Do AI Models Link Differently?
API-level comparison of direct links in hotel recommendations across ChatGPT (GPT-5.1 → 5.4) and Gemini models. With web search enabled, where do the links actually point?
TL;DR
We called ChatGPT (GPT-5.1, 5.2, 5.3, 5.4) and Gemini (2.0 Flash, 2.5 Pro) via their APIs with web search enabled, using identical hotel queries. The hypothesis: direct links should be model-agnostic — because links come from the search tool, not the language model. The model decides what to say; the search infrastructure decides where to link.
Executive Summary
When ChatGPT recommends "Hotel Le Marais" and includes a link, that link comes from the web search tool (SerpAPI, Google Places) — not from the language model's training data. This means switching from GPT-5.1 to GPT-5.4 should not change where links point, only which hotels get mentioned and how they're described.
The same logic applies across providers: ChatGPT and Gemini use different search backends, so their link distributions may differ — but within each provider, model version shouldn't matter.
Hypothesis
H1: Within-provider consistency
Different ChatGPT model versions (5.1 → 5.4) should produce the same link distribution when web search is enabled. The model generates text; the search tool provides links.
H2: Cross-provider differences
ChatGPT vs Gemini may show different link distributions because they use different search backends (SerpAPI vs native Google Search). The search infrastructure matters more than the model.
The Architecture Argument
The language model controls which hotels to mention and how to describe them. But the URLs come from the search tool. Changing the model changes the text — not the links.
Models Tested
Models included in the study
| Model | Provider | API Access | Web Search |
|---|---|---|---|
| GPT-5.1 | OpenAI | ||
| GPT-5.2 | OpenAI | ||
| GPT-5.3 | OpenAI | ||
| GPT-5.4 | OpenAI | ||
| Gemini 2.0 Flash | |||
| Gemini 2.5 Pro |
web_search tool in the Responses API. For Gemini, this means enabling Google Search grounding. This ensures all responses use fresh data, not training-data knowledge.Link Distribution by Model
Where do the links in hotel recommendations actually point?
Link destination distribution by model (% of all links)
Link destination breakdown by model
| Model | Hotel Direct Website | OTA (Booking, Expedia, etc.) | Review Site (TripAdvisor, Yelp) | Editorial / Media | Google Maps / GBP | Other |
|---|---|---|---|---|---|---|
| GPT-5.1 | 0% | 0% | 0% | 0% | 0% | 0% |
| GPT-5.2 | 0% | 0% | 0% | 0% | 0% | 0% |
| GPT-5.3 | 0% | 0% | 0% | 0% | 0% | 0% |
| GPT-5.4 | 0% | 0% | 0% | 0% | 0% | 0% |
| Gemini 2.0 Flash | 0% | 0% | 0% | 0% | 0% | 0% |
| Gemini 2.5 Pro | 0% | 0% | 0% | 0% | 0% | 0% |
ChatGPT vs Gemini
Different search backends, different link patterns?
ChatGPT Search Stack
- • SerpAPI → Google Search results
- • Google Places API → entity data
- • Bing Images → secondary images
- • Yelp → US cities + Berlin
- • OpenStreetMap → map tiles
Gemini Search Stack
- • Native Google Search integration
- • Google Maps / Places (direct access)
- • YouTube (Gemini-native)
- • Google Hotels / Travel
- • No intermediary (SerpAPI) needed
Frequently Asked Questions
Methodology
Setup
- • All queries run via official APIs
- • Web search enabled on every request
- • Identical prompts across all models
- • Same set of hotel queries per model
- • Links extracted and categorized from responses
Why Web Search Matters
- • Ensures fresh, real-time data
- • Links come from search results, not training data
- • Without web search, models may hallucinate URLs
- • Isolates link behavior to search infrastructure
- • Reflects real-world ChatGPT/Gemini usage
Link Classification
Each link in a hotel recommendation is classified into one of these categories:
Continue Reading
Explore more Nicolas Sitter research on AI hotel search.