Generative Engine Optimization

62 %

of LLM mentions go to the top 15 % of restaurants

62 %

The top 15 % of restaurants get 62 % of all LLM mentions

How it was measuredMention counts per restaurant, summed across all prompts, model slots and runs
On what450 LLM answers, 182 restaurants, 1,351 mentions
Keep in mindOne domain (Tunisian restaurants), French prompts, 30 prompts — every restaurant was labelled unstable

Built with Emna Kaaniche · supervised by Mrs. Sonda Ammar (ENET'Com)

The problem

Generative Engine Optimization · before

Chatbots decide
People now ask chatbots where to eat; a restaurant that never appears in those answers doesn't exist for that user.
Small means invisible
Small local businesses with few reviews are structurally invisible to LLMs.
One answer proves nothing
One query proves nothing: LLM answers are random, so visibility needs many prompts, runs and models.

Generative Engine Optimization · after

Measure mentions
Measure how consistently LLMs mention local businesses.
Map the web footprint
Collect each business's real web presence.
One clean dataset
Link both in one clean dataset. Pilot: Tunisian restaurants, French prompts.

Generative Engine Optimization

Step 1 of 6 Agent 0 writes 30 French prompts over 6 user intents and checks they make models name real places.
Step 2 of 6 Agent 1 sends each prompt to 3 models, 5 times each (450 answers); a registry picks free models within quota.
Step 3 of 6 A cleanup turns 2,020 raw mentions into 234 restaurants, each with a mention rate and a stability score.
Step 4 of 6 For 213 visible restaurants, Agent 2 searches Google Maps, TripAdvisor, Wikipedia and more through MCP tools.
Step 5 of 6 When an API fails, a supervisor switches model, waits, retries or skips; each result is saved at once.
Step 6 of 6 Agent 3 merges everything, judges restaurants with no web trace, and keeps 182 in the final dataset.

Sources