The Territory Simulator.

If you have thirty seconds.

You can understand this page without reading Part I. That first study was simple: show an AI-generated shopper eight nearby clinics, make it choose one, and ask why the other seven lost. Do that 300 times.

This study asks the next question. What happens if the shopper sees a different version of the market? Make the new clinic easier to understand. Change its prices. Add stronger reviews and safety information. In another version, improve only the competitors. For each version, generate 750 choices and count how often the new clinic wins.

The answer did not rise neatly. A clearer story barely helped. Clear prices and stronger trust information did better. The final version combined both with stronger ratings, 220 additional reviews, warm consultations, and visible aftercare. It reached 64.4%. But that version was tested against the starting competitors, not the improved ones. The study found a compelling offer. It did not yet prove a moat.

Before there is a strategy, there is a person with a phone.

Imagine the moment. You are thinking about Botox, filler, or a skin treatment. You search nearby. A handful of clinics appear. You scan distance, reviews, photos, prices, and anything that makes the place feel safe.

The first study tried to recreate that moment with AI. It generated 300 separate shoppers. Each shopper saw eight clinics, chose one, named two backups, and explained why the other seven lost. Four of the clinics were made-up versions of the same new medspa idea in different neighborhoods.

One neighborhood version did well. The other three barely registered. Clinics with visible core prices were chosen much more often than clinics with no prices. The written answers kept returning to the same concerns: reviews, trust, distance, and whether the place felt right.

That was useful, but it was only one view of the market. This second study asks the question again after changing what the shopper sees. The same 750 modeled shopper profiles appear in all six versions, using the same booking task. That lets us compare each profile's modeled answer across versions. It still does not turn those answers into real customer behavior.

The study is simpler than its name.

Write down one version of the market. Ask the AI shopper which clinic it would choose. Repeat that choice 750 times. Then rewrite the market and do it again.

There are six versions. The first is the starting point. The second gives the new clinic a clearer story. The third changes its price menu. The fourth strengthens its ratings, reviews, and care information. The fifth leaves the new clinic alone and improves nearby competitors. The final version resets those competitors to the starting point and gives the new clinic the full bundle.

The original project contains all 4,500 answers, the 4,500 prompt records, the same 750 shopper profiles in every version, and the code that created each version. The short summary omitted the recipe. The underlying files did not.

There is still a catch. These are six separate “what if?” versions, not a clean ladder where only one detail changes at a time. The final version changes prices, ratings, review counts, wording, and a review excerpt together. We can describe the bundle and compare its result. We cannot give one ingredient all the credit.

Think of the same street rewritten six times.

The shopper is always choosing a clinic. What changes is the information around that choice.

Start with the plain version of the market. Then make the new clinic’s story clearer. Change its prices. Add stronger reviews and safety information. In a separate version, improve the competitors instead. End by giving the new clinic the full bundle while putting competitors back at the starting point. Click through slowly or play the sequence.

These versions are not cumulative steps. Each starts from its own copy of the market. They are six separate “what if?” scenes. That is why the final version does not tell us whether the full bundle can survive a competitor response.

The answer did not climb neatly.

If every improvement helped, the bars would rise from left to right. They do not.

The new clinic starts at 12.9%. A clearer story moves it to 13.5%, almost no change. Changed prices reach 31.7%. Stronger reviews and safety information reach 28.9%. In a separate version, only the nearby competitors improve and the new clinic falls to 11.3%. The final bundle reaches 64.4% against the starting competitors.

The underlying setup files explain the jump. In the final version, every new-clinic option received clear $12–14 pricing, a 0.2-star rating lift, 220 additional reviews, warm-consult and aftercare language, and a new five-star review tying all of those signals together. Nearby competitors kept their starting profiles.

The 64.4% result is not a mystery. It is a stack of aligned evidence.

The final version did more than tell a better story. It made nearly every visible signal support the same low-risk choice.

The new clinic showed a $12–14 injectable price, a low consultation fee credited toward treatment, stronger star ratings, 220 more reviews at every modeled location, warm consultations, visible care proof, and aftercare. A new five-star review then said those exact pieces “all matched” and called the clinic “the safest easy choice.”

That makes the result useful, but not magical. The model was shown a clinic with clearer prices and materially stronger reputation evidence. In two locations, the advertised injectable price also became lower, so this was not merely a formatting change. The final version tells us the full bundle is compelling inside this simulation. It does not tell us how much credit belongs to any one ingredient.

Most of the change lands in one bucket.

The same 750 AI shopper profiles appear in the starting and final versions. The study compares each profile's two modeled answers and sorts them into four groups.

387 were labeled “another clinic at the start, new clinic at the end.” 96 had the new clinic in both versions. Only 1 went the other way. The remaining 266 had another clinic in both versions.

The big number is 387. Most of the reported difference between the first and final versions sits there.

Do not picture 387 real people changing medspas. No customers were followed over time. These are paired AI answers for the same modeled profiles, not observed switching.

The final version was designed to feel safer. The answers noticed.

Among the 483 choices of the new clinic, 354, or 73.3%, named Trust / Safety as the single top reason.

Across all 750 final-version choices, including choices of competitors, Trust / Safety was the top reason in 441. Reviews and brand aesthetic account for the 39 choices omitted from the earlier short summary.

But trust was not independently discovered after the run. The final input explicitly described the clinic as high-trust, added safety and aftercare language, and inserted a review calling it “the safest easy choice.” The result shows the model responding to a deliberately trust-heavy scenario. It does not prove what real shoppers value.

Why did the new clinic still lose?

The written answers also explain why another clinic won. The same objections keep coming back: too few reviews, unclear prices, too far away, not enough trust, or a poor fit for the treatment.

In the starting version, those objections appeared 1,281 times alongside 653 choices of another clinic. In the final version they appeared 508 times alongside 267 choices of another clinic.

The totals are larger than the number of choices because one answer can give several reasons. For example, a clinic can be too far away and have too few reviews. The important point is not the raw total. It is that deeper reviews remained the most common objection in both versions.

No. By the final version, all four locations were contributing.

The “new clinic” in this study is actually a group of four unnamed locations. The study calls them A, B, C, and D.

At the start, Location A accounted for 70 of the group’s 97 choices. Location D received only one. In the final version, the four locations received 177, 139, 105, and 62 choices.

This does not prove that four real locations should open. It tells us only that the final total was not being carried by one simulated location alone.

A simulation is useful only if it tells you what to try next.

Do not open a clinic because an AI-generated result reached 64.4%. Use the result to design three small tests with real people.

First, make price easier to understand. Publish common service ranges, explain what a consultation costs, and say what happens after someone clicks. Then measure completed bookings and completed visits, not just page views.

Second, make care easier to trust. Strengthen reviews, provider explanations, safety information, and aftercare details while keeping the offer and booking flow the same. If bookings do not improve, the trust idea gets weaker.

Third, test whether the result survives competition. Keep the full final bundle on the new clinic and give nearby competitors the clearer prices and easier booking from the competitor-response version. Then repeat the run. That is the missing round.

This was AI role-play, not customer research.

No person booked a clinic. No patient was interviewed. No real shopper changed providers.

An AI model was repeatedly asked to choose among clinic options. The original project retains all 4,500 answers, all 4,500 prompt records, and the code that built the six versions. This page publishes totals and selected examples rather than every row.

The final recipe is known. The weakness is experimental: several ingredients changed together, the advertised price itself changed in some locations, and the final bundle was not tested against the improved competitors. The right use is to choose the next test, not forecast demand.

The shopper is not choosing your strategy. They are choosing what they can see.

The first study put eight clinics on a phone and forced one choice. This study changed what appeared on that phone and asked the question again.

A clearer brand story barely moved the answer. Price changes and stronger trust evidence each did more. The largest result appeared only when price, rating, review volume, consultation, safety, and aftercare all told one coherent low-risk story.

So the takeaway is not “the new clinic will capture 64.4% of Chicago.” It is this: language alone is weak; aligned proof is powerful. The model chose the clinic when the visible evidence made “safe and easy” feel true rather than merely claimed.

But the study stopped one round early. The final bundle faced the starting competitors, not the improved ones. Part II found an offer worth testing. It did not show that competitors cannot copy it.

Distance gets you considered. Aligned proof gets you chosen. Defensibility is the next test.