Competitive analysis with AI: the map versus the coordinates

Most competitive analysis produces a feature matrix. Twelve competitors down the side, fifteen features across the top, a grid of checkmarks in the middle. It's a deliverable, and it's almost useless. Nobody makes a product decision because a competitor has a checkmark you don't.

The point of the exercise is a strategic implication — one sharp “therefore” that changes a product conversation. AI can help you get there much faster than reading twenty landing pages by hand, but only if you're clear about what it's good at and, more importantly, where it will confidently lie to you.

The rule I work by: Claude gives you the map. You spot-check the coordinates.

Descriptive versus strategic

Here's the difference the whole thing turns on. Descriptive output sounds like: “Enterprise players ignore the collector, AI-native players are automating the collector out.” True, tidy, and inert. The strategic version: “No player is competing on making collectors better at their job. If we own that position, we're not competing with anyone directly.”

Same underlying facts. The second one is a bet you can take into a roadmap meeting. The landscape map is the evidence; the implication is the headline. If your analysis ends with the map, you stopped at half the work.

Where the tool is actually strong — and where it isn't

I'm specific about this because the failure mode is trusting it evenly across the board.

Claude is strong at mapping player categories and segments quickly, reading positioning and messaging patterns across a set of competitors, synthesizing secondary sentiment — G2 reviews, Reddit threads, App Store comments — into patterns, and acting as a thinking partner when I bring my own observations back from actually using the products.

It is weak at knowing what a product feels like to use, at anything recent — updates, pricing, new features all sit past the training cutoff — at niche regional players, especially Eastern and Southern EU fintech, and at specific statistics or market-size claims. It will produce all of these fluently, and some of them will be wrong. The wrong ones read exactly like the right ones.

The three phases

Phase 1 — Landscape, Claude-heavy. Start by defining the decision you're making, then use Claude to map the player set, segment it, and surface the dominant positioning patterns. The output isn't a list — it's a landscape map with a gap hypothesis attached.

Phase 2 — Deep dives, you-heavy. Go use the actual products. Take screenshots, document the flows, feel where they're clumsy. Then bring that material back to Claude to stress-test your read. This is the phase that makes the analysis real, and Claude can't do it for you. I don't hand it “here's what I found” — I hand it a specific hypothesis and ask it to confirm or break it.

Phase 3 — Synthesis, Claude-heavy again. Bring everything together and use Claude to articulate the insight and the strategic implication, not just the observation. This is where “here's the landscape” becomes “here's the therefore.”

How to write the landscape prompt

Every strong landscape prompt has three parts: scope (what slice of the market matters, not the whole thing), lens (the specific axis you're analyzing against), and output goal (what a useful answer looks like — a gap signal, not a list). Drop any one and Claude defaults to broad and generic.

The weak version: “I'm designing a B2B debt collection tool for a mid-size EU bank. Help me map the competitive landscape.” The strong version: “I'm designing a B2B debt collection tool for collectors at a mid-size EU bank with about 50k active clients. Map the landscape to show how existing players frame the problem they solve — specifically whether they position around the collector's daily workflow or around executive and compliance outcomes. I want to know if there's a gap in how the market talks about the collector's experience.” Scope, lens, output goal — all three present. The answers you get back barely resemble each other.

Verify by trust level, not everything equally

After Phase 1 I triage. Some of it Claude is reliable on and I move fast; some of it I always check.

High-trust, take it and go: category structure (who the player types are), positioning patterns (how competitors talk about themselves), and the gap logic itself (if X and Y, then Z). Low-trust, always verify: specific features attributed to a named tool, pricing and tiers and recent updates, niche regional players and their real segments, and any hard statistic or market-size number.

The verification method is cheap. Take the player names, open their homepage and pricing page, read the hero copy. Five minutes a competitor. Check whether Claude's positioning read matches what they're actually saying about themselves. And one hard rule: never ask Claude for source links. It will generate URLs that look completely real and may not exist or may be years stale.

What I bring to the PM

Not a competitor breakdown. A strategic implication with the breakdown as supporting evidence, in that order. One sharp insight — “no player in this market is positioning around the collector's daily experience.” The strategic therefore — “which means the user is underserved by design, and that's a defensible position if we take it.” Then the evidence: the landscape, segmented by how each player approaches the problem.

The gap-detection move underneath all of it: if nobody in a market is talking to a specific user, that's signal, not coincidence. Positioning analysis is the tool. You're not hunting for missing features; you're hunting for missing people in the conversation. A feature matrix can't show you that.