ChatGPT search works by deciding whether a prompt needs fresh information, sending search queries to third-party search providers and partner content, reading the pages that come back, and writing an answer with inline links to the sources it used. OpenAI confirms the broad pipeline. It does not publish a ranking formula, so anything beyond the pipeline is informed inference.
For a website, the practical part is short. OpenAI’s own crawler documentation says that sites which block OAI-SearchBot will not be shown in ChatGPT search answers. Blocking GPTBot is a separate choice that only covers model training. ChatGPT-User is a third agent that fetches pages when a person asks ChatGPT to open them. Mixing these three up is the most common technical mistake we see when a brand says ChatGPT “never cites us”.
Once access is right, citations depend on what any search system rewards: pages that are indexed, that answer the exact question clearly, and that a model can quote with confidence. Shopping results follow their own rules, based on product metadata and feeds rather than ads.
This guide walks through each stage, separates what OpenAI has documented from what the industry has inferred, and ends with a checklist you can run on your own site.
When does ChatGPT search the web instead of answering from memory?
ChatGPT searches automatically when it judges that web information would improve the answer, and users can also force a search. OpenAI’s Help Center says ChatGPT “will automatically search the web if your question might benefit from information on the web”. Users can trigger it manually with the web search option in the message box.
In practice, a prompt is more likely to trigger search when it involves:
- Recent events, prices, availability or anything that changes often
- Named local businesses, opening hours or “near me” intent
- Product comparisons and shopping intent
- Niche facts the model is unlikely to have learned well in training
Stable, general knowledge (“what is a canonical tag”) is often answered from the model’s trained knowledge without a search. That list of triggers is our observation from testing, not an OpenAI specification. OpenAI documents that the model decides; it does not publish the decision rules.
This split matters for brands. If an answer comes from training data, the only way to influence it is to be well represented across the web before a model’s training cutoff. If an answer comes from search, a well-structured page published last week can be cited today. Our guide on GEO vs SEO explains how the two kinds of work divide.
What happens between your prompt and a cited answer?
The pipeline has five stages, and OpenAI has confirmed the outline of each.
- Decide. The model decides whether to search, or the user forces it.
- Query. ChatGPT turns the prompt into one or more search queries. The Help Center says it may share “disassociated search queries with third-party search providers such as Bing”, along with a general location inferred from the user’s IP address.
- Retrieve. Results come from those providers and from “content provided directly by our partners”, in the words of OpenAI’s October 31, 2024 launch post. Named launch partners included the Associated Press, Axel Springer, Condé Nast, Financial Times, Hearst, Le Monde, News Corp, Reuters and Vox Media, among others.
- Read. ChatGPT reads the retrieved pages. Pages crawled by OAI-SearchBot can supply summaries and snippets.
- Answer and cite. The response includes inline citations and a Sources button that lists cited sources and other relevant links.
At launch, OpenAI said the search model was a fine-tuned version of GPT-4o, post-trained with synthetic data including outputs distilled from o1-preview. Models change often, so treat that as a historical detail rather than a description of today’s system.
What OpenAI does not document is how many queries are run per prompt, how results from different providers are merged, or how passages are scored. Industry studies suggest assistants often run several related “fan-out” queries for one prompt. That is a reasonable inference from observed behavior, not a published fact.
Which OpenAI crawlers matter, and what does each one control?
OpenAI documents four user agents, and only one of them decides whether your pages can appear in ChatGPT search answers. Each robots.txt setting is independent of the others.
| User agent | Purpose | What blocking it does | Published IP list |
|---|---|---|---|
| OAI-SearchBot | Surfaces websites in ChatGPT’s search features | Your site will not be shown in ChatGPT search answers, though it may still appear as a plain navigational link | openai.com/searchbot.json |
| GPTBot | Crawls content that may be used to train OpenAI’s foundation models | Signals that your content should not be used for model training; does not remove you from search | openai.com/gptbot.json |
| ChatGPT-User | Visits a page when a user asks ChatGPT or a custom GPT to do so; not an automatic crawler | OpenAI says robots.txt rules “may not apply” because the action is user initiated; not used to decide search inclusion | openai.com/chatgpt-user.json |
| OAI-AdsBot | Checks the safety of landing pages submitted as ads on ChatGPT | Only visits pages submitted as ads; data is not used to train foundation models | openai.com/adsbot.json |
Two details from the same documentation are easy to miss. First, robots.txt changes can take about 24 hours to affect search results. Second, if you allow both OAI-SearchBot and GPTBot, OpenAI may use the results of one crawl for both purposes to avoid duplicate visits.
How do you allow ChatGPT search but block model training?
Allow OAI-SearchBot, disallow GPTBot, and leave ChatGPT-User alone. A minimal robots.txt looks like this:
User-agent: OAI-SearchBot
Allow: /
User-agent: GPTBot
Disallow: /Before you rely on that, check four things that robots.txt alone will not fix:
- Firewalls and CDNs. Bot protection at Cloudflare, Akamai or your host can block OAI-SearchBot even when robots.txt allows it. Verify requests against OpenAI’s published IP ranges instead of trusting the user agent string, which anyone can fake.
- Noindex, not disallow, to hide a page. OpenAI’s publisher FAQ says a disallowed page may still show up as a link and title. To keep a page out, use a noindex meta tag, and let the crawler reach the page so it can read that tag.
- Rendering. If your main content only appears after heavy JavaScript runs, assume some crawlers will see an empty shell. Server-render the text that answers the question.
- Bing visibility. Because OpenAI names Bing as one search provider, a site that is poorly indexed in Bing is at a disadvantage. Check Bing Webmaster Tools. OpenAI does not say Bing is the only provider.
Deciding whether to allow training is a business choice, not an SEO requirement. Blocking GPTBot does not remove you from ChatGPT search, according to OpenAI’s documentation.
How does ChatGPT choose which sources to cite?
OpenAI has not published a citation algorithm, so this section is about evidence and inference, clearly labeled.
What is documented: sources come from third-party search providers and partner content, pages must be reachable by OAI-SearchBot to supply snippets, and the answer links to the sources it used.
What independent data suggests: an Ahrefs study published on August 11, 2025, covering 15,000 long-tail prompts, found that only 12% of links cited by ChatGPT, Gemini and Copilot appeared in Google’s top 10 for the same prompt. For ChatGPT specifically, about 8% of in-text citations ranked in Google’s top 10 and about 8% in Bing’s top 10. Ahrefs sells SEO software and the data was collected in early July 2025, so treat it as directional. Its own authors suggest the gap is partly explained by fan-out queries: a cited page may rank well for a related sub-query, not for the original prompt.
The academic evidence points the same way. The GEO paper by researchers from Princeton University and IIT Delhi (Aggarwal et al., accepted at KDD 2024) tested content edits on a generative engine benchmark. Adding quotations improved a visibility measure by about 41% over baseline, adding statistics by about 31% and citing sources by about 27%. Keyword stuffing performed below baseline. This was a research benchmark, not ChatGPT itself, so it supports a direction rather than a guarantee.
Taken together, the practical inference is:
- Rank for the specific sub-questions, not only the head keyword
- Answer each question in the first two sentences of its section
- Back claims with named sources and dates that a model can quote
- Keep entity facts (name, products, locations, prices) consistent across your site and third-party profiles, which our entity SEO guide covers
How does ChatGPT shopping choose products and merchants?
OpenAI states that product results “are selected independently by ChatGPT and are not ads, nor influenced by any OpenAI partnerships.” Ads, where shown, are separate from product results.
According to OpenAI’s shopping Help Center article:
- A product appears in the carousel when ChatGPT judges it relevant to the user’s intent, using the query and, where enabled, Memory and custom instructions
- Merchant lists are built from merchant and product metadata from third-party providers or directly from merchants
- Merchants are ranked on factors such as availability, price, quality and whether they are the maker or primary seller
- Shopify merchants are integrated through Shopify Catalog; other merchants can apply to send a product feed to OpenAI
- Some eligible products show an Instant Checkout option inside ChatGPT
OpenAI’s commerce documentation adds that the product feed is a CSV or JSON file with identifiers, descriptions, pricing, inventory, media and fulfillment options, refreshed regularly. Instant Checkout runs on the Agentic Commerce Protocol and is “currently available to approved partners”; OpenAI is not the merchant of record, and payment processing happens on the merchant’s systems.
For stores, that means clean product data matters as much as copy: accurate prices, stock, GTINs and Product schema on the page. Our ecommerce SEO guide covers the on-site side.
What can ChatGPT search get wrong?
ChatGPT search can still misread, misattribute or over-simplify sources, and OpenAI says so in its own shopping documentation: “the model can make mistakes.” The failure patterns we see most often when auditing brands:
- Stale facts. An old price or a discontinued product survives because a third-party page still lists it.
- Wrong entity. Two companies with similar names get merged into one answer.
- Citation without support. A link is attached to a claim the page does not actually make.
- Aggregator preference. A directory or marketplace is cited instead of the brand’s own page, because it answers the comparison more directly.
You cannot submit corrections through a publisher console; OpenAI’s publisher FAQ does not describe one. The fix is upstream: correct the source pages, update your own page so it states the fact plainly, and check the answer again after recrawling.
What does OpenAI confirm, and what is only inferred?
| Topic | Documented by OpenAI | Inferred or unknown |
|---|---|---|
| When search runs | Automatically when web info would help, or manually | Exact trigger rules |
| Where results come from | Third-party providers “such as Bing” and partner content | Full list of providers and how they are weighted |
| Search inclusion | Blocking OAI-SearchBot removes you from search answers | Crawl frequency and freshness thresholds |
| Training | GPTBot controls training use, independent of search | Whether training presence affects search citations |
| Citation choice | Answers link to sources used | Scoring factors; the role of structure, statistics and authority comes from outside studies |
| Shopping | Not ads; based on metadata and feeds; ranked by availability, price, quality | Relative weight of reviews versus feed data |
| Measurement | Referral URLs carry utm_source=chatgpt.com | Clicks from answers that show no link |
How do you optimize a site for ChatGPT search?
Start with access, then structure, then evidence, then measurement.
- Confirm access. Allow OAI-SearchBot in robots.txt, check that your firewall does not block OpenAI’s IP ranges and confirm that key pages are not noindexed by accident.
- Get indexed in Bing. Submit your sitemap in Bing Webmaster Tools and fix any coverage errors.
- Map the sub-questions. For each money topic, list the five to ten questions a buyer asks before choosing. Give each one a heading and a direct answer.
- Answer first. Put the answer in the first sentence under each heading, then the detail. Use tables for comparisons and lists for steps.
- Add verifiable evidence. Named sources, dates, specifications and prices. Avoid vague claims a model cannot attribute.
- Align third-party sources. Make sure directories, marketplaces, review sites and partner pages state the same facts as your site.
- Feed product data. Stores should keep Product schema accurate and consider OpenAI’s product feed if they are not on Shopify.
- Measure. OpenAI’s publisher FAQ confirms ChatGPT adds utm_source=chatgpt.com to referral URLs. Build a GA4 channel for AI assistants and track it monthly, as described in our SEO tracking guide.
This matters at scale. OpenAI CEO Sam Altman said ChatGPT had 800 million weekly active users at OpenAI’s DevDay on October 6, 2025, as reported by TechCrunch. Not every user searches, but the audience is large enough that being absent is a real cost.
Who should work with Marketer Coffee on ChatGPT visibility?
Marketer Coffee is an AI SEO and GEO agency that fits brands which sell across several markets or languages and want to be cited in ChatGPT, Perplexity and Google AI Overviews. Our evidence comes from different industries, which we state plainly:
- Planika, a design fireplace brand. Seven US-localized articles took a site Google had not indexed to 1,052 monthly US visits. Ahrefs recorded 401 Google AI Overview mentions, 136 Perplexity mentions and 20 ChatGPT mentions. Read the Planika case study.
- Landingi, a SaaS company. Roughly 5,000 sessions a month referred by AI assistants, tracked in GA4. See the Landingi case study.
- IDG-Online, CCTV services and affiliate. From 500 to more than 30,000 monthly clicks in 16 months against Ceneo, MediaExpert and MediaMarkt, using ranking pages, Q&A content and schema. Read the IDG-Online case study.
Who it does not fit: a business that needs a one-off robots.txt fix, or one that wants guaranteed placement in ChatGPT. Nobody can guarantee that, and an agency that promises it is guessing.
The first step is a free AI visibility audit that checks crawler access, Bing indexing and how ChatGPT currently describes your brand. More detail on our AI SEO service page.
Frequently asked questions about how ChatGPT search works
Does blocking GPTBot remove my site from ChatGPT search?
No. OpenAI’s documentation says GPTBot only controls whether content may be used for model training. Search visibility depends on OAI-SearchBot, and the two settings are independent.
Does ChatGPT search use Bing?
OpenAI’s Help Center says ChatGPT may share disassociated search queries with third-party search providers such as Bing. It also uses content provided directly by partners. OpenAI does not say Bing is the only source.
How can I tell if ChatGPT searched the web or answered from memory?
Answers that used search show inline citations and a Sources button listing the pages used. An answer without links was most likely generated from the model’s trained knowledge.
Can I pay to appear in ChatGPT shopping results?
No. OpenAI states that product results are selected independently and are not ads. Merchants can improve eligibility with accurate product metadata or, outside Shopify, by applying to send a product feed.
How do I track traffic from ChatGPT?
ChatGPT adds utm_source=chatgpt.com to referral links, according to OpenAI’s publisher FAQ. Create a GA4 segment or channel for that source and compare it with conversions.
What is the short version for marketers?
ChatGPT search is a retrieval system on top of a language model: let OAI-SearchBot in, be indexed where OpenAI’s search providers look, answer the exact sub-questions buyers ask, and back every claim with a source. Training data and search are separate doors, and you control each with a different crawler.
If you want to know how ChatGPT describes your brand today, which pages it cites instead of yours and what to fix first, send us your site and your main markets. Book a free consultation.
Sources: OpenAI, Overview of OpenAI Crawlers (developers.openai.com/api/docs/bots, accessed November 5, 2026); OpenAI Help Center, ChatGPT search; OpenAI Help Center, Publishers and Developers FAQ; OpenAI Help Center, Shopping in ChatGPT; OpenAI, Commerce key concepts and Agentic Commerce Protocol documentation; OpenAI, Introducing ChatGPT search (October 31, 2024, with updates of December 16, 2024 and February 5, 2025); TechCrunch, Sam Altman says ChatGPT has hit 800M weekly active users (October 6, 2025); Ahrefs, Only 12% of AI Cited URLs Rank in Google’s Top 10 for the Original Prompt (August 11, 2025); Aggarwal et al., GEO: Generative Engine Optimization, KDD 2024 (arXiv 2311.09735); Marketer Coffee, Planika, Landingi and IDG-Online case studies (other industries).
Pingback: How AI Remembers Brands Across Sessions (2026)