Randomized smoothing provides certified robustness guarantees, but its direct application to text remains challenging due to the discrete, sparse, and semantically structured nature of textual inputs. We propose CRICO, an input-dependent covariance optimization framework for certified robustness in the text embedding space. CRICO estimates token-level decision influence in a black-box manner, making it readily extensible to LLMs. By adapting the certified ellipsoid to local decision structure, CRICO significantly enhances certified robustness. Experiments on multiple tasks demonstrate its effectiveness and broad applicability.