Online support chat systems seem to me an interesting layer of generally unused data about site visitors. The main object of such systems, obviously, is to sell more goods. On most retail or service sites that try live support chat, the human agent is trained to move site visitors towards the product they really need. But at the same time these interactions must offer a fascinating window into what visitors are thinking while at their site.

And so I had my own chat session recently with Joel Granoff, President and CEO of Be Greeted, a provider of these services to sites like Hubspot and the American Institute of Gastric Banding. The company works with consumer and B2B sites primarily to produce better leads, to move people towards goods or capture contact information from visitors that can be passed on to sales staff. It turns out that these chat sessions produce transcripts that themselves become a kind of conversational "cloud," a database of keywords, needs and desires that a mere clickstream cannot expose or imply.

Granoff tells me that a chat session on a client's site is itself a response to a set of defined behaviors. If an incoming user follows a particular navigational path or uses a specific interactive tool, or simply comes into the site as a result of a particular keyword search, he or she might be invited into a live chat session. The conversation produces a real-time transcript that goes to the client's sales staff and includes the context of the exchange, where the person was on the site, where in the buying cycle they seem to be and whether they are demo-ing a product.

Often the terminology that users employ in chat conversations is quite different from the nomenclature on the site and suggests that marketers are viewing their own market position differently from their own customers. Data mining these conversations also exposes mismatches in search strategies. Granoff recalls one client who discovered that some of the key terms they were optimized around on search engines never came up in the course of their customer conversations. "You can look at the discrepancies in how [a client] views themselves," he says.

One of the interesting things about data-mining conversation is that it is such an open-ended field of possibilities. When companies tag and track user patterns at a site, they really are viewing how Web clickers behave within a set of predefined possibilities (the site's architecture). Conversation employs the richer terrain of language and so can expose ideas the retailer or publisher never anticipated. Granoff recalls one chat exchange in which a female customer asked "how many dress sizes you could lose in a month." It was a way of thinking about the product the provider never anticipated and certainly couldn't be gleaned from clicks. That simple query can be spun into white papers, site navigation prompts or marketing messages.

The Internet is a tech-driven platform, invented by engineers. The bias here is technological, with a tendency to overlook the rich "data" human interactions can render. When BeGreeted aggregates conversations for the month, the keyword cloud pulls up the kinds of responses users have to being helped in conversational English rather than by navigating a site that struggles to anticipate user needs through algorithms and engineering. The sense of satisfaction the user gets by interacting with a human rather than a machine comes in two people being able to zero in, through language, on what a user really wants or needs. It is so telling that the most common word in the word cloud from these exchanges, according to Granoff, is "Exactly!"

This commentary is insightful. I recommend it to others.