Disclosed are various embodiments for locating product references in community content. Character sequences (n-grams) are extracted from a page of text content. Each n-gram is evaluated as a potential product reference using a product catalog search for the n-gram or a conditional probability for the n-gram. The conditional probability is obtained from behavior-based search data. When the search was used for the evaluation, each n-gram is found to be a potential product based on results from the product catalog search. When the behavior-based search data was used for the evaluation; each n-gram is found to be a potential product based on the conditional probability exceeding a threshold.