Can You Actually Change What AI Says About You?

You can change public evidence and observe later answers. Establishing what caused a different answer takes more work.

Editorial note. This revision narrows earlier claims about AI behavior and competing services. The July and August entries below are the author's retrospective field notes. Their full prompts, responses and session records are not attached to this page, so they cannot serve as a reproducible baseline or verified improvement case.

A mention and a recommendation answer different questions

A name appearing in an answer is observable. It does not, by itself, prove that a particular page was retrieved. A shortlist identifies candidates; an explicit recommendation advises a user toward an option under stated conditions. These distinctions help us describe the answer without claiming access to the model's internal selection process.

Independent support can matter when a buyer evaluates an option. I use corroboration gap for the distance between a published claim and independent confirmation. Whether that gap explains a particular AI answer remains a hypothesis until the evidence supports it.

The subject is me

I started with my own public identity because I could change it, keep a record and accept inconvenient results. SKINBOT belongs in that history as experience building an AI decision system. It should not replace my current identity as founder of Legibi and legibility strategist.

My original notes described four points: on July 16, name confusion and an outdated business description; on August 7, better identity resolution across several systems; and on August 16, two answers that appeared to select or recommend me. These are self-reported observations, not independently verified findings. The underlying files would need to be published and checked before anyone could evaluate those claims as a study.

What one answer can tell us

One preserved answer establishes what was returned in that recorded session. It does not establish how often other people receive the same answer. Repeated runs can estimate outcomes within a defined test set, with the surface, dates, language, market, prompts and exclusions reported. That estimate is not the frequency of private customer questions or a market-wide recommendation rate.

My historical notes also described answers that recommended me while qualifying the evidence as first-party claims and suggesting questions for a buyer to verify. That is a useful distinction to investigate: being named and having sufficient evidence are separate observations. It is not a basis for saying that competing audits or platforms never assess evidence or current facts.

Identity and current activity need separate checks

A system can identify a person and still describe their work using outdated information. That may suggest a source, freshness or interpretation problem. The answer alone does not identify which mechanism is responsible. A useful diagnostic preserves the mistaken statement, visible sources and current evidence before proposing a correction.

What a change would need to demonstrate

A before-and-after difference is an observation. To evaluate an intervention, preserve the starting answers, document exactly what changed, repeat the same core scenarios under comparable conditions and keep failed attempts. Unchanged comparison scenarios can help detect broader shifts. None of this guarantees a recommendation or establishes causality by itself.

A descriptive audit can still be useful without a control group. It can identify wrong facts and practical gaps. A stronger causal claim requires a stronger design. The Synthetic Term Control protocol addresses one kind of unsupported attribution; it is not a universal control for all changes in recommendations.

The September repeat mentioned in the original essay was a plan. No completed September report is attached here. A scheduled date is not evidence that the run happened or that an earlier result persisted.

A public record you can inspect now

On October 17, ChatGPT named other consultants. That separate record includes the exact question, full answer and observed links. It also records failed attempts and why the answer is excluded from a strict T0 baseline. I was not named. We measured ourselves first; the purpose is to make the evidence inspectable, including the result we would not have chosen.