How Perplexity and ChatGPT Choose Sources: An AEO Deep Dive

The Basic Architecture

User query
    │
    ▼
┌────────────────────────────┐
│   Query Understanding      │
│   - Intent classification  │
│   - Entity recognition     │
│   - Topic identification   │
└─────────────┬──────────────┘
              │
              ▼
┌────────────────────────────┐
│   Source Retrieval         │
│   - Web search/index       │
│   - Knowledge base         │
│   - Trusted sources        │
└─────────────┬──────────────┘
              │
              ▼
┌────────────────────────────┐
│   Source Evaluation        │
│   - Authority scoring      │
│   - Relevance matching     │
│   - Freshness check        │
└─────────────┬──────────────┘
              │
              ▼
┌────────────────────────────┐
│   Answer Synthesis         │
│   - Information extraction │
│   - Citation assignment    │
│   - Response generation    │
└────────────────────────────┘

The Key Insight

AI systems don't just find information—they evaluate which sources deserve to be cited. This evaluation happens before the answer is generated.

Step 1: Candidate Retrieval

AI first gathers potential sources:

Step 2: Authority Evaluation

Sources are scored on trust signals:

Authority Score = 
  Domain reputation (30%)
  + Topic expertise (25%)
  + Content freshness (20%)
  + External validation (15%)
  + Technical quality (10%)

Step 3: Relevance Matching

How well does the source answer the query?

Step 4: Citation Assignment

Deciding what to cite:

• Factual claims → Cite authoritative source
• Statistics → Cite original research
• Definitions → Cite recognized authority
• Opinions → Cite expert or don't cite

How Perplexity Works

Perplexity is explicitly designed as an "answer engine" with visible citations:

Query: "What is the best CRM for startups?"

Perplexity response:
"For startups, HubSpot CRM is often recommended as the 
best option due to its free tier and scalability [1]. 
Pipedrive offers simpler pipeline management [2], while 
Salesforce provides enterprise features but at higher 
cost [3]."

[1] 
hubspot.com/crm/startups

[2] 
pipedrive.com/blog/startup-crm

[3] 
salesforce.com/solutions/startups

Perplexity Citation Factors

Getting Cited by Perplexity

What works:

• Clear, direct answers to specific questions
• Recent publication dates
• Authoritative domain (not new/unknown)
• Original data or insights
• Well-structured content with headings

What doesn't work:

• Generic, rehashed content
• Outdated information
• Thin content without substance
• Paywalled content

How ChatGPT Uses Sources

ChatGPT with web browsing:

1. Searches the web for relevant pages
2. Reads and evaluates content
3. Synthesizes answer
4. Cites sources (when using browsing)

ChatGPT's Implicit Authority Model

Even without browsing, ChatGPT's training reflects source authority:

Optimization for ChatGPT

For training influence (long-term):

• Publish content that gets linked/cited widely
• Appear on Wikipedia and Wikidata
• Get mentioned in authoritative sources

For browsing citations (immediate):

• Rank well in search (ChatGPT uses Bing)
• Have clear, extractable answers
• Maintain current, accurate content

How AI Overviews Select Sources

Google's AI Overviews draw from:

1. Search index - Already-ranked pages
2. Knowledge Graph - Verified entity data
3. Structured data - Schema markup
4. Quality signals - E-E-A-T indicators

The AI Overview Algorithm

Source Selection for AI Overview:

1. Identify top-ranking pages for query
2. Evaluate E-E-A-T signals
3. Check for direct answer content
4. Prefer pages with:
   - Schema markup
   - Clear heading structure
   - Answer-first content
   - Recent updates
5. Synthesize from 3-5 top sources
6. Display with source links

What Gets Featured

Domain-Level Signals

Content-Level Signals

Entity-Level Signals

The Citation-Worthy Content Formula

Citation Likelihood = 
  Uniqueness (Is this information found elsewhere?)
  × Authority (Should readers trust this source?)
  × Extractability (Can AI easily use this?)
  × Relevance (Does this answer the query?)

What Makes Content Extractable

Good (easy to extract):

Schema markup is a standardized vocabulary of tags
that you add to HTML to help search engines and AI
systems understand your content. It uses JSON-LD
format to define entities, relationships, and
properties in machine-readable format.

**Bad (hard to extract):**


```markdown
You might be wondering about how to make search 
engines understand your content better. Well, 
there's this thing called schema markup, which is 
basically like... let me explain with an example...

Content Patterns That Get Cited

Content That AI Skips

Red Flags for AI Systems

Content signals that reduce citations:

• No author attribution
• No publication date
• Generic stock photos
• Thin word count
• Keyword stuffing
• Excessive ads
• Poor mobile experience
• Slow load times

Strategy 1: Be the Definitive Source

Create content that AI must cite because no alternative exists:

• Original research and data
• First-party case studies
• Expert interviews
• Industry surveys
• Proprietary analysis

Strategy 2: Structure for Extraction

[Direct answer in first 40-60 words]

[Supporting section]

[Elaboration with specifics...]

[Related topic]

[Additional context...]

### Strategy 3: Build Entity Authority

1. **Claim your entity** - Knowledge panel, Wikipedia
2. **Implement schema** - Organization, Person, Product
3. **Build consistency** - Same info everywhere
4. **Get corroboration** - Third-party mentions

### Strategy 4: Maintain Freshness


| Action            | Frequency     |
| ----------------- | ------------- |
| Review and update | Quarterly     |
| Add new data      | As available  |
| Update dates      | After changes |
| Check accuracy    | Ongoing       |


### Strategy 5: Target Specific Query Types


| Query Type     | Content Strategy    |
| -------------- | ------------------- |
| "What is X"    | Definition-focused  |
| "How to X"     | Step-by-step guide  |
| "X vs Y"       | Comparison content  |
| "Best X for Y" | Recommendation list |
| "X cost/price" | Pricing page/table  |

AI System Documentation

• Perplexity AI
• OpenAI ChatGPT
• Google AI Overviews

Research

• How LLMs Select Sources
• AI Search Citation Patterns

Industry Analysis

• Marie Haynes on AI Search
• Lily Ray on E-E-A-T