How ChatGPT search picks and cites sources
ChatGPT search pulls from partner indexes and OpenAI’s own crawl, then cites the pages that answer cleanly. You can be one of them. This guide shows how sources are chosen, how citations look, what to fix on your pages, and how to measure it for a small site.
By Théophile Louvart, founder of Porteur · Updated 14 September 2026 · Markdown
Where ChatGPT search gets results
ChatGPT search mixes a partner index, Bing among them, with pages crawled by OAI-SearchBot. It decides when to search, or you can force it in the UI. It rolled out to paid users on 31 October 2024, to all users in December 2024, and without an account in February 2025.
When it answers, it shows inline citations and a Sources panel with links. Shopping questions may show product cards. It cannot execute JavaScript, so content rendered only after scripts run is invisible to it.
There is no submission form. Your route in is to let the right crawlers in, write pages that answer directly, and appear on pages that are already being cited for your topic.
How a citation appears in ChatGPT
ChatGPT places small inline links at the claim it attributes, and repeats sources in a panel. The label is the site or page title. Links go to the cited URL.
An example: you ask “how to set up billing in YourProduct”. ChatGPT quotes a step from yourproduct.com/guides/getting-started, adds a link after the sentence, and shows that URL in the Sources panel.
It prefers pages that answer in the first paragraphs. Clear headings, consistent product and person names, lists, comparisons, and figures it can quote make a page easier to cite. A visible date and an author help it choose a recent, credible source across assistants as of 2026.
Make your pages retrievable by ChatGPT search
- Allow OAI-SearchBot in robots.txt. Blocking it makes you invisible in ChatGPT search.
- Keep content in the HTML. Do not rely on client-side rendering for key text.
- State the answer early. Put the gist in the first 1 to 3 paragraphs.
- Use headings that match the questions buyers ask. Write one page per question.
- Show a date and an author on guides and comparisons.
- Name products and people consistently across pages.
- Add lists and simple tables for steps, options and figures.
- Use descriptive, permanent URLs like /pricing and /guides/getting-started.
A fixed page looks like this: /guides/getting-started opens with “Set up YourProduct: create a workspace, add a card, invite your team.” Then a bulleted list, dated August 2026, with an author line and no content hidden behind scripts or tabs.
Write answers assistants can cite
Generative engine optimisation is about being named and cited by assistants: ChatGPT search, Perplexity, Google AI Overviews and AI Mode, Claude, Gemini, Microsoft Copilot. It is the same aim as answer engine optimisation from the featured-snippet years. The practice is the page work you control.
Match the question in the H1 and H2
If buyers ask “How to export invoices from YourProduct”, use that as the H1, and repeat it in a scannable H2 near the answer.
Lead with the answer, not context
First sentence: “Go to Billing, click Export, choose CSV.” Follow with steps and edge cases. Keep setup and rationale later.
Use numbered steps and lists
Assistants quote bullet and number lists. Write steps 1 to 5. Keep each step one action.
Include a small comparison or figure
Add a table like Basic vs Pro limits or a one-sentence price note if relevant. Assistants lift crisp facts.
Stamp a recent date
Show a visible updated date. Assistants often prefer fresher pages when many are similar.
Keep naming consistent
Use the same product and feature names on /pricing, /docs and /blog. Avoid nicknames the index will not match.
On /pricing, make the plan limits a clean table, not an image. On /integrations/slack, place a three-step setup at the top, dated. On /blog/compare-yourproduct-vs-the-legacy-tool, conclude with a clear winner and a table of differences. This is the copy that gets quoted.
Check whether ChatGPT already cites you
List buyer questions
Write the top 20 questions from support and sales. Example: “yourproduct pricing”, “yourproduct alternatives”, “how to connect yourproduct to QuickBooks”.
Ask ChatGPT in a clean session
Open chatgpt.com. Ask each question as buyers phrase it. If it does not search, press the search button. Screenshot the answer with the Sources panel.
Log every named site
Record the names and URLs ChatGPT cites. Tag them by question and by page type, like “docs”, “comparison”, “list of tools”.
Repeat on the other assistants
Run the same list on Perplexity, Claude search, Gemini, and Copilot. Compare overlaps and gaps.
Segment by referrer in analytics
In your analytics, build segments for chatgpt.com, perplexity.ai, copilot.microsoft.com, gemini.google.com. Watch them weekly.
No assistant publishes a citations report as of 2026. Your methods are asking on a schedule, reading referrers, and using trackers that automate the asking. Tools in the market include Profound, Peec AI, Otterly, Scrunch, the AI toolkits of Semrush and Ahrefs, and DataForSEO’s LLM mentions data. Pick one if you have many questions to monitor.
Control OpenAI access with robots tokens
Three tokens control how OpenAI uses your site as of 2026. Set them explicitly in robots.txt. Treat search, on-demand fetching, and training as separate decisions.
| Token | Used for | Allow if you want | Block if you want |
|---|---|---|---|
| OAI-SearchBot | Indexing for ChatGPT search results | To appear as a source in ChatGPT search | To stay out of ChatGPT search |
| ChatGPT-User | Fetching a page when a user asks about it | To let ChatGPT fetch your page on demand | To stop on-demand fetches |
| GPTBot | Crawling for model training | If you are fine with training use | To keep content out of OpenAI training |
# keep in ChatGPT search, allow on-demand fetches, block training
User-agent: OAI-SearchBot
Allow: /
User-agent: ChatGPT-User
Allow: /
User-agent: GPTBot
Disallow: /
Robots is voluntary. Blocking a training crawler does not remove data already collected. If you block answering and search crawlers like OAI-SearchBot, Bingbot or Googlebot, assistants cannot retrieve and will not cite your pages.
Fix common blockers for ChatGPT search
- JavaScript-only content. Move critical copy into the HTML source. Server render or hydrate after content is present.
- Hidden answers. Avoid accordions and tabs for the core steps. Place a concise answer above them.
- No date. Add an updated date under the title of help and comparison pages.
- Ambiguous names. Use the same feature names across the site and title tags.
- Thin product pages. Give /pricing and /features plain-language bullets with figures assistants can cite.
- Robots rules. Confirm you allow OAI-SearchBot and your CDN does not block it by IP or rate limit.
Check a page by pasting its URL into your browser’s View Source. If the key sentence is not there, ChatGPT cannot read it. For /pricing, put plan limits such as “Projects: 3, Seats: 5” in text near the top. For /guides/export-invoices, list steps 1 to 5 before screenshots.
Win citations through roundups and comparisons
Assistants name the pages they retrieve. Being on the lists, roundups and comparisons that already get cited is the straightest line to being named. This matters if your own site is small and new.
Find the cited roundups
When you ask “best invoice tools for freelancers”, list the roundup URLs ChatGPT cites. Repeat on Perplexity and Copilot. Rank by how often they appear.
Pitch for inclusion
Contact the authors with a short note and a clear reason to add YourProduct. Offer a test account and the one feature that stands out.
Publish your own comparison
Write /alternatives/legacy-tool and /compare/yourproduct-vs-legacy-tool. Use a table and a date. Assistants often cite both third-party lists and vendor comparisons.
Keep names aligned
Make sure your product is named the same way on your site and in outreach. Consistency helps assistants and editors match mentions.
A fixed example: /alternatives/legacy-tool opens with “The 7 best Legacy Tool alternatives in 2026” and a numbered list. Editors can cite it. Assistants can quote it. Your brand name appears in the same form across the page, the title tag and headings.
Measure impact and set a simple cadence
You will not get a native report of where your site was cited. Build a loop. Ask, log, ship fixes, and ask again. Treat it like keyword tracking, but for names and links in answers.
Weekly checks
Run your 20 buyer questions in ChatGPT and Perplexity every week. Log names, links and positions in the Sources panel. Flag new entrants.
Source-focused edits
When a rival page is cited for “yourproduct pricing”, compare structures. Rework /pricing to lead with the figures and plan limits in text.
Analytics segment
Report sessions with referrers chatgpt.com, perplexity.ai, copilot.microsoft.com, gemini.google.com. Tag pages they land on. Compare trend lines.
Quarterly content refresh
Update dates and figures on your key guides and comparisons. Keep the first paragraphs stable and precise so they stay quotable.
Clicks from Google’s AI Overviews and AI Mode arrive as google.com in Search Console and analytics. They are mixed with web search. Track rankings and clicks for your question pages to see if your changes help. Use your /search-console Performance report to monitor queries and pages together.
How this differs from classic SEO work
You still need crawlable pages, titles under about 60 characters, and content that matches intent. The difference is what you optimise for: being quoted, not just ranked. That shifts layout and wording choices on key pages.
- Front-load the answer, not a teaser paragraph.
- Write one page per buyer question, not a broad hub that buries answers.
- Prefer text and tables to images of text.
- Show dates and authors on all help and comparison content.
- Use headings that restate the question in the buyer’s words.
- Avoid JS-only content for anything you want cited.
Questions
There is no submission tool. Allow OAI-SearchBot in robots.txt, keep the content in the HTML, and link to the page from places ChatGPT already retrieves. If speed matters, get listed on a roundup that is already cited so your brand is present while your own page is discovered.
No. Allow OAI-SearchBot and ChatGPT-User to appear and be fetched. You can disallow GPTBot if you do not want your pages used for model training. This keeps you in search results and out of training.
They pick pages that answer early and cleanly. If your docs bury the steps under JS, tabs or long preambles, you lose. Fix the structure, add a date and an author, and make sure your product is named consistently. Also, get your product onto third-party lists that are already cited.
Look at referrers. ChatGPT sends chatgpt.com, Perplexity sends perplexity.ai, Copilot sends copilot.microsoft.com, Gemini sends gemini.google.com. Clicks from Google AI Overviews and AI Mode arrive as google.com, mixed with regular search. Build segments for each and watch the landing pages.
No. The same structure helps both: a direct answer at the top, clear headings that match the question, lists and small tables, consistent naming, a date, and an author. Keep content server-rendered. Tailor examples to your buyers’ phrasing, not to a brand-new keyword list.
Client-side rendering of the answer, missing dates, and vague headings. Also, robots rules that block OAI-SearchBot or rate limits that look like a block. Put the key text in the HTML, add a date under the title, and confirm your CDN allows OAI-SearchBot and ChatGPT-User.
Check my site, free
Want to know if ChatGPT can retrieve and cite your key pages, and who it names now? Paste your URL into the free check and read three findings in about thirty seconds.
- Free check, no card
- Read-only, your own accounts
- Readable by your agent
Read next
- Guiderobots.txt for AI crawlers: GPTBot, ClaudeBot, PerplexityBot and what to allow
- GuideGenerative engine optimization (GEO): how to be named by AI answers
- GuideSEO for ChatGPT: how assistants choose what to cite
- GuideHow to appear in Google’s AI Overviews
- GuideHow to track your AI visibility without a subscription
- GuideHow Perplexity picks its sources, and how to become one
- GlossaryChatGPT search
- GuideMicrosoft Copilot cites from Bing: the way in is Bing Webmaster Tools