# Where Does ChatGPT Get Its Information? A Brand Source Audit

> Where does ChatGPT get its information about your brand? Audit the pages, third-party sites and licensed feeds it cites, step by step, with a free tool.

- Author: Sheetal Shakshi, Founder's Office, Nurdd
- Published: 2026-11-12
- Publisher: Noma by Nurdd (https://nomahq.app)
- Original page: https://nomahq.app/blog/where-does-chatgpt-get-its-information

Where does ChatGPT get its information about brands? From four places: its training data, live web searches, content licensed from publishers, and the user's own context, such as prompts, files and memory. For a brand, the live search and licensed layers are the ones you can audit, by reading the sources ChatGPT cites.

## Key takeaways

- Audit by **source type, not by ranking**: your own pages, review and comparison sites, community threads, licensed publishers and reference sources each need a separate check.
- Record cited **domains**, not just URLs. [Ahrefs, November 2026](https://ahrefs.com/blog/where-does-chatgpt-get-its-information/), reports that only about a quarter of the pages cited for a prompt were cited again the next day.
- Expect the source mix to change with the question. For brand questions, Ahrefs found 59% of citations went to company websites; for "best X" questions, review sites led with 27%.
- Check crawl access first. OAI-SearchBot, GPTBot and ChatGPT-User are three separate agents with three separate jobs.
- Google and Bing rankings do not tell you whether ChatGPT cites you. Read the answers themselves.
- Run the audit on a fixed prompt set, repeat it, and compare like with like.

## Where does ChatGPT get its information about a brand?

Ahrefs' explainer, [published in November 2026](https://ahrefs.com/blog/where-does-chatgpt-get-its-information/), splits ChatGPT's sourcing into four layers. The table below keeps that structure and adds the column most explainers leave out: what a brand can actually do with each layer.

| Layer | What it is | Can a brand audit it? | What to check |
| --- | --- | --- | --- |
| Training data | Public web content, partner data, and material from users, human trainers and researchers (as described in [Ahrefs' summary of OpenAI's statements, November 2026](https://ahrefs.com/blog/where-does-chatgpt-get-its-information/)) | Not directly. You cannot see what a model learned | Whether GPTBot is blocked, and what ChatGPT says about you with no search |
| Live web search | Searches ChatGPT writes and sends to search providers, plus OpenAI's own crawler | Yes, through the sources cited in answers | Cited domains, crawl access, which of your pages appear |
| Licensed publisher content | Summaries from publishers with OpenAI deals, shown with attribution and links | Partly: you can see whether partner sites are cited for your prompts | Whether partner sites that cover your category name you |
| The user's own context | Prompt, files, memory, chat history, connected apps and approximate location | No. It differs for every user | Nothing to fix; account for it as noise in manual tests |

That last row matters for testing. Ahrefs showed the same question, "what is a backlink", answered differently in a logged-out account and in a Business workspace, with the second mentioning Ahrefs. A manual check on your own logged-in account is therefore not a neutral sample.

### What the training layer means for you

Per [Ahrefs' reading of OpenAI's statements, November 2026](https://ahrefs.com/blog/where-does-chatgpt-get-its-information/), OpenAI's models are developed from public internet information, information from third-party partnerships, and material provided or generated by users, trainers and researchers. Ahrefs notes that the last detailed breakdown OpenAI published was for GPT-3 in 2020, so nobody outside the company knows the current mix. If a public site does not block GPTBot, its pages may reach future models. The effect is slow and invisible, which is why the audit below concentrates on the layers that leave a trail.

## How does ChatGPT use live search to answer?

OpenAI's [ChatGPT search help page](https://help.openai.com/en/articles/9237897-chatgpt-search) (as cited by Ahrefs, November 2026) says ChatGPT searches based on your prompts and "may share disassociated search queries with third-party search providers such as Bing". The page also says responses that use search carry inline citations, and a Sources button lists every cited link. That button is what a manual audit reads.

Two findings from Ahrefs shape how you should audit:

- **Fan-out.** In a study of 66.6 million logged-out ChatGPT responses (August to September 2026), responses with fan-out queries held about 1.6 each, and 6.8% of those queries used the `site:` operator. Three in four of the `site:` searches went to brand websites.
- **Weak overlap with rankings.** In Ahrefs' 2025 study of 15,000 prompts, about 8% of ChatGPT's in-text citations ranked in Bing's top 10, and about 8% in Google's top 10.

The second point means a rank tracker cannot answer the question. A page can rank first and never be cited; a page you have never tracked can be cited every day.

### Crawl access: three agents, three jobs

OpenAI's [crawler documentation](https://developers.openai.com/api/docs/bots) separates GPTBot (training), OAI-SearchBot (ChatGPT search) and ChatGPT-User (visits triggered by a user, such as a pasted link). Blocking GPTBot does not remove a site from ChatGPT search results. Blocking OAI-SearchBot, or a CDN rule that rejects it, can. Check this before anything else with the free [AI crawler checker](https://nomahq.app/tools/ai-crawler-checker), which shows whether each OpenAI bot can reach your site and generates a corrected robots.txt. The [AI Crawler Policy Kit](https://nomahq.app/resources/ai-crawler-policy-kit) covers the training-versus-search decision.

## How can I find out which sources ChatGPT cites about my company?

Run a fixed set of buyer questions, open the Sources list for each answer, and log every cited domain by type. Repeat on different days, then count domains rather than URLs. The steps below are the audit; the existing [citation sources study method](https://nomahq.app/blog/ai-search-citation-sources-study) covers one-question research for publishing, whereas this is a brand-level coverage check.

1. **Write 15 to 25 prompts across question types.** Include brand questions ("what is [brand]"), category questions ("best [category] for [use case]"), comparison questions and, if you sell products, shopping questions. The [prompt selection guide](https://nomahq.app/blog/choose-prompts-to-track) explains the mix.
2. **Fix the conditions.** Use a logged-out session or a clean workspace, the same country, and write down the date. Memory and location change the searches ChatGPT writes, as [Ahrefs, November 2026](https://ahrefs.com/blog/where-does-chatgpt-get-its-information/), reports from OpenAI's documentation.
3. **Run each prompt at least three times on different days.** Ahrefs' reading of Glen Allsopp's four-week tracking of 1,300+ prompts is the reason: domains cited on 80% or more of days changed about 14% day to day, against 74% for all other domains.
4. **Log every cited source.** Columns: prompt, date, domain, URL, source type, whether your brand is named on that page, whether the description is accurate.
5. **Group by domain and count.** The domains that recur are your real source set. Ignore one-off URLs.
6. **Score coverage and accuracy** using the matrix below, then assign each gap to an owner.

To skip the spreadsheet for ChatGPT alone, the [ChatGPT visibility tracker](https://nomahq.app/features/chatgpt-visibility) from Noma by Nurdd shows how often ChatGPT names your brand, which sources it cites and which rivals appear instead. For the manual method, see [how to track brand mentions in ChatGPT](https://nomahq.app/blog/track-brand-mentions-in-chatgpt).

## Which source types should a brand check, and in what order?

Ahrefs sorted cited sources by question type from 4,000 US ChatGPT prompts pulled from its Brand Radar in October 2026. Those figures give you a public reference for what a normal source mix looks like per prompt type. They are one vendor's sample, not a universal benchmark, so use them to spot anomalies rather than set targets.

| Prompt type | Most typical cited source | Share of that type's citations | What it tells a brand to check |
| --- | --- | --- | --- |
| Brand questions | Company websites | 59% | Is your own site crawlable, current and clear on facts? |
| "Best X" and comparisons | Review and advice sites | 27% | Are you listed, and described correctly, on those sites? |
| Shopping | Retailers | 28% | Is product data complete on retailer pages and feeds? |
| Definitions | Dictionaries and encyclopedias | 63% | Is your category term linked to you on reference sources? |
| Health | Government, medical and academic sites | 82% | Regulated brands: is your approved claim on authority sources? |

Source: [Ahrefs, November 2026](https://ahrefs.com/blog/where-does-chatgpt-get-its-information/).

Read your own log against it. If your brand prompts cite mostly third-party pages and few of your own, the first fix is on your site. If your "best X" prompts cite review sites you are absent from, the fix is earned coverage, not a content rewrite.

### The source-by-source coverage matrix

This is the scoring sheet. For each recurring domain in your log, mark one status and one owner.

| Source type | How to find it in your log | Status to record | Typical owner |
| --- | --- | --- | --- |
| Your own pages | Your domain in the cited list | Cited / not cited / wrong page cited | Web and content |
| Review and comparison sites | Third-party "best of" lists, directories | Listed and accurate / listed and wrong / absent | Product marketing, PR |
| Community sources | Reddit, forums, YouTube | Named / not named / negative | Community, PR |
| Licensed publishers | Domains of OpenAI's publisher partners | Covers you / covers category without you | PR |
| Reference sources | Wikipedia, Wikidata, dictionaries | Entity exists and correct / missing | Brand, SEO |
| Product and structured feeds | Retailer and marketplace pages | Complete / incomplete attributes | E-commerce |

A worked example of the logic, with illustrative numbers rather than real data: a software brand logs 20 prompts and finds 12 domains that recur. It appears accurately on 4, is named wrongly on 2 and is absent from 6. Its coverage is 4 of 12 and its accuracy gap is 2 of 12. The 6 absences go to PR and product marketing; the 2 errors go to whoever can request a correction. That split is the output the audit exists to produce.

## How do licensed publishers change the picture?

Ahrefs reports that OpenAI has signed content deals with more than 20 publishers since 2023, and that licensed content appears as summaries with attribution and a link to the original. In its analysis of 85 partner domains, ten were among the 50 sites ChatGPT cites most, and partner sites' share of citations among the 200 most-cited domains rose from 7.5% in March 2026 to 15.8% in July, then 12.5% in late September.

The same analysis stresses a limit. Vogue was cited 42,108 times; The Atlantic, also a partner, 724. Ahrefs' summary: "A partnership makes a publisher eligible to show up in ChatGPT, but it doesn't guarantee it." Cite-worthy partner sites skew to review and lifestyle titles, so for product categories the audit step is to ask: which partner sites already appear in my log, and do they mention me?

If they do not, an [earned media plan for AI answers](https://nomahq.app/resources/earned-media-for-ai) starts from that list. If you are weighing a pitch, remember that a mention helps only on pages ChatGPT actually retrieves for your prompts. [Citation versus brand mention](https://nomahq.app/blog/citation-vs-brand-mention) explains why the two metrics differ.

## What if ChatGPT cites a source that gets your brand wrong?

Treat it as a source problem. Reproduce the answer, note the prompt, date and cited page, and find the claim in that page. Correct it where you control it, request a correction where you do not, then re-run the prompt. The [correction runbook](https://nomahq.app/resources/ai-correction-runbook) sets out the sequence, and [fixing brand confusion in AI answers](https://nomahq.app/blog/fix-brand-confusion-in-ai-answers) covers cases where another company shares your name.

Two preparations reduce errors at the source. A [brand fact block](https://nomahq.app/blog/brand-fact-block-for-llms) puts your name, category, audience and pricing in one quotable passage. A consistent [Wikidata entry](https://nomahq.app/blog/wikidata-brand-entity-ai) helps engines tell you apart from lookalikes. Ahrefs found that adding structured data to 1,885 already-cited pages did not change ChatGPT citations in any statistically meaningful way, so use schema for clarity, not as a citation lever.

## What does Noma track here?

Noma asks 13 AI engines, including ChatGPT, Gemini, Perplexity and Claude, the questions your buyers ask, every day. For each answer it lists the sources cited, so you can see which pages engines rely on, where competitors are named instead of you, and which topics you are absent from. It also reports which AI crawlers fetch your pages. It does not show what a model learned in training.

[Book a demo](https://nomahq.app/demo) to see it on your own brand, or [start the 3-day free trial](https://platform.nomahq.app/signup). A card is taken at signup so billing can start on day 3; cancel before then and nothing is charged.

## How should you read the results?

- **Mostly your own pages cited, but you are not named:** the content is retrieved but not persuasive. Rework the passage that answers the prompt.
- **Competitors on review sites, you absent:** earned coverage gap. Start with the three most-cited domains.
- **Own site never cited on brand prompts:** check OAI-SearchBot access, rendering and page titles. The [technical SEO for AI search checklist](https://nomahq.app/blog/technical-seo-for-ai-search) helps.
- **Different sources every run:** expected. Compare domains and use repeated runs, as in [AI citation stability](https://nomahq.app/blog/ai-citation-stability).

On Ahrefs' evidence, title match matters too: cited pages' titles were more similar to the user's question than retrieved-but-uncited ones (0.60 against 0.48). Check that the page you want cited has a title phrased like the question.

## How we researched this

We read Ahrefs' explainer on ChatGPT's information sources in full, scanned two other ranking articles on the same query (The Digital Bloom and LLM Pulse) for what they cover, and read OpenAI's crawler documentation and ChatGPT search help page as cited by Ahrefs. Study figures are Ahrefs' own and come from its stated samples; we have not re-run them. Facts, figures and links were checked on 2026-11-12. We did not run tests of our own for this guide.

## What to do next

Pick 20 buyer questions this week, log the cited domains for each and fill in the coverage matrix. Fix crawl access first, then your own pages, then third-party gaps. To automate the daily runs, [book a demo](https://nomahq.app/demo) or [start the free trial](https://platform.nomahq.app/signup).
