Most review programmes are built to raise a number. Get more reviews, keep the average above 4.7, respond to the bad ones. That is sound, and it addresses only half of what reviews now do.
Two systems, two appetites
Conventional local ranking cares about count, average, recency and velocity. Those are signals, and they are largely quantitative.
A retrieval system reads the words. When someone asks which company handles a particular job in a particular place, the assistant is looking for text that connects a business to that job and that place. A hundred reviews saying "great service, highly recommend" supply almost nothing to match against.
How to get more useful text without coaching
You cannot tell customers what to write, and you should not try. You can ask better questions.
- Ask at the moment the outcome is concrete, not a month later when it has blurred into a general good feeling.
- Prompt with a question rather than a request. "What did we do for you?" produces specifics. "Leave us a review" produces adjectives.
- Ask the technician or project lead to request it, because the review then tends to describe the actual job.
- Never script it, incentivise it or gate it by sentiment. Beyond being against every platform's rules, it produces uniform text that reads as synthetic to both systems.
And watch where they are
Different assistants lean on different platforms, and the mix varies by category. Run your prompt set, read which sources get cited in the answers about your category, and concentrate the programme there rather than spreading it across every directory that asks.