HowtogroundChatGPTinverifiedfinancialdata

ChatGPT answers from memory unless you connect it to a data source, so its financial answers can be outdated or invented. Grounding links it to sources you choose, such as Bigdata.com's 13,000+ financial sources, so every figure comes with a citation you can open and check.

Key takeaways

• Ungrounded, ChatGPT answers from memory and memory has no date on it.

• Grounding is not one setting, it depends which part of ChatGPT you are using.

• An answer without a source you can open is an answer you cannot check.

• Test it with a fact you already know before trusting it with one you do not.

What does ungrounded ChatGPT actually look like?

An ungrounded answer can sound right even when there is nothing to back it up.

Ask a model for a company’s latest quarterly revenue without giving it access to current financial data, and it may still return a confident number. The model is trying to produce the most likely answer from the information it learned during training. It is not automatically checking the company’s latest filing or earnings release. The figure could therefore come from an older quarter, last year’s results or a reasonable-sounding guess.

Here is a simple example using made-up figures:

You: What was Acme Corp’s revenue last quarter, and how did it compare with the year before?

ChatGPT: Acme Corp reported revenue of $4.2 billion, up 8% year on year.

The answer looks complete, but several important details are missing. It does not identify the quarter, show the previous year’s revenue or explain how the 8% growth was calculated. There is also no filing, earnings release or transcript to support the figure.

Without those details, you cannot tell whether the answer is current or even correct. A grounded answer, for example one drawn from the filings and earnings call transcripts indexed by Bigdata.com, would name the reporting period, show both figures, link to the original company source and make the calculation easy to check. In financial research, a believable number is not enough; you need to know exactly where it came from.

What does grounding actually change in a ChatGPT answer?

Grounding changes what ChatGPT bases its answer on. Instead of relying on what it remembers, it works from a source you have chosen.

When ChatGPT is connected to financial data, it first looks for the relevant information and then uses what it finds to answer the question. This is the basic idea behind RAG: retrieve the information first, then generate the response.

A few things change:

• The information can be more current, as long as the source is kept up to date.

• You can check where the answer came from, whether it is a filing, transcript or news article.

• The model can admit when the information is not there, rather than trying to fill in the missing pieces.

The source you choose decides how much of this you get. Bigdata.com’s Bigdata Search Service, for example, indexes millions of documents in real time from more than 13,000 trusted financial sources. It detects the companies, people, organizations, products and places in each document and the events they are involved in. Unlike a general search engine built for broad context and advertising, it is focused on retrieving the newest, most accurate information for financial research.

Grounding does not mean the model will always reach the right conclusion. It still has to read, summarise and interpret the information it finds. The difference is that it has real evidence to work with, although its reasoning still needs to be checked.

Which parts of ChatGPT can reach outside data, and which cannot?

ChatGPT can only access outside data when it has the right tools, plugins or API access. The base model cannot do this on its own.

Different versions of ChatGPT have different capabilities, and what is available may depend on your plan. These features can also change over time, so use this as a general guide and check OpenAI’s latest documentation for the current details.

Web search is a form of grounding, but it is not always the best fit for financial research. The quality of information can vary, and there is no guarantee that the answer comes from the original filing, transcript or company announcement. A dedicated financial source such as Bigdata.com closes that gap by retrieving from vetted financial publishers rather than the open web.

Another option is ChatGPT for Financial Services, a finance-focused version of ChatGPT Work with built-in data from providers such as Daloopa, PitchBook, LSEG News and Crunchbase. OpenAI’s Help Center notes that these sources are only available with that plan, not with standard ChatGPT Enterprise. If you use a different plan or are building your own application, the connector and API options below, including Bigdata.com’s MCP server, will be more useful.

A quick note on the word “grounding”: you may also see it in Azure OpenAI documentation, where it often refers to Bing Search or Azure AI Search. In this post, it simply means giving ChatGPT or the OpenAI API access to a financial data source you have chosen.

How do you set up ChatGPT to ground its answers in financial data?

Choose a source you trust, connect it, and tell the model to use only that information.

The process is similar whichever version of ChatGPT you use.

1. Start with the source. Decide what counts as reliable for the question you are asking. That might be regulatory filings, earnings transcripts, a licensed news feed or your own research library. Bigdata.com brings filings, transcripts and news from more than 13,000 financial sources into one searchable index. The answer can only be as reliable as the information behind it.

2. Connect the source. In ChatGPT, this usually means adding a plugin, if your plan supports it. If you are using the API, you can connect the Responses API to a remote MCP server, such as Bigdata.com’s at https://mcp.bigdata.com, or use a function linked to a search service. Bigdata’s guide to grounding explains how this retrieval step works.

3. Be clear about how the source should be used. Tell the model to answer only from the retrieved material, cite each claim and say when the information is missing. For example: “Use only the retrieved sources. Cite the document name and page for every figure. If a figure is not available, write ‘not found in sources’ and do not estimate.”

4. Ask for citations you can open. A direct link or clear document reference is much more useful than a vague phrase such as “according to company reports.”

Here is what step 2 looks like with the Responses API and Bigdata.com. The model queries Bigdata automatically until it has enough to answer:

from dotenv import load_dotenv
load_dotenv()

from openai import OpenAI
client = OpenAI()

resp = client.responses.create(
    model="gpt-4.1",
    input="List the name and companies of the people who asked a question during the last earning call of NVIDIA",
    tools=[
        {
            "type": "mcp",
            "server_label": "bigdata",
            "server_url": "https://mcp.bigdata.com",
            "headers": {
                "x-api-key": "YOUR_API_KEY"
            },
            "require_approval": "never",
        }
    ]
)

print(resp.output_text)

You will need a Bigdata API key. Full setup details are in the Bigdata Search Service for OpenAI docs.

How do you tell whether the answer came from your data or from memory?

A simple way to test grounding is to check whether the answer includes a source you can open and says when the information is not available.

Do not rely on how confident the answer sounds. Grounded and ungrounded answers can sound very similar.

1. Start with a question you can check. A figure from the latest filing is a good example, especially if an older figure could easily be used by mistake.

2. Look for a citation. The answer should include a link or clear document reference, such as the Bigdata.com document it retrieved. If there is no source, treat the claim as unverified.

3. Check the claim in the source. Make sure the figure is actually there and relates to the right company, period and context.

4. Ask something the source does not cover. The model should say that it could not find the information. If it gives a confident answer anyway, it may be relying on memory.

5. If you are using the API, check the response output. This will show whether the MCP tools were called and whether retrieval took place. With the setup above, look for tool calls labelled bigdata.

Run this test when you set up the connection and again whenever you change the source or instructions.

What will grounding still not fix?

Grounding controls what the model reads, but it does not guarantee that the information will be understood correctly.

• The reasoning can still go wrong. The model might retrieve the right figures but misread them or make a calculation error.

• The source may be outdated or incorrect. If the model is grounded in poor information, that problem will carry into the answer. This is why the choice of source, such as Bigdata.com’s curated financial sources, matters as much as the connection itself.

• The question may be unclear. Terms such as “last quarter” or “revenue” can mean different things depending on the context.

• The information may not be available. If the source does not cover a particular company or period, the model cannot create the missing data.

This is why citations are so important. They allow you to check the answer against the original source instead of simply trusting the model.

Ready to try it? Get started with the Bigdata Search Service for OpenAI.