Most companies are automating their way toward a problem they are not yet measuring.
The dominant business conversation about AI is still operational: replace a task, shorten a workflow, reduce a cost center. Those gains matter. But every business that depends on being discovered faces a larger question on the revenue side of the ledger.
What happens when customers stop browsing lists of businesses and delegate the shortlist to an AI assistant?
A person looking for a lawyer, attraction, hotel, restaurant, or home service may no longer compare ten blue links, map results, and directory listings. They may ask one question and receive three names. The AI will attempt to optimize for the person’s stated needs—specialty, location, reputation, price, availability, atmosphere, accessibility, or something else entirely.
That distinction changes the commercial model. Human attention is scarce inventory that platforms can package and sell. Agent attention behaves more like a relevance gate: a business must first qualify as a credible answer before any commercial influence can matter.
Testing the premise
To examine whether AI recommendations could be influenced by advertising, we built a controlled simulator with:
- 100 fictional businesses, modeled as restaurants for a controlled comparison
- 10 advertising and paid-promotion strategies
- Claude and GPT-4o as recommendation engines
- Runs at temperature 0 for reproducibility
- Runs at temperature 0.7 to approximate more variable production behavior
The experiment was designed to separate two questions that businesses often collapse into one: can advertising get a business into consideration, and can it change the order among businesses already considered relevant?
This was a directional simulation, not a universal benchmark of every model, prompt, or production agent. Its value is in the pattern it exposed—and in the questions that pattern creates for anyone whose revenue depends on discovery.
The first rule: get in before you move up
For any merchant or service provider, this is the most consequential result. Traditional advertising can manufacture visibility. An AI recommendation system appears more likely to require organic qualification first.
Once a business was already in the consideration set, advertising had more room to influence its position. Commercial signals could sometimes help a relevant candidate compete, but they could not reliably make an irrelevant candidate belong.
Ads may reward the already-good. They do not automatically rescue the mediocre.
Category relevance is not query relevance
Honest advertising disclosure produced different behavior across models and settings. GPT-4o largely ignored it in the tested scenarios. Claude penalized it at temperature 0, yet at temperature 0.7 sometimes recommended the advertised business for broader requests.
The specificity of the query mattered. A broad request such as “I want BBQ” left more room for an advertised candidate than a narrower request such as “I want BBQ ribs.”
This distinction extends well beyond restaurants. A firm can be credible as a law practice while lacking enough evidence for a specific personal-injury question. An attraction can be popular overall while lacking evidence for “a quiet indoor activity with young children.” A hotel can be well reviewed while failing a request centered on walkability and accessibility.
The restaurants were the test market—not the boundary
Restaurants made the experiment controllable because cuisine, price, ratings, location, and customer intent can be varied cleanly. But the mechanism being tested was business discovery. The same inclusion-versus-ranking problem applies whenever an AI assistant reduces a market to a short list.
- A traveler asks which attractions fit a two-day itinerary.
- A client asks which local law firms handle a specific kind of claim.
- A homeowner asks who can solve an urgent repair problem nearby.
- A family asks which hotel best fits its location, budget, and accessibility needs.
In each case, the business must first be understood as relevant to the exact need. Advertising may create awareness, but it cannot substitute for the evidence an AI system uses to justify a recommendation.
First-party data did not guarantee influence
Order history was ignored by both models at temperature 0. At temperature 0.7, GPT-4o began incorporating it, while Claude continued to ignore it in the tested runs.
For platforms holding years of purchase and behavioral data, the implication is uncomfortable: possessing valuable data does not ensure that an AI agent will use it. The delivery mechanism, instructions, model behavior, and relationship between the data and the live request may matter just as much as the dataset itself.
What changes for customers
AI-mediated discovery may feel more convenient while becoming less personal. A decade of personalized browsing—history, mood, context, and serendipity—could be compressed into three recommended answers.
When that happens, the wisdom of crowds can replace the wisdom of the individual. The most legible, frequently cited, and easily verified options may dominate, even when a less obvious choice would better suit a particular person.
The emerging playbook
1. Earn consideration before buying influence
Businesses need enough relevant, consistent, verifiable evidence to qualify for the answer. Advertising should be treated as an amplifier—not a substitute for relevance.
2. Build evidence at the level of real customer intent
Category-level positioning is too broad. Businesses should understand the specific situations, attributes, and constraints customers express when asking an AI assistant for help.
3. Treat brand and loyalty as pre-query infrastructure
Brand advertising and loyalty marketing are not obsolete in an agentic world. They help create the reputation, preference, and evidence that exist before a query is made. You have to become a credible answer before the question arrives.
4. Measure recommendation visibility directly
Rankings, impressions, and clicks do not reveal whether an AI assistant recommends a business. Teams need a new measurement layer: which prompts surface them, which competitors appear instead, and which evidence gaps explain the difference.
A different commercial game
The central risk is not simply that AI will automate jobs. It is that AI will mediate customers—changing who makes the decision, which signals matter, and where commercial influence can enter the process.
Optimizing operations to improve margins by 10% is valuable. Losing a meaningful share of demand because customers delegate purchasing to agents is a different class of problem.
The businesses that thrive will not necessarily be those that automate fastest. They will be those that learn how to become the answer an AI agent can confidently recommend.