I opened my ChatGPT agentic browser (Atlas) recently and asked it to find me travel insurance for an upcoming trip to Chengdu. I gave it my dates, my age and my budget.

It browsed the web, looked through multiple links, and returned one option in about 50 seconds. It came up with a recommendation for a travel insurance policy and was ready to purchase. It even asked if I wanted it to use my stored payment details to complete the transaction right there. I said no, because I prefer to type my own credit card details. But then I stopped.

I had just made a financial decision without browsing or even comparing options. The agent had collapsed my consideration set to zero. And that transaction got me thinking about what this means to the future of consumer research.

For close to a century, consumer research has obsessed about one base truth that the consumer is a human being. Everything we have built, be it focus groups, ethnography, surveys, conjoint analysis, assumes we are trying to understand biological individuals making emotional, messy decisions. The ‘black box’ we have been trying to open is a human brain choosing Coke over Pepsi.

But something structural is breaking.

We were promised the ‘Year of Agents’ in 2024, when Satya Nadella predicted a world of agents acting on our behalf. The reality arrived a little late, but in 2025, infrastructure caught up to the early promise, with agentic browsers like OpenAI’s Atlas and Perplexity’s Comet delivering the reliability needed to autonomously execute some (not all) complex workflows rather than just chat.

While we spent 2025 adjusting to AI-powered answer engine optimisation (AEO) as a mainstream marketing tactic vs the classic search engine optimisation (SEO), 2026 is bringing the next phase. We are moving from business-to-consumer (B2C) to business-to-consumer’s-agent (B2CA).

However, I'm not certain our industry has worked out what this shift actually requires.

Why does being second place now mean you’re invisible? 

Ben Thompson’s aggregation theory provides us with a framework to analyse the shift from B2C to B2CA. Thompson says that in the digital age, power shifts to those who control demand, not supply. With AI, agents become the ultimate aggregators of demand – their job is to search the web, find the answer and, increasingly, take the action.

In a B2C world, a consumer browses five insurance providers, weighs the trade-offs, and picks one. In a B2CA world, the consumer states a goal (eg "best coverage under $200") and the agent analyses multiple plans in less than a minute and presents one choice to the consumer for sign-off.

That’s a new competitive reality. Being the second-best option in a Google search used to mean you still got traffic. Being the second-best in an agent’s logic means you are basically invisible.

For researchers, top-of-mind awareness may now matter less if the human doesn't need to recall brands. Share of voice starts to lose its meaning if the human never hears the voice (especially in a new category). What matters is whether the agent’s training data weights your product as optimal.

Can agentic logic kill the consumption ‘vibe’?

Marketing professor Scott Galloway calls AI ‘corporate Ozempic’ – a force of pure calculative thinking that strips away inefficiency. With this logic, a ‘Galloway agent’ is expected to buy paper towels by evaluating price per sheet, absorption rate and delivery speed. It will return the logical winner. It won’t be swayed by a cute mascot, by nostalgia, or by the feeling that Bounty is "the good one" because your mom used it.

However, Ogilvy’s Rory Sutherland argues in Alchemy that consumption value is often irrational, psychological and constructed. We pay extra for wine with prestigious labels not because the chemistry differs, but because the context creates worth.

This presents a massive risk for high-engagement brands. If agents optimise purely on specs and price, commerce becomes a race to the bottom and differentiation disappears. While I trusted Atlas to pick my travel insurance (a grudge purchase), I would never trust it to pick a fragrance, a luxury watch or a family car. The agent can process the data, but it cannot process the ‘vibe’.

The challenge for researchers is figuring out how to quantify ‘special’ into parameters a machine can weight. I suspect the answer isn't about teaching the agent ‘feelings’ because we cannot edit the frozen model of LLMs but it can about signal engineering, a discipline increasingly codified as generative engine optimisation (GEO). We must ensure that softer consumption variables (like ‘trustworthiness’ and ‘uniqueness’) are structured as machine-readable functional attributes.

How do we start simulating agents?

So, what changes operationally for consumer research? The survey is giving way to the simulation which can be clearly seen in the rise of synthetic consumer research. Now, it might be time to also simulate the gatekeepers.

However, in the B2CA world, we also face a ‘black box’ problem. We cannot interview Atlas or Comet to ask why they ranked our competitor first; their algorithms are closely-guarded and proprietary. The only way to ‘research’ a non-human buyer is to probe it.

Like synthetic consumer responses, we can also use LLMs to run adversarial simulations to generate thousands of synthetic consumer queries, mimicking different user intents, to bombard search agents and map their responses. Here, we aren't asking people what they think but we are stress-testing the AI to see what it does.

This adds a new dimension of auditing outputs to consumer research and it is driven by new questions:

• Auditing presence: Does the AI model know my brand exists (AEO at the survival level)?
• Semantic proximity: What sentiment is associated with my product in the Common Crawl data?
• Structural health: Is my product information structured in machine-readable formats (plumbing AEO)?
• Preference weighting: When Atlas or Comet compare my category, which attributes do they prioritise: price, speed, or prestige?

These questions will obviously feel strange right now but won't in the coming year. Gartner projects that machine customers will eventually participate in trillions of dollars of transactions. To capture that value, we have to stop researching the user and start researching the filter.

Are we now marketing to a team of two?

We need to retire the idea that we are marketing to either a human or a machine. We are marketing to a team.

• The agent handles the informational middle: sorting specs, prices, and reviews to collapse the search set based on the human’s question and their historical preferences/context stored in the LLM memory.
• The human handles the initial asks and experiential verification: checking if the brand ‘feels’ right before signing off.

The future of consumer research lies in the translation layer between the two. Success won't come from abandoning traditional methods, but from layering them with adversarial testing. We cannot ‘audit’ the black box of Gemini or OpenAI directly, but we can ‘red team’ it. We must constantly simulate agent-led purchases to see where our brand fails.

We still need to know why the consumer craves a specific wine label, but we now also need to ensure the digital signal structure tells the agent that this label justifies the price.

If we ignore the agent, we become invisible. If we ignore the human, we become a commodity. The winners will be the brands that can speak poetry to the person and logic to the agent.

Felicia Hu is founder and managing director at Assembled