Why AI tool counts disagree by an order of magnitude
Published AI tool totals differ by a factor of ten or more because there is no registry and no shared definition. Four decisions explain nearly all the variance: whether features count as tools, whether wrappers count separately from models, whether dead links stay in, and whether renames create new rows.
By the The AI Tools Index team
If you look up how many AI tools exist, you will find totals separated by an order of magnitude, all published with equal confidence and none with a stated method. The spread is not caused by different research quality. It is caused by four unstated definitional choices, and once you know them you can read any published count in about a minute.
Decision one: is a feature a tool?
Every large product now ships AI features. Count each feature and the total explodes; count only standalone products and it collapses. Neither is wrong, and the difference between them is larger than any other single factor. A catalog that never states which convention it uses has published a number that cannot be compared to anything.
Decision two: does a wrapper count separately from the model?
Thousands of products are interfaces over a handful of foundation models. Counting each interface as a tool is defensible, because each is a separate product with its own pricing and its own support. Counting the model separately as well, and then counting each hosting platform that serves it, produces three rows for one thing.
The distinction is live: ByteDance's Seedance 2.0 launched on 10 February 2026 and is served through third-party platforms including Higgsfield, which is a separate company. One model, several product surfaces, and a real editorial decision about how many rows that should be.
Decision three: do dead links stay in the total?
This is the quietest inflator, because nothing visibly breaks. A catalog that never re-checks accumulates entries for products that closed, and the total grows while the useful content shrinks. Kevin Indig's Growth Memo analysis of AI citations found 64 of 365 cited URLs were broken, costing 203 citations, which is the same rot measured from the other direction: dead links do not just mislead readers, they cost the publisher.
Any total published without a re-check cadence should be read as a high-water mark rather than a current figure.
Decision four: does a rename create a new row?
It should not, and frequently it does. The project now known as OpenClaw was previously named Clawdbot and then Moltbot; the Google image model widely called Nano Banana shipped as Gemini 2.5 Flash Image in August 2025, with Nano Banana 2 following on 26 February 2026. A catalog matching on names rather than identities can carry each of those families two or three times over.
How to read any count in one minute
- Find the stated definition. No definition, no comparison.
- Find the date the total was computed, not the date the page was updated.
- Check whether the category counts sum to the headline total. If they do not, the taxonomy is multi-filed and the total is not what you think.
- Open three random entries. If one is dead, assume the rot rate applies to the whole catalog.
What this index claims, precisely
That on 28 July 2026 it held 209 live listings across 12 categories under the method linked above, and that the twelve category counts add up to that number. It claims nothing about how many AI tools exist in the world, because that figure is not knowable and anyone offering it is describing their own database.