A company uploads a product manual to ChatGPT and asks which customers it suits. The answer accurately names the brand. That can show that the supplied material helps the response. It does not establish that an unfamiliar buyer, who supplied no document, would receive the same recommendation.
Internal tests often conceal this difference. The tester already knows which document to provide; a buyer may only describe a need.
Record the files supplied to the model
OpenAI's file upload guidance describes extraction, synthesis, and analysis of supplied documents. Record the filename, version, and product context provided before evaluating the response.
Uploading a manual is appropriate when the question is whether the manual communicates clearly. Review whether the model finds limitations, reads a table correctly, or misses an exception. That evaluates understanding under supplied conditions.
A discovery test asks something different. Giving the system brand material changes those conditions. Its result should not automatically enter a report of public search exposure.
Test public discovery separately
OpenAI's ChatGPT search guidance describes answers using web information and sources. A monitoring team should inspect which sources a particular answer used rather than stopping at a brand mention.
Maintain separate tasks if both questions matter. One can supply files to evaluate document understanding. Another can avoid tester-supplied brand files and record answers and visible sources in the intended discovery setting. Each can be useful, provided the input conditions remain explicit.
Previous conversation content also matters. Removing the brand from the latest question does not remove it from an existing conversation. Retain relevant context so another analyst can understand the observation.
State what the test is meant to establish
Suppose a sales team provides an approved product description and receives an accurate account of migration limits. That can support a document-assisted sales workflow. It cannot estimate how often buyers encounter the brand in search. Conversely, a search answer not citing the file does not demonstrate that the file is hard to understand.
Before a result reaches a report, write down the objective, supplied inputs, and observed output. That keeps a successful document test from being presented as growth in independent market discovery.
References
-
OpenAI Help Center, “File Uploads FAQ”, for document input functionality.
-
OpenAI Help Center, “Searching the web with ChatGPT”, for web search and sources.