Artificial Audiences: Can AI Really Stand In for Your Buyers?

Updated October 29, 2026: revised to match the current version of the paper.

You have probably heard the pitch by now. Run your ad past a panel of AI buyers before you spend a dollar, and they will tell you who clicks and who scrolls past. If your first reaction is a raised eyebrow, good. You should be skeptical of any tool that claims to know your customer. Many of them do not.

So we are not going to ask you to take our word for it. Two independent research teams have put the core idea to the test: can an AI actually stand in for a real human buyer? Here is what they found, and where the whole thing holds up or falls apart.

A Harvard and Microsoft study put AI in the shopper's seat

In a working paper by researchers at Harvard Business School and Microsoft (the April 30, 2026 version; page numbers below refer to it), the team asked a simple question: can an AI answer market research surveys the way real shoppers do? They asked large language models the same product-choice questions used in real consumer studies and compared the answers with what people said.

The answer was mixed. The AI picked up real signal: in a pilot, its estimate of what people would pay for fluoride in toothpaste came within 50% of a real consumer survey (page 19, Section 3.1). But on its own it was often wrong. It overestimated what people would pay by two to three times, got some brand preferences backwards, and changed its answers between AI versions (page 34, Section 4). What helped most was giving it real survey answers from the same product category first (Abstract).

Here is why that matters for your ad. It is the difference between a magic 8-ball and a wind tunnel. A magic 8-ball gives you a random answer that feels like guidance. A wind tunnel gives you a reading that tracks reality closely enough to make a decision before you build the real thing. The Harvard work says an AI on its own sits closer to the magic 8-ball, and moves toward the wind tunnel only when it is given real information about the people it stands in for. The authors' own conclusion: use AI as a supplement to studies with real people, not a substitute for them (Abstract).

The second study named the thing that makes it work

A separate study, published in the journal Political Analysis, went further. It asked whether an AI could stand in for specific groups of people, not just people in general. The researchers gave the model thousands of short backstories built from real survey respondents: their age, gender, race, income, party and political views. Then they compared the AI's answers with how those actual groups had answered real surveys about politics and voting (Studies 1 to 3).

The match was close, and not just on the surface. The model reproduced the way different groups genuinely responded. The researchers gave this property a name: algorithmic fidelity. Their finding, in their own words, is that with "proper conditioning" the model will "accurately emulate response distributions from a wide variety of human subgroups" (Abstract). They are careful about its reach: the study is "restricted to a specific domain," American politics, and they call for more work on the limits in other fields (Discussion).

Read that quote twice, because the whole thing turns on the first two words. Proper conditioning.

The entire game is in the conditioning

Here is the part the hype merchants skip. Neither study found that AI magically knows people. Both point the same way: an AI does better when it is given real information about the people it stands in for. The Harvard team also warns that, even then, it did not reliably capture differences between customer segments (page 35, Section 4). Left generic, an AI asked "would you buy this?" is close to worthless.

That one word, conditioning, is the difference between a party trick and a tool. Ask a blank chatbot to rate your ad and you get a confident, agreeable, meaningless answer, because it wants to please you. It does not want to scroll past your ad the way a tired homeowner does at 9pm with a dead AC. The research points the same way: the value is not in the AI itself, but in what it is given before it ever sees your ad.

There is more than one way to do that building. One route conditions the AI on who the buyer is: age, income, region, the demographic facts. That is the route the second study took, and for political surveys it worked. The other conditions it on how the buyer thinks and feels: what they say, what they quietly worry about, what they are trying to get done, the pains that stop them and the gains that move them. Marketers already have a name for that second approach. It is the empathy map, a way of getting inside a customer's head that predates AI by decades. Hand an AI a real empathy map and you are not describing a demographic. You are handing it a working model of a decision.

That second route is the one CouncilAds is built on. Before you spend a dollar, CouncilAds runs your ad past a council of 12 buyer personas, each conditioned not on a demographic profile but on a full empathy map of a real small business buyer, from the price-sensitive skeptic to the ready-to-buy decision maker. They are not one generic AI asked to be nice. They are twelve specific buyers built to react the way those buyers would, and they tell you who would click, who would scroll past, and the exact objection that would stop the sale. The research shows why what the AI is given matters. The empathy map is how we give it. Neither study tested empathy maps or ads, so treat the Council's verdict as a first read before real clicks, not a replacement for them. To see the panel at work, here are the 12 personas scoring a real pair of Google ads.

What this means before your next campaign

You no longer need a focus group and three weeks to get a first read on an ad. Two research teams found that an AI given the right information about people can approximate how they answer, and the Harvard and Microsoft team also showed where it falls short. Neither tested ads, so a Council read is a check before you spend, not a promise of results. The reading is only ever as good as the buyers behind it.

We have written before about why ads fail, and how to fix the copy, the match, and the aim. This is the layer underneath all of it. It is a way to test the fix against real buyer reaction while the ad is still free to change.

If you want more breakdowns like this, you can Add CouncilAds as a Google Preferred Source on Google - https://google.com/preferences/source?q=councilads.com - so our Council Notes show up when you are searching for real answers about your ads.

The skeptic's question is the right one to ask. Can an AI really stand in for your customer? The honest answer from the research is: partly, and only when the AI is given real information about the people first. That is the part we obsess over, so you do not have to.

Ready to see what 12 conditioned buyers think of your ad before you pay for a single click? Test your ad free at councilads.com.