How business context changes AI recommendations

Every AI success story seems to end the same way. Someone shares a remarkable output, and almost immediately someone asks, “Will you share the prompt?”

It’s a reasonable request. We often give prompts too much credit. By the time someone writes a prompt, they’ve already defined objectives, gathered context, weighed tradeoffs, and decided what success looks like.

Prompts capture the result of conversations, assumptions, revisions, and editorial judgments that came long before anyone started typing.

To see how much that process matters, I gave the same strategic assignment to ChatGPT, Claude, and Gemini, changing only the information that came before the prompt.

The experiment

The assignment was straightforward. A company had invested in SEO for years, and its website was well established. Leadership knew search behavior was changing as AI-generated answers became more common, but no one knew what, if anything, should change about the marketing strategy.

I gave the same assignment to ChatGPT, Claude, and Gemini. The goal was to create a strategic roadmap identifying:

• What information should be gathered.
• Which opportunities deserved attention.
• Which assumptions required validation.
• What should happen before content creation began.

I wasn’t trying to identify the “winning” model. I wanted to see how the responses changed as the assignment evolved.

First run: The models filled in the missing intent

At first, I believed I’d crafted a perfectly reasonable prompt. Then I realized what I’d actually written was a perfectly reasonable assignment. Rather than reproduce it line by line, here’s what it included:

• A request for strategic guidance rather than marketing copy.
• An explanation that it was for an established business concerned about its visibility in AI-generated search.
• Instructions to identify missing information to prevent hallucinations and invented facts.
• An expectation that the response would prioritize planning over execution.

The complete opening assignment is shown in Figure 1.

It defined the problem, established the type of help I needed, and discouraged unsupported assumptions. It also left room for interpretation.

Then I ran the assignment through all three models. The responses came back quickly. They were thoughtful, well-organized, and complete. 

Each would have seemed reasonable on its own. Side by side, however, they revealed something important: They weren’t really answering the same question.

• Claude treated the assignment as the beginning of a discovery project. 
• Gemini interpreted it primarily as an AI search optimization project. 
• ChatGPT expanded it into a formal consulting framework with governance, measurement, and phased implementation.

None of the interpretations were unreasonable. The issue was that the models hadn’t been told which business problem mattered most.

Was the problem declining traffic? Brand visibility? AI citations? Lead generation? Competitive pressure? Future readiness? Each requires a different strategy. 

Without enough information to understand the underlying intent, each model supplied its own interpretation. When an assignment leaves room for interpretation, models don’t simply fill in missing facts. They fill in the missing intent.

I stopped thinking about how to improve a prompt and started thinking about how to improve the brief.

Second run: Better context, better recommendations

Experienced marketers begin with conversations. They learn what success looks like, which customers matter most, where the business makes its money, what constraints exist, and which compromises are acceptable. Then they turn that information into a functional brief.

That’s what my process intentionally skipped for the experiment.

For the second run, I kept the original assignment largely intact and added the information that would normally emerge during those conversations.

The revised brief established the client as a regional HVAC company with a mature website, a limited budget, and a preference for improving existing assets before creating significant new content. It also clarified the objective of increasing qualified service inquiries during the upcoming season, with maintenance agreements representing the highest-value long-term outcome.

The revised brief is shown in Figure 2.

I didn’t add an elaborate prompting formula or ask the models to reason in a particular way. I supplied the information a strategist would ordinarily gather before recommending a course of action.

The difference was immediate. All three models retained their original approaches, but their recommendations were grounded in the same business.

• Claude remained discovery-oriented, focusing on seasonal demand, maintenance agreements, existing assets, and budget prioritization.
• Gemini remained more technical, tying its recommendations more closely to local service questions and commercial constraints.
• ChatGPT retained its framework-driven approach, prioritizing repair-versus-replacement decisions, maintenance-plan development, and protecting the company’s existing SEO environment.

Additional business context transformed the assignment into a brief. The responses became more detailed, more commercially relevant, and more closely aligned with the same business challenge.

The quality of the response depended less on the sophistication of the prompt than on the quality of the business context behind it.

Final run: Where human judgment still matters

A better brief brought the models closer together. It didn’t automatically make their recommendations useful.

All three generated more possibilities than a regional business with a limited budget could reasonably pursue before the upcoming season. They identified opportunities, organized them, and explained why each might matter. Choosing among them still required strategic judgment.

Choosing among the recommendations required strategic judgment. I evaluated them by a different standard:

• Would I present this to a client? 
• Could they act on it? 
• Was it supported by available evidence? 
• Did it align with the objectives, or was it simply a reasonable SEO activity?

Any recommendation could be technically valid and still be a poor investment. A content opportunity might appear useful but lack evidence of customer interest or demand. A measurement framework could be thorough but unrealistic, given available staffing or systems.

A few priorities consistently survived across model revisions:

• Establishing a baseline for how the company appears in commercially significant AI-generated answers.
• Measuring qualified leads, booked work, and maintenance agreement growth alongside citations and AI referrals.
• Defining what counts as a qualified inquiry before treating visibility as a success.
• Auditing existing service pages, local business information, and conversion paths before commissioning substantial new content.
• Creating new assets only when existing pages couldn’t address a validated information gap.
• Emphasizing urgent seasonal decisions and maintenance agreement opportunities rather than a sitewide transformation.

The models expanded the field of possibilities, but strategic judgment still needed to narrow it.

Prompts are evidence, not explanations

A prompt screenshot typically conceals how the assignment framed the task, how the brief added business context, and how human judgment established priorities.

Exact prompts are useful because they show how a task was framed and what information the model received. They don’t reveal what the user already knew, what the first run exposed, what assumptions were rejected, or why the next revision changed.

That’s why a prompt that produced an impressive result in one situation delivers something merely adequate in another. The words are the same, but the thinking behind them isn’t.

A prompt makes discovery, research, business judgment, and editorial review visible to the model. It doesn’t replace them.

What the prompt leaves out

The experiment started with an assignment. It evolved into a brief. Human judgment established the priorities.

A prompt is the first visible artifact of that workflow. We ask others for prompts because they’re easy to share. They fit into a screenshot, a social post, or a slide.

The conversations, changing assumptions, false starts, and judgment calls behind them are harder to package. They’re also far more instructive.

The next time someone shares an impressive AI output, asking for the prompt may still be worthwhile. Just don’t mistake it for the full explanation. The more useful question is how they got there.