Media and market intelligence company, CivicScience, today announced the deployment of a new platform for auto-generating poll questions, matched to content people are reading online. As traditional polling and paid survey research suffer rapidly declining response rates, the newly-launched technology, known as QGen, has increased CivicScience’s internal efficiency by over 10X, while producing millions more daily responses from a larger and even more representative sample of Americans.
“We believe this is one of the first practical, end-to-end applications of large language models in full commercial use,” said John Dick, CivicScience’s founder and CEO. “The business outcomes have more than validated the power of this amazing technology.”
CivicScience reaches over 11 million U.S. teens and adults daily through scientific polling applications delivered inside articles of hundreds of premium news and content sites. After answering a question related to the topic of the page where the poll is served, respondents answer four randomized questions about thousands of possible topics before seeing results.
Poll response rates are highly correlated with the relevancy of the initial question to the article in which it appears. Previously, these ‘engagement’ questions were written by a team of editors, who skimmed articles across the CivicScience network and manually wrote upwards of 200 new questions daily. Remaining articles received generalized questions based on a topical feed library, yielding lower overall response rates.
QGen uses OpenAI’s GPT-4 large language models to create over 20,000 editorial-quality poll questions every day. In the first week of deployment, CivicScience collected over 6.5 million responses in a day, nearly triple the prior average, while reaching a larger and more diverse audience of Americans.
“This increased engagement is a big boost for our media partners too,” said Rachel Clayton, CivicScience’s EVP of Strategy & Partnerships, “as more compelling, brand-safe poll questions improve time-on-site, return traffic, and the overall user experience.”
QGen also delivers on AI’s promise of increasing human productivity, not replacing it. Proposed questions are fed into a simple web interface, where editors can quickly approve, edit, and score them for quality and brand safety, continuously training QGen to write better questions. This system has reduced editors’ workload by over 50%, enabling them to focus on higher-value analysis and content development for CivicScience’s numerous publications and reports.
“While AI by itself can be compelling, we’ve found that it’s even more powerful when used as a force multiplier for human expertise,” said Chief Product Officer Joseph Galarneau, adding that generative AI models still suffer from periodic inaccuracy or inappropriateness. “The machine can have a surprising, sometimes shocking, level of creativity.”
CivicScience is now incorporating AI throughout its product portfolio, such as building machine-learning models to predict economic indicators and allowing marketers to interact with its database of billions of polling responses via plain-English queries. The multilingual nature of GPT is also enabling CivicScience to expand its library of Spanish language questions to reach Hispanic consumers, as well as branch into other languages in the future.
“The mission of CivicScience is to power the world’s opinions, enabling people of all backgrounds and walks of life to influence the things they care about,” continued CEO John Dick. “The scale and efficiency of AI is getting us closer and closer to that dream.”
Stay Ahead of the Game with MTC Podcast