Mapping the Real Competitive Set Behind One Pitch

You have a deck that claims a huge market and almost no competition, and a partner meeting on the calendar. This guide is for early-stage investors, platform and talent partners, and angels who need to independently find who else is building this thing and judge whether the founder's "no real competitor" line holds. It carries one worked deal - a US-based AI voice agent company - all the way through the actual searches, the intermediate counts, and the wrong turns, ending with a named competitive map and a defensible verdict.

Published diligence checklists tell you to "list direct and indirect competitors" and stop. That is not a method. What follows is the method: the queries across code, professional profiles, and the open web that turn a blank slide into a named set with the people behind each entry.

Why "no competition" is a signal, not a finding

"We have no competition" is a documented red flag, not a claim to be taken at face value. Seasoned investors treat it as evidence the founder either has not done the research or is naive enough to believe there is nothing comparable, and multiple named publications list it outright.

The reasoning is simple. There is always a substitute, even if it is a spreadsheet, a call center, or the customer doing nothing. Roughly 20% of startups fail because competitors outperform them, per a 2024 Stripe figure cited in an investor guide, so a founder who cannot name a rival has usually not looked hard enough to survive.

The line does not settle anything. It reframes your job: prove the set is genuinely empty, or find the names the founder left out. An empty market is not automatically white space. It can be a closed opportunity - a market too small to generate venture-sized returns - or a sign no one has the pain point, because if no one else were addressing it, there might not be a real problem to solve.

The worked case: an AI voice agent deck

Our example deal is a seed-stage company building an AI voice agent, pitching a large market with, in their words, "no direct competitor." I will carry this exact case through every pass below, with the real counts from Refolk's index.

The first fork appears immediately. If you take the founder's category coinage and search it exactly, you get almost nothing - which reads as confirmation. That is the trap.

The exact phrase looks like proof of an empty field. But the same index, queried for the broad stealth signal, returns a different universe entirely. The category is not empty. Your query was narrow.

The roughly 220x ratio is derived (1,757 divided by 8). It is the single most important number in this guide, because it quantifies exactly how a naive search manufactures the illusion the founder is selling. Founders name their category with novel words. Vary the terms and the "no competition" read evaporates.

The competitive spectrum before you search

Define four buckets before running a single query, so nothing gets filed as "not a competitor" by default. The four buckets are direct, indirect, substitute, and future entrant.

Map the complete spectrum first because overlooking indirect competitors or substitutes is the critical blind spot. Direct competitors offer a product similar to the deal's, aimed at the same audience. Indirect competitors address the same customer need through a different approach. Substitutes are what the customer uses today instead - the four fiercest competitors most products face include ignorance of your product and the alternates a buyer already trusts. Future entrants are incumbents with the assets to move in.

• Direct: same product, same audience. For our case, another startup shipping an AI voice agent.
• Indirect: same need, different shape. A contact-center automation tool that routes calls without full voice AI.
• Substitute: what they use now. Human call centers, IVR menus, or doing nothing.
• Future entrant: an incumbent with distribution. A contact-center platform incubating an internal voice product.

The inclusion test two analysts can share is customer-overlap AND need-overlap. If a candidate serves a similar audience and solves the same problem, it is in. This is not a published single formula - the sources converge on classification, not one shared rubric - but it is reproducible enough that two people applying it reach the same map.

The procedure, one deal end to end

Here is the full pass for the AI voice agent deck. It runs from deck extraction to a written verdict and takes roughly one to two working days for a first map. The code pass and the profile pass can swap order: run code first if the product is developer-facing.

The wide capture, and the first wrong turn

Start with a wide capture of different players, then zoom in over time. Keyword and SEO overlap tools are indispensable here because they surface who ranks for the same search terms the deal is targeting. My first attempt used only the deck's phrase, "AI voice agent," and returned the 8-founder set. Wrong turn: I varied to "real-time voice agent," "speech-to-speech," and "conversational AI," and the long-list tripled. The lesson is in the numbers above - the exact phrase under-counts by design.

The code pass

Because this product is developer-adjacent, the code pass runs early. Seed one known open-source voice-agent repo, then expand by shared GitHub topics and stargazer overlap. GitHub topics are the strongest similarity signal, and one lookalike-repo tool scores candidates on topic overlap (40 points), language (20), recency (20), and popularity (20).

The catch: about 30% of GitHub repos carry no topics at all, so a topic-only pass silently drops real competitors. Stargazer overlap and README text recover them. Two open-source approaches - a lookalike repo finder that turns a seed repo into a ranked table, and a similar-repo discovery tool that works via shared stars - are the standard way to run this. Watch the rate limit: GitHub's API allows 5,000 calls per hour, so large crawls need pacing.

Reading talent to find what has not launched

Hiring is the earliest public signal of a competing product, leading announcements by 6 to 18 months. Marketing materials reveal only what a company wants you to know, but job postings must attract qualified candidates, so they specify the role, the technologies, and the product being built.

This is where the profile and hiring passes earn their place. In the AI voice agent case, the exact-phrase profile search gave 8 US founders. The stealth signal gave 1,757. And the stealth set is not noise - it is where the unannounced efforts live.

Track people leaving top companies. When someone leaves and lists no new job, a gap, or a "self-employed" label, it often means they are building something new. Stealth founders are typically ex-big-tech or early-startup employees, not first-timers, and stealth clusters in AI, cybersecurity, biotech, and deep tech - exactly the categories where a clean web search is least trustworthy.

That 14.4x skew cuts both ways. A US-only scan can credibly and wrongly return "no one else is doing this" for a market crowded in Europe. Query at least two geographies. And note a useful quirk from the UK set: "Stealth Startup" appears as the top listed employer, meaning people self-label the company name. What looks like noise is a positive signal.

Running two markets, several category synonyms, and stealth signals by hand is slow, and it is exactly the work that gets skipped the night before a partner meeting. Asking in plain English across professional profiles collapses those passes into one.

Refolk is built for this: ask for the people behind a category the way you would describe them to a colleague, across professional profiles, GitHub, and the open web, and get a named set instead of a keyword count. Where a single exact query manufactures a false "no competition" read, describing the job-to-be-done in plain language recovers the set the founder omitted.

Reading the hiring surge without misreading it

Hiring signals can reveal build intent 60 to 90 days before formal research begins, and a jump from 5 to 40 open engineering roles in a quarter signals a product push. But that same jump can mean attrition or fresh funding. Read the role content - the stack, the product direction named in the post - not just the count. A contact-center incumbent posting for "real-time speech-to-speech" engineers is building your deal's competitor whether or not they have announced it.

Where this goes wrong

The failure modes below are the reason a generic checklist misleads. Each one makes "no competition" look true when it is not, and each has a specific check.

• Exact-keyword search returns near-zero. The deck's coined phrase yields 8 profiles; the real set is larger under synonyms. Check: re-run with three or more adjacent phrasings and the job-to-be-done, not the founder's coinage.
• Topic-blind repo search. About 30% of repos have no GitHub topics, so a topic-only pass drops real competitors. Check: expand via stargazer overlap and README text, not topics alone.
• Stealth employer masks the entry. A profile listing "Stealth Startup" or a generic name looks like noise. Check: trace the prior employer and location, which reveal the team behind the label.
• Hiring surge misread. A 5-to-40 role jump can be attrition or funding, not a competing product. Check: read the role content and product direction, not just the count.
• Founder's list taken at face value. It excludes substitutes and future entrants by construction. Check: force-fill all four spectrum buckets before deciding "none."
• Big-market absence ignored. A genuinely empty market may be a closed opportunity, not white space. Check: prove TAM and willingness to pay before crediting first-mover status.
• Regional blindness. A US-only scan misses UK and EU builders, and the index shows a 14.4x US:UK skew. Check: query at least two geographies.

A single exact-phrase search does not confirm an empty market. It manufactures one.

The verdict and how to defend it

State plainly whether "no competition" holds, and if it does, why. If the company really is first to a large market, say so, justify why they are first, and describe the moat that keeps them there. The bar for displacing an incumbent is high: a "10X better" product is the cited threshold for a new entrant to move a buyer off what they already use.

For the AI voice agent case, the verdict writes itself once the passes are done. The exact phrase returned 8 founders; the varied search and stealth signal returned a crowded field, including ex-big-tech founders now in stealth and incumbent contact-center vendors hiring for speech-to-speech products. The claim does not hold. The map, not the slide, goes to the partner.

Verdict on the "no competition" claim for [Company]: The claim does NOT hold / holds with caveats.
Named direct competitors: [list, with in-reason: customer + need overlap].
Named indirect and substitute players: [list].
Stealth or unannounced efforts: [named people or teams, with the departure gap or job post as evidence].
Incumbent future entrants: [vendor + the job post or title signaling the internal effort].
Geographies scanned: [at least two].
Basis for any residual "first mover" credit: [TAM and willingness-to-pay evidence, or state that the market may be a closed opportunity].

Keeping the map current

A competitive map is accurate the day you build it and decaying the next, because stealth efforts announce and incumbents ship. Initial capture takes one to two days; after that, keep adding and run regular scans. Set a re-check cadence tied to the deal's stage: monthly while the deal is live, quarterly once it is on the watchlist.

Re-run the two cheapest passes on that cadence. First, the hiring pass, because it leads announcements by 6 to 18 months and will surface incumbent efforts before they become public. Second, the stealth-profile pass across your two geographies, because the people in the 1,757-strong US set are actively forming companies and their labels change from "Stealth Startup" to a real name when they launch. When a profile's bio flips from "building something new" to a named product, that is a new entry for your map and a data point for the founder's original claim. The verdict you wrote should carry a date and a next-scan trigger, so the partner always knows how stale the map is.