A buyer asks ChatGPT about your category. The answer names three companies and cites a handful of websites, and none of them is yours.
That doesn’t automatically mean your content is weak. ChatGPT may have answered without searching, its crawler may not have been able to reach your page, or another source may have matched the exact question more closely. Those are different problems with different fixes.
One limitation up front: OpenAI documents its crawlers and how it rewrites search queries, but it doesn’t publish how sources are ranked or chosen. Everything below about selection comes from independent studies with their own samples and methods. Read it as evidence to test, not a description of how ChatGPT works internally.
A search-backed ChatGPT answer happens in stages. OpenAI describes the first three in its help documentation; the fourth is where it says the least.
A page can pass one stage and still lose at the next.
Searching doesn’t happen every time. One clickstream analysis estimated ChatGPT used web search on 34.5% of prompts in February 2026, down from 46% in late 2024. That’s one dataset’s estimate, not a fixed rate for every query type.
Rewriting usually produces a couple of searches. In an April 2026 study of 16,851 queries, 88.6% produced exactly two follow-up searches, and 8.8% produced none.
Selecting is where most retrieved pages drop out. In a March 2026 study of 15,000 prompts, ChatGPT retrieved 548,534 pages and cited only 15% of them.
Imagine a pharmaceutical distributor asks ChatGPT: “Which cold-storage warehouses near Navi Mumbai can hold vaccines at 2–8°C and meet GDP standards?”
ChatGPT might search for phrases like “pharma cold storage Navi Mumbai 2–8°C” or “GDP-certified warehouse Mumbai”. Those are illustrations, not the exact searches it would run. The results might include three kinds of source:
All three are about warehousing. Only the second and third give specific evidence for this request.
Now suppose the first company actually meets every requirement. But its temperature range appears only inside an image, its GDP certificate is a scanned PDF with no text summary, its CDN challenges automated crawlers including OAI-SearchBot, and the directory lists it under “dry storage” only.
The company doesn’t need a new strategy. It has four fixable gaps across three stages: access, readable facts, and an outdated third-party listing.
Crawler access is the part of ChatGPT search OpenAI documents most clearly, and the easiest for a site owner to check. OpenAI lists separate user agents, and each setting is independent of the others.
| User agent | What OpenAI says it’s for | What blocking it means |
|---|---|---|
| OAI-SearchBot | Surfacing websites in ChatGPT’s search features | Your pages won’t be shown in ChatGPT search answers, though they can still appear as navigational links |
| GPTBot | Crawling content that may be used to train OpenAI’s models | Signals your content shouldn’t be used for training; it doesn’t opt you out of search |
| ChatGPT-User | Some page visits triggered by a user’s request | Not used to decide search inclusion; robots.txt rules may not apply |
| OAI-AdsBot | Checking pages submitted as ChatGPT ads | Only visits ad landing pages |
Four details are easy to miss:
The findings below come from two independent studies published in March and April 2026. Each used its own prompts, dates and methods, and the April study collected data from ChatGPT’s interface. Treat them as associations to test, not rules that cause citations.
In the April study, pages in the first position of ChatGPT’s search results were cited 58.4% of the time, against 14.2% at position ten. In the March study, pages ranking first on Google were cited at 3.5 times the rate of pages outside Google’s top 20.
In the April study, pages whose headings closely matched the query were cited 41% of the time, against about 29–30% for weak matches. In the March study, pages with strong title-to-query overlap were cited at more than twice the rate of pages with little overlap.
Among pages that already matched the query well, those covering a quarter to half of ChatGPT’s follow-up searches were cited slightly more often (38.2%) than pages covering all of them (34.0%).
In the March study, 32.9% of cited pages that appeared in Google’s top 20 were found only through a follow-up search, not the original prompt.
The April study ran each query three times. Only 2.3% of page and query combinations were cited in all three runs. One check tells you very little.
None of this is a ChatGPT trick. It points back to familiar work: rank well, answer a specific question directly, and state the facts plainly.
It’s tempting to look at one ChatGPT answer and announce why you weren’t cited. We can see the outcome much more easily than the decision behind it, so use this table as a starting point.
| What you observe | Possible explanations | What to check next |
|---|---|---|
| Your logs show no OAI-SearchBot requests | It hasn’t visited yet, logs are incomplete, or access is blocked | robots.txt, CDN and firewall rules, OpenAI’s published IP ranges |
| The answer shows no sources at all | ChatGPT answered without searching | Repeat with search turned on, or ask a more current or specific question |
| Competitors are cited and you aren’t | A better-matched or higher-ranked page, or your facts aren’t readable | Your search ranking for the likely queries, and whether key facts are in text |
| Your brand is named but your URL isn’t | The answer leans on a third-party source | Which source was cited, and whether it describes you accurately |
| You’re cited one day and missing the next | Normal variation between runs | Repeat checks before treating it as a trend |
The first row deserves extra care. No logged visit isn’t proof of a block. Logs are often incomplete, and crawlers don’t fetch every URL on demand. Treat missing hits as a reason to test access directly.
Pick five to ten buyer questions, including a few with specific requirements. For each one, record:
For visits, OpenAI’s publisher FAQ says ChatGPT adds utm_source=chatgpt.com to referral links, so you can filter for it in analytics. That counts clicks only; it can’t show answers where you appeared and nobody clicked.
Then decide whether there’s a fix worth owning: a crawler-access problem, a fact locked in an image or PDF, a wrong directory listing, or a missing page. Sometimes the right call is to leave a low-value question alone.
ChatGPT search is well documented at the start and much less transparent at the end. OpenAI tells you how to be eligible: let OAI-SearchBot reach your pages. After that, independent studies point to familiar strengths: strong search rankings, pages that answer a specific question, and facts in plain text.
Measure over repeated runs, not single answers.
ApexaRank.ai helps teams monitor brand mentions, citations and competing sources across selected AI prompts, then turn the gaps into work for content, SEO and development.
CHECK YOUR AI VISIBILITYSee where you stand, uncover visibility gaps, and discover what to improve next.
Get your first visibility score in minutes.