The Reviews That AI Actually Cites (And the Ones It Ignores)

ChatGPT crossed 1 billion web searches in a single week in April 2025. Perplexity launched a shopping hub with Shopify integration and PayPal checkout. Google AI Overviews now appear on 14% of shopping queries — up 5.6x in four months.

But before you reorganize your entire strategy around AI search: only about 9% of ChatGPT prompts actually trigger the shopping feature, and 95% of Americans still use traditional search monthly. AI search is growing — especially in product discovery — but it hasn’t replaced anything yet.

So why does this matter now?

Because the reviews that work for AI search also work better for traditional conversion. Detailed reviews with specific use cases, honest comparisons, and real context help both a shopper scanning your product page and an AI deciding which products to recommend. Optimizing for AI discoverability isn’t a separate initiative — it’s the same thing as collecting better reviews.

This piece covers what the data shows about how AI search engines use review content, why most reviews give them nothing to work with, and what you can do about it.

---

How AI search engines use reviews today

Three platforms dominate AI-powered product discovery: ChatGPT, Perplexity, and Google AI Overviews. Each uses review content differently, but they share a common requirement.

ChatGPT

A study of 43,000+ product carousels found that 83% of ChatGPT’s product recommendations are sourced from Google Shopping results. So the data pipeline starts with Google’s product index, not direct crawling.

But when ChatGPT explains why it recommends a product — the text that accompanies the carousel — it draws from review content, editorial mentions, and third-party analysis. Research from Onely found that review volume accounts for roughly 16% of brand selection, with AI-recommended products averaging 3.6x more reviews than non-recommended ones. Critically, ratings above 4.4 have minimal additional impact — what matters is whether the reviews say something an AI can extract and paraphrase.

One thing worth noting: Amazon blocked ChatGPT’s web crawlers in November 2025, coinciding with OpenAI’s Shopping Research launch. This means ChatGPT can’t access Amazon reviews. Independent review content — on your own product pages — is now the primary source of review data available to ChatGPT for non-Amazon products.

Perplexity

Perplexity’s shopping hub runs sentiment analysis on reviews to extract recurring pros, cons, and themes. Its ranking model includes a review trust score alongside intent match, schema completeness, and price freshness. Unlike ChatGPT, Perplexity cites sources explicitly — pulling from review sites, Reddit threads, YouTube comparisons, and editorial roundups.

Perplexity also has a direct Shopify integration for product data access, making Shopify-hosted review content particularly accessible.

Google AI Overviews

Google’s AI Overviews now appear on 14% of shopping queries (up from 2.1% in November 2025), based on analysis of 20.9 million search results. On informational shopping queries — “best running shoes for flat feet,” “espresso machine under $500” — the presence rate jumps to 83%.

Pages with structured data (Schema.org Product + AggregateRating + Review) are cited 3.1x more often in AI Overviews. And here’s the counterintuitive finding: 80% of products appearing in AI Overviews don’t rank in the traditional top 10 organic results. AI Overviews surface different products than standard search — and what they surface is based partly on what review content is available to summarize.

The common thread

Star ratings are metadata. Review text is content.

All three platforms need something they can extract from — specific claims about use cases, comparisons to alternatives, mentions of product attributes, numbers and measurements. A review that says “5 stars, love it” is invisible. A review that says “I switched from [competitor] because [specific reason], and after 3 months the [specific attribute] has held up better than expected” gives an AI something to cite.

---

What makes a review “AI-discoverable”

Based on the available research, AI search engines preferentially cite review content that contains:

Specific use cases. Not “great product” but “I use this for [specific activity] and it [specific outcome].” LLMs extract these as evidence that a product fits a particular need.

Comparisons. “Better than [alternative] because [specific reason]” gives an AI a basis for recommendation. When someone asks ChatGPT “which is better, X or Y?” it needs comparison content to answer.

Attribute mentions. Size, fit, durability, taste, setup time — concrete details about the product experience. This is what Okendo tries to capture with structured attribute forms. Conversations can draw out the same information naturally.

Numbers and measurements. “Lasted 6 months of daily use,” “took 10 minutes to set up,” “runs 20% quieter than my old one.” Quantifiable claims are extractable and verifiable.

Emotional context with specifics. Not just “I’m happy” but “I was worried about [concern] but [specific experience resolved it].” This maps to purchase objections that AI can address.

What review content is invisible to AI

• “Great product!” (no extractable information)
• “5 stars, fast shipping” (logistics, not product experience)
• “Love it!!!” (sentiment without substance)
• Star rating only, no text (metadata only)

These reviews count toward your aggregate rating and review count — both of which matter for social proof and basic AI signals. But they contribute nothing to the detailed content that AI uses to explain why it recommends a product.

The technical layer matters too

Bazaarvoice’s analysis found that AI crawlers (including OpenAI’s and Anthropic’s) don’t currently render JavaScript. Reviews hidden behind JS-rendered widgets are invisible to AI crawlers — they might as well not exist for AI search purposes.

Server-side rendered reviews with proper structured data are readable. Client-side-only JS widgets are not. This is a technical consideration most merchants don’t think about when choosing a review app.

---

The collection method is the bottleneck

Here’s where the discourse around “optimizing reviews for AI search” misses the point. Most advice focuses on what to do after reviews are collected — add schema markup, use structured data, ensure reviews are crawlable. That’s all necessary. But it skips the upstream problem: what your reviews say is determined by how you collect them.

Traditional review collection works like this: customer gets an email, clicks through to a form, sees a star picker and a blank text box, types the minimum, submits. The form itself defines the ceiling. No follow-up questions. No prompting for specifics. No drawing out the comparison, the use case, or the measurement.

The result is a corpus of reviews where 80-90% are thin one-liners. You can add perfect schema markup to a review that says “Works great!” and it still gives AI nothing to work with.

There’s evidence that the collection method matters. Stamped’s Smart Assist — which adds AI-powered prompts to traditional forms — saw a 40% improvement in review quality compared to standard forms on the same products. That’s with prompts layered on top of forms. A full AI-guided conversation — where the system asks natural follow-up questions adapted to each customer’s responses — draws out even more.

The point isn’t that one tool is better than another. It’s that the method of collection directly determines the substance of what you collect, which directly determines whether AI can use it. Schema markup and structured data are necessary infrastructure. But they’re infrastructure for content that needs to exist first.

---

What to do about it

Whether or not you change your review app, these steps make your reviews work harder in both AI search and traditional conversion.

1. Audit what your reviews actually say

Pull your 50 most recent reviews. Count how many contain: a specific use case, a comparison to an alternative, a measurable claim, or a described photo. If fewer than 10% contain any of these elements, your review corpus is thin — regardless of volume or star rating.

This isn’t about AI search exclusively. Detailed reviews convert 68% higher than thin ones on your own product pages. The AI benefit is a second-order effect of solving the primary problem.

2. Check whether your reviews are crawlable

View your product page source (View Source, not browser inspector — the inspector shows the rendered DOM after JavaScript runs). Search for your review text. If you can find it in the raw HTML, it’s server-side rendered and AI-crawlable. If it only appears after JavaScript executes, AI crawlers likely can’t see it.

Most review apps render client-side. Some offer server-side rendering options. This is worth checking regardless of your AI search strategy — it also affects traditional SEO and page speed.

3. Verify your structured data

Use Google’s Rich Results Test on a product page. You should see Product markup with AggregateRating and individual Review items. Pages with structured data are cited 3.1x more often in Google AI Overviews.

If your review app doesn’t add this automatically, it’s a significant gap.

4. Search for your products on AI platforms

Open ChatGPT, Perplexity, and Google (look for the AI Overview box). Search for the product category you sell — “best [your product type] for [common use case].” Are your products recommended? If so, what review content is cited? If not, look at what products are cited and what their review content looks like compared to yours.

This takes 10 minutes and gives you a concrete picture of your current AI visibility.

5. Evaluate your collection method

If your reviews are mostly one-liners, the form itself may be the limiting factor. Some options, in order of effort:

• Add prompt questions to your existing forms. Most review apps support custom questions. Even a single “What do you use this product for?” prompt improves substance.
• Use AI-assisted prompts. Tools like Stamped’s Smart Assist add contextual prompts that guide customers toward detail. This is a middle ground between forms and conversations.
• Switch to conversational collection. AI-guided conversations (like BetterReviews) replace the form entirely with a dialogue that adapts to each customer’s responses. This produces the most detailed reviews but requires changing your collection tool.

The right choice depends on how much your business relies on review content for marketing and conversion. If reviews are primarily social proof (star count matters, text doesn’t), forms are fine. If reviews are a marketing asset — used in ads, email, landing pages — the substance of each review matters more.

---

The timing question

Is AI search big enough to matter right now?

Honestly: for most stores, not yet. SparkToro’s research suggests AI tools could rival traditional search in 6-10 years, not 1-2. ChatGPT’s shopping feature activates on a small percentage of prompts. Google AI Overviews are expanding rapidly but still appear on a minority of shopping queries.

But the argument for acting now isn’t that AI search has already arrived. It’s that:

1. The reviews you collect today are permanent. A review written this week lives on your product page for years. If it says “Great product, 5 stars,” that’s what AI will have to work with when AI search does scale. If it contains specific use cases and comparisons, it’s ready for both paradigms.

2. Review quality compounds. A product with 30 detailed reviews is more useful to both human shoppers and AI than a product with 300 one-liners. The corpus you build now is the corpus you’ll have when it matters most.

3. There’s no trade-off. Reviews that are detailed enough for AI citation are also better for traditional conversion, better for marketing reuse, and better for product development insights. You’re not optimizing for a hypothetical future — you’re collecting better reviews, period.

The stores that start now build a compounding advantage. The ones that wait will eventually need to retrofit a review corpus that was built for star-count social proof, not substance.

---

Sources

All data points cited in this article with their original sources and confidence levels:

• 1B+ weekly ChatGPT searches — OpenAI announcement, April 2025 (primary source)
• 84M U.S. shopping queries/week — Stackline retail analytics (single credible vendor, widely cited but proprietary)
• ~9% shopping feature activation — TryProfound, 2M prompts tracked Sept 2025–Jan 2026 (primary research)
• 95% of Americans still use traditional search monthly — SparkToro (primary research)
• 83% of ChatGPT carousels from Google Shopping — Search Engine Land, 43K+ products analyzed (primary research)
• Review volume = 16% of brand selection, 3.6x more reviews — Onely (primary research, methodology partially opaque)
• Amazon blocked ChatGPT crawlers, Nov 2025 — Modern Retail, eMarketer, TechRadar (multiple credible secondaries, robots.txt publicly inspectable)
• 14% of shopping queries trigger AI Overviews (up 5.6x) — Visibility Labs, 20.9M SERPs analyzed (primary research)
• 3.1x citation rate for pages with structured data — Visibility Labs (same study)
• 80% of AI Overview products not in traditional top 10 — Alhena.ai (single source)
• AI crawlers don’t render JavaScript — Bazaarvoice (vendor with commercial interest, but technically accurate)
• 40% review quality improvement with AI prompts — HeadWest Guide, citing Stamped Smart Assist data
• 68% higher conversion with detailed reviews — KudoBuzz
• 6-10 year timeline for AI to rival traditional search — SparkToro/Rand Fishkin (primary research)