AEO for Reviews: Getting Your Ratings Quoted by AI Agents

AEO for reviews means structuring your customer ratings so AI agents can read them, verify they are real, and quote them in an answer. The mechanics: mark up genuine reviews with Review and AggregateRating JSON-LD, link every rating to an identifiable author and date, and surface photo and video proof to crawlers. Agents weight verified, attributable reviews far higher than anonymous star averages, so the evidence has to be machine-readable, not just visible.

The gap is not credibility. A 4.7 from thousands of buyers is plenty credible to a human. The gap is that an agent cannot see your rendered star widget the way a shopper does. It reads the structured data underneath, cross-checks it against the rest of the open web, and decides whether the number is safe to repeat. If the markup is missing, thin, or looks padded, the agent quietly routes around you and cites someone else.

Why do agents weight verified reviews higher?

An AI agent is doing something a search ranking never had to: it stakes a claim. When it tells a buyer "this serum averages 4.7 from over 4,000 reviews," it is putting a number in its own mouth. That raises the bar for what counts as evidence. A bare star average with no provenance is exactly the kind of thing a model has been trained to distrust, because it is also exactly what a manipulated listing looks like.

Verification is what closes the gap. A review the agent can tie to an identifiable author, a date, and ideally a corroborating signal elsewhere on the web is one it can repeat without much risk. An anonymous "5 stars, great product" floating in a carousel is not. We covered the weighting gap in detail in why AI engines weight verified reviews up to 14x, and the short version holds here: provenance is the currency. Without it, the rating is noise the model filters out.

• Attribution: a name or handle the engine can resolve, not "Anonymous" repeated 4,000 times.
• Recency: a datePublished that proves the reviews are current, not a frozen snapshot from three years ago.
• Corroboration: the same rating echoed on Google, Trustpilot, or Feefo so the number is not your word against nobody.
• Specificity: review bodies that mention the actual product and use case, which read as human rather than templated.

What review schema actually gets quoted?

Two Schema.org types carry the load: Review for an individual rating and AggregateRating for the summary. Agents read both. The individual reviews give the model attributable evidence it can sample and even quote verbatim; the aggregate gives it the headline number. Emit one without the other and you hand the engine half a story. The aggregate alone looks unsubstantiated; the individual reviews alone leave the engine to do its own (often wrong) math.

The example below is illustrative schema, not data to copy. The author names, dates, and bodies stand in for your real customers. The single hard rule: reviewCount and ratingValue must be computed from reviews that genuinely exist. Do not pick a flattering number and back-fill. If you have 412 real reviews averaging 4.6, that is what ships.

{
  "@context": "https://schema.org",
  "@type": "Product",
  "name": "Bodycon Sunscreen SPF 50",
  "aggregateRating": {
    "@type": "AggregateRating",
    "ratingValue": "4.6",
    "reviewCount": "412",
    "bestRating": "5",
    "worstRating": "1"
  },
  "review": [
    {
      "@type": "Review",
      "author": { "@type": "Person", "name": "<real customer name>" },
      "datePublished": "2026-05-18",
      "reviewBody": "<verbatim review text from a real customer>",
      "reviewRating": {
        "@type": "Rating",
        "ratingValue": "5",
        "bestRating": "5",
        "worstRating": "1"
      }
    }
  ]
}

How do you avoid fake-rating penalties?

Google's review-snippet guidelines are explicit: the aggregate you mark up must reflect reviews users can actually see on the page, and self-serving markup (a brand rating its own product, or invented reviews) is disallowed. Agents apply the same logic with less mercy. They cross-check your stated aggregate against what they can independently find, and a number that floats free of evidence gets dropped.

The failure modes are boringly common. A store mass-imports reviews and stamps them all with today's date, so 4,000 reviews appear to have landed in a single afternoon. Another points AggregateRating at the brand itself rather than the product, inflating a storewide average onto every page. A third ships reviewCount: 5000 while only 300 reviews are visible. Each of these is a tell, and each one teaches the engine to ignore your review block on every future query.

How do you make photo and video reviews citable?

Photo and video reviews are the strongest trust signal a human sees and the weakest one a text-first crawler reads, because most agents do not watch your clips. The fix is to translate visual evidence into text the crawler can index: descriptive alt text on review photos, written captions that name the product, and transcripts or VideoObject markup on video reviews. The proof a shopper trusts becomes proof an agent can quote.

This is also where UGC and reviews stop being separate problems. A customer video that shows the product in use, captioned and transcribed, doubles as a photo review and as the kind of specific, human evidence agents prefer. Idukki captures that content with consent through Rights Management, then ships it to your PDP and to your structured feed at the same time, so the evidence a shopper sees and the evidence an agent reads come from one source of truth.

How do you measure whether reviews get cited?

You measure it the way a buyer would: by asking. Build a short list of the questions a real shopper types into ChatGPT, Gemini, or Perplexity about your category ("best SPF 50 for oily skin," "is X serum worth it"), run them on a schedule, and log whether your brand and its rating get named. Track which source the engine cited, whether the number it quoted matches your real aggregate, and whether photo or video evidence got referenced. That log is your AEO scorecard for reviews.

• Citation rate: of N buyer questions, how many name your brand and rating.
• Number accuracy: does the quoted rating match your true aggregate, or a stale/third-party one.
• Source: is the engine citing your domain, or a Reddit thread and a roundup site.
• Evidence type: text rating only, or photo/video proof referenced too.

The same discipline scales across the whole AEO surface, not just reviews. If you want the full rubric, the 42-signal AEO scorecard covers schema coverage, crawlability, and citation tracking as one system; reviews are one high-value slice of it.