The research

How we approach a market

1. Map the market first. Before we write one question we build a map: who the players are, who they are up against, who is asking and what each of them is trying to decide. Facts about a player are checked against its own site. For Atlanta the map holds 63 organizations and eight publications, 26 cities it gets compared with, and 24 kinds of people with a reason to ask.
2. Become the buyer. We turn the map into people: a founder raising a seed round, a student choosing a campus, a company weighing a move to Georgia. Where a client has sales calls and customer interviews, we build from their customers' own words. Where there are none, as here, we build synthetic buyers from the map, each with a job to do, a decision in front of them and something in the way.
3. Ask what they would really ask. Each buyer produces the questions that come right before a decision: who should I talk to, which of these two, is this worth it. Every question records who is asking, what they are deciding and what kind of answer it calls for, and none of them names an organization, so nobody is handed the answer. Then we ask all of them in ChatGPT, Claude, Gemini and Perplexity and keep every word and every source.
4. Find the gap, and the reason for it. We read who was named, who was recommended, what was said about the ones passed over, and which pages the answer was built from. What comes out isn't a score. It is a short, ranked list of pages to fix, publish and link, and the same questions asked again ninety days later to see what moved.

For Atlanta: 13 organizations each have a full study of their own buyers' questions, and everyone is measured on two shared studies, 150 questions about US startup cities and 190 about Atlanta's ecosystem. Collected September 2026.

2,290 questions, 9,160 answers, 63,532 citations traced, 4 AI engines.

Every question, every answer

Both shared studies are public in full: 340 questions and 1,360 answers, exactly as the engines gave them in September 2026. Open the query explorer.

How we count

• The unit is one answer: one question asked of one engine.
• A city or organization named in the question itself isn't counted on that question.
• Cities are compared on the questions that aren't head-to-heads.
• A publication isn't scored on being named or recommended. It is placed by how many of the 1,360 shared answers cite it as a source.
• Every answer was read twice: an exact scan for every tracked name, then a model that coded how each was described. Names and quotes it returned had to appear in the answer.
• "Picked" means the answer singles it out as a top recommendation. Reasons are our coding; quotes are the engines' own words.
• On the map, a line joins two organizations when the same answer names both; the more answers, the stronger the line.
• Rows built on fewer than 30 answers are marked as direction.

What this can't tell you

• One snapshot. Answers vary from run to run and the engines update their models; the re-measure repeats the same questions.
• A citation shows an engine listed a page as a source, not that the page produced a particular sentence.
• City results are being checked against answers collected from outside Atlanta.