How to use these prompts
The value isn't the summary, it's the grouping: how many times the same problem comes up, and how much it weighs. An isolated complaint isn't a priority; the same criticism in twenty reviews is. The prompts below count and group, so you can tell the single case from the real signal.
Two cautions. If you paste reviews with customer names or personal data, you're handling other people's data: remove the names first, work on the content. And remember that the AI reads what is written, not what the customer really meant: irony, sarcasm and context escape it. You validate the conclusions, because you know your customer.
The prompt library
Group by theme and count
I'm pasting you a block of reviews. Group them by recurring theme and
for each one tell me: how many reviews mention it, whether it's a strong point or
a problem, and one example sentence taken from the reviews. Order the
themes from most mentioned to least. Don't invent themes that aren't there.
Reviews:
What to change, in order of priority
From these reviews, tell me the three things worth changing
first, based on how often the problem recurs and how much it weighs for the
customer. For each: what the problem is, how many times it comes up, and a
direction for a solution. Distinguish the real problems from the isolated
complaints.
Reviews:
Analyze the tone and the emotion
Read these reviews and give me the emotional picture: how many are
positive, neutral, negative, and above all where the customers are disappointed
relative to an expectation (flag the "I expected X and instead got Y"),
because that's where trust is lost. Flag the ambiguous or ironic reviews
that could be misread.
Reviews:
Extract the usable sentences (with caution)
From these positive reviews, extract the most sincere and concrete
sentences that I could use as a testimonial, quoting only what is
actually written. Discard the generic ones like "great service".
Don't change the customer's words.
Positive reviews:
Draft a reply to a negative review
Write me a draft reply to this negative review: acknowledge
the problem without defensive excuses, say what you'll do concretely, and
invite them to a private contact to resolve it. Human and calm tone, no
call-center formulas. 80 words maximum. I'll reread and adapt it myself.
The review:
A concrete example
Marco runs a bed and breakfast and has sixty reviews scattered across two platforms. He reads them now and then, gets depressed over the negative one and forgets the rest. He copies them all (removing the names) into the grouping prompt. The AI tells him that the most mentioned theme is positive (the breakfast, in thirty-two reviews), but that twelve complain about the same problem: the weak wifi in the back rooms.
Marco hadn't caught it: each complaint taken on its own seemed like a one-off. Twelve together are a signal. He uses the priorities prompt, which confirms wifi as the first thing to fix. He calls a technician for a repeater. Then he uses the reply draft to respond to the negative reviews about the wifi, honestly saying he's sorting it out. The reviews didn't just give him a rating: they told him what to fix. The AI turned sixty scattered opinions into a list of three things to do.
When it does NOT work (and how to fix it)
If it invents a problem nobody raised
The AI sometimes "infers" themes that aren't present. Constrain it: "group only what is written in the reviews, quote the exact sentence you derive each theme from; if a theme recurs fewer than three times, flag it as isolated". That way every conclusion is traceable to a real sentence.
If it misreads irony or sarcasm
An ironic review ("congrats on the hour-long wait") can be read as positive. Ask the AI to flag the ambiguous reviews instead of force-classifying them, and reread those yourself. The human tone under the words is the blind spot of automated analysis.
If there are too few reviews to draw conclusions
With five reviews there's no pattern, there's noise. The AI will tell you so only if you ask it to: add "if the data is too scarce for a reliable trend, tell me instead of forcing conclusions". On small numbers, each review is an anecdote, not a data point.
A tip from someone who actually uses it
Look for the "I expected X and instead got Y": that's where a customer is lost. A review that says "all good" teaches you nothing; one that says "I thought it was included and instead it cost extra" points you to a betrayed expectation, which is the most common cause of dissatisfaction. Ask the AI to isolate exactly those: they're worth more than ten full ratings, because they tell you where the promise and the reality don't match.
Frequently asked questions
Can I use sentences from the reviews on my site or on social media?
With prudence. A public testimonial can usually be quoted, but the rules on privacy and the use of other people's content vary by platform and context. Don't change the customer's words, don't attribute names without permission, and if you have doubts about a commercial use, check the conditions of the platform they come from. The AI extracts the sentences; you assess the right to use them.
Does the AI analysis replace reading the reviews?
No, it makes them more productive. The AI gives you the map of themes and priorities, but the serious negative reviews are worth reading in full: the detail that makes the difference is often in a sentence that the summary flattens. Use the analysis to understand where to look, then look at the hot spots with your own eyes.
If I have few reviews, is the analysis still useful?
It's useful for reading them well, not for faking trends. With little data, the AI can help you understand each individual review and reply, but don't ask it for percentages or rankings: on small numbers they're made up. The value of grouping grows with the quantity; below a certain threshold, every review is a single conversation to tend by hand.