GEO is being sold as a new trade. Some people sell “AI rank.” Some sell “special markup.” Some promise you will be inside the answer in seven days. The sentences sound like the loudest, most brand-damaging years of early SEO.

The work that can be done is narrow.

First, the entity must be recognizable. A person, a company, a product — how the name is written, how the English is written, who it relates to — has to be fixed on pages you control, and those pages have to point at each other. Engines and models are afraid of mixing you with a similarly named org. Yisuan is not Xiaoyi and not a chip firm with a similar Chinese name. That has to appear in visible text, not only in ads.

Second, pages have to open. A homepage, a store listing, a published post — after you send it out, open it. If it fails, or if your name is not on the page, the work is not done. Xiaoyuan encodes this as a product rule because it is the only action you can check on the spot.

Third, questions need answers. If people ask what a company does, the page needs a paragraph that still makes sense when lifted out of context. Headings, lists, dates, and sources beat adjectives. Google said in 2026 that AI Overviews need no special Schema and that Search ignores llms.txt. Betting the business on a magic file for robots is the wrong bet. I still publish llms.txt, as a description for systems that read it, not as a ranking switch.

Fourth, you have to ask. Tested stays tested. Untested stays untested. Writing “might be mentioned” as “already recommended” is a fake result.

What cannot be done is equally clear. You cannot guarantee that a model will mention you tomorrow. You cannot mix agency work and a software fee into one large number. You cannot stack entities with generated look-alike pages — that is spam, not GEO.

So can GEO be done? Yes. It is verification, plus some writing. It is not a spell. Anyone who delivers a spell will be asked through by the first serious client.

Can GEO actually be done