite your website, first make sure its search agent can fetch and read the page: allow OAI-SearchBot and ChatGPT-User, confirm your firewall lets them through, and serve the answer in the initial HTML. Then make the page worth citing: one answerable claim under each question heading, honest dates, and formats that match how buyers ask. Finally, count citations properly, or you will not know whether any of it worked. Those are the seven steps below, grouped into three phases and meant to be done in order.
None of this guarantees a citation. ChatGPT chooses its sources answer by answer. These steps remove the reasons it could not choose you, and give you a way to see the result. If you first want to know whether ChatGPT cites you today, our post on checking whether ChatGPT cites your website covers that test.
Each of the first six steps clears one gate a page has to pass before it can be cited at all. The gates fail differently, and only one of them produces an error you would ever notice:
Work down the gates rather than starting with the writing. A beautifully structured page behind a firewall rule is invisible, and no amount of rewriting will tell you so.
Phase 1: let ChatGPT reach and read the page
Step 1: allow the agents that lead to citations
OpenAI documents three agents a typical site will see. OAI-SearchBot builds the index that surfaces websites in ChatGPT's search features; this is the one a citation depends on. ChatGPT-User fetches a page because a person in ChatGPT asked for it. GPTBot crawls content that may be used to train OpenAI's models. Blocking GPTBot is a training decision and does not remove you from ChatGPT search results. Blocking OAI-SearchBot does.
User-agent: OAI-SearchBot
Allow: /
User-agent: ChatGPT-User
Allow: /
# Optional: opt out of training only
User-agent: GPTBot
Disallow: /
How you change robots.txt depends on the platform. On Shopify you add a robots.txt.liquid template to the published theme and add your rules on top of the Liquid robots object rather than replacing the file, because Shopify updates the defaults. On self-hosted WordPress the file is virtual and comes from do_robots(); a real robots.txt in the web root is served before WordPress runs and overrides it. One more rule catches people out: an agent that has its own group ignores the User-agent: * group entirely, so private paths have to be repeated inside the OAI-SearchBot group. Our working robots.txt for AI crawlers has a fuller file with that handled.
Step 2: make sure the firewall agrees with robots.txt
robots.txt states intent; it does not grant access. A CDN or bot-management rule that returns 403 to these agents is a block in practice, and the file will keep saying otherwise. OpenAI publishes an IP list for each agent: openai.com/searchbot.json for OAI-SearchBot and openai.com/chatgpt-user.json for ChatGPT-User. If your edge has an allow list for verified bots, build it from those lists rather than from the user-agent string, which anyone can send.
Also note what robots.txt cannot do. OpenAI's documentation notes that ChatGPT-User, because it acts on a person's request, may not follow robots.txt the way a crawler does. Treat the file as a statement for crawlers, and use the edge for anything you truly need to keep out.
Step 3: serve the answer in the initial HTML
Most AI crawlers do not execute JavaScript. Text that appears only after client-side rendering does not exist for them, and a 200 response with an empty body is a failure, not a pass. Each platform has its own gaps:
| Platform | Usually in the HTML | Usually not |
|---|---|---|
| Shopify | Normal storefront pages, composed in Liquid on the server | Content loaded through the Ajax Section Rendering API; app-injected reviews and Q&A |
| WordPress | Classic and block theme output | Load-more listings, builder-injected tabs and accordions, headless front ends |
| Webflow | Page copy and CMS collection content bound in the Designer | Site search results, Embeds built at runtime, third-party scripts |
| HubSpot | HubL page and blog body content | Text inside embedded forms, CTAs and chat |
Check a key page with curl, once as OAI-SearchBot and once as an ordinary browser, and search the output for the sentence you want cited. If the browser copy has it and the agent copy does not, you have found your gate. JavaScript rendering and AI crawlers covers the fixes for each common setup.
Phase 2: make the page worth citing in ChatGPT
Step 4: put one answerable claim under each question heading
Citations attach to passages. A passage is easy to cite when it answers one question completely in its first one or two sentences, names the product or condition in full, and states the fact with units: a price, a dimension, a time frame, a compatibility. A heading phrased as the question a buyer would ask tells a retriever what the passage under it answers. A heading such as "Our approach" tells it nothing.
Test each section by reading its first two sentences out of context. If they stand on their own as an answer, the passage is citable. If they lean on "it", "this plan" or "as mentioned above", rewrite them so they name the thing.
Step 5: date the page honestly
Many questions imply a current answer: pricing, comparisons, policies, "best" lists. Pages that state a real published and modified date have an advantage on those questions over undated pages. Pages whose modified date is bumped by every deploy are worse than undated ones, because a crawler can check the date against the content. Stop the build pipeline from touching dates, and change the modified date only when the content changes. Our guide to content freshness and dates shows how.
Step 6: publish the shapes that get cited, and cover the pages that compete with you
Comparison and round-up pages match the shape of a buying question directly, which is why they dominate citation lists in commercial categories. A round-up that states its method and criteria is far more quotable than one that just lists products. If you publish one, include yourself honestly, including where you are not the best fit.
Your page is also not the only candidate for the citation, and the field is wide. In Profound's analysis of 680 million citations, the single most-cited domain in ChatGPT, Wikipedia, held 7.8% of them, and Reddit, in second place, 1.8% (data collected August 2024 to June 2025, so the shares will have moved since). Each kind of page in that field needs different work:
Two things follow. Wikipedia and Reddit are third-party pages that describe products and companies, so what they say about you matters as much as what you say. And because no domain dominates, a page that answers a specific question well has real room to be chosen. In product categories a large share of citations go to marketplaces, review sites and publisher round-ups rather than brand sites, so make sure those pages carry your current facts, prices and availability. When they are cited, what is said about you should at least be right.
Phase 3: count ChatGPT citations the way they actually appear
Step 7: count only what the answer marks
Citation numbers are easy to inflate by accident, and a flattering number points the work in the wrong direction. An assistant typically retrieves many more pages than it cites, and the difference between the two is where most counting errors live.
Four rules keep the count honest, and the table shows what each one prevents:
| Rule | Counting it wrong | Counting it right |
|---|---|---|
| Count only sources marked in the answer text | Every page retrieved while the answer was composed. Before we fixed this, it overstated our own citation figures by two to four times. | Only the sources the answer itself marks |
| Discard markers that resolve to nothing | A citation recorded for a marker with no real URL behind it | One citation fewer, rather than one that is not real |
| Report breadth next to volume | One answer citing twelve of your pages read as twelve times the reach | The number of distinct questions that cite your domain, alongside the raw count |
| Keep cited and recommended apart | Your page counted as a win while the answer recommends a competitor | Two separate measures, so the gap between them stays visible |
The last row is the one most teams miss. Being cited means your page supported the answer; being recommended means the answer named you as the pick. The second is a different problem with different fixes, covered in how to get recommended by ChatGPT.
Run the same questions on a fixed schedule, without your brand name in them, and trust a change only when it appears in two consecutive runs. A question that names you will of course produce an answer that names you, which tells you nothing about the buyers who did not.
Which step to start with
If you already have some evidence, you do not have to start at Step 1. Match what you see to the gate it points at:
| What you see | Most likely gate | Start at |
|---|---|---|
| ChatGPT never cites any page on your domain, even for questions only you answer | Access: robots.txt or the edge | Steps 1 and 2 |
| Some pages are cited, but your product or pricing pages never are | Rendering on those templates | Step 3 |
| Your pages are cited for background facts, while competitors are recommended | Passages and formats | Steps 4 to 6 |
| Citation counts jump around from month to month | Measurement, not the site | Step 7 |
Whichever row you are in, finish with Step 7. It is the only step that tells you whether the others worked. If you would rather not run the counting by hand, InsightWonder's citation tracking applies these four rules automatically; a spreadsheet and a fixed question set will do the same job.
Sources
- OpenAI: overview of OpenAI crawlers (OAI-SearchBot, ChatGPT-User, GPTBot, and their IP lists)
- Shopify: robots.txt.liquid
- WordPress: do_robots()
- Profound: AI platform citation patterns (680 million citations, August 2024 to June 2025)
- Yext: 2026 Consumer Search Behaviors Report