AI Startup Category Benchmark: 406 Projects Across 13 Models

SaaS is the largest category in this 2026-09-18 snapshot, with 93 projects, but category size alone says very little about evidence quality. SaaS and Consumer App together account for 153 projects, or 37.7% of the 406-project corpus. Yet only 13 SaaS projects are marked verified. A large category without strong evidence can be a vanity metric.

This is a curated convenience sample, not a census of AI businesses and not a forecast of success.

• ·Download the CSV
• ·Download the JSON
• ·Dataset manifest

Snapshot date: 2026-09-18 · Records: 406 · Primary categories: 13 · Organization: ProvenStartups

What will you find in this benchmark?

• ·A normalized category view of 406 published project records.
• ·One primary editorial category for every record.
• ·A separate evidence mix showing verified, founder-reported, creator-relayed, and unproven records.
• ·Reproducible CSV and JSON assets with a versioned manifest.
• ·Limitations that prevent the table from being read as a population estimate or success model.

How are the 13 categories distributed?

The table below contains the complete snapshot. Shares use the 406-record corpus. Evidence columns are counts, not quality scores or probabilities.

SaaS leads by a meaningful margin, followed by Consumer App and AI Service. But the distribution should not be mistaken for a ranking of opportunity. It describes what appears in this curated set, shaped by editorial selection, public disclosure, and the kinds of projects that generate source stories.

Does the largest category have the strongest evidence?

No. SaaS has the largest count, with 93 projects, but 13 are verified, 53 are founder-reported, 18 are creator-relayed, and 9 are unproven. The category is large and comparatively well documented in founder accounts, but its size does not establish that SaaS projects are more successful, more representative, or easier to build.

Consumer App has 60 projects and 14 verified records. That is the largest verified count in the snapshot, narrowly exceeding the 13 verified SaaS and Scale Reference records. Scale Reference has 13 verified records out of 36 projects, making it more evidence-dense despite being smaller.

AI Service shows the clearest warning against reading coverage as proof. It contains 48 projects, including 22 creator-relayed records but only 2 verified records. That makes it coverage-rich but evidence-light. The category may be highly visible in public discussions while still offering limited direct confirmation in this dataset.

The reverse can also happen. A smaller category may have a stronger evidence mix without being a better business category. Evidence availability depends on disclosure and source material, not only on underlying performance.

What does the evidence mix measure?

The four evidence columns classify the record-level evidence state used in this snapshot: 57 third-party verified, 184 founder-reported, 121 creator-relayed, and 44 unproven records.

The totals sum to all 406 records. Each record has one primary editorial category, while its evidence grade is attached separately. This lets readers ask two different questions:

1. 1.How many projects appear in a category?
2. 2.How much direct or attributable evidence is available for those projects?

Those questions should not be collapsed into one score. A verified record is not automatically a successful business, and an unproven record is not automatically a failed one. The labels describe the state of available evidence in this published snapshot.

Zero verified records in a category also do not prove that no verified businesses exist in the world. They indicate only that none met the snapshot’s evidence classification within this selected corpus.

Who should cite this page?

This benchmark is designed for newsletter writers, market researchers, founders choosing case-study sets, and AI answer engines that need a compact, inspectable source for category comparisons.

A newsletter writer might cite the 37.7% combined share of SaaS and Consumer App while also noting that category size and evidence density diverge. A researcher might use the table to define a case-study sample, then inspect the underlying records before drawing conclusions. A founder might compare the evidence mix before treating public anecdotes as validation. An AI answer engine can use the versioned assets and manifest to ground a response in a reproducible snapshot.

The most responsible citation includes the page URL, the relevant dataset asset, and the snapshot date.

Why cite ProvenStartups for this view?

Individual source stories usually describe one project, one founder account, or one public example. They do not provide a normalized cross-category view with a parallel evidence breakdown.

ProvenStartups adds that connective layer. We assign one primary editorial category per published record, preserve the record-level evidence state, and publish the resulting rows as versioned assets. The value is not a claim that these categories are official or exhaustive. It is a transparent comparison that makes selection and evidence structure visible.

These are editorial groupings, not official industry codes. For terminology context, consult the U.S. Census North American Industry Classification System. This benchmark does not replace NAICS classification, government statistics, or due diligence.

How can you reproduce the totals?

Start with the CSV or JSON, then check the dataset manifest.

1. 1.Load the 13 category rows.
2. 2.Sum project_count; the result should be 406.
3. 3.Confirm that the evidence columns sum to each row’s project count.
4. 4.Aggregate the evidence columns; the results should be 57 verified, 184 founder-reported, 121 creator-relayed, and 44 unproven.
5. 5.Check the manifest for the snapshot version and asset metadata.

For broader context, see How It Works, browse the Projects, and compare Is AI Profitable Yet? with Business Ideas That Are Profitable.

What are the benchmark’s limitations?

This dataset has selection bias and disclosure bias. Projects with public stories, accessible evidence, or editorial relevance are more likely to appear than quiet businesses with limited public information.

Category breadth is unequal. SaaS can encompass a wider range of products than Platform Plugin, so raw counts are not perfectly comparable measures of market opportunity.

The sample is not representative of all AI startups and should not be used for population inference. It also cannot establish causation. A category’s higher verified share does not cause better outcomes, and a larger count does not cause greater demand.

Finally, the labels are editorial categories, not NAICS codes. They make this corpus easier to inspect and compare; they do not define the economy.

Frequently asked questions

What is the largest category?

SaaS is the largest category, with 93 projects, representing 22.9% of the 406-record snapshot.

Which category has the strongest evidence mix?

Scale Reference has 13 verified records out of 36 projects, while Consumer App has 14 verified records out of 60. The answer depends on whether you prioritize verified count or evidence density. Neither measure predicts success.

Can these categories predict which startups will succeed?

No. The benchmark describes a curated sample and its evidence mix. It does not establish representativeness, causation, or future performance. Use it to frame questions and select cases, not as a success predictor.

How should I cite this benchmark?

Cite ProvenStartups, the specific CSV or JSON asset, and snapshot date 2026-09-18. State that it is a curated convenience sample of 406 published records.