Markov chained content has no grammatical error but read terrible. Anyone can tell it is machine generated garbage.
The best way to generate content is to dissect each sentence, manually craft synonym and phrases for each part of the sentence, and then use these as the seed to generate hundred of variations. The content generated this way is readable and understandable at least.
Now if you can automate the generation of phrase synonym, put some intelligence on dissect the the sentence, that would be a terrific achievement.