ChatGPT training data vs live web search
Current answers about who to use in a town usually come from live web search, not from what ChatGPT memorised when it was trained. Training data helps the model write and recall general knowledge. It is a poor source for who still trades in your town, who has moved, or which hours and reviews are current. OpenAI says ChatGPT can search the web for current information and links to sources. A clear, crawlable page helps. It does not guarantee a mention. AI results vary.
A lot of advice still treats ChatGPT as a closed book that only updates when a new model is trained. That leads people to do the wrong work, or to wait for a retraining that will not fix a blocked site or an outdated listing. The mechanism is also covered in how ChatGPT decides which businesses to name.
What is training data, in plain terms?
Training data is the text and other material used to teach a model language and a wide range of facts. It can include public web pages that existed at the time, among other sources. Once training finishes, that snapshot does not update itself when you change your opening hours, move office, or add a service page.
For a history question or a well known public fact, that stored knowledge may be enough. For "best emergency plumber in [town]" or "family solicitor near me", it is not. Practices close, websites change, and review counts move. An answer that cannot search will lean on whatever the model already knows, which may be stale, incomplete or simply silent about a small local firm.
I am not going to invent a knowledge cut-off date for a particular ChatGPT model. OpenAI changes those details. Training data is a past picture, and a local recommendation needs a current one.
What does live web search add?
OpenAI's help article on searching the web with ChatGPT says ChatGPT can search the web to answer questions with current information and links to relevant sources. Web search is available on ChatGPT's consumer and business plans, and to people who are not signed in. ChatGPT may search automatically when a question would benefit from current information. You can also turn Search on yourself, or regenerate an answer with a web search when that option is shown.
Responses that use web search may include citations. OpenAI says you should open those sources, check that they support the answer, and treat results as something that can be incomplete, outdated or incorrect. Placement is not guaranteed. You can usually tell that live search was used because the answer includes citations, and on desktop you can open Sources when that control is shown. If you asked a local question and the answer has no sources and feels generic, turn Search on and ask again in a fresh chat.
When the question is local, ChatGPT may use an approximate location from the person's IP address. Device location sharing is optional and off by default. OpenAI also says ChatGPT may share location with trusted third-party providers that help with nearby business listings or weather reports, and that it does not share the IP address itself, or ChatGPT account information, with third-party search providers in order to run the search.
A naming test should put the town, neighbourhood or postal code in every question. Otherwise you may be testing a different area from the one your customers mean.
Where does the search go?
OpenAI says ChatGPT search sometimes partners with other search providers, and typically rewrites the question into one or more targeted queries before sending them. After the first results, it may send more specific queries to other providers. The help article points to Microsoft's privacy statement and Shopify's privacy policy for how those providers may process queries.
That is not a published, exclusive contract with any one search engine, and I do not claim one. It is a reason to keep the site crawlable, the sitemap current, and the facts consistent on more than one platform, including Bing Webmaster Tools if you use them, and the listings in listings and NAP.
To be eligible for inclusion in ChatGPT search answers, OpenAI says you should allow OAI-SearchBot to crawl the site, and confirm that the host or content delivery network allows traffic from OpenAI's published search-bot IP addresses. Sites that block that crawler will not be shown in those search answers. The robots.txt steps are in robots.txt for AI search.
What work does this change?
If you believe ChatGPT only uses training data, you might wait for a new model, stuff keywords into a page in the hope they were "learned", or ignore a robots.txt block because "the site already ranks on Google". None of that fixes what live search can see this week.
Work on what is current and public:
- Let the search crawlers read the pages that matter.
- Publish a clear service page for each main service, in ordinary text, with the areas you cover.
- Keep Google Business Profile and other listings consistent.
- Ask every customer for an honest review, offer nothing, and never tell them what to write, as set out in reviews and AI search.
- Test with Search on, in a fresh chat, and count whether you are named. Being cited is not the same as being named. That difference is in named or cited? What ChatGPT results actually mean.
If the usual reasons you are missing are what you need next, they are in why is my business not showing up on ChatGPT?.
Frequently asked questions
Can I turn web search off and still get a useful local test?
You can, but you will not be testing the way most current local answers are produced. For a naming test, use a fresh chat, turn Search on, and put the town in the question.
Does ChatGPT use Google or only Bing?
OpenAI says ChatGPT search sometimes partners with other search providers, and its help article points to Microsoft's privacy statement and Shopify's privacy policy. That is not a published, exclusive contract with any one search engine, and I do not claim one.
If my site is already in the training data, am I covered?
No. Training data is not a reliable picture of who trades in your town this month. Current local answers depend on what search can crawl and cite now.
Will allowing the search crawler get me recommended?
No. OpenAI says allowing OAI-SearchBot makes a site eligible for inclusion, and that placement is not guaranteed. AI results vary.
Want someone to run the checks for you?
A Who-Gets-Named Check asks the customer questions for your town and shows who is named and which pages are cited. The counting rule is on the method page. If you would rather hire someone, how to choose an AI visibility service lists the questions to ask, including of me. AI results vary.