The 6 Rules for Deciding What to Build When Competitors Copy You Every Few Weeks

Stop deciding what to build by watching what competitors ship. When an engineer with Claude Code can rebuild a rival's feature in a weekend, features are no longer the moat. Understanding your users faster than anyone else is. Decide what to build with six rules: ignore competitor launches, listen to what users say about competitors, build where unmet needs run deepest, invest in what compounds, measure the retention effect of every release, and speed up the evidence loop, not just the build loop.

The short answer: your roadmap should come from user signals, not competitor changelogs. The team that hears switching talk, comparisons and unmet needs first, and then proves what moved retention, makes better bets than the team that simply ships fastest.

Why features stopped being the moat

Every AI product category moves the same way right now. Someone ships on Tuesday. Two rivals have their version within weeks.

Engineering got cheap. A thin team with agents can ship a credible version of almost anything in days. Building is no longer the bottleneck. Deciding is.

Most moat advice reads like a strategy deck and never tells you what goes into next sprint. Reacting to every launch is worse: a roadmap written by someone else's PM, chasing parity that lasts about two weeks.

Features are copyable. The speed at which you understand your users is not. For AI-native products, that is the moat.

The 6 rules for deciding what to build

1. Stop chasing competitor launches

A competitor launch tells you what they decided to build. It says nothing about whether their users wanted it, or whether yours do.

Treat every launch as a question, not a ticket. Did your users ask for this before the rival shipped it? Are they asking now? If not, let it go.

2. Watch what users say about your competitors

The useful competitive signal is not in a rival's changelog. It is in their users' posts. Switching talk ("moved over because..."), side-by-side comparisons and the requests users make of a rival are the clearest map of where your category is underserved.

Feedback Bench, built by Enterpret, shows what this looks like at category scale. It ranks AI agents in a category, currently coding agents and personal agents, by what their users say in public on channels such as Reddit, X, YouTube, app stores, G2 and Trustpilot, depending on the category. Each agent gets a Feedback Score that weighs popularity and customer love equally, across more than 50 criteria per category, and every page lists the posts behind its numbers.

One of its published analyses is titled "Hermes Agent users choose their models, and Grok Bot users ask to do the same." That is a roadmap signal hiding in a rival's feedback: a need one product meets and another product's users are asking for. Do the same with your own lost users and track which competitors customers switch to when they churn, and why.

3. Build where unmet needs run deepest

Loud is not the same as deep. A heavily upvoted request in your community can be a nice-to-have. A complaint that a handful of enterprise accounts raise in tickets and sales calls can be the reason they don't renew.

Score every need on three things: how many users hit it, how badly it blocks their work, and who those users are. Thin teams can only afford to build where all three line up.

4. Invest in what compounds

Some bets get more valuable every week. Those beat any single feature.

• Workflow depth: the more of a user's real job your product owns, the more painful switching becomes.
• Data and context: usage, corrections and history that make the product better for each user over time.
• Trust: reliable behavior, honest reporting of what the agent did, and controls users can see. In AI products, trust builds in small moments and breaks in one bad run.
• Integrations: every connection into the tools a team already uses makes you harder to rip out.

A competitor can copy your feature in a week. They cannot copy a year of compounding.

5. Measure the retention effect of what you ship

Shipping is not the result. Retention is. For every release, check whether adopters stay longer, complain less or expand, and whether the complaints it targeted actually dropped.

Compare churned vs retained customer feedback before and after each release. If a feature moved neither retention nor the related complaint volume, it was parity, not a moat. Stop investing in it.

6. Speed up the evidence loop, not just the build loop

Your build loop is already fast. The evidence loop is the slow one: collect, categorize, spot what changed, decide. At many AI startups it runs on someone pasting a Discord export into an LLM on Friday.

That breaks in a specific way. In Enterpret's study of run-to-run instability across eight frontier models, a raw model changed more than half of its top theme set between two runs on the same data at 100 records, and nearly three quarters at 5,000. If your top five pain points reshuffle every time you ask, you cannot tell whether last week's release moved anything.

86 to 88% Reduction in theme churn when an AI agent is grounded in a persistent taxonomy instead of re-clustering raw feedback on every run, with the model held constant. Source: Enterpret Research, "Same Feedback, Different Answer" (arXiv, 2026)

The same study found grounded runs reported identical counts for matched themes every time, while raw models' counts moved 7.5 to 9.8%. Stability is not accuracy, but you cannot measure change on numbers that wobble by themselves.

Where Enterpret fits

Enterpret is the customer intelligence layer for the evidence loop. It sits upstream of your roadmap and brings user signals into one place: community channels like Discord, App Store and Google Play reviews, support conversations, sales calls and surveys.

Adaptive Taxonomy categorizes feedback consistently and updates as you ship, so categories don't go stale every sprint. The Customer Context Graph joins feedback to accounts and revenue, which matters the moment enterprise deals arrive. Wisdom answers questions in natural language, and the MCP server brings the same answers into Claude, Claude Code, Cursor, Codex, Glean and Notion. If your community lives on Discord, see the MCP servers for Discord community feedback.

Once sources are connected and processed, AI agents and alerts flag new or spiking issues after a release and deliver them to Slack, Jira, Linear or Notion. Enterpret does not replace evals or tracing tools like Braintrust and Langfuse: they show how the model behaved, Enterpret shows how users experienced it.

What this looks like in a real week

Picture an AI writing assistant. On Monday a well-funded rival ships brand voice profiles. The team's instinct is to match it by Friday.

Instead, the Head of Product asks Wisdom two questions. Have our users asked for brand voice? A few, mostly free users. When users mention the rival, what are they praising? Its docs integration, not voice profiles.

So the team skips the copy, ships a deeper docs integration, and watches two numbers: complaints about copy-pasting between tools, and retention for adopting accounts. The evidence decides the next sprint.

The new moat is understanding

The next generation of AI products won't be won by whoever ships the most features. It will be won by teams that hear their users, and their competitors' users, first, and act with conviction.

Where your team sits on the customer feedback maturity model tells you how fast that loop can run today. Move your moat upstream. Own the evidence loop. Then build.

FAQ

Should we ever copy a competitor's feature?

Yes, when your own users are asking for it and its absence shows up in churn, lost deals or switching talk. Table-stakes features that block adoption are worth matching. Copying only because a rival shipped it usually wastes a sprint.

What is a defensible moat for an AI startup if features aren't?

The bets that compound: depth in the user's workflow, data and context, trust earned through reliable behavior, and integrations into tools teams already use. Underneath them all is understanding users faster than competitors do.

How do I track what users say about competitors?

Watch the public channels where users compare products, such as Reddit, X, app stores and review sites, and pull competitor mentions out of your own tickets and sales calls. Feedback Bench publishes this kind of analysis for AI agent categories, ranking agents by public user feedback and listing the posts behind each number.

Can I just use Claude or ChatGPT to analyze our feedback?

For a quick look, yes. For tracking change release over release, it is risky: Enterpret's stability study found raw frontier models changed more than half of their top themes between runs at 100 records, and grounding the agent in a persistent taxonomy cut that churn by 86 to 88%.

How do I know if a feature we shipped actually mattered?

Compare retention, expansion and related complaint volume for users who adopted it against those who didn't, before and after the release. If the complaints it targeted didn't fall and retention didn't move, treat it as parity, not differentiation.

If you want your evidence loop to run as fast as you ship, see how Enterpret works.

Research cited

1. Bagal, V., Ganta, R., & Chellingi, P. (2026). Same Feedback, Different Answer: Measuring Run-to-Run Instability in Frontier-Model Customer Feedback Analysis. arXiv:2610.08036.