How ChatGPT and Gemini Choose Which Brands to Recommend: A Pipeline Walkthrough

When you ask ChatGPT which payment gateway to use, the language model does not know the answer itself. Instead, it follows a pipeline to find the information. First, the model searches the web, picks the top results, reads passages, reasons over them, and then writes its response. The language model is just one part of the process, not the deciding factor for whether your brand is included.

In the background, the model receives tools to complete its task. For product questions, it performs at least two tool calls: a web search and a fetch of pages with relevant information. Your prompt is transformed into several search queries, which the model then reads passages from and evaluates against its existing knowledge. To be included in the final answer, a brand must be indexed, fetchable, quotable, and backed up by other sources. The model alone cannot know about your brand.

Language models are trained in two phases: pretraining and fine-tuning. Both phases finish on a fixed date, after which the model's weights do not change. This date is called the knowledge cutoff, and it usually sits further back than people realize. For example, GPT-5 was trained up to September 2024, while Gemini 2.5 Pro was trained up to January 2025.

As a result, newer models have a narrower gap in knowledge, but there is still always a gap. A model released this month still has no information about pricing pages you might have added last week.

Moreover, the model only has a statistical impression of your brand within the knowledge cutoff, not a definitive record. It has seen your name associated with some words more often than others but cannot reliably differentiate between your 2024 and 2025 feature sets. Asking a model without tools for the best CRM for a 10-person agency will give a confident answer based on patterns that are a year old and will not flag that it lacks up-to-date information.