GEO Articles
Which companies does GPT cite? — Public criteria and unpublicized aspects
We have organized the access paths and control measures actually documented by OpenAI and Google, distinguishing them from the undocumented selection sequences.
Quick Conclusion
- The fixed formula and selection order by which GPT chooses vendors have not been disclosed.
- What has been disclosed are the roles and control methods for OAI-SearchBot, GTBot, and ChatGPT-User.
- Blocking model training and ChatGPT search exposure can be separated from each other.
- Collection is controlled via robots.txt, and noindex is used to block search exposure as well.
- Google AI features do not have separate requirements or dedicated schemas. The key is to place unique information in accessible HTML.
First, we distinguish — what is public and what is not
The fixed formula by which GPT cites specific vendors has not been disclosed. What OpenAI and Google have disclosed is the role of the bots, access paths, and means of display control. The scores and selection order of sources are outside the scope of disclosure.
What is documented and what is not (Checked on 2026 8 23) (1, 2, 3, 4)
| Category | OpenAI | |
|---|---|---|
| Documented | The types and purposes of bots accessing the page, robots.txt control methods and reflection times, IP lists per bot, how sources are displayed via inline citations and the Sources area, and the fact that display can be blocked with noindex | The fact that AI Overviews and AI Mode use existing search indexing and quality systems, that no additional requirements or dedicated structured data are needed, and that snippet control applies to AI functions |
| Undocumented | How pages are evaluated based on criteria and scores, and in what order. The only publicly available explanation is a single line: 'signals the page is relevant to a user's query' | The order and weight in which specific pages are selected in AI responses |
| Guaranteed | None. Public websites can only 'appear' in ChatGPT search | None |
GPT citations cannot be determined solely by applying schema markup or ranking first in Google. On the other hand, settings that block access and display can be checked in the official documentation and modified immediately.
Three paths through which ChatGPT reaches pages
OpenAI operates separate bots with different purposes, and each setting operates independently. If you handle model training blocking and ChatGPT search display with the same setting, you can block desired exposure as well.
OpenAI Bot Purposes and Controls (Checked on 2026 8 23) (1)
| Bot | Purpose | robots.txt applied | If blocked | IP list |
|---|---|---|---|---|
| OAI-SearchBot | Collection for displaying websites in ChatGPT's search function | Applied | The possibility of appearing in ChatGPT search answers is removed. However, if the URL is obtained through a third-party search provider, the link and page title may be displayed | openai.com/searchbot.json |
| GPTBot | Collection of content that may be used to train Generative AI Foundation models | Applied | Excluded from training. Regardless of search visibility | openai.com/gptbot.json |
| ChatGPT-User | Page visits when requested by a user in ChatGPT | May not apply | Difficult to control via robots.txt as it is a user-initiated action | openai.com/chatgpt-user.json |
| OAI-AdsBot | Policy compliance verification for pages submitted as ChatGPT ads | — | Visits only pages submitted as ads, and collected data is not used for model training | openai.com/adsbot.json |
Because these actions are initiated by a user, robots.txt rules may not apply.
Since this action is initiated by a user, robots.txt rules may not apply. — OpenAI's description of ChatGPT-User (1)
In practice, there are three points
- If you want to deny training but still appear in search results, block only GPTBot and allow OAI-SearchBot. These two settings are independent of each other.
- It may take approximately 24 hours for robots.txt changes to be reflected in ChatGPT search.
- Control collection access via robots.txt, and use noindex if you want to block search visibility as well. (1, 2)
# Exclude collection for model training
User-agent: GPTBot
Disallow: /
# Allow collection for ChatGPT search display
User-agent: OAI-SearchBot
Allow: /
Sitemap: https://example.com/sitemap.xml
How Sources Appear
Sources for ChatGPT search answers are displayed in inline quotes and in the Sources area. Clicking a quote takes you to the original text, and a preview is available on desktop. The order in which sources are selected is not disclosed. (3)
Google AI Features Do Not Have a Separate Schema Requirement
Google states that no separate requirements or special optimizations are needed for exposure in AI Overviews and AI Mode. The following two sentences are the standards for Google AI features.
Therefore, llms.txt is not a separate signal for Google AI feature exposure. Google explicitly states that a text file dedicated to AI is not required.
Means to Actually Control AI Response Exposure
While there are no separate entry requirements, there are means to limit exposure and usage scope. You must apply settings that are distinct to suit your purpose.
Controls and Scope (Checked on 2026 8 23) (2, 4, 5, 6)
| Means | What it does | Range |
|---|---|---|
| noindex | Excludes the page from both search results and AI features. This is a measure to prevent it from being displayed in ChatGPT search as well. | Page |
| nosnippet | Does not display snippet text for the page. | Page |
| max-snippet:[n] | Specifies the upper limit for the length of snippets displayed. | Page |
| data-nosnippet | Excludes only specific areas within a page from snippets. | Element |
| Search generative AI control (Search Console settings) | Excludes the site from generative AI features in AI Overviews, AI Mode, and Discover. It is not used as a signal for inclusion in rankings or other search areas, and takes 1 to 2 days to reflect. Currently rolling out to select site owners. | Site |
| Google-Extended | Controls Gemini model training and grounding usage. It does not affect search inclusion and is not used as a signal for ranking. | Site |
First, check that no exclusion settings such as noindex, nosnippet, or data-nosnippet remain on the cited page.
Three Common Misconceptions
- Blocking Google-Extended does not affect Google Search inclusion or ranking. (5)
- llms.txt is not a separate exposure signal for Google AI features. (4)
- There is no dedicated schema.org structured data for AI feature exposure. (4)
Recent Changes in Structured Data
Google stopped displaying FAQ rich results on 2026 5 7 and removed related documentation on 6 15. FAQPage markup is no longer a qualifying element for Google rich results. (7)
The visible question and answer body remains valid. Question-type subheadings and the answers immediately following them serve as citation units linking user questions and documents.
Structured data must match the on-screen body. Including information that is not in the body only in the JSON-LD violates the Structured Data policy.
What Remains After Access
After organizing access and display settings, verify whether the document directly answers the question. The selection explanation released by OpenAI is 'signals that are relevant to the question,' and a separate list of ranking factors has not been disclosed. (2)
Document Requirements for Source Candidates
| Requirements | Check on the page | Why it is needed | ||||
|---|---|---|---|---|---|---|
| Accessibility | 200 response, robots allowed, canonical and sitemap point to the representative URL | Inaccessible documents do not become candidates | ||||
| Response to Question | The title and first paragraph answer a single question | The unit used for the answer is a paragraph, not a site | Source Source | Proprietary data, procedures, measurement conditions, time of verification, limitations | Reasons to cite only information that cannot be taken verbatim from other documents | |
| Publisher | The operator, author, contact information, and official domain must be consistent across multiple pages | It is difficult to use as evidence if the claim cannot be verified | ||||
| Timeline Attribution | Publication and revision dates, clean up discontinued services and broken links | Documents with unknown timestamps are difficult to cite in their current state | ||||
| Location in Text | Core information is in HTML text, not in images or scripts | Information that does not exist as text cannot be cited |
The items below are the minimum requirements for becoming a candidate source, not company rankings.
View Related Standards
- Check robots.txt·canonical·sitemap·JSON-LD: Website Diagnosis for AI Search Exposure 25 Checklist
- Comparison of Naver Search and Generative AI Citations: Differences between GEO Agencies and Naver SEO Agencies
- GEO Work Scope and Costs: How much do SUMMITFEED Basic, Standard, and Premium GEO plans cost?
Four Checks You Can Run Today
By checking the 4 steps below in order, you can identify which stage—collection, display, or citation—the site is being blocked at.
1 Step · robots.txt
- Is OAI-SearchBot allowed?
- Is the GPTBot policy separated from OAI-SearchBot? (Is it not blocking both learning denial and search display together?)
- If changed, check again after approximately 24 hours — this is the reflection time guided by OpenAI (1)
2 Step · Server Log Comparison
Check the OAI-SearchBot user-agent and the requesting IP together in the access logs. The IP must be compared against the public range in openai.com/searchbot.json to determine it as a genuine visit. GPTBot, ChatGPT-User, and OAI-AdsBot are verified in the same way using their respective JSON addresses. (1)
3 Step · Check Appearance Controls
- Is no index remaining on the page you want to be cited?
- Is nosnippet or max-snippet excessively shortening the snippet?
- Is data-nosnippet applied to the core body area?
- If your account exposes Generative AI controls in Search Console, are the settings correct as intended? (4, 6)
4 Step · Check Citations
Run with the same set of questions and check the original response and citations It stores the URL and collection time. It also verifies whether the cited URL is actually connected. The determination procedure can be found in GEO Practice Checklist and Measurement Criteria, and metric calculations in GEO Performance Indicators 5.
Google impressions can be viewed in the Generative AI performance report in Search Console. This report provides link impressions for AI Overviews and AI Mode, but does not include clicks, click-through rates, or average ranking. (8)
Conclusion — No formula, but much to verify
The fixed formula and selection order by which GPT cites vendors have not been disclosed. Instead, the following four criteria can be verified:
- The paths ChatGPT takes to reach pages are categorized by purpose.
- Different controls are used for collection access and search display.
- Google AI features do not have additional requirements or dedicated structured data.
- The disclosed selection description is a 'signal that it is relevant to the question.'
Two things to do now
- Remove unintended blocking from robots.txt, noindex, and snippet settings.
- Place your own data, actual procedures, measurement conditions, checkpoints, and limits in the public HTML body.
After fixing the settings, repeat the measurement of citation changes with the same questions and conditions.
Frequently asked questions
Does ranking first in Google mean a company will also be cited by GPT?
No. Search exposure increases the likelihood of a page being discovered, but it does not guarantee citation in ChatGPT. OpenAI only states that public websites can appear in ChatGPT search and has not disclosed the formula for which sources are selected and in what order. Google rankings and GPT citations must be managed using different metrics.
Does allowing OAI-SearchBot increase GPT citations?
Allowing it is a necessary condition, but not a sufficient one. OpenAI advises against blocking OAI-SearchBot to appear in ChatGPT search, but states that simultaneous display is not guaranteed. If you have modified robots.txt, it may take approximately 24 hours for the changes to be reflected in search results, so please check again after that.
If I block GPTBot, will it disappear from ChatGPT search as well?
No. GPTBot is the collection for model training, and OAI-SearchBot is the collection for search display; the two settings operate independently of each other. If you want to appear in search results without using it for training, you simply need to block GPTBot and allow OAI-SearchBot. However, even if you block OAI-SearchBot, links and page titles may still be displayed if the URL is obtained through a third-party search provider; to completely block display, you must use noindex.
Does adding a lot of structured data make GPT cite companies better?
There is no dedicated structured data for AI feature exposure. Google states in its documentation that "there is no special schema.org structured data to add." Structured data is intended for rich result eligibility and conveying document information; if included, it must match the content of the body text displayed on the screen. For reference, Google stopped displaying FAQ rich results starting from 2026 5 7 and deleted related documents on 6 15, so there is no longer any reason to maintain FAQPage markup for the purpose of rich results.
Does creating llms.txt help with Google or GPT citations?
According to Google standards, it does not help. Google states in its documentation that "there is no need to create new machine-readable files, AI-specific text files, or markup to appear in these features." The priority is open HTML content, crawl accessibility, unique information, and clear internal links.
Conversely, how do I prevent our page from appearing in AI answers?
The method varies depending on the purpose. To exclude it from both search and AI features, use noindex; to reduce the exposure of snippets only, use nosnippet, max-snippet, or data-nosnippet. To exclude only Google's AI Overviews, AI Mode, and Discover Generative AI features, use the Generative AI controls in Search Console; this setting is not used as a ranking signal and takes 1 to 2 days to be reflected. To block only Gemini training and grounding, use Google-Extended, which does not affect search inclusion or ranking.