ChatGPT Search is OpenAI’s integration of live web search into ChatGPT. Instead of answering only from what the model learned during training, it retrieves current web pages, synthesizes an answer from them, and shows the sources it used. Ask it who provides a service in your city, or what something costs, and it will search, read, and reply in prose with citations rather than handing back a list of links.
It is available to everyone. OpenAI introduced it on 31 October 2024, opened it to all logged-in users on 16 December 2024, and made it available without a login on 5 February 2025. Any description that still calls it a ChatGPT Plus feature is out of date, and that matters for planning, because it means the audience is general rather than a subscriber niche.
How it works
ChatGPT Search is a retrieval system wrapped around a language model. When a question needs current information, it issues searches, fetches pages, and generates an answer grounded in what it retrieved, displaying the sources alongside. OpenAI has described it as using a fine-tuned model together with third-party search providers and content supplied directly by partners, which include major news organizations.
The practical consequence is that your pages can be read and quoted at the moment someone asks, without your content being in the training data at all. This is the same retrieval-augmented pattern used in business chatbots, applied to the open web, and it is why updating a page can change what ChatGPT says about you relatively quickly, where changing a model’s training knowledge is not something you can do at all.
The three crawlers, and why the distinction matters
OpenAI operates several crawlers with different jobs, each controllable separately in robots.txt:
GPTBot gathers content that may be used to train models.
OAI-SearchBot indexes content so it can appear in ChatGPT Search results.
ChatGPT-User fetches a page when a user’s question causes ChatGPT to visit it directly.
This separation is the important part. A publisher who does not want their work used for training can block GPTBot while still allowing OAI-SearchBot, and remain visible in search answers. The reverse mistake is more common and more costly: a blanket rule, an over-broad security plugin setting, or a firewall configuration that blocks all three, quietly removing the business from ChatGPT’s answers without anyone noticing. If AI visibility matters to you, confirm what your robots.txt and your security layer actually allow rather than assuming.
What gets cited
There is no published ranking formula, but the patterns are consistent with how retrieval systems work and with what we observe.
Content that directly answers the question. A page that states the answer plainly near the top is far easier to quote than one that builds to a conclusion in paragraph nine. Answer-first structure is the single highest-leverage change.
Clear structure. Descriptive headings, self-contained sections, and semantic HTML make a page easy to parse and extract from.
Specific, checkable facts. Concrete detail gives a retrieval system something to quote. Vague positioning language gives it nothing.
Content in the HTML. Text assembled by JavaScript may not be seen at all. This is a recurring theme across AI crawlability.
Corroboration. Claims supported by several independent sources are treated more confidently than claims that appear only on your own site. We cover this dynamic in winning the consensus layer.
Freshness. For anything time-sensitive, recently updated pages tend to be preferred.
Structured data. Schema markup states facts about your business unambiguously, which reduces the chance of being described incorrectly.
What it means for a business
Three effects are worth planning around.
Some research now ends without a click. A prospect comparing providers may read a synthesized paragraph and never visit anyone’s website. Being named and described accurately in that paragraph becomes a form of visibility in itself, even when it produces no session in your analytics.
The traffic that does arrive tends to be further along. Someone who clicks a citation after reading a summary has usually already narrowed their options, which is a different visitor from a broad search click.
Accuracy is now a marketing problem. If ChatGPT describes your services, location, or pricing incorrectly, that description reaches prospects directly. You cannot edit the answer, but you can change the sources it draws on, which is the whole practical response.
Improving your chances
Answer the real questions directly. Use the questions customers actually ask as headings and answer each one in the first sentences beneath.
State facts plainly. What you do, who you serve, where you work, how engagements run. Specifics are quotable; adjectives are not.
Keep information consistent everywhere. Your site, your Google Business Profile, directories, and professional profiles should agree. Contradictions produce vague or wrong descriptions.
Earn independent coverage. Mentions, reviews, and third-party references are what turn your claims into corroborated facts.
Make the site fast and readable. Clean markup, content in the HTML, structured data.
Check what it says about you. Ask ChatGPT the questions your customers would ask, and record the answers. This is ordinary competitive research now, and it is the only way to know whether any of the above is working.
This work is the substance of generative engine optimization and of our AI search optimization services. Very little of it is specific to ChatGPT, which is the good news: the same changes improve how you appear in AI Overviews, Perplexity, and Gemini.
Measuring it
Measurement is thin, and honesty about that is more useful than a dashboard implying otherwise. Referrals from chatgpt.com do appear in analytics, so you can see sessions and conversions from that source, though volumes are usually small relative to search. What you cannot see is the question that produced the citation, how often you were mentioned without a click, or how you were described. The practical substitutes are tracking referral traffic and its conversion quality, and running the same set of prompts on a regular schedule and recording what comes back, which is what our AI visibility plans do.
Should you block it?
For most businesses that want customers to find them, no. The calculation is different for publishers whose revenue depends on people reading their content on their own site, and OpenAI’s separate crawlers exist precisely so that decision can be made with some nuance: block training, allow search. Whatever you decide, decide it deliberately and write it down, rather than discovering later that a plugin default made the choice for you.
Limits worth knowing
ChatGPT Search reduces but does not eliminate error. It can misread a page, attribute a claim to the wrong source, blend two companies with similar names, or present an outdated figure confidently. Citations make this easier to catch than it was, since you can check the source, but the underlying tendency described in our entry on hallucination has not gone away. If you find it describing your business incorrectly, the fix is upstream: clearer facts on your site, corrected third-party listings, and more corroboration. If you want to know what it currently says about you and what would change it, book a discovery call.
