How to Read an AI Product Page and Check Its Claims
How to choose - - Priya Raman
Two statements can both be true and still mislead. OpenAI's API documentation says data sent to the API is not used to train its models by default, as of March 1, 2023. The same page says abuse-monitoring logs, which can include prompts and responses, are kept for up to 30 days, and that removing them needs Zero Data Retention, which requires approval (OpenAI API docs). A landing page that keeps only the first half is not lying. It is selecting.
Reading an AI product page is the skill of finding the second half. A claim is checkable when it names a unit, a condition, a date or a method. Marketing is the same sentence with those four removed.
Phrases and what they leave out
| Phrase on the page | What to check at once | Where proof usually sits |
|---|---|---|
| "AI-powered" | Which task the AI does, and what happens when it is off | Feature list, docs |
| "Unlimited" | The unit being counted and any fair-use wording | Footnotes under the pricing table, terms |
| "Included in all plans" (for an AI feature) | The quota or credit amount per plan | Pricing page, billing FAQ |
| "We do not train on your data" | Which plans and endpoints, and how long logs are kept | Privacy policy, data docs |
| An accuracy figure | Measured on what data, by whom, and when | Method note or linked report |
| "Works with your tools" | Whether the integration is native, via an API, or a roadmap item | Integration directory |
Source: the OpenAI and Formaloo pages named in this guide, as of November 2026.
The Formaloo case fits the third row. Its pricing page lists AI-powered surveys and quizzes on all plans, and says every AI module uses credits, but states no credit amount per plan. "Included" is true. How much is included is a question the page leaves open, and the answer decides whether the free plan is a tool or a demo.
Checkable in five minutes, and not
Some claims can be settled from the page itself or one click away. A price, a plan limit, a date on a changelog, the export formats in a comparison matrix. These carry a number or a name, so they can be wrong in a way anyone can see.
Others cannot be settled from a vendor page at all. Accuracy, speed on your data, and "saves hours" depend on the task, and a figure with no stated method is a description of the vendor's own best case. That does not make it false. It makes it unusable for a decision. The honest move is to say so in the evaluation notes: figure given, method not published. The catalog uses the same wording for the cards whose facts it could not confirm, marked "not independently verified".
The order matters because the first four steps cost nothing and can end the evaluation early. A tool whose unit is undefined, as with messages and resolutions in the chatbot checks guide, cannot be priced however good the demo looks.
When the page is silent
Silence is information. If the pricing page lists no credit amount, no retention period or no method, that fact belongs in the notes next to the tool's name, and the question goes to the vendor in writing. Treat a vague reply as the answer. A form builder such as Formaloo shows how this plays out: the page names credits, so the amount is the first question to send.
The last step is the only one the page cannot do for you. A handful of real inputs from your own work say more about fit than any figure on a landing page.