What our columns can and cannot tell you

We publish about a dozen columns for every tool in the catalog, and they are not equally useful. Some separate products sharply; others barely move. This report grades our own measures on how much they actually discriminate, and says which ones we would not build a decision on. It is the least flattering thing we publish and the most useful for anyone citing the rest.

The rating column barely works

Of 5,916 rated live products, 4,232 sit in a single half-point band between 4.0 and 4.4. Exactly 0 score below 3.5.

This is the well-known compression in every public review system, and we inherit it because our ratings are aggregated from places where people mostly rate things they chose to keep using. A product nobody liked is not rated badly; it is not rated at all. The distribution is a filter, not a measurement.

The best-rated market averages 4.4 and the worst 4.2. Free, freemium and paid products are separated by four hundredths of a point. Cut the data almost any way and the answer is 4.3.

There is one cut where it moves. Products scoring 80 or above on opportunity average 4.5, against 4.33 for the 50–64 band. That is a real difference, and it is the only defensible use of the rating column: as weak corroboration at the very top, never as a way to rank two products against each other.

Wrapper risk works, but not as a scale

The opposite problem. 3,962 products score under 20 and 1,182 score 60 or above, with 53 in between. Ninety-eight percent of the population sits at one end or the other.

We think that reflects reality — there is no gradual way to half-own a model — but it has a practical consequence for anyone quoting the number. The difference between 72 and 81 is noise. The difference between 15 and 70 is the whole finding. Treat it as two buckets with a confidence attached, which is how our wrapper-risk report uses it.

Saturation is a property of the market, not the product

Saturation is inherited from a tool’s category by construction. Two products in the same market have the same saturation score, always. That makes it a good instrument for comparing markets and a meaningless one for comparing products, and it is the single most common way we see our own data misread.

Opportunity is the product-level counterpart and does vary within a market, which is why every ranking we publish inside a category uses opportunity rather than the heat score that blends the two.

Coverage is the real limit

More of our numbers are constrained by missing data than by weak measurement. The table below is every major column with its actual coverage.

Headquarters is the worst at 58%. Growth signal and wrapper risk are both missing for the same 328 recently added products, which is why those two columns agree so precisely on their gaps.

Every report we publish computes rates against the populated subset rather than the full catalog, and says so. That is the right choice, but it carries an assumption worth stating: that the products missing a column are not systematically different from the ones that have it. For headquarters that assumption is probably fine. For growth signal it is definitely not — the missing products are overwhelmingly recent additions, which is why our cohorts report refuses to draw conclusions about the 2025 cohort.

And the catalog itself is a sample

The largest limit is the one that sits underneath all of the above. A catalog of 7,030 products is not the AI market; it is the part of the AI market that became visible enough to be listed. Products that launched and failed without ever being written about are absent, which means every failure rate we publish is a floor.

Our coverage is also not uniform across geographies. Singapore was swept exhaustively in 2026 and now appears as the second-largest hub in the catalog. That is a fact about our effort, not about the world, and we would rather repeat it in every report than let it be quoted as a finding.

How to use our numbers

Market-level figures — saturation, wrapper density, mortality, free-tier rate, API rate — are the strongest thing we publish. They rest on large populations and the differences between markets are far larger than the measurement error.

Product-level figures are weaker. Opportunity is meaningful and worth reading. Wrapper risk is meaningful as a bucket. Rating is close to useless, and saturation on a product page is just its category’s number. Anyone building on our data should weight accordingly, and we would rather say that than have someone discover it later.

Photo: Ono Kosuki / Pexels. Colour-graded for Falcoscan.

← All reports