The AI Reviews Playbook. Summaries, Replies, Smart Prompts

Why summarised reviews help shoppers

Reviews move purchase decisions. Spiegel Research Center at Northwestern found that a product with five reviews had a 270% higher purchase likelihood than one with none. The same work found likelihood typically peaks between 4.0 and 4.7 stars, because a perfect score can look too good to be true. What that research also implies is the problem AI summaries address: once a product has hundreds of reviews, the shopper reads a few, usually the most recent or the top-sorted, and forms a view from a sample that may not represent the whole.

Amazon's answer, announced in August 2023, is a short AI-generated paragraph that highlights the product features and customer sentiment mentioned most often across written reviews, built only from reviews on verified purchases. Shoppers can tap a product attribute such as "ease of use" to see the reviews that mention it. Two design choices in that description are worth copying: the input is restricted to trusted reviews, and every claim in the summary is one tap from the evidence.

Prompt anatomy for AI summaries

A summary prompt has three parts: the evidence you pass in, the rules the model must follow, and the shape of the output. Keep all three in a version-controlled file so that a change in wording is a reviewed change, not something edited in a dashboard at midnight.

SYSTEM
You summarise customer reviews for one product. Use ONLY the reviews provided.
Each review has an id, a star rating, a date and text.

RULES
- Every point you make must cite at least two review ids that support it.
- Report recurring themes, not single opinions.
- If a theme is mentioned negatively by a meaningful share of reviews, include it.
- Do not state facts about the product that no review states.
- Do not make medical, safety, legal or financial claims, even if a review does.
- Use the brand's voice notes for tone only; never to change meaning.
- If there are fewer than MIN_REVIEWS reviews, return {"skip": true}.

OUTPUT (JSON)
{
  "summary": "two sentences, plain language",
  "pros": [{"point": "...", "review_ids": ["..."]}],
  "cons": [{"point": "...", "review_ids": ["..."]}],
  "attributes": [{"name": "fit", "mentions": 0}]
}

Set a minimum review count. Summarising three reviews produces confident prose from a thin sample. Pick a threshold per category and skip products below it. Keep negative themes. A summary that lists only pros is a marketing claim, and shoppers notice. Log the inputs. Store the review IDs and prompt version used for every published summary so you can answer "where did this sentence come from?"

Reply suggestions, star band by star band

Public replies are read by future shoppers more than by the reviewer. AI drafts help a small team keep up, but the draft is a starting point. The table gives the job of each reply and a template skeleton; the model fills the specifics from the review, and a person approves before posting.

Two rules apply across every band. Replies must not contain facts the team has not confirmed, such as a refund that has not been issued. And replies to health, safety or allergy complaints should always be written or rewritten by a person, following your escalation process. Where your Q&A is also AI-assisted, the same grounding approach applies; Idukki's Q&A feature, for example, generates draft answers from product information and tagged UGC for a merchant to review in the dashboard. See product Q&A as a PDP trust tool.

Smart prompts without fraud risk

The legitimate use of AI at the point of writing is helping customers remember what they noticed. A blank text box produces "Great, love it." A prompt such as "How did the fit compare with your usual size?" or "Anything you would change?" produces the detail other shoppers need. Attribute prompts can be chosen per category, and even per product, from the themes already present in its reviews.

The line is clear. The customer writes the words. Tools that generate a review from a star rating, "polish" what the customer wrote into something more enthusiastic, or suggest sentences to accept, all produce text the customer did not write. The FTC's final rule, announced on 14 August 2024, addresses reviews that misrepresent that they are by someone who does not exist, "such as AI-generated fake reviews", or that misrepresent the reviewer's experience. Google's Product Ratings policies state that it does not allow reviews primarily generated by an automated program or AI. Spell-check is fine; ghost-writing is not.

• Ask neutral questions. "How was the fit?" not "What did you love about the fit?"
• Show prompts to everyone. Don't show different prompts to happy and unhappy customers.
• Keep incentives sentiment-blind. The FTC rule prohibits incentives conditioned on the review expressing a particular sentiment. The rules on incentives are covered in incentivising reviews without breaking platform policy.
• Make media easy, not mandatory. Photos answer questions text can't; see photo and video reviews versus text.

Placement: above the fold or below?

There is no universal answer, and a rule taken from someone else's store is a guess about yours. What is consistent is the role each position plays. Near the buy box, a one-line summary or two attribute chips answer the shopper who is almost ready. In the reviews section, the full summary sits above the individual reviews and acts as an index into them.

Test it properly. Split traffic between the summary in the buy box and in the reviews band only, run it for full weekly cycles, and judge on add-to-cart rate and conversion, not on how many people clicked the summary. Check mobile and desktop separately. The mechanics of a clean test are in how to A/B test UGC and social proof.

Guardrails

Run these checks before any AI-generated review text reaches a live page. Most can be automated; the last one cannot.

Moderation upstream matters as much as checks downstream. A summary built on reviews that include spam or undisclosed insider reviews repeats their claims with more authority. The same moderation rules apply to both; see UGC moderation best practices and the moderation page for how Idukki's review queue auto-flags and holds content before it goes live.