Ron research desk

Show the work.

Original AI-search experiments with the method, limits, and summary data left in the open. Use the numbers, challenge them, or rerun the studies yourself.

Each download is a compact summary of the published findings, with the study date and limitations kept beside the numbers.

An independent researcher comparing printed grids of AI recommendation results

180 grounded answers · July 2026

Five household-name brands, audited across three AI engines

Notion, HubSpot, Mailchimp, Webflow, and 1Password were absent from 48% of buyer-intent answers. The summary file includes every brand-level score, appearance count, citation rate, and leading rival.

Method in brief

✓Five brands, each assigned one clear buyer category.
✓Twelve prompts per brand across awareness, consideration, comparison, and purchase intent.
✓ChatGPT, Claude, and Gemini, each grounded with live web search.
✓One July 2026 snapshot. Results can move as models and source indexes change.

Read the study →Download the 5-brand summary CSV ↓

96 recommendation runs · June 2026

Four AI engines rarely agreed on the product to recommend

Across eight categories, ChatGPT, Claude, Gemini, and Perplexity agreed on one top pick in only a single category. The summary file contains the leading pick by engine and category.

Method in brief

✓Eight common software categories.
✓One natural-language recommendation prompt per category.
✓Three independent runs on each of four web-connected engines.
✓The published CSV is a top-pick summary, not the verbatim model responses.

Read the study →Download the engine study summary CSV ↓

A note on evidence

These are probes, not censuses. AI answers vary by model, date, location, prompt wording, and repeat run. That volatility is part of the finding. We publish dates and limitations so a useful directional result does not get dressed up as permanent truth.