How to Use an Analyst for Competitive Deep Dives

The first time a PMM asks ChatGPT to "do a competitive analysis of Datadog vs. New Relic," they get back four paragraphs that read like the first page of a G2 listing. The second time, they paste in a pricing page and ask the same question. The output gets sharper but still misses what they actually need: the part that maps to their deal, their ICP, their last lost opportunity. By the third try most PMMs decide AI isn't useful for competitive work and go back to a Google doc.

The problem isn't the model. It's that competitive deep dives are a context-heavy task being asked of a context-starved tool. An analyst — by which I mean an AI working session built on top of your own positioning, win/loss notes, and strategic assumptions, not a chat window with no memory — produces a different kind of output. This guide is how to set one up, what to ask it, and which questions to stop asking.

Generic prompts produce generic competitive analysis. Specificity is the entire job.

What a deep dive is, and isn't

A deep dive is not a feature comparison. Feature grids are useful artifacts and a terrible analysis. They flatten the question of why a buyer would choose one tool over another into a list of checkboxes, most of which the buyer doesn't care about.

A deep dive answers four questions about one competitor:

1. Where do they win, and why?
2. Where do they lose, and why?
3. What story are they telling the market right now, and how is it shifting?
4. What does this mean for our next three deals, our next quarter's messaging, and our next product bet?

The first three are research. The fourth — the so what — is the part most analysis skips, and the part an analyst can actually help with if you've given it enough context.

Step 1 · Load the context before you ask the question

The single biggest determinant of whether an analyst gives you useful competitive output is what you've put in front of it before the first prompt. A blank-slate model knows about your competitor from public web data. It knows nothing about you. So its analysis defaults to symmetric — "here's what each tool does well" — when what you need is asymmetric: "here's where they beat us specifically, given what we sell and to whom."

Context to load before the first prompt

This is twenty minutes of prep work. It's also the difference between analysis that surprises you and analysis that summarizes the competitor's marketing site back to you.

Step 2 · Ask the question that has a wrong answer

Vague questions produce vague answers. "Compare us to Acme" is not a question — it's a topic. The model has no way to know what shape of answer you want, so it produces the safest, broadest version: a balanced overview that helps no one.

The fix is to ask questions that could be answered wrongly. Falsifiable questions force the model to commit.

Each of these can be answered badly, which means the model has to actually think about the specifics rather than retreat to general statements. Each also references context you've already loaded — the funnel, the lost deals, the homepage. The question is doing half the work.

Step 3 · Push back on the first answer

The first answer to a competitive question is almost always too clean. It will list three reasons the competitor wins and three reasons you do, in parallel structure, with similar levels of specificity. Real competitive dynamics are not parallel. One side usually has a sharper edge in two specific scenarios; the other has a structural advantage that's harder to name.

The way to get past the parallel-structure answer is to keep asking. A few prompts that consistently surface better thinking:

• "What's the strongest version of the counterargument to your last point?"
• "If you had to bet which of these three claims is wrong, which would it be and why?"
• "What would have to be true for our positioning to be the right one in this comparison? What evidence would falsify it?"
• "Read the win/loss notes again. Which of your earlier conclusions doesn't match the texture of those calls?"

C

The first response is always the McKinsey answer. The third or fourth response is when it starts saying things I haven't already thought of. The trick is not stopping at the first answer that sounds smart.

Step 4 · Force the so-what

The output of a deep dive is not a document. It's a set of decisions someone on your team will make differently next week. If the analysis doesn't end with concrete shifts to messaging, sales enablement, or product bets, the work didn't land.

The closing prompts of a deep dive session should always be specifications, not summaries:

The fourth step is the one most teams skip. Without it, every deep dive ends with a list of reactive moves and the team's roadmap quietly drifts toward parity. Naming what you won't change is how positioning stays a choice rather than a reflex.

Questions that consistently fail

A short list of prompts that produce thin output, in case the failure modes are useful:

• "Do a SWOT analysis of [competitor]." SWOT is a 1960s frame applied to public data. The model will produce four boxes of generalities.
• "What's [competitor]'s strategy?" Too abstract. Ask instead: what's their pricing strategy, what's their ICP shift in the last twelve months, what's their messaging emphasis in Q1 vs. Q4.
• "Who wins, us or them?" The answer depends on the deal. Ask about specific deal shapes.
• "Write a battle card." Battle cards built from generic prompts get ignored by sales. The deep dive should feed a battle card the AE actually uses — the card itself is a separate craft (covered in piece #4).

We stopped asking the model "what should we do about Acme" and started asking "which of these three lost deals would have been winnable, and what would we have had to say differently." The second question gives us something to ship.

What to do Monday

Pick one competitor — the one named most often in your last quarter's losses. Spend twenty minutes loading context: positioning brief, five recent win/loss notes, their pricing and homepage, two G2 reviews. Then run a single forty-five-minute analyst session built around the four questions above, ending with the specification step.

You'll get one homepage change, one or two AE talk-track shifts, and a clearer sense of which competitive gaps actually matter. That's a deep dive. Everything else is summary.