How SaaSafras data is made

01What this is

One curated cohort, not a market. The database holds 140 companies (150 rows before deduplication) chosen for being interesting bootstrapped or mostly bootstrapped software businesses, assembled by one researcher. Funding is a per-row field (bootstrapped, partly-vc, vc, undisclosed): notion and calendly are VC-funded, carrd-co partly, and they stay in as the contrast cases. Every statistic on the site describes this list. None is a claim about SaaS in general.

Each company record has two layers. The prose profile (growth playbook, distribution moat, builder's takeaway, growth timeline) is written analysis. The structured fields (channels, category, growth stage, revenue, launch year, audience) are what cross-company analysis runs on. The prose came first. The structured fields were derived from it in August 2026 under written rules, and this page names those rules.

02Where the information comes from

The v1 dataset was collected in February 2026 from public material: founder interviews and podcasts, company blogs and about pages, pricing pages, acquisition announcements, open-metrics pages, and founder posts on X, Indie Hackers, and Hacker News. Candidates were found by hand and through Product Hunt and Indie Hackers listings. Extraction and first-pass classification used language models with human review; the analysis prose was written by the researcher.

Every record carries a Sources field. Its current state, measured on 2026-08-27:

That is the honest starting point. The refresh described in section 11 replaces prose sources with dated links, field by field, starting with revenue.

Growth timeline steps already carry a per-step tag: documented when a source states the event, inferred when the researcher deduced it. Across all 150 rows: 585 documented steps, 8 inferred, 1 documented with a stated caveat.

03How companies are selected

By interest, not by sampling. A company entered v1 because it was a bootstrapped or mostly bootstrapped software business with enough public material to write a growth playbook from. Launch years run 1996 to 2025. There is no random sample, no size threshold, and no attempt to represent a category in proportion to its size.

Three consequences follow, and they are stated again in section 12: no dead companies, no denominator, and the researcher's taste is in the selection.

Deduplication

Separate collection passes picked up 8 companies more than once, as 17 rows under different slugs: Ghost, Carrd, Transistor.fm, Lemon Squeezy, Plausible, Nomad List, Remote OK, and Fathom. The rule: group rows by company name, keep the row carrying a numeric MRR; if more than one does, keep the row with more channels, then the higher MRR. The mapping lives in a versioned file and every analysis reads it. The source CSV is never edited to remove a duplicate, so provenance survives.

One company, one id, kept across renames and pivots. Old slugs resolve as aliases.

04Provenance labels

Four labels, applied per field, not per company. A company can have a founder-reported price and an estimated MRR in the same record.

Status 2026-08-27: classification fields (channels, category, growth stage, audience) are labeled by construction. Revenue, pricing, and status fields carry a label only after the refresh reaches them; until then they are unlabeled v1 values and should be read as Estimated.

05How revenue figures are handled

The revenue column is Normalized MRR (USD): a monthly figure in US dollars. 111 of the 140 companies carry a numeric value; 29 have none and are excluded from every revenue table rather than filled in. All figures are USD; the cohort contains no non-USD reporting.

v1 mixed metric types in that one column. The same cell can hold current MRR, an annual figure divided by twelve, lifetime revenue divided by months in business, or an acquisition price. Example: Laravel Forge shows $333,333 per month, which is $40M lifetime over 120 months, not an observed MRR. This is the single largest quality problem in the dataset and the first thing the refresh fixes.

Rules going forward

• Every revenue figure gets a revenue_metric_type: current MRR, stated monthly revenue (a per-month figure not called MRR), ARR, stated annual revenue (not called ARR), launch-window revenue (a total inside 30 days of launch), lifetime revenue, acquisition price, or run-rate estimate. Figures of different types are never compared or averaged. Funding (bootstrapped, partly VC, VC, undisclosed) is its own field, kept out of the status column.
• Every figure gets a date (as_of) and a provenance label. A figure with no date is a v1 figure and reads as "at some point before February 2026".
• USD is assumed where the source writes "$" and names no currency. Where the source names one, the record says so; where the figure is spoken in a non-US interview with no currency stated, the record carries a currency-unstated note.
• Revenue history is append-only. A new figure adds an observation; it never overwrites the old one. Old figures stay visible with their dates.
• Until a company's figures have been through the refresh, treat every dollar amount on the site as an order of magnitude.

06What verified means

Status 2026-08-27: the test has been applied to 30 companies (the top-30 refresh, 229 dated observations, every cited revenue source opened by a human). 15 of the 30 have at least one revenue field that passes: Profitwell, Doist, Ghost, Nomad List, Formula Bot, Carrd, TinyPilot, Statuspage, Baremetrics, Plausible, Tally, Photo AI, Balsamiq, TypingMind, Bannerbear. The other 15 have no source that states a current figure, or the only sources repeat each other with no founder or filing behind them; their v1 figures are withdrawn, not replaced. No company outside the 30 has been tested, so the v1 figure on any other row is unverified by this definition.

Verified is not "true". It is "we can show you where it came from and when we looked". A founder can be wrong on a podcast; the label tells you the chain, not the truth.

07What an estimate is

An estimate is a value SaaSafras produced when no source states it outright. Every estimate names its method in the record. The methods in use:

• Pricing-page math: stated customer count multiplied by the entry price, or a stated ARR divided by twelve.
• Range midpoint: a founder says "$80K to $110K"; the record stores 95,000 with the range noted, never the midpoint alone.
• Comparable: a company of similar size and model with a reported figure. Weakest method; used only when the alternative is a blank and the record says so.
• Historical carry-forward: a figure reported in an earlier year with no update since. The date on the figure is the reported date, not today.

Estimates are never presented with more precision than the method supports. A comparable-based MRR reads as a band, not a number.

08How channels are classified

Acquisition channels are a 12-value, multi-valued field: seo, paid, outbound, community, social, product-led, partnerships, marketplace, audience-led, content, word-of-mouth, launch-platform. Multi-valued because the data forced it: the v1 prose named 593 channel mentions across 150 companies, 3.45 per company, and a single value would have discarded three quarters of the column.

Each value has a written definition and eight disambiguation rules (community is a place with other people's threads; social is broadcasting to your own followers; content is the artifact pulling, SEO is the ranking pulling; and so on). The 307 distinct phrases in v1's channel prose were mapped against that taxonomy, cross-checked by an independent keyword mapper, and 96.5 percent of mentions resolved. Unmappable phrases were recorded as unmapped, not guessed: one company (a customer count where a channel should be) has zero channels on purpose.

The taxonomy grows only when measured data shows a value with no honest home, not when a category sounds plausible. Three of the twelve (content, word-of-mouth, launch-platform) were added that way, each a top-5 phrase by frequency with no place in the original nine.

Channels are recorded after success, from what founders said about it. That is a description of what winners report, with the hindsight that carries. See section 12.

09Categories and growth stage

Category

Nine buckets plus Other: Marketing & Growth Tools, Analytics & Monitoring, AI & Creative Content Tools, Developer Tools & Infrastructure, E-commerce & Shopify Tools, No-Code Builders & Content Publishing, Productivity & Collaboration, Vertical/Niche Business Software, Consumer Apps & Games. The 128 distinct v1 category strings map to these through a published table. AI & Creative holds generative products only; AI-powered support, productivity, and infrastructure tools sit in the bucket their buyer would look in. Other holds one company (hardware). Every assignment is a SaaSafras classification.

Growth stage

Four values: build, launch, operate, scale. v1 described each company's stage in a paragraph; the leading word of that paragraph is the stage. Exited and acquired companies fold into scale, and the exit itself is kept in a separate exit_note (acquired, sold, at exit) so the signal is not lost. Current distribution across 140: build 8, launch 15, operate 50, scale 67.

Audience

b2b, b2c, mixed, or unclear, by a keyword rule over the audience prose. Unclear is a real value; it is not resolved by guessing.

The AIBuilder flag

25 v1 rows carry an AIBuilder flag whose original definition was not recorded. 23 of the 25 launched 2021 to 2025, and the flag behaves like a recency marker. It is not the AI-native cohort. That cohort is being built separately with its own criteria, and no before-and-after comparison is published until it exists.

10Conflicting sources

When two sources disagree, both values are kept, each with its date, metric type, and label. SaaSafras does not average them and does not pick one silently. The record shows the conflict; the analysis uses the figure whose metric type matches the question being asked.

Four genuine conflicts were known in v1; each was resolved on 2026-09-01 as shown below:

A formal precedence rule (which label wins when types match) is deliberately not published yet, because until every figure has a metric type there is nothing to arbitrate. It ships with the refresh.

11When records were last updated

A record with no per-field date has not been refreshed since v1 collection. Once a company has been refreshed, its profile shows the date beside each figure, and the old figure stays in its history.

12What this data cannot say

• Nothing about the market. 140 companies chosen for being interesting successes. "Paid ads at 1 percent" describes the list, not SaaS.
• Nothing causal. Channels were recorded after the outcome, from what founders said. Associations here are descriptions of what winners report.
• Nothing about failure. No dead companies. There is no denominator.
• Nothing precise about revenue. Metric types are mixed and unlabeled in v1; 29 companies have no figure. Order of magnitude only, until the refresh.
• Nothing about the AI-native era yet. The flag in v1 is a recency marker with an unrecorded definition. The comparison this database exists to make waits for a cohort collected under stated criteria.

Every published table reproduces from scripts in the dataset repository against the frozen snapshot. Changing a classification means editing the published mapping and re-running, never editing a row by hand.

13Corrections

If a figure about your company is wrong, say so. A founder correction becomes a Founder reported observation with today's date; the old value stays in the record's history with its own date, so readers can see both. Corrections that come with a link get the Publicly reported label instead.

Send corrections to [email protected].