AI Search Operations
Who Does the Choosing Now?
A New Model of Consumer Choice in the Age of AI
- Author
- Anton SopovLinkedInantonsopov.com
- Published
- August 31, 2026
- Length
- 7 sections, 15 min
- Sources
- 13 references
For the better part of a century, marketing has been obsessed with a single, foundational question: how do people make decisions? We constructed funnels, A/B tested copy & visuals, and built elaborate brand equity models 1. While these traditional marketing theories often clashed over the details, they all anchored themselves to one indisputable assumption.
The chooser was always human.
Today, that assumption is breaking apart. Consumers have by no means stopped making decisions. Instead, they are increasingly delegating the cognitive work and decision making to AI agents and Large Language Models (LLMs). Millions of buying journeys now initiate entirely within the architecture of artificial intelligence.
The data confirms this massive behavioral shift. More than half of American adults under thirty now use AI chatbots to search for information 23.
Adoption, though, is the least interesting part of the story. The consequential change is that the answer now ends the search. Roughly 60% of searches conclude without the user clicking through to any destination 4: the answer is the destination. And the answer does not merely inform the decision, it redirects it. Among 1,076 B2B decision-makers surveyed, 69% said an AI chatbot led them to a different software vendor than the one they had planned on 5. McKinsey puts the stake at $750 billion in US revenue flowing through AI-powered search by 2028 6.
A consumer choosing an OTC healthcare product can still reach for the brand they’ve always used. A marketing leader picking an analytics platform can still default to the tool their last three agencies swore by. Both can navigate traditional search, scroll through results, synthesize the top three, and make a decision. But increasingly, they hand the cognitive lift to AI models, which sift through hundreds of sources in seconds and surface the most relevant answer, laying out the alternatives while the human reserves the final evaluation. Sometimes they go further, asking AI to declare the “best” option and pausing only to nod in agreement. And occasionally, they skip deliberation altogether with a simple command: “Pick the best one.”
Eventually, the AI agent will no longer be a recommendation tool. It will be the decision-maker and the buyer. The consumer or enterprise buyer may never reconsider the category, and in many cases, will never need to.
These scenarios do not represent new stages in a traditional funnel. They are entirely distinct, competing systems of choice. This realization demands a radical shift in how organizations conceptualize brands, consumer behavior, and Answer Engine Optimization (AEO). At The Prompt Group, our thesis is built on this exact premise. There is a smarter way to win in AI Search, but it requires a rigorous understanding of how AI discovers, interprets, evaluates, recommends, and crucially, how that machine logic connects back to human choice.
Section 01
The Delegation Curve
The initial question for any company is no longer a matter of simple brand awareness. The vital question is: how much of the purchasing decision is delegated to AI agents, and how much remains with the purchaser? They must also consider how quickly this will change over time.
The Delegation Curve, introduced and governed by The Prompt Group, suggests that all purchasing decisions, both B2C & B2B, are probabilistically distributed across five degrees of agentic delegation, heavily influenced by the stakes of the decision, and the subjective experiences & emotions of the purchaser.
Share of purchase decisions at each degree, from The Prompt Group’s model of the curve. An individual category sits differently from the aggregate.
| Degree | Decision made by | Purchaser supplies | AI agent supplies | Stated example |
|---|---|---|---|---|
| 01 Human Choice | The purchaser | Experience, emotion, habit, trust | Nothing. Not present or referenced | A bottle of wine the family likes; the printing paper the company has used for 10 years |
| 02 AI-Assisted Choice | The purchaser | The weighing, and the decision | Retrieval and synthesis of information | Comparing laptops, then buying one two weeks later |
| 03 Dual Choice | The purchaser | Equal parts experience | Equal parts evidence and authority; ranking and recommendation | A new personal vehicle; an HR lead evaluating a benefits platform |
| 04 AI-Delegated Choice | The purchaser, on approval | Criteria and initial constraints | Most of the cognitive load, and a recommendation | Booking hotels, personal or corporate travel |
| 05 Agentic Choice | The AI agent | Criteria and constraints | The decision, the approval, and the purchase | Reordering toothpaste for home; printer ink for the office |
Every cell is drawn from the definitions below. Where the model states no value, the cell is left empty.
The five degrees of agentic delegation are:
- 01
Human Choice
In a human choice, the AI agent is not present or referenced in the purchasing moment. The purchaser makes their decision based on subjective experience, emotion, habit, and trust. You buy a specific bottle of wine because you know your family likes it, or buy the same printing paper the company has used for 10 years.
- 02
AI-Assisted Choice
In an AI-Assisted choice, the AI agent may be present, or at least referenced during the purchaser’s decision journey. In this stage, the AI agent is reserved only for retrieval and synthesis of information. The purchaser weighs all the information & criteria presented, and makes a decision based on their subjective experiences and emotions, either then, or at some future purchasing moment. You might ask AI to compare laptops, and then buy one of the recommendations two weeks later because it is reaffirmed by the expertise of a sales person at the department store you are shopping in.
- 03
Dual Choice
A dual-choice decision is equal parts the purchaser’s experience and the AI engine’s evidence and authority. At this stage, the purchaser is asking the AI agent to rank their options and provide recommendations based on deeper analysis or research criteria. Examples of these decisions include consumers purchasing a new personal vehicle, or an HR lead evaluating a new corporate benefits platform.
- 04
AI-Delegated Choice
In this stage, the decision making is now mostly being outsourced to the AI agent. The purchaser uses their subjective experiences to set criteria & initial constraints, and leaves the rest of the cognitive load to the AI agent. The purchaser then approves or denies the AI agent’s recommendation, and makes a decision. Booking hotels, both for personal or corporate travel, are great examples of AI-delegated decisions.
- 05
Agentic Choice
The main difference between an agentic choice and an AI-delegated choice is that in an agentic choice, the purchase decision is not approved by the human; the decision is made by the AI agent. The purchaser’s subjective experiences are still responsible for setting criteria and constraints, but after that is done, the decision is made, approved, and purchased all by the AI agent. Reordering toothpaste for home, or printer ink for the office, would be good early examples of these decisions.
The decisions that fall under every segment of the curve will shift right over time as AI agents improve and purchasers grow more comfortable trusting them. The movement will not always be orderly. Decisions made by human choice today can skip delegation entirely and land in agentic choice, with no interrogation phase in between.
Three questions become mission critical for companies & marketers. Where do my purchasers sit today? How fast will they move? And what does that require of my presence inside AI answers?
Section 02
The Dual Engines of Choice
The vast majority of high-value commercial decisions will occur in the messy middle ground between full human control and full cognitive delegation. This is the Dual Choice System.
| The Human Engine | The AI Engine | |
|---|---|---|
| Runs on | Experience, emotions, trust, values, habits | Information retrieval, synthesis, pattern recognition, criteria weighting |
| Operates on | Availability and convenience | Claims evaluated and evidence assessed |
| Drives | Awareness and deeply held preferences | Visibility and cold, logical recommendation |
Understanding the interplay between these forces is the new frontier of marketing science. When does algorithmic authority override human loyalty? When does an unfamiliar brand gain instant credibility simply because a machine vouched for it? The Delegation Curve maps exactly how this plays out in commercial reality.
Section 03
The Impact of Decision Stakes
Behavioral theory offers a profound lens here. Daniel Kahneman’s exploration of fast and slow thinking popularized the reality that humans deploy varying degrees of cognitive effort depending on the situation 7. AI introduces a fascinating new variable: we are no longer merely asking how much thinking a consumer is willing to do 8. We must now ask: who will actually do the thinking?
Where a decision lands on The Delegation Curve is governed largely by what the purchaser has at stake.
The stakes of a decision are influenced by four factors:
- 01Financial Magnitude
- 02Irreversibility
- 03Identity Attribution
- 04Decision Scrutiny
The weight of each of these factors in the total “Decision Stake” is influenced by the business and the subjective experiences of the purchaser.
Counterintuitively, higher decision stakes does not mean less AI involvement, it often means more AI, but significantly less delegation. Where a purchaser sits on The Delegation Curve determines whether they are asking AI models to help them understand their options, versus asking AI to tell them what to buy, and eventually buying it for them.
A consumer buying a fifty-dollar appliance might accept a single AI recommendation without question. A B2B purchaser responsible for migrating the entire company to a new logistics platform might interrogate their AI model of choice for a week, demanding evidence, comparing sources, & creating visualizations, while fiercely guarding the right to make the final call. High stakes elevate both the depth of machine involvement and the intensity of human oversight simultaneously.
For marketers, this distinction becomes everything. This creates a dilemma; companies must suddenly persuade a decision-maker that marketing theory has not had to consider for a century: the machine itself.
The AI-Mediated Choice Model
To navigate The Delegation Curve, The Prompt Group developed the AI-Mediated Choice Model, which maps the anatomy of a decision once AI enters the loop.
Before a machine can choose a brand, a purchaser chooses the machine. There is no universal AI consensus. Each model stocks a different shelf. Across 3,750 responses to 250 unbranded category queries, three leading models agreed on the single top-recommended brand in only 41.6% of cases 9. Ask the same unbranded question across platforms and the responses diverge not only in which brands appear but in how many: one model returns a shortlist of three, another returns twelve. That list length sets the odds of inclusion before a single word of brand messaging is read.
Question one: do I appear in AI?
Inclusion is the first question a brand must answer: does the AI mention us? In order to measure inclusion, companies will track synthetically generated “unbranded questions” across all the AI models they care about, and measure:
- 01What % of those questions are they included in?
- 02When included, what is the average ranking of their mention?
- 03What sources are being crawled in order to answer these questions?
Everything else is basically just filters on top of those 3 core data points. These questions are top-of-funnel and awareness based questions.
Current solutions in the market answer the inclusion question well. But the inclusion question only solves for the top of the funnel. Purchasers are not making decisions here.
Question two: how do I appear in AI?
There is a second question. One that is significantly more important for many of the world’s companies, and has much more influence on how purchasers make decisions: “How do I appear in AI?”
Many of these prompts are branded questions. They mention the company name in the prompt. If you want to try & track branded questions with the market’s current early solutions, you must include them in the same prompt set as all of your unbranded prompts, which measure: do you appear in AI?
But of course you will appear in AI for a question that mentions your company name in the prompt. The AI model is actively searching for your company’s context across the web. But now your inclusion data is skewed because your branded prompts are contributing to the visibility score & ranking measures of your unbranded questions.
This creates a dilemma that the industry has not attempted to solve yet. How you appear in AI requires different tracking methodologies than measuring do you appear in AI.
The Agent Promoter Score
The Prompt Group introduces the “Agent Promoter Score” to solve this dilemma and enable brands & enterprises alike to control how they appear in AI.
Akin to the familiar “Net Promoter Score” 10, the “Agent Promoter Score” is an AI agent loyalty measure. The Agent Promoter Score represents how accurately, enthusiastically & loyally an AI agent will recommend your company. It is composed of 4 types of user questions:
- 01
Knowledge Accuracy & Factuality
How accurately is the AI agent representing my company inside of AI answers?
Example promptWhat is the price of [company]?
- 02
Subjective Perceived Value
How favourably & positively does an AI agent speak about my company, when purchasers ask subjective or opinionated questions about the brand?
Example promptIs [company] a good choice for me?
- 03
Competitive Loyalty
When the purchaser’s questions are comparative, how loyal is the AI agent to your company, or does it recommend competitors more often?
Example promptShould I hire [company X] or [company Y] if I care about quality of service? Which one should I choose?
- 04
Strategic Outcome Effectiveness
Does the AI strategically guide the user toward the intended outcome, and cite the highest value sources to accomplish this effectively? Does the AI agent intentfully use language, sources & links that guide the purchaser towards the next business step and intended outcome (booking a demo, resolving a ticket, an upsell)?
Choice Survivorship
What connects question 1, “Do I appear in AI?”, with question 2, “How do I appear in AI?”, is a metric The Prompt Group has coined “Choice Survivorship”. Choice Survivorship measures how often a company that gets included in AI answers retains that recommendation as purchasers begin asking follow-up questions.
Consider the following two scenarios:
Case 01 · B2B
The airline that was recommended and still lost
An office administrator is booking flights for a company-wide Christmas party. The administrator responsible for coordinating flights is using AI to help her: she is in a dual choice delegative state during this purchasing process. AI recommends a carrier for the employees flying out of Canada. She asks it to confirm the fare, and it returns a figure from an authoritative source without noting the price is in USD. Converted to CAD, the flight is over budget. She moves on.
The airline was recommended and still lost. No competitor outsold it. It was beaten on the second question by a missing currency label.
Case 02 · B2C
The SUV that lost a question it had never answered
A buyer asks for family SUVs under forty thousand and gets four names. One catches her eye. She has two kids in car seats and a ninety minute winter commute, so she asks the obvious follow-up: is this a good choice for me? The agent says no. It tells her the model reviews well but is generally positioned as a city crossover, and points her toward two other names on the same list for family use.
Nothing the agent said was false. It just could not find a source describing who this vehicle is for, so it answered from the safest read available. She books a test drive with one of the others.
The brand was on the list, in her budget, and right for her. It lost on a question about fit it had never answered anywhere.
This is Choice Survivorship. Not whether a brand appears, and not how it appears, but how far into a conversation it holds. Inclusion measures whether a company appears in AI answers; the Agent Promoter Score measures how favourably and accurately it appears. Choice Survivorship connects the two, measuring conversational retention: how likely a company is to hold its recommendation once included, as the conversation progresses down the funnel to more specific follow-up questions.
Section 04
A Working Model of the Complete System
The full architecture of AI-mediated choice can be drawn as a decision tree. Every purchase occasion enters at the top. The Delegation Curve routes it, sorting it into one of the five degrees of delegation based on the stakes of the decision and the subjective experience of the purchaser.
Human Choice exits right away. The machine is never consulted. Nothing below that point runs, and the decision is settled by habit, emotion, and trust.
The other four degrees all enter the machine-choice sequence, and they run it in the same order.
Human choice exits before the sequence begins. The remaining four degrees run the same four nodes in the same order; the degree governs how much weight lands on each.
- 01
Model Selection
Before a machine can choose a brand, a purchaser chooses the machine. Each model stocks a different shelf, and each shelf holds a different number of slots.
- 02
Inclusion
Does the brand make that shelf, and where does it land in the order?
- 03
Agent Promoter Score
Once the brand is on the shelf, how is it described? This covers how accurately its facts are stated, how favourably it is characterized, how loyally it holds up against a named competitor, and how well the answer moves the purchaser toward the next step.
- 04
Choice Survivorship
Does that description hold as the questions get specific? This is where the purchaser stops asking what the options are and starts asking about price, fit, terms, and value.
Where a decision sits on the curve does not change which nodes it passes through. It changes how much each node matters.
Take an agentic choice, like an AI agent reordering the office printer ink. The agent moves through all four nodes in seconds. Inclusion settles it. If the brand is on the shelf when the agent looks, it gets bought. There is no follow-up question, because there is no human in the loop to ask one.
A dual choice evaluation passes through the same four nodes, but almost all the pressure lands on the last two. Inclusion barely matters here, because a dozen brands made the shelf. The purchaser then spends a week asking follow-up questions, and what decides it is the Agent Promoter Score and Choice Survivorship: how the brand gets described, and whether that description holds up under questioning. Every path, fast or slow, ends at Choice.
Look at the two examples again.
The airline was included. It failed on Knowledge Accuracy and Factuality, the first pillar of the Agent Promoter Score, when the agent quoted a fare without labeling the currency.
The SUV brand was included and ranked near the top. It failed on the second pillar, Subjective Perceived Value, when the agent was asked whether the vehicle suited this particular buyer and said no. The two failures are not the same kind. One brand had a fact stated wrongly. The other had no story about itself for the agent to find, so the agent wrote one.
Both broke at Choice Survivorship. Both cleared the one node the market knows how to measure, and lost at a node almost nobody is measuring. That is where the decision was actually made.
Section 05
For the CMO: Reframe the Metrics of Success
A company can dominate human choice and still lose the category. The two engines are measured differently, and the metrics most marketing organizations report today describe only one of them.
- 01
A category has to be located on The Delegation Curve before anything else can be budgeted.
What share of its decisions are still human choice, what share has already moved to dual choice, and what share is drifting toward agentic? That distribution governs budget, channel & urgency, and marketing organizations have not yet been enabled to describe it for their own category.
- 02
Branded and unbranded prompts cannot share a prompt set.
When they do, the visibility score is inflated by the very questions that named the company, and the inclusion data stops describing anything real. “Do I appear in AI?” is measured against competitors; “How do I appear in AI?” is measured against the company’s own facts. Two questions, two prompt sets, two benchmarks.
- 03
Inclusion buys a place on the shelf; it does not win the decision.
That is settled several questions later, which is what makes Choice Survivorship the measure that matters: how often a company holds its recommendation once the purchaser begins asking about price, fit, and terms.
Section 06
For the Marketing Team: Engineer the Answer
SEO earns a place in the list. AEO earns the right to be the answer 11.
- 01
The knowledge base comes before the content calendar.
Every fact a purchaser might ask about, pricing, currency, availability, terms, and who the company serves, approved and held in one place. Until that exists, the accuracy of an AI answer is left to whatever the model can assemble on its own. The airline lost on a missing currency label, which is the least expensive failure in this entire model to prevent.
- 02
The fit question has to be answered in public.
The SUV lost because no source described who the vehicle was for, so the agent inferred an answer, and inferred it unfavourably. A company that publishes who its product is for, and who it is not for, gives the model something to defend when a purchaser describes their own circumstances. Naming the non-fit is what makes the fit credible.
- 03
Corroboration has to come from outside the domain.
Traditional SEO still supplies the fundamentals, and a well-built, well-structured website remains the base an engine reads from, but only around 15% of a brand’s presence in AI answers traces back to its own domain 1213. The rest is what independent sources say about it, which puts most of the work outside the pages a company controls.
Section 07
Conclusion
The chooser is no longer only human, and not yet only the machine. Choice now sits distributed between the two, and every day that distribution moves further right along The Delegation Curve.
This makes AI brand strategy a measurement problem before it is a marketing problem. Inclusion tells a company whether it is on the shelf. The Agent Promoter Score tells it how it is described once it is there. Choice Survivorship tells it whether that description survives the questions that actually decide the purchase. Most organizations can answer the first today, and have not started on the other two.
That gap is the opportunity. The companies that win their category in AI will not be the ones mentioned most often, but the ones that are still the recommendation several questions later.
References
- [1]Aaker, D. A. (1991). Managing brand equity: Capitalizing on the value of a brand name. New York: Free Press. ISBN 978-0-02-900101-1.openlibrary.org
- [2]AEO Updates. (2026). State of AEO 2026 research.aeoupdates.com
- [3]Pew Research Center. (2026, June 17). Americans and AI 2026: Chatbots, smart devices and views on impact. Survey of 5,119 U.S. adults, American Trends Panel, fielded February 17–23, 2026.pewresearch.org
- [4]Bain & Company. (2025, February 19). Consumer reliance on AI search results signals new era of marketing. Bain & Company research brief. Sample size and field dates are not disclosed in the published release.prnewswire.com
- [5]G2. (2026, March). The answer economy: How AI search is rewiring B2B software buying. G2 2026 AI Search Insight Report. Survey of 1,076 global B2B decision-makers, supplemented by 39 qualitative interviews.learn.g2.com
- [6]McKinsey & Company. (2025, October 16). New front door to the internet: Winning in the age of AI search. By E. Silliman, J. Boudet, & K. Robinson. Figure is a McKinsey projection for U.S. revenue; accompanying survey data from the McKinsey AI Discovery Survey (n = 1,927 U.S. consumers, August 2025).mckinsey.com
- [7]Kahneman, D. (2011). Thinking, fast and slow. New York: Farrar, Straus and Giroux. ISBN 978-0-374-27563-1.openlibrary.org
- [8]Chiriatti, M., Ganapini, M., Panai, E., Ubiali, M., & Riva, G. (2024). The case for human–AI interaction as system 0 thinking. Nature Human Behaviour, 8(10), 1829–1830. Correspondence.doi.org
- [9]Żatuchin, D. (2026). Who owns the AI recommendation? A multi-industry empirical map of brand category ownership across large language models (arXiv:2606.23057) [Preprint]. arXiv. 3,750 responses to 250 unbranded queries across five industries on GPT-5.2, Gemini 3 Flash, and Perplexity sonar-pro.doi.org
- [10]Reichheld, F. F. (2003). The one number you need to grow. Harvard Business Review, 81(12), 46–54.hbr.org
- [11]Aggarwal, P., Murahari, V., Rajpurohit, T., Kalyan, A., Narasimhan, K., & Deshpande, A. (2024). GEO: Generative engine optimization. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (pp. 5–16). Association for Computing Machinery. Open-access preprint: arXiv:2311.09735.doi.orgarxiv.org
- [12]AirOps. (2025, October 17). The influence of offsite signals in AI search: Third-party sources drive 85% of brand discovery. Analysis of 21,311 brand mentions across 500+ commercial-intent queries on ChatGPT (GPT-5), Claude Sonnet 4.5, and Perplexity Sonar across six B2B verticals.airops.com
- [13]Ash, I. (2026). Beyond domain authority: The five domains of AI visibility. AEO Updates.aeoupdates.com
Applying the model
The three measures in this paper are the ones we run as engagements: locating a category on the curve, scoring the Agent Promoter Score against a company’s own facts, and testing Choice Survivorship through the follow-up questions that decide the purchase.