I've been testing the brand voice training in Anyword for a few campaigns now. The core idea is solid—feed it your past high-performing copy so it stops generating generic fluff. But the implementation has some sharp edges that aren't obvious from their docs.
The main pitfall is assuming "winners" are all in the same voice. If your dataset includes a snappy social media post, a formal case study, and a mid-funnel ebook promo, the trained profile becomes a confused average. You get watered-down suggestions that try to be everything. My process now is:
* Segment winners by channel and funnel stage before uploading. Train separate profiles: "Social_Voice," "BottomFunnel_Whitepaper," etc.
* Manually prune any outliers from the dataset. That one top-performing post that used heavy irony? It'll skew everything if left in a general batch.
* The "weight" sliders for tone attributes feel like guesswork. "More confident" vs. "less confident" lacks a concrete reference point. I track output against my actual past winners in a spreadsheet to see if the slider move actually moved the needle.
Biggest gripe: a recent UI update moved the "apply brand voice" toggle to a less obvious spot during campaign creation. I accidentally generated a whole set using the default voice because the workflow changed. Wasted credits. Pay attention to the review step before generation.
For those who've done this, how are you validating the output? I'm comparing similarity scores, but I'm not convinced that's the best metric for actual performance prediction.
-- CRM Surfer
Your CRM is lying to you.