The FTC has been settling cases against AI marketing claims that nobody checked before publishing them — $18,000,000 against one AI phone-agent business in March 2026, a permanent ban on a chatbot marketed as replacing a lawyer, several more in between. None of it needed a new AI law. It is ordinary advertising law: a claim needs evidence behind it before you make it.
This is the register that holds that evidence. One row per claim, with the testing, the date, the sample, the method and where the file lives — and a status that never quietly promotes “nobody has checked” into “fine”.
Somebody wrote “99% accurate” on slide six eight months ago. It came from a benchmark run against a model version you no longer ship. The person who ran it has left. Nobody can now say what was measured, on what sample, or when — and that sentence is still in the enterprise deck.
That is the exact shape of the published cases. A chatbot was marketed as performing comparably to a licensed attorney and the company had simply never tested whether it did. An AI phone system was sold on replacing human sales agents and on the income buyers would make. In both, the failure was not weak evidence. It was the absence of any.
A spreadsheet does not fix this, because a spreadsheet lets a blank cell mean anything. This tool refuses to let “we haven’t looked” sit in the same column as “we checked and it holds” — they are different states, banded differently, and neither one is a pass.
One row per promise, pasted exactly as published. A landing page making three different claims is three rows, because each one stands or falls separately.
Who tested it, when, on what sample, by what method, and where the file sits. Miss any one of those and the claim bands as weak — because that is the part you would be asked for.
A dated substantiation pack with a detail sheet per claim. It is the answer to “show us what you had, and when you had it”.
Not because they are illegal, but because they are the two that produced the largest published settlements. The app marks them on the form, in the register and on the printed sheet — and it holds them to a higher record standard: a named reviewer, and evidence that is not simply a vendor’s own figures passed on.
An AI chatbot marketed as performing comparably to, or replacing, a licensed attorney. The company had not tested whether it did. Permanent ban on the claim without supporting evidence.
An AI phone-agent business marketed on replacing human sales agents and on the substantial income buyers would earn, plus a refund guarantee that did not operate as advertised.
The app carries five of these as a read-only reference tab, each tagged with the claim type it turned on, so you can see which of your own rows sit in the same shape. Summarised from public records, not embellished, and not legal advice.
The example register that ships with the app, exactly as the tool scores it. Worst first, so whatever needs a decision is always at the top.
| Claim | Type | Status | Evidence dated |
|---|---|---|---|
| Replaces your entire first-line support team — no human needed. Reply Engine (AI inbox assistant) · R. Okonjo | Replaces a human professional Elevated |
Unsubstantiated | None |
| Produces a finished blog post that needs no editing. Copy Studio · R. Okonjo | Capability | Evidence does not cover the claim | 13 Mar 2026 |
| Writes publish-ready campaign copy in under a minute. Copy Studio · D. Marsh | Capability | Not yet reviewed | None |
| Agency owners using the platform earn an extra $12,000 a month within 90 days. Reply Engine growth programme · D. Marsh | Earnings or income potential Elevated |
Weak evidence | 13 Jun 2026 |
| 99% accurate on customer intent classification. Reply Engine (AI inbox assistant) · K. Aluko | Accuracy or quality | Weak evidence | 13 Nov 2025 |
The first five of eight rows, in the order the tool puts them. The three below the fold band as substantiated, substantiated-needs-attention and withdrawn. Note the blog-post claim: it has a real study behind it — eleven of thirty drafts needed no edit — which supports a narrower sentence than the one published. That is a gap, not a pass, and the tool says so.
Being straight about the edges, because a $19 tool that overpromises would be an odd thing to sell alongside this subject.
There does not need to be. Every action referenced in the app was brought under ordinary consumer-protection authority against unfair or deceptive practices — the same law that has always required advertisers to have a reasonable basis for a claim before making it. For objective performance claims that basis is competent and reliable evidence. AI framing does not change the analysis, which is precisely why these cases did not wait for an AI statute.
Business-to-business marketing is not outside the frame. In May 2026 three marketing companies were reported as facing action over how they described an AI-powered advertising tool’s capabilities to business customers. A plain capability claim was enough.
The tool records the kinds people actually rely on: internal testing, a user study or customer trial, a published benchmark, an expert assessment, or a vendor’s figures passed through. It does not rank them — except for the two elevated claim types, where vendor figures alone are treated as a pass-through rather than substantiation, since you would be the one making the claim.
In your browser, on your machine, in local storage. There is no account, no server and no sync. Export a JSON backup if you want a copy, because clearing your site data erases it.
Each browser holds its own register. Export JSON from one and import into another to hand a copy over. It is a single-operator tool by design, not a shared workspace.
A single-file web app you keep, plus its printable pack. One payment, no subscription and no upsell. It installs to a phone or desktop and runs offline afterwards.
Load the example register, then paste in the first three sentences from your own landing page. Most teams find one they cannot back within ten minutes.
Get the log — $19 one-time