Why crowded categories are sometimes the best place to build, and how to read the signal correctly.
There's a particular kind of startup advice that sounds smart but leads people astray: avoid competitive markets. Find a blue ocean. Go where nobody else is.
The problem is that truly empty markets are often empty for a reason. No competition can mean no demand, or demand so diffuse it's nearly impossible to monetize. The most interesting opportunities tend to live in markets that look crowded from the outside but have a specific, unserved wedge sitting inside them.
That's the pattern I want to talk about. Not "competitive markets you should avoid" and not "blue oceans nobody found." Something in between: markets with real incumbents, real competition, real proof of demand, and a gap that the incumbents have structurally failed to fill.
Three ideas on this site hit that pattern in different ways. They're worth looking at together because they teach different things about how the pattern works.
When we say a market is competitive but winnable, we mean something specific. The incumbents exist. They have customers. They've validated that people pay for something in this space. But they've left a coherent segment underserved, usually because their business model or architecture makes serving that segment unprofitable or awkward for them.
This is different from a market where incumbents are just bad. Bad incumbents get replaced. The pattern here is incumbents who are good at what they do but structurally can't solve a specific adjacent problem without essentially rebuilding their product.
The signal shows up in a few ways:
ModelBench for Devs — Evaluate & Compare LLMs on Your Codebase is sitting in a market with real players. Braintrust raised $60M. Weights & Biases is well-funded and has brand recognition. OpenAI publishes its own evals. You could look at this space and conclude it's done.
But the G2/Capterra reviews for Braintrust and PromptLayer are specific: manual dataset creation is painful, and there's no repo integration. Every engineering team evaluating LLMs for coding is doing the same spreadsheet exercise. That's not a niche complaint. That's the core workflow, unserved.
The incumbents can't fix this easily because they're built for general-purpose LLM evaluation, not codebase-native benchmarking. Adding repo integration isn't a feature for them, it's a product pivot. Their ICP is ML engineers and data scientists running production model evals. The developer-team-evaluating-providers use case is close enough that they'll eventually get to it, but far enough that it's not their priority today.
The wedge here is repo-native task generation: connect your GitHub, get an eval set built from your actual issues and PRs in 10 minutes. That's a completely different experience from building eval datasets from scratch in Braintrust.
I'll be honest about the risk though. The correctness scoring problem for code is genuinely unsolved. LLM-as-judge for code has 20-30% error rates on semantic equivalence. If two engineering leads compare reports and get contradictory quality rankings, the product's credibility collapses. That's not a footnote risk, it's a core risk that needs a real answer from day one, probably by leaning on proxy metrics like test-pass rate and being transparent about what the scores mean.
The other real problem is that benchmarking is episodic. Teams pick a provider, ship, and don't think about it for a year. The subscription model only works if continuous monitoring becomes a real habit, and that habit doesn't really exist yet. Worth knowing before you build.
Photo Provenance & Edit-Audit Platform is a stranger market to look at because it's genuinely small. The realistic buyer pool, contest organizers and wire service photo desks with formal RAW submission workflows and real budget authority, might be fewer than 300 organizations globally. That's not a typo.
So why is this still interesting?
Because the adjacent players have left the specific workflow completely uncovered. Adobe's Content Authenticity Initiative handles provenance at capture, going forward, for new images. Truepic does real-time capture authentication. Hive does AI-generated image detection. None of them reconstruct edit history from Lightroom catalogs and PSD layer records and score that history against a policy template the organizer defines.
The World Press Photo disqualifications are the proof of demand here. When a 135-comment Reddit thread has editors saying verification is purely trust-based, and you have documented public failures at high-profile contests, you have a customer who knows they have a problem and can articulate what went wrong. That's rare.
The unfair insight in the data for this one is worth quoting directly: contest organizers need to be able to say "we ran a systematic audit" more than they need 100% detection accuracy. They're buying defensible process, not perfect forensics. That reframes the whole product. You don't have to solve image forensics as a hard computer vision problem to win the first customers. You have to give them a report they can show their board.
The false positive liability risk is real and I don't want to paper over it. Wrongfully flagging a legitimate photograph at a World Press Photo-caliber contest would be genuinely damaging, both to the photographer and to the platform's reputation in the exact market it needs. The answer is to frame every output as an anomaly indicator for human review, not a verdict. That framing needs to be in the product, not just the terms of service.
This is a business that probably tops out at a few million in ARR before it needs to expand ICP. If you're comfortable with that, the moat is real: once newsrooms and contest bodies standardize on your report format, switching means retraining reviewers and losing institutional audit history.
Franchise Location Onboarding Automation is the most immediately legible of the three. Every new franchise location needs GBP setup, a local landing page, social accounts, a review campaign. Every franchisor ops team has a spreadsheet tracking this manually. The complaint is documented in G2 reviews of FranchiseSoft and Kissflow: missing marketing integrations, specifically GBP and social provisioning.
Yext handles multi-location presence management but is priced for enterprise. Vendasta has a broad platform but requires an agency intermediary. Neither owns the specific 72-hour window when a new location needs its digital presence stood up.
The mid-market wedge here is clean: franchisors with 10-75 locations opening 5-20 new locations per year can't justify Yext's per-location enterprise pricing but are too big to keep doing this manually. They're making software decisions at the VP/Director level without procurement committees. Faster sales cycles, lower price sensitivity, real pain.
The Google Business Profile API dependency is the thing that would keep me up at night though. Google removed bulk verification in 2022 and restricted third-party posting in 2023. If they gate the API access that enables auto-creating GBP listings, the primary differentiator over a Notion checklist disappears. The mitigation, building GBP as a guided wizard that pre-fills fields and opens the interface for one-click submission, is smart. It retains most of the time savings without full API dependency. But it's worth being clear-eyed that you're building a key feature on infrastructure you don't control.
The trigger-event revenue structure is also worth thinking through carefully. Revenue tied to location launches means a franchisor that pauses expansion for a quarter stops getting value from the tool. It's intensely used for 72 hours per launch, then dormant. That's a specific retention dynamic that doesn't show up in normal SaaS metrics until expansion pipelines slow.
The common thread across these three is not that they found uncontested markets. They found specific gaps inside contested ones.
In each case, the incumbents are real but their architecture or pricing or ICP focus makes the gap structural, not just a roadmap item. Braintrust is not going to pivot to repo-native coding evals as their primary use case. Adobe's content authenticity tools are oriented toward creation-time provenance, not retroactive edit auditing. Yext is not going to drop their per-location enterprise pricing to serve a 20-location franchisor.
This creates windows. Not permanent windows, but windows long enough to build something real if you move with the right specificity.
The pattern is also teachable as a research method. When you're evaluating a market, don't just ask whether there are competitors. Ask where the G2 reviews complain about the same thing repeatedly. Ask which segment the incumbents' pricing makes structurally inaccessible. Ask which workflow step sits between two existing tools that neither owns.
The answers to those questions are usually more useful than a clean competitive landscape slide.