How ChatGPT Search Decides Citations: What OpenAI Documents and What Tests Show

So OpenAI documents the plumbing but not the scoring. Independent tests fill part of the gap. One of the largest, Ahrefs' April 2026 study of 1.4 million prompts, found ChatGPT cited only about half of the pages it retrieved, and the cited ones had titles closer in meaning to the sub-queries it searched. That's a correlation, not a rule, and this post keeps the two kinds of evidence apart.

The same pipeline sits behind other doors. When Siri hands a question to ChatGPT on an iPhone, it goes to this ChatGPT, as our guide to how Apple Intelligence routes Siri queries to ChatGPT explains. For crawler setup, see our ChatGPT SEO guide.

When ChatGPT Search Runs at All

A citation can only appear if ChatGPT searched. OpenAI's search help page says ChatGPT "may search the web automatically when your question would benefit from current information." You can also force it: open the tools menu and choose Search, type "/" and pick Search, or regenerate an answer with "Search the web." Web search works on every plan, including Free and logged-out use, and in voice conversations when you ask for it.

That first decision explains a lot of missing citations. A question about a stable fact ("what is double-entry bookkeeping?") may be answered from the model's training, with no sources at all. A question about prices, reviews, news or "best X in 2026" needs current facts, which is the case OpenAI's wording describes. Ahrefs noted in its August 2025 overlap study that ChatGPT and Gemini "don't always link out" and that their citations depend on the query. In the 2026 paper "Don't Measure Once", just 42.2% of repeated ChatGPT runs produced any citation at all, the lowest of the four engines tested. The authors link that to ChatGPT skipping web search for definition-style prompts.

Evidence grade: OpenAI documents that search is automatic or manual. It doesn't document the threshold for "would benefit from current information."

How ChatGPT Search Rewrites Your Question

ChatGPT doesn't send your words to a search engine as typed. OpenAI's help page says that when ChatGPT search partners with other providers, it "typically rewrites your query into one or more targeted queries." Its example turns a researcher's question about drugs that target CCR8 into "CCR8 immunotherapy drug development 2025," then a narrower follow-up, "CHS-114 conference 2025."

Three inputs shape the rewrite, per OpenAI:

1. Your words, turned into short keyword-style searches.
2. Your rough location, estimated from your IP address. "Good restaurants near me" becomes "top restaurants San Francisco." OpenAI says it doesn't send your IP address or account details to the search provider.
3. Your saved memories, if memory is on. A user who once said they're vegan gets "good vegan restaurants San Francisco."

How many sub-queries run has changed a lot over time. Nectiv measured an average of 7.61 per prompt in August 2026, up from 2.17 a year earlier. Our ChatGPT rank tracker guide covers what this fan-out means for measurement, and our glossary defines query fan-out. The point here is simpler: you're competing for the sub-queries, not for the prompt the user typed.

Where ChatGPT Search Gets Its Candidate Pages

Once the queries exist, ChatGPT needs results. OpenAI names four kinds of input across its pages, each with a different control.

Two details matter most. First, OpenAI's crawler page says sites that opt out of OAI-SearchBot "will not be shown in ChatGPT search answers, though can still appear as navigational links." OpenAI's publisher FAQ adds that in its Atlas browser, a disallowed page found through a search provider or another page may still show as a bare link and title. A noindex tag stops that, but only if the crawler is allowed to read it. Robots.txt changes take about 24 hours to take effect.

Second, OpenAI keeps some kind of index of its own. Its API web search guide offers a mode that uses "only cached/indexed results" instead of live fetching. That's the developer API, not the ChatGPT app, but it shows OpenAI's web search can run from stored results, not only live lookups.

Studies suggest the pool is uneven. In Ahrefs' 2026 data, URLs that came through ChatGPT's general search channel were cited 88.5% of the time, while Reddit URLs pulled through a separate channel were cited 1.9% of the time. Reddit made up 67.8% of all retrieved-but-uncited URLs. ChatGPT reads Reddit a lot, it seems, but rarely credits it.

How Citations and the Sources Panel Are Shown

"Cited by ChatGPT" can mean three different things, and trackers don't always separate them.

• Inline citations sit next to a claim in the answer. OpenAI says you can select one to open the source, and on desktop web you can hover to preview it.
• The Sources panel opens from a Sources button under the answer. OpenAI says it shows "cited sources and other relevant links," so it lists more than the inline citations.
• Images and maps can link to their source pages. On iOS and Android, local answers may include a map.

OpenAI's API docs make the split explicit. Inline citations "show only the most relevant references," while a separate sources field "returns the complete list of URLs the model consulted." The docs add that "the number of sources is often greater than the number of citations." The ChatGPT app may differ from the API, but the principle matches what the help page describes.

One change in 2026 muddies the panel further. Since 5 May 2026, the Sources icon can also show "memory sources": saved memories, past chats and custom instructions that shaped a personalized answer. Those aren't web pages, and a shared chat doesn't show them.

Ads and shopping sit outside all of this. OpenAI's ads help page says "Ads do not influence ChatGPT's answers" and that ads appear below the response, labeled as sponsored. Its shopping help says product results "are selected independently" and aren't ads.

What Independent Tests Found About Which Pages Get Cited

OpenAI publishes no citation factors beyond "relevant, reliable information." Everything below is correlation from third-party studies. Each used different prompts, dates and methods, so read them as patterns, not as rules. Our guide to the seven ChatGPT ranking factors grades each claimed factor by the strength of its evidence.

Titles and URLs do early work

The Ahrefs 2026 study leans on outside research suggesting each search result reaches the model with a title, a short snippet and a URL, and the model picks which pages to open from those. OpenAI hasn't confirmed this. But it fits the numbers: cited pages scored 0.602 on title-to-prompt similarity against 0.484 for skipped ones, and 0.656 against the best-matching sub-query. Ahrefs measured similarity with open-source embedding models as a stand-in, since ChatGPT's scoring is private.

Ranking helps less than you'd think

Fewer than 1 in 10 ChatGPT citations in the 2025 overlap study ranked in Google's or Bing's top 10 for the user's prompt. One likely reason is the rewriting step: ChatGPT looks for pages that match its own sub-queries, not your keyword. That's an inference, not something OpenAI states.

Fresh, but not brand new

Across all its citations, ChatGPT skews toward newer pages than Google's results do. Yet within a single prompt, Ahrefs found the newest retrieved pages were often the ones left uncited, and the median cited page was about 500 days old. For news questions, where relevance scores were close, newer pages won.

The Citation Path Check: Trace One Page Through Five Checkpoints

When a page you expected isn't cited, don't guess. Walk it through the pipeline in order and stop at the first checkpoint it fails. We call this the Citation Path Check. It's a diagnosis tool; for the fixes themselves, our playbook on how to get cited by ChatGPT goes step by step.

A worked example with Tallyfold

Tallyfold is a fictional invoicing app for agencies; the pages and numbers below are made up for the example. A buyer asks ChatGPT, "How much does Tallyfold cost for a five-person agency?" Across 10 runs, ChatGPT cites a review site 7 times and Tallyfold's own pricing page twice. The team traces the pricing page:

1. Searched: pass. All 10 answers show a Sources button.
2. Eligible: pass. Logs show OAI-SearchBot fetching tallyfold.example/plans every few days.
3. Retrieved: pass. The page ranks third in Bing for "Tallyfold pricing."
4. Shortlisted: fail. The page title is "Plans," and the URL says /plans. The likely sub-query is "Tallyfold pricing per user 2026." The review site's title is "Tallyfold Pricing 2026: Cost Per User Explained."
5. Cited: fail. The page opens with "Simple plans for growing teams," and the per-user price sits in a table image with no text version.

The fix targets checkpoints 4 and 5: retitle the page "Tallyfold Pricing: Plans and Cost Per User," add a /pricing URL that redirects from /plans, and open with one sentence stating the price for five users in text. Then rerun the same 10 prompts after a few weeks. A move from 2 to 5 citations out of 10 would be encouraging, but 10 runs is a small sample; our guide to measuring GEO explains how many runs you need before a change is real.

Rankbox can help with checkpoint 5 on new pages. Its Citation-Ready Writer researches the live web and drafts source-backed articles that open with a direct answer. It doesn't track ChatGPT citations or fix crawler access, so pair it with the free robots.txt tester. The Business plan is $49.50 a month; see pricing.