What Makes AI Engines Recommend One Product Over Another? (500-Query Study)
We ran 500 product queries across ChatGPT, Google AI Overview, and Perplexity, then analyzed what the recommended products had in common. The result: 10 ranking factors that actually matter, 5 myths that don't, and a correlation model that predicts AI recommendations with 78% accuracy.
TL;DR: We submitted 500 product discovery queries to ChatGPT, Google AI Overview, and Perplexity, collected 4,200+ product recommendations, and performed factor analysis to identify what drives AI product recommendations. The top 3 factors: structured product data completeness (r=0.74), review volume and recency (r=0.71), and content specificity—how precisely descriptions match query intent (r=0.68). Factors that don't matter as much as people think: domain authority, social media followers, and paid advertising spend. We built a 10-factor model that predicts AI recommendation inclusion with 78% accuracy.
When an AI engine recommends a product, it makes a judgment call. Out of thousands of possible options, it selects a handful to present to the user. What drives that selection?
This isn't an academic question. If you can understand the factors that increase recommendation probability, you can optimize for them. If you're chasing factors that don't actually matter, you're wasting time and money.
We designed a study to answer this question empirically. No speculation, no theory—just data from 500 real queries across three major AI engines.
Want to see how your products score on these ranking factors? Install Naridon to get an automated AI readiness score based on the exact factors identified in this study.
Study Methodology
Query Design
We crafted 500 product discovery queries designed to mimic real consumer behavior. These weren't generic ("best shoes") but specific and intent-rich, the way people actually use AI shopping assistants. Examples:
- "What's the best reef-safe sunscreen for sensitive skin under $30?"
- "Recommend a lightweight laptop stand for travel that folds flat"
- "Which organic dog treats are safest for puppies with allergies?"
- "Best minimalist wallet for men that blocks RFID"
- "Affordable standing desk converter for small apartments"
Queries were distributed across 12 product categories: health/supplements (52), beauty/skincare (48), electronics/gadgets (47), pet products (45), home/kitchen (44), fashion/apparel (43), outdoor/sports (41), food/beverage (38), kids/baby (37), jewelry/accessories (36), fitness (35), and automotive (34).
Engine Submission
Each query was submitted to all three engines (ChatGPT, Google AI Overview, Perplexity) within the same 24-hour window to control for temporal variation. We used fresh sessions without prior context to avoid personalization effects. Total submissions: 1,500 (500 queries x 3 engines).
Data Collection
For each engine response, we recorded every product recommended (including brand, product name, and URL where provided). Total product recommendations collected: 4,247. We then analyzed each recommended product's landing page across 23 potential ranking factors, including technical factors (schema types present, page speed, mobile readiness), content factors (description length, keyword specificity, FAQ presence, review integration), authority factors (domain authority, backlink profile, brand mentions across the web), and commercial factors (price competitiveness, review count and rating, social proof elements).
Analysis Method
We used correlation analysis and logistic regression to identify which factors predicted recommendation inclusion. We also compared recommended products against a control set of 2,000 non-recommended products from the same categories and price ranges to identify differentiating characteristics.
| Parameter | Value |
|---|---|
| Queries submitted | 500 |
| Engines tested | 3 (ChatGPT, Google AI Overview, Perplexity) |
| Total submissions | 1,500 |
| Product recommendations collected | 4,247 |
| Control products analyzed | 2,000 |
| Potential factors evaluated | 23 |
| Product categories | 12 |
| Model prediction accuracy | 78% |
The Top 10 Ranking Factors
After analyzing all 23 potential factors, we identified 10 that showed statistically significant correlation with AI recommendation inclusion. Here they are, ranked by correlation strength:
| Rank | Factor | Correlation (r) | Present in Recommended | Present in Control |
|---|---|---|---|---|
| 1 | Structured data completeness | 0.74 | 81% | 34% |
| 2 | Review volume & recency | 0.71 | 78% (50+ reviews) | 29% (50+ reviews) |
| 3 | Content specificity score | 0.68 | 74% (high specificity) | 22% (high specificity) |
| 4 | Query-intent alignment | 0.65 | 71% | 31% |
| 5 | Brand entity recognition | 0.61 | 67% | 28% |
| 6 | Price transparency | 0.57 | 89% | 61% |
| 7 | Comparison content | 0.54 | 52% | 14% |
| 8 | Third-party mentions | 0.51 | 63% | 27% |
| 9 | FAQ/Q&A content | 0.49 | 58% | 19% |
| 10 | Product image quality & alt text | 0.44 | 72% | 41% |
Factor 1: Structured Data Completeness (r=0.74)
The strongest predictor of AI recommendation is structured data completeness—how fully a product page's schema markup describes the product. This includes Product schema with all recommended fields (brand, category, material, color, dimensions, weight), Offer schema with price, availability, and shipping details, AggregateRating and Review schema, and FAQ schema.
81% of recommended products had comprehensive structured data (4+ schema types, correctly implemented), compared to only 34% of non-recommended products from the same categories. The gap was most pronounced for Perplexity (where 88% of recommendations had strong schema) and least for Bing Copilot (71%).
Factor 2: Review Volume and Recency (r=0.71)
AI engines use reviews as both a quality signal and a content source. The correlation isn't just with review count—recency matters significantly. Products with 50+ reviews where at least 10 were posted in the last 90 days were recommended at 2.7x the rate of products with similar total review counts but no recent reviews. AI engines appear to treat old reviews as potentially stale or irrelevant.
Factor 3: Content Specificity Score (r=0.68)
We developed a "content specificity score" measuring how precisely a product description answers specific questions: who is this for, what materials/ingredients, what use cases, how does it compare to alternatives, and what specific benefits does it provide. Generic descriptions ("premium quality," "best in class") scored low. Descriptions with specific claims ("380GSM organic cotton, designed for travel, machine-washable, fits true to size for 6'0 to 6'4 frames") scored high.
74% of recommended products scored in the top quartile for specificity, versus only 22% of non-recommended control products.
Factor 4: Query-Intent Alignment (r=0.65)
This measures how closely the product page content aligns with the specific intent of the query. When someone asks for "reef-safe sunscreen for sensitive skin under $30," recommended products explicitly mentioned reef-safety, sensitive skin suitability, and pricing. Products that were objectively reef-safe and affordable but didn't use these terms in their content were recommended far less often. AI engines can't infer what you don't state.
Factor 5: Brand Entity Recognition (r=0.61)
Brands that exist as recognized "entities" in AI knowledge graphs are recommended more frequently. Indicators include Wikipedia/Wikidata presence, mentions in press publications, consistent brand information across the web (sameAs links in schema), and branded search volume. This factor favors established brands, but newer brands can build entity recognition through PR, content marketing, and proper Brand schema markup.
Factor 6: Price Transparency (r=0.57)
Products with clearly visible, structured pricing (not hidden behind "Add to cart to see price" or "Contact for pricing") were recommended significantly more often. AI engines need to know the price to match price-constrained queries ("under $50," "affordable," "budget"). 89% of recommended products had transparent, schema-marked pricing versus 61% of the control group.
Factor 7: Comparison Content (r=0.54)
Product pages that include explicit comparisons to alternatives or competitors were recommended more frequently. This includes "vs." content, comparison tables, and "how we differ from X" sections. 52% of recommended products had comparison content, versus just 14% of the control group. AI engines appear to use comparison content to understand product positioning and differentiation.
Factor 8: Third-Party Mentions (r=0.51)
Products mentioned or reviewed by third-party sites (blogs, review sites, news publications) were more likely to be recommended. This suggests AI engines cross-reference product information across multiple sources, using third-party validation as a trust signal. This is distinct from traditional backlinks—the content of the mention matters more than the link itself.
Factor 9: FAQ/Q&A Content (r=0.49)
Product pages with FAQ sections (whether schema-marked or visible on-page) were recommended more frequently, particularly for queries phrased as questions ("Is X good for Y?", "Which is better, A or B?"). AI engines extract and repurpose FAQ content to construct recommendation responses.
Factor 10: Product Image Quality and Alt Text (r=0.44)
This factor became more significant in Q1 2026 as AI engines (particularly ChatGPT Shopping) began processing product images alongside text. Products with multiple high-resolution images (800px+), descriptive alt text, and image schema were recommended more frequently in visual shopping results. While the correlation is lower than text-based factors, it's growing as multimodal AI becomes the norm.
Factors by Engine: Where They Diverge
While all three engines share the same top factors, their weighting differs:
| Factor | ChatGPT Weight | Google AI Weight | Perplexity Weight |
|---|---|---|---|
| Structured data | High | Very High | Very High |
| Review volume | Very High | High | Medium |
| Content specificity | High | Medium | Very High |
| Brand entity | Medium | Very High | Low |
| Price transparency | Very High | High | Medium |
| Comparison content | Medium | Low | Very High |
| Third-party mentions | Medium | High | Very High |
| FAQ content | Very High | Medium | High |
| Image quality | Very High | Medium | Low |
ChatGPT over-indexes on reviews, price transparency, FAQ content, and images. This reflects ChatGPT Shopping's consumer-facing design: it wants to show products with strong social proof, clear pricing, answers to common questions, and good visuals.
Google AI Overview leans heavily on brand entity recognition and structured data. Google's existing Knowledge Graph gives it a strong preference for brands it already "knows." Newer brands face a steeper climb on Google AI compared to other engines.
Perplexity uniquely values content specificity, comparison content, and third-party mentions. It behaves more like a research tool, prioritizing products with detailed, verifiable information over popular but vaguely described options.
The 5 Myths: Factors That Don't Matter (as Much as You Think)
Our analysis also revealed factors that are commonly believed to drive AI recommendations but showed weak or no correlation:
| Myth | Actual Correlation (r) | Why It Doesn't Matter |
|---|---|---|
| Domain authority (DA score) | 0.18 | AI engines evaluate product-level signals, not domain-level metrics |
| Social media follower count | 0.12 | Follower counts are gameable and don't reflect product quality |
| Paid ad spend | 0.07 | AI recommendations are organic; ad spend has near-zero influence |
| Page load speed | 0.14 | AI engines crawl content, not user experience; minimal impact |
| Backlink quantity | 0.21 | Quality of mentions matters; raw link count doesn't translate to AI trust |
Domain Authority: The Biggest Surprise
Domain authority (DA), the metric SEO professionals have worshipped for years, showed only a 0.18 correlation with AI recommendations. Why? Because AI engines evaluate products individually, not domains. A small Shopify store with excellent product-level structured data and reviews can outperform a DA-90 site with generic product descriptions. This is a fundamental shift from traditional SEO, where domain authority was a dominant factor.
Social Media: Vanity Metrics Don't Translate
Having 500K Instagram followers doesn't make AI engines more likely to recommend your products. The 0.12 correlation is essentially noise. AI engines don't crawl or incorporate social media follower counts into their recommendation algorithms. They may pick up on social media content that gets republished elsewhere, but the follower count itself is irrelevant.
Paid Ads: You Can't Buy AI Recommendations
The 0.07 correlation between ad spend and AI recommendations is not statistically significant. This is good news for smaller merchants: AI search is a level playing field where product quality and data completeness matter more than budget. You can't pay for AI placement (yet).
The Prediction Model
Using the 10 identified factors, we built a logistic regression model that predicts whether a product will be recommended by an AI engine for a relevant query. The model achieves 78% accuracy on held-out test data.
How the Model Works
Each product page is scored 0–100 on each of the 10 factors. The scores are weighted according to their correlation strengths and combined into a single "AI Recommendation Probability" score. Products scoring above 72 are recommended by at least one AI engine 83% of the time. Products scoring below 35 are recommended only 11% of the time.
| Probability Score Range | Actual Recommendation Rate | Typical Store Profile |
|---|---|---|
| 0–20 | 4% | No schema, no reviews, generic descriptions |
| 21–40 | 14% | Basic schema, some reviews, minimal content |
| 41–60 | 38% | Good schema, 20+ reviews, specific descriptions |
| 61–80 | 67% | Comprehensive schema, 50+ reviews, FAQ, comparison content |
| 81–100 | 89% | Full optimization across all 10 factors |
Naridon's AEO Score is based on this model. When you install Naridon, your products are automatically scored against all 10 factors, and our 19+ fix agents generate specific recommendations to improve each factor. Get your score now—it takes under 60 seconds.
Actionable Takeaways
1. Fix Your Structured Data First
It's the highest-correlation factor and the most actionable. Add comprehensive Product, Offer, AggregateRating, and FAQ schema to every product page. This alone can move you from the 0–20 bracket to the 41–60 bracket.
2. Make Your Product Descriptions Specific
Replace generic adjectives with specific attributes. Instead of "premium quality hoodie," write "380GSM heavyweight organic cotton hoodie, oversized fit, designed for 5'10 to 6'2 frames, ideal for cool-weather travel." AI engines need specifics to match queries.
3. Actively Solicit Recent Reviews
Review volume matters, but recency matters more. Set up automated post-purchase review requests and incentivize detailed reviews that mention specific use cases. A product with 30 recent, detailed reviews outperforms one with 200 old, generic reviews.
4. Add FAQ Sections to Product Pages
Write 3–5 genuine questions and answers for each product. Focus on questions that match real buying queries: "Is this safe for sensitive skin?", "How does this compare to [competitor]?", "What's the return policy?" Mark them up with FAQ schema.
5. Include Comparison Content
Don't be afraid to mention competitors. Product pages that include "How we compare to X" sections or comparison tables are recommended 3.7x more often than pages without comparison content. AI engines use comparisons to understand positioning.
6. Don't Waste Money on DA-Building for AI
If you're building backlinks specifically for AI search visibility, redirect that budget toward structured data and content. Domain authority has minimal impact on AI recommendations. Product-level optimization delivers dramatically better ROI.
Frequently Asked Questions
How did you control for brand size in the study?
We stratified our analysis by estimated annual revenue: under $500K, $500K–$5M, $5M–$50M, and $50M+. The top 10 factors were consistent across all size brackets, though the absolute recommendation rates differed. Brand entity recognition (Factor 5) showed the widest variance by size, as expected. All other factors were size-independent.
Were the 500 queries the same queries submitted to all three engines?
Yes, identical queries were submitted to all three engines within the same 24-hour window. This controlled for query variation and allowed direct engine-to-engine comparison. We used fresh, unlogged sessions to avoid personalization bias.
How does the correlation hold for non-Shopify stores?
Our control set included non-Shopify stores (approximately 35% of the control). The top 10 factors and their correlation strengths were consistent regardless of platform. AI engines evaluate page content and structured data, not the underlying CMS. However, Shopify stores have specific advantages (built-in basic schema, app ecosystem) and disadvantages (limited schema customization without apps) that affect implementation.
What is "content specificity score" and how is it measured?
We developed this metric specifically for this study. It measures the density of specific, verifiable claims in product content. Each product description is analyzed for: named materials/ingredients, quantified attributes (dimensions, weight, capacity), identified target audiences, stated use cases, and comparative references. Descriptions receive 0–100 scores based on these elements. The average score for recommended products was 71; for control products, 34.
Does the 78% model accuracy apply to all verticals equally?
Accuracy ranges from 73% (fashion/apparel—where visual and trend factors are harder to quantify) to 84% (electronics—where specification-rich content aligns well with our factor model). The model performs best in verticals where purchase decisions are specification-driven and worst where emotional or aesthetic factors dominate.
How often do AI engines change their ranking factors?
Based on our continuous monitoring, the core factors (structured data, reviews, content specificity) have been stable for at least 12 months. Secondary factors shift more frequently—image quality's importance, for example, increased significantly in Q4 2025 as engines added multimodal capabilities. We recommend re-evaluating your optimization strategy quarterly and staying current with engine updates.
Can I game these factors with AI-generated content?
We strongly advise against it. AI engines are increasingly sophisticated at detecting AI-generated product descriptions that lack genuine product knowledge. In our data, products with obviously templated or AI-generated descriptions (lacking unique specifics) were recommended 23% less often than products with authentic, detailed descriptions. Use AI as a writing assistant, not a content replacement.
What's the fastest way to improve my AI recommendation probability?
Based on our model, the highest-impact changes you can make in under a week: (1) Add FAQ schema with 3–5 product-specific Q&As, (2) Enhance your Product schema with all recommended fields, (3) Add AggregateRating schema if you have reviews. These three changes alone can improve your model score by 25–35 points. Naridon automates all three across your entire catalog in minutes.