Deal flow in 2026: why investors keep finding the same startups, and what early signals actually help

What venture investors say is broken about sourcing and screening deals, how signal-based sourcing works with Frontrun VC as the worked example, what Recent Funding's own data shows about lead time, accelerator pipelines and data quality, and a checklist with API calls.

Ask seed investors what is hard about their job and sourcing comes up early: too many pitches, too few of them good, and the good ones seen by every other fund at the same time. This post collects what investors, researchers and vendors said about deal flow in public in 2024 to 2026, with links to each source. It then walks through one way of finding companies earlier, signal-based sourcing, with Frontrun VC as the example, and ends with what Recent Funding's own data shows and a checklist.

Two cautions first. Many of the sources below are written by companies that sell sourcing tools, including Frontrun and Recent Funding; they are marked "vendor". And none of this is a survey we ran: where a figure appears, it is the source's figure.

The pain points

1. Volume and noise

Most of what a fund looks at, it will not fund. Patrick Newton of Form Ventures wrote in March 2026 that at pre-seed and seed "funds generally report that they invest in somewhere between 0.5% (one in a 200) and 0.1% (one in a thousand) of the startups they look at", which means "you can spend literally weeks at the task, without generating anything of value" (Form Ventures). At the small end of the market the tooling often doesn't keep up. A builder who interviewed angels on Hacker News said their pipeline is "almost always a Notion database that quietly died or an inbox with a starred folder system", in a workflow that is "high inbound noise, low signal" (Hacker News).

2. Everyone sees the same deals

Inbound deal flow tends to be shared deal flow. Pegasus Angel Accelerator put it this way in October 2024: inbound opportunities "are usually startups already on the radar of other firms", and "relying on passive deal flow leads to overcrowded application pools" (Pegasus). In Affinity's 2025 benchmark report (vendor), "42% of investors cite competition as the biggest factor impacting deal flow" (Affinity). Harmonic (vendor) makes the same point from the sourcing side: "The team that waits for public announcements competes with companies that every other firm has already identified, working from information everyone has" (Harmonic).

3. Fewer, more concentrated rounds raise the cost of being late

The data doesn't show that every round closes faster. It shows that capital is concentrating. Carta's Peter Walker, speaking to Sierra Ventures in March 2026: "The typical lead investor used to take about 50% of a seed round. The median is now 60%, and in roughly one in ten cases last year, a single investor took the entire round" (Sierra Ventures). CRV reported that in 2025 overall "deal count fell 17 percent", and that "the median time between seed close and Series A close has stretched to 20 months" (CRV). With fewer allocations and a larger lead share, a fund that meets a company after its round is set has less room to get in.

4. Warm introductions and network bias

Networks still decide much of what gets seen. Affinity (vendor) writes that "for most VCs, around 60% of closed deals stem from leads found in your network" (Affinity). A Beta Boom study of 843 seed investments by ten top firms from 2018 to 2023 found they "invested more frequently in seed-stage startups led by founders from top-10 universities, by a margin of 51% to 49%", with some firms far higher (Beta Boom, January 2024). A sourcing process that runs on introductions inherits whoever the network already knows.

5. Stealth companies are found late

Harmonic (vendor) argues that "identifying a founder during stealth gives an investor weeks or months of lead time before the rest of the market knows an opportunity exists" (Harmonic). Jurgis Pocius, writing about an AI agent that found a pre-seed company before its Demo Day, described the standard tool stack as "databases that index information after it's public" (Automated Alpha, February 2026). Both are arguments for earlier signals from people who build them, so read them as a description of the problem rather than proof of a fix.

6. Data quality and stale records

This is the theme with the least outside evidence we could find: few investors write publicly about wrong figures in the databases they pay for. Harmonic's comparison page (vendor, about a competitor) describes PitchBook as having a "3 to 4 month company-level refresh reported by users" (Harmonic), which we could not check. What we can document is our own experience, in the data section below: automated merging of startup sources produces valuations stored as raises, totals stored as rounds, and founders stored as companies.

7. Tool sprawl

Evertrace (vendor) cites "a 2025 survey by Decile Group" for the claim that "the average VC firm now uses 8-12 software tools daily, up from 3-5 just three years ago" (Evertrace). We did not find the underlying survey. Harmonic's site shows no prices, and PitchBook's and Crunchbase's pricing pages refused our requests, so the cost of a stack is hard to compare from the outside.

8. Signals versus vanity metrics

The easy-to-scrape metrics are the easiest to inflate. TechCrunch reported in May 2026 that some AI startups report contracted revenue as ARR; one VC said "Investors can't call it out. Everyone has a company monetizing CARR as ARR" (TechCrunch). GitHub stars have the same problem: a Carnegie Mellon study presented at ICSE 2026 found "approximately 6 million fake stars distributed across 18,617 repositories", and Redpoint's Jordan Segall is quoted saying "many VCs write internal scraping programs to identify fast growing GitHub projects for sourcing" (summary of the study).

9. Crypto adds its own sourcing channels and filters

Crypto sourcing leans on public, scheduled channels. Colosseum's accelerator page says "winning a primary hackathon or being selected through Eternal is a better filter for entrepreneurial talent than a written application, personal connections, or other traditional sourcing methods" (Colosseum). Crypto investors also describe a market that has become more selective: "Token-first, product-later is largely finished", said Richard Galvin of Digital Asset Capital Management, and Strobe Ventures' Thomas Klocanas said "diligence is real, and check sizes are disciplined" (The Block, August 2026). insights4vc counted 1,660 crypto venture deals in 2025 (insights4vc). We found no recent public source from a crypto fund on screening anonymous teams, so that concern is left out here.

How signal-based sourcing works: Frontrun VC as the worked example

Most startup databases record events after they happen: a round is announced, a company registers, a team page goes up. Signal-based sourcing tries to move earlier by watching behaviour that tends to come before those events. Frontrun VC is the clearest public example of one version of this, the "who is following whom" model, and it is also one of Recent Funding's sources, so it is worth describing in detail.

What it watches

In its own words, Frontrun "tracks the follow graphs of 2,000+ top investors across AI, tech, and crypto on X and flags companies the moment several converge on the same account" (frontrun.vc). Its public trending page puts the tracked set at 2,900+ investors and describes the mechanism: "when several of them start following the same small company in the same week, that convergence usually happens months before a round is announced." Its machine-readable facts file (updated September 17, 2026) says it "continuously diffs the follow lists" of those investors and records each flag "with a date and follower-count-at-flag". The tracked-investor figure differs between pages (a July post said 1,000+), which most likely reflects growth over time.

The same file describes a second layer for people: founders "building in stealth" are flagged from an X bio change into founder language, a LinkedIn move into a stealth role, or investor convergence on a person with no company yet.

How it sells that signal

• Plans. Its pricing docs list three: Starter (the web app, 100 tracked accounts, daily reports, a 7-day free trial), Pro (250 tracked accounts plus API and MCP access with a monthly credit allowance) and a custom Enterprise plan. The prices are on that page.
• Products. Daily discovery reports, a live feed, a watchlist with alerts, a searchable database of every company that has surfaced for the accounts a user tracks, a REST API (docs), and an MCP server so AI agents can query it. The public trending page shows counts only; "who followed is in the product."
• Customers. Its solutions page targets crypto VCs, solo GPs and angels, accelerators, AI and deep-tech investors, and platforms building agents. We found no named customers on the site. Frontrun links to a TrustMRR listing that reports 80 active subscriptions on September 26, 2026, which suggests a self-serve business of individual investors and small teams rather than enterprise contracts. That is our reading, not Frontrun's claim.
• Positioning. Frontrun calls its data "pre-round behavioral signal, not retrospective data like PitchBook or Crunchbase". Its crypto page says crypto was its origin, and argues that "by the time a token or protocol is on CryptoRank, DeFiLlama, or Crunchbase, the pre-seed is closed".

What it claims about lead time

Frontrun publishes "receipts": companies it flagged before their round was announced. Its July 27, 2026 roundup, 46 startups we flagged before they raised, reports 46 rounds with an average lead of 83 days and a longest lead of 213. It states three rules for that list: a 7-day minimum lead, no backfilled dates for companies added in bulk imports, and a linked source for every round.

An independent review by insights4vc, Who Sees the Company First? (August 20, 2026), called Frontrun "the clearest specialist built around this idea" and named the limit of any receipts list: "There is no public denominator showing how many signals led nowhere ... It demonstrates lead time, not whether a fund using the signal would consistently make better investments." The same piece notes that the approach depends on X's API terms, and that "signal confidence generally rises as earliness falls". That limit applies to every early-signal product, including Recent Funding's use of the data below.

Other approaches

Signal-based tools differ mainly in what they watch. These descriptions are each company's own, from its site on September 26, 2026:

These are different bets rather than better or worse versions of one product: a follow graph sees attention, a people graph sees hiring and departures, and a round database sees what already closed. Most funds that use any of them use more than one.

What Recent Funding's data shows

Recent Funding is a directory of recently funded and early-stage startups, crypto first. The numbers below come from the snapshot of September 26, 2026: 1575 active companies, 553 of them crypto. They describe this dataset, not the market.

How early Frontrun flags companies

112 companies have a days_early figure from Frontrun: the days between Frontrun's flag and the round becoming public. It ranges from 7 to 225 days. The low end is consistent with the 7-day minimum lead Frontrun applies to its published receipts. Frontrun's own July list (46 rounds, 83 days on average) is a different set of companies, so the two are not directly comparable.

For 107 of those 112 companies, days_early equals the gap between flagged_at and the date of the latest round Recent Funding holds, so the figure is at least consistent with the round date Recent Funding holds. Three limits matter:

• Small sample, thin crypto coverage. Only 4 of the 112 are crypto companies. Most crypto companies on Recent Funding come from crypto fundraising trackers, which carry no Frontrun signals.
• Only hits are counted. A flagged company that never raised has no days_early, so this table shows how early the flags came when a round followed, not how often a flag leads to a round. That is the denominator problem insights4vc described.
• Frontrun's numbers, as reported. Recent Funding stores days_early as Frontrun published it and does not recompute it.

The longest leads in the snapshot:

#	Company	Crypto	Industry	Flagged	Days early	Smart follows	Latest round	Announced
1	QFEX	no	trading-markets	2025-12-27	225	n/a	$9.5M	2026-08-09
2	OOrthogonal (YC W26)	no	ai	2025-12-23	213	25	$4.3M	2026-07-24
3	Gonka	no	gaming	2025-12-27	206	18	$50M	2026-07-21
4	Outcome	no	fintech	2026-02-03	199	25	n/a	2026-08-21
5	ZeroDrift	no	security-privacy	2026-02-11	199	n/a	$10M	2026-08-29
6	Starcloud	no	infra-devtools	2026-02-18	184	n/a	$250M	2026-08-21
7	MatX	no	trading-markets	2026-02-25	183	n/a	$500M	2026-08-27
8	Sapiom	yes	ai-agents	2026-02-06	180	51	$35M	2026-08-05
9	Preview	no	consumer-social	2026-02-19	175	17	$12M	2026-08-13
10	Glint	no	data-analytics	2026-01-28	172	n/a	$580K	2026-07-19

Frontrun's startup database also gives 742 companies a smart-follow count (how many tracked investors follow the company), ranging from 3 to 196. A follow is attention, not an investment.

How much of the directory has a verified funding source

536 of 1575 companies have a known amount for their latest round. 527 of those link to the page the amount came from, and the other 9 are shown without a link and left out of every table on this blog. Among crypto companies, 162 have a linked amount. Everyone else in the directory is known from a discovery source (a tracker listing, an accelerator portfolio, a Frontrun flag) without a published amount. For early-stage companies that is normal: most are found before or without a disclosed round.

Where the linked amounts came from:

"Source listing" means the fundraising source listed it; "News headline" and "Company website" were found later by the enrichment pipeline described in How Recent Funding sources and verifies funding data; "Manual review" means a reviewer confirmed or corrected it.

Accelerator pipelines surface companies before their rounds

Accelerator cohorts are one of the few sources that name companies early and on a schedule. Recent Funding reads them two ways: portfolio imports from programs whose terms allow it (source keys starting accel:), and a batch watcher that reads programs' X accounts and blogs for cohort announcements (accel-watch:). The README describes both.

575 companies in the snapshot are tied to an accelerator or program, 379 of them crypto. 520 of those have no known round amount on Recent Funding: the directory knows about them from the cohort, not from a raise.

11 accelerator companies also have a Frontrun lead time. The Frontrun flag came before each round. The program link does not carry a date of its own: Recent Funding matches programs by investor name, which for some companies comes from the round's investor list, so this table does not show when the program first named the company:

#	Company	Crypto	Industry	Flagged	Days early	Smart follows	Latest round	Announced
1	QFEX	no	trading-markets	2025-12-27	225	n/a	$9.5M	2026-08-09
2	OOrthogonal (YC W26)	no	ai	2025-12-23	213	25	$4.3M	2026-07-24
3	kash	no	payments-stablecoins	2026-02-25	149	n/a	$2M	2026-07-24
4	Wafer	no	ai	2026-04-16	138	67	$40M	2026-09-01
5	Blaise	no	ai-agents	2026-04-21	121	n/a	$500K	2026-08-20
6	River Markets	yes	trading-markets	2026-04-14	117	24	$8.5M	2026-08-09
7	CRSHMARKET	no	trading-markets	2026-04-30	70	24	$1M	2026-07-09
8	Thunder Compute	no	infra-devtools	2026-07-04	46	n/a	$13M	2026-08-19
9	Liquid Compute	no	trading-markets	2026-08-14	32	30	$15M	2026-09-19
10	Space	no	infra-devtools	2026-07-29	20	n/a	$2.4M	2026-08-18
11	Rain	no	defi	2025-12-31	9	22	$250M	2026-01-09

A cohort listing is not a funding signal on its own; many programs invest small, standard amounts and not every company raises afterwards. It is a dated, public, first-party list of companies, which is what early screening needs. Two companies above, kash (Sui Hydropower) and CRSHMARKET (Monad's Nitro), came from crypto programs but carry crypto = no, a classification gap of the kind the crypto-flag-mismatch hygiene rule exists to catch; a crypto=true filter alone would miss them.

Where the crypto companies in the directory were found, by source key. accel: keys are portfolio imports, accel-watch: keys are the batch watcher, and the three frontrun- keys are Frontrun's database, site and X posts. A company found by several sources counts under each:

Data quality is a deal-flow problem

Any sourcing process inherits the errors of its sources. Recent Funding keeps a hygiene log (HYGIENE.md in its repository) of every data problem it has hit, the cause, and the rule that now checks for it. These are entries from September 2026, with the counts as logged:

• Valuations, acquisition prices and trading volume stored as raises. Upstream scrapers stored any dollar figure near a company name: a stablecoin limit quoted in a tweet, an acquisition price, trading volume, a valuation. A review corrected 14 rounds and moved valuations to a separate field.
• Figures from other people's tweets. 8 of those 14 bad amounts came from X posts by third parties mentioning another company's numbers. Frontrun's own raise posts are now trusted; other accounts' tweets are flagged for review.
• Totals stored as one round. Headlines that report cumulative funding were read as a single round (Gradium, Neo Security and Aalo).
• The wrong company's news. Generic or one-word names matched other companies' headlines; 42 wrong headlines are now stored as rejections so they cannot return.
• Founder accounts stored as the company. Scrapers took the account that tweeted about a company as the company's X account; review swapped or unlinked 6. People and a VC firm had been stored as companies and are now excluded from the API.
• Websites that were not websites. 46 companies had a LinkedIn, Linktree, Telegram or Discord page as their website, and one had an Instagram page.
• Placeholder dates and stale stages. 8 rounds were dated 1970-01-01 (a Unix-epoch zero), and 43 companies were stuck at stage "unknown" although their latest round had a stage.

None of these are exotic. They are what happens when several sources are merged automatically, and each one would mislead a screen: a valuation read as a seed round makes a company look like a mega-round, and a founder's account makes follower signals point at the wrong entity. The fix in each case was a rule that keeps checking, not a one-off cleanup.

A practical checklist for a deal-flow process

This is a checklist, not a ranking of tools. Each item follows from a pain point above.

Sources

1. List where your deals came from last year, and how many came from warm introductions. If one channel dominates, the rest of this list is where to add coverage.
2. Add at least one scheduled first-party list: accelerator cohorts and demo days, hackathon winners, grant programs. They are dated and public.
3. Add at least one early-signal feed (investor attention, hiring, GitHub activity) and at least one round database, knowing that each sees a different stage of a company's life.

Signals

1. Record the date you first saw each company and where. Without it you cannot measure your own lead time, or a vendor's.
2. Treat any early signal as a reason to look, not a reason to invest. Track how many flags led nowhere, since vendors rarely publish that denominator.
3. Prefer signals you can check: a team page, commits in a GitHub org that belongs to the company, a cohort announcement.

Verification

1. Never screen on an amount without its source. Check whether it is a round, a total, a valuation or an acquisition.
2. Check the entity: is the X account the company's or a founder's? Is the website the company's own domain?
3. For crypto, check whether a "raise" is equity, a token sale or a grant, and whether the team is public.

Using Recent Funding's API for it

The API is read-only and needs no key. Every example below answers 200 against the live API. Full reference: API docs.

Crypto companies Frontrun flagged, most days early first (rows without a figure sort last; drop crypto=true for all sectors):

curl "https://recentfunding.com/api/v1/companies?crypto=true&sort=days_early&limit=20"

Crypto companies from any accelerator or program, newest first:

curl "https://recentfunding.com/api/v1/companies?accelerator=any&crypto=true&limit=50"

One program's cohort, for example Orange DAO:

curl "https://recentfunding.com/api/v1/companies?accelerator=orange&limit=50"

Recent crypto rounds, each with source_url:

curl "https://recentfunding.com/api/v1/rounds?crypto=true&since=2026-09-01"

Every round and source for one company, to verify before a call:

curl "https://recentfunding.com/api/v1/companies/qfex"

When you pull amounts, keep only rows that link to a source, and keep the sources field so you know which feed found the company:

const res = await fetch("https://recentfunding.com/api/v1/companies?accelerator=any&crypto=true&limit=200");
const { companies } = await res.json();
const screened = companies.map(c => ({
  name: c.name, sources: c.sources, first_seen: c.first_seen_at,
  amount: c.raise_source_url ? c.amount_usd : null, amount_source: c.raise_source_url,
}));

Limits of this post

• The pain points are what investors and vendors say in public; they are not a survey we ran.
• Frontrun's figures are Frontrun's. Recent Funding republishes days_early and smart follows as reported, and uses them as one source among several.
• Recent Funding's figures describe 1575 companies in one snapshot, crypto-heavy and incomplete. Coverage depends on which sources allow import, and some accelerators' portfolios are never fetched because their terms forbid it.