I've run into this exact thing when trying to get help with longer projects. The issue is that these models have a context window - basically a limit to how much conversation they can actually "see" at once. Think of it like a short-term memory that only holds the last N messages or tokens. Once you go past that limit, the earlier stuff just drops out of view, even if you refer back to it. So yeah, it's not you doing something wrong with your prompts - it's a genuine limitation of how the system works.
The workaround is to keep reminding the model about key details as you go. If marketing context is important, slip it back into your prompts every few messages instead of assuming it remembers. Something like "Remember, I work in marketing, so..." helps anchor it back to what matters for your question. Some people also find it helps to be really explicit about constraints upfront - putting all your key context in one solid prompt before you start asking follow-up questions, rather than sprinkling it across multiple exchanges. It's a bit annoying to manage, but it definitely reduces the confusion.
For really long documents, another option is to break the work into chunks and handle each section in its own conversation, or paste a summary of your requirements at the start of each new batch of questions. It's not perfect, but it beats watching it drift into irrelevant advice halfway through.