Grounding vs training data: where an AI answer comes from
Grounding vs training data explained: how AI answers mix model memory with a knowledge cutoff and live web search, and how to tell which one produced an answer.
By Crawlito · 8 min read
In short
- An AI answer comes from one of two places, or both: what the model learned in training, up to a published knowledge cutoff, or pages it retrieved through a search tool while answering.
- Anthropic, OpenAI and Google all document that, in their API search tools, the model itself decides whether a prompt needs a search, so the same assistant can answer one question from memory and the next from the live web.
- Grounded answers carry evidence: the OpenAI, Anthropic and Gemini search tools all return citations that link a part of the answer to a source URL.
- An answer with no citations and no search step most likely came from training data, which is why it can repeat details that were true before the cutoff and are wrong now.
- Blocking GPTBot or ClaudeBot affects future training data, not grounding; Google-Extended is the exception, because Google says it also covers grounding in Gemini Apps and Vertex AI.
On this page
Every AI answer comes from one of two sources, or a mix of both. Training data is what the model learned before its knowledge cutoff, and it is stored in the model itself. Grounding is content the assistant retrieves from a search tool while it answers, and it can be as fresh as the last crawl of the page. If an assistant gets your opening hours wrong or has never heard of your product, the first question is which of the two produced the answer, because they fail for different reasons.
This post explains both sources from the vendors' own documentation, and how to tell them apart in a given answer. It is a concept post. If you already know an answer is wrong and want to correct it, go to fixing wrong business information in AI answers.
Two sources behind one answer
The glossary of AI search terms defines training data, knowledge cutoff and grounding in a line each. What matters in practice is how differently they behave:
| Training data | Grounding | |
|---|---|---|
| Where it lives | Inside the model's weights | In a search index, fetched per request |
| How fresh | Fixed at the model's knowledge cutoff | As fresh as the indexed copy of the page |
| Changes when you edit your page | Only when a later model is trained | After the page is recrawled and reindexed |
| Shows a source | No | Yes, as citations or source links |
| Can be checked | Hard: you cannot see what the model read | Yes: open the cited page |
| Collected by | Training crawlers and datasets | Search crawlers and user-triggered fetchers |
The last row is why the two are often confused. Different crawlers feed each source, as covered in training vs search vs user crawlers. This post is about the other end: what the assistant does with that content when someone asks a question.
What a knowledge cutoff tells you
Model vendors publish a cutoff date for each model, and it is the clearest public fact about training data.
- Anthropic lists two dates per model in its models overview. The "reliable knowledge cutoff" is the date through which the model's knowledge is most extensive and reliable. The "training data cutoff" is the broader range of data used. At the time of writing, both are June 2026 for the current Claude models.
- OpenAI shows a knowledge cutoff for its flagship models in its model catalog. At the time of writing, for example, GPT-6 Astra lists April 30, 2026.
Three things a cutoff does not tell you:
- Whether your page was in the data. A cutoff says the data stops at a date. It does not say which pages were collected, and no vendor publishes a per-site list.
- Which version of your page the model saw. If you changed your prices in March and the model's data runs to June, it may have seen the old page, the new one, or both.
- When the answer will change. Editing your page today does nothing to a model that is already trained. Its knowledge only moves when a newer model replaces it.
Anthropic describes the reliable knowledge cutoff as the date through which the model's knowledge is "most extensive and reliable". Vendors do not publish how well a model knows any particular business, so a business that launched shortly before a cutoff should not assume the model knows much about it.
How grounding works in the vendors' search tools
Each of the three big model vendors offers a web search tool to developers, and their documentation describes the same basic loop. The consumer apps are not documented at this level of detail, so the API documentation is the most specific public description of how grounding behaves.
Google: Grounding with Google Search. Google's grounding documentation says the tool connects Gemini to real-time web content, and that the model "determines if a Google Search can improve the answer." If it does, the model writes one or more search queries, runs them, and builds the answer from the results, with citations to sources beyond its knowledge cutoff.
OpenAI: the web search tool. OpenAI's web search guide says the tool lets models "access up-to-date information from the internet and provide answers with sourced citations." The guide says the model can choose to search or not based on the prompt, and that with tool_choice: "auto" search is optional. A developer can set tool_choice: "required" when a search must run.
Anthropic: the web search tool. Anthropic's web search tool page says the tool gives Claude access to real-time web content so it can answer "with up-to-date information beyond its knowledge cutoff." It lists when Claude searches: recent events, current prices, organizations, people or products that might have changed, and explicit requests to look something up. It answers directly, without searching, for stable knowledge such as established facts, coding concepts and creative writing.
That last list matters for businesses. Questions about a specific company, its products or its prices sit in the category Anthropic says tends to trigger a search. A general question about your industry may not.
The evidence each tool returns
Grounded answers leave a trail. All three tools return citations that tie part of the answer to a source:
| Vendor | Search step in the response | Citation | Extra evidence |
|---|---|---|---|
| Google Gemini | A search call listing the queries the model ran | url_citation annotations linking a text span to a source URL and title | Search suggestions HTML for display |
| OpenAI | A web_search_call item, with actions such as search and open_page | url_citation annotations with URL, title and location in the text | Optional list of every source consulted, often longer than the cited list |
| Anthropic | A server_tool_use block holding the query | web_search_result_location citations, always enabled | Each result's page_age and up to 150 characters of cited_text |
Here is the shape of an Anthropic citation, trimmed from the example in its documentation, with placeholder values:
{
"type": "text",
"text": "The shop opens at 9am on weekdays.",
"citations": [
{
"type": "web_search_result_location",
"url": "https://example.com/contact",
"title": "Contact us | Example",
"cited_text": "Opening hours: Monday to Friday, 9am to 6pm..."
}
]
}The cited_text field is the useful part: it shows the exact passage the answer relied on. Both OpenAI and Anthropic also tell developers to show citations to end users. OpenAI asks for inline citations as "prominent, clickable links", and Anthropic says citations to the original source must be included when API output is shown directly to users.
How to tell which source produced an answer
You rarely see the API response, but the consumer apps show the same evidence in their interface. Work through these signs:
- Look for citations. Claude's help centre says that with web search, every response includes citations. Other assistants with a search mode also show source links beside or below a searched answer, though the exact interface varies and changes. No citations usually means no search.
- Look for a search step. In the APIs, the search is a separate step in the response that holds the queries. If an app shows the assistant searching or lists its queries, the answer was at least partly grounded; if nothing suggests a search, treat it as memory.
- Check whether search was on. In Claude, the standard interface has a Web search option in the "+" menu, while Anthropic's help article says the new Claude experience has no toggle and Claude "searches the web when it helps". On Team and Enterprise plans, an Owner must enable web search for the workspace first. An answer from a session without search can only come from training data.
- Compare the facts with the cutoff. If the answer describes your business as it was a year ago, and the model's cutoff is recent, it is probably repeating old training data, or an old copy of a page that is still indexed somewhere.
- Open the cited page. A grounded answer is only as good as its source. If the citation points to an outdated directory listing rather than your site, the problem is the source, not the model.
A quick test is to ask the same question twice: once plainly, and once with "search the web and cite sources" added. If the plain answer shows no citations and differs from the searched one, the plain answer probably came from training data.
Why an assistant can be wrong, or know nothing, either way
Each source fails in its own way, and the evidence above usually tells you which one you are looking at.
| What you see | Likely source | Why |
|---|---|---|
| Old prices or hours, no citations | Training data | The model learned the old page before its cutoff |
| Old details with a citation to a third-party page | Grounding | The search found a stale copy elsewhere and trusted it |
| Correct details with a citation to your site | Grounding | Search found and read your current page |
| "I don't have information about that business" | Training data, with no search | The model did not search and has little or nothing in memory |
| A search ran but your site is not cited | Grounding | Your pages were not in the results, or could not be fetched |
The last row is where crawler access comes in. Grounding can only use pages a search system has indexed or can fetch. If search crawlers are blocked, or a firewall challenges them, the assistant may search and still find nothing of yours. That is a technical barrier you can check; whether an assistant then chooses to cite your page depends on many things outside your site. For one operator in detail, see how ChatGPT search finds sources and how Claude web search finds sources.
The other practical conclusion: keeping your own pages accurate and dated honestly helps the grounding path, because that is the path that reads current content. Content refreshes for AI search covers how to do that without fake freshness.
Frequently asked questions
Does blocking GPTBot or ClaudeBot stop AI answers from using my site?
Not by itself. Those are training crawlers, and blocking them is a signal about future training data, while grounded answers rely on search crawlers and user-triggered fetchers with separate controls. Google-Extended works differently: Google says it also controls grounding in Gemini Apps and Vertex AI, though not Google Search.
Will updating my website change what ChatGPT or Claude says?
It can change grounded answers once search systems recrawl the page. It does not change what an already trained model remembers, which only moves when a newer model is released.
How do I find a model's knowledge cutoff?
Check the vendor's model documentation. Anthropic lists a reliable knowledge cutoff and a training data cutoff for each current Claude model, and OpenAI shows a knowledge cutoff for its flagship models in its model catalog.
Does the model always search when it can?
No. Google, OpenAI and Anthropic all document that the model decides whether a search would help. Anthropic says Claude tends to search for current or changing information and answers stable questions directly.
Is a grounded answer always correct?
No. It is only as accurate as the page it cites. Open the citation and check that the source says what the answer claims.
Check that search tools can read your pages
Grounding is the source you can influence, and it starts with search crawlers being able to fetch your pages. Run the free check on an important page to see whether the main AI crawlers can reach it and read its content in the HTML.
Filed under Strategy and audits. Topics: ai-search, ai-training, strategy, content-freshness