Techniques are disclosed for presenting recommended marketing content that can be dynamically varied (e.g., in real time or near-real time) based on the particular content of a web page that a user is currently reading, the content that the user has previously read, or a combination of these factors. For example, the current interactions of a user with content of a web page, such as scroll position or eye fixation, can be used to identify which portion or portions of the web page content the user is currently reading and which portions of the web page content the user is skimming or skipping over. Keywords can then be generated based on the content in those portions of the web page that the user is currently reading. Next, recommended marketing content corresponding to the keywords can be retrieved and presented to the user.