ivanoooBook a demo

Does your marketing stack compound, or does it just produce?

The Agentic Compounding Index — a buyer-side instrument for the one question vendor demos never answer.

The question

A tool that produces gives you output. A tool that compounds gives you an asset: it is better in month twelve than in month one because of what happened in between. Almost every vendor claims the second. We measured which ones evidence it.

Why it needed measuring

Gartner estimates only about 130 of the thousands of agentic-AI vendors are real, and predicts over 40% of agentic-AI projects will be cancelled by the end of 2027. Its word for the gap is agent washing: rebranding assistants, RPA and chatbots as agents. You cannot buy your way past that with a demo.

Five vendor pages · “AI agents that learn”Example
VendorClaims agentsShows the mechanism
Vendor Ayesno
Vendor Byesno
Vendor Cyespartly
Vendor Dyesno
Vendor Eyesno
What the Compounding Index scores: whether the loop that learns can be seen working.

The claim is everywhere. The mechanism is nowhere.

Searching the category for a platform that learns from campaign outcomes and updates its own rules returns the claim in almost every vendor's copy and in no vendor's dated changelog. Marketing copy cannot establish a capability. Release notes can.

How it scores

Eight dimensions, each a five-rung ladder: Memory, Evolution, Distinctiveness, Production, Direction, Outcome, Receipts, Eval. The ladders are monotone, so rung four requires rungs one to three. A platform's grade is its weakest dimension, not its average. Systems fail at their weakest link, so that is where the grade is read.

What counts as evidence

Every rung is scored from a class: vendor marketing, vendor documentation, third-party testing, or first-party observation. Marketing copy can never establish a rung that requires observed behaviour. Evidence expires and is re-verified on a date, in public.

Where the wall sits, between storing and learning

Every platform we scored can show you outcomes. Several can attach an outcome to the asset that caused it. None reads a rule derived from those outcomes at generation time. The industry has built the scoreboard and not the loop.

Approval is not a standard

Given the chance to build a rule that can refuse, four independent vendors each built a person who approves. One ships a review step that pauses the run to fetch a human. Two tell you to fine-tune the instructions yourself when a test fails. One asks you to review before publishing. A person can be tired, rushed, or agreeable. A written standard cannot. That difference shows up on a busy Friday.

The strongest testing in the category cannot stop a bad release

The two largest platforms have real evaluation: named graders measured against reference answers, multi-turn tests, version-to-version regression comparison, test cases held as version-controlled metadata. Neither documents anything that blocks a publish when those tests fail. The verdict ships. The block is left to a pipeline you assemble.

What we are not claiming

Thirty registered platforms are still unscored, and unscored means unscored, never absent. Where our own evidence is internal and unpublishable, the cell says so and does not count toward our grade. The coverage gap is printed beside the board, not buried under it.

Is this a ranking of the best marketing tools?

No. It measures one property — whether a system improves from its own outcomes — and it is deliberately harsh about it. A platform can be excellent to use and still sit low on this instrument, because producing well and compounding are different capabilities.

How is this not rigged in Ivanooo's favour?

Three rules. Every criterion is anchored to an external source that existed before we scored anyone. Strike our name from the rubric and it still reads as a fair test. And Ivanooo is held to the same bar as the platforms it scores.

Why is the grade the weakest dimension rather than an average?

Because an average hides the break. A platform with perfect memory and no evaluation will confidently repeat a mistake forever. Systems fail at their weakest link, so the instrument reads the grade there.

What counts as proof of a rung?

A dated vendor surface, third-party testing, or observable behaviour. Marketing copy can never establish a rung that requires observed behaviour, and an aggregated summary is a lead to check, never evidence. Twice in one pass a summary would have handed a vendor a rung its own release notes did not support.

How current is the evidence?

Evidence expires. Every scored platform carries the date it was last re-verified against a dated vendor surface, and a re-verification pass records what moved and what held — including the entries that did not change.

Compounding is a mechanism, not an adjective. Ask any vendor which rung they meet, and on what dated evidence.