Before you hire a GEO agency: seven questions and three red flags

We build a tool, not an agency, and some of the companies that use it also pay an agency. So we see both the good engagements and the invoices for the other kind. This is what I would ask if I were buying.

Seven questions

1. Which questions will you measure, and can I see them before you start? Generative engine optimisation is about whether AI answers cite you for the things your customers ask. If the agency cannot name those questions on day one, in your customers' words, in your language, there is nothing to optimise towards. A good answer is a list of twenty to fifty, reviewed with you.

2. How do you measure, and how often? The honest methods are the same ones you could do yourself: ask the tools, through their interfaces or APIs, and record who gets cited. Monthly at least, because answers drift. Be suspicious of a proprietary "AI visibility score" with no explanation of what was asked and what came back. Ask to see one raw run.

3. What will you change on my site, specifically? The work that moves citations is unglamorous: first paragraphs that answer the question, named authors with real profiles, structured data that identifies the company, the discovery files, fixing robots.txt, turning PDFs into pages. If the proposal is mostly "content strategy" and "AI-optimised copy" with no page-level list, it is an SEO retainer with a new label.

4. What will you do off my site? Third-party mentions carry more weight in AI answers than your own pages. Ask what the agency will do about directories, trade press, partner pages, your Wikidata entry. If the answer is "we focus on on-page", half the lever is missing.

5. Who writes, and under whose name? Assistants weigh named humans with a track record. Agency copy published under your managing director's name, without her having read it, is a liability waiting to be quoted. Agree in writing who the byline is and who approves.

6. What happens to my robots.txt and my CMS access? You should know every change made to the technical side of your site, and own the accounts. An agency that wants exclusive admin access to "manage the AI files" is building a lock-in, not a result.

7. What does a good quarter look like, and what does a bad one? A real answer sounds like: "for these twelve questions you are cited by none of the four tools today; in three months we expect movement on the four where your page already ranks well in Google, and probably none on the rest." A bad answer sounds like a percentage.

Three red flags

Guaranteed citations, guaranteed rankings in ChatGPT, guaranteed traffic. Nobody controls what an answer engine says. The agencies that know this say so. The ones that promise otherwise either do not know or are counting on you not to check.

"We submit your site to the AI models." There is no such thing. There is no submission form at OpenAI or Anthropic or Google for being included in answers. Anyone selling this step is selling a form that does not exist.

A price with no scope. "GEO package, CHF 3,000 a month" without a list of questions, pages and deliverables is a retainer for attention. It may still be worth it if you trust the people; just know what you are buying.

What a fair engagement looks like

For a Swiss SME, in my experience: a one-off audit and measurement baseline, a defined list of pages to fix with the fixes applied or specified for your developer, a short off-site plan, and a monthly re-measurement against the same questions. Three to six months, then a decision on whether to continue, based on whether the citations moved. Anything that cannot be stopped after a quarter should be a bigger conversation.

Or do the first step yourself

The baseline measurement is the part you can do for nothing before any meeting: ask the tools your customers' questions and write down who gets named. The free analysis on this site runs that across four tools in a minute. Walk into the agency conversation with that printout, and the quality of what you are offered changes immediately.