5 reasons your GEO strategy sucks.

“Why is Acme the best CRM for agencies?” is a friendly prompt. You've supplied the company and the conclusion.

That can be useful when you want to inspect how a system explains a named brand. It tells you very little about whether an unfamiliar buyer will discover you.

Separate your questions by what the buyer knows. Someone investigating a problem might ask how to manage client handoffs. A buyer making a shortlist might ask which CRMs suit a ten-person agency. Someone close to a decision could ask whether Acme supports a particular integration.

You need room for all of those situations. Label branded questions separately, and write down why each question belongs in the panel, the set of questions you track. For a deeper question-design workshop, see Gist’s prompt architecture guide.

Use sales calls and customer interviews to find the language. If you generate candidate questions with AI, treat them as research suggestions. A list of 80 plausible questions does not establish that 80 questions are popular.

Fix this today: inspect the questions behind your headline score. If almost every question gives the model your name, build a separate discovery panel before calling the result market visibility.

Your visibility score rose. Good. Now open an answer that contributed to it.

Was your company recommended? Mentioned in passing? Used as a comparison? Was the linked source your website, or a third-party article about you?

Those are different outcomes. Define them before you combine anything.

Consider a hypothetical panel of 20 completed answers. Your company is mentioned in eight, recommended in three, and its own website is linked in two. Reporting “40% visibility” describes only the first of those measures. That tells a colleague how often your name appeared. It doesn't tell them how often you were recommended or your website was linked.

Keep the denominator visible. Keep completed answers with no citations in the panel. Log failed requests separately and explain how they affect the denominator. If you change the question panel, add an engine, or switch collection methods, annotate the change before comparing the new score with last month's.

And keep observed answers separate from audience data. Google's Generative AI performance report measures impressions for links shown in supported Google AI features. A tracker sampling answers to your questions measures a different thing.

Fix this today: put one plain-English definition and the denominator beside every metric in the client report.

A fifth article explaining that AI search is growing will struggle to justify its existence.

Start with a real decision. For an agency choosing a GEO tool, the useful detail might be whether clients get separate workspaces, what an export includes, or how the bill changes when another engine is added. Those answers can affect a purchase.

Your company should have material a generic rewrite cannot supply: a documented implementation, an original measurement, a product limitation explained honestly, or a useful comparison built from current sources. Pick the evidence that helps the reader finish the job.

Google's people-first content guidance asks whether a page contributes original information or analysis and adds value beyond rewriting other sources. Apply that test before approving a draft.

Fix this today: ask what your article helps a reader understand or decide that the source material alone doesn't. If the answer is “the same facts in our tone,” keep working.

“Allow AI” is a dangerously vague ticket.

For ChatGPT search, OpenAI distinguishes OAI-SearchBot from GPTBot, which is associated with potential model-training use. Those controls are independent. ChatGPT-User serves certain user-initiated actions and is not the control for search eligibility. See OpenAI's crawler documentation.

Google has its own requirements. Its AI-search guidance says a page must be indexed, eligible for a search snippet, and included in Search generative AI features. It also says special AI markup or an llms.txt file is not required for Google Search.

None of that gives you a guaranteed citation.

Ask the website owner to check the actual page and the relevant crawler. Record which system was tested, when it was tested, and the observed result. A failed fetch from one audit tool is a reason to investigate. It does not establish that every AI system is blocked.

Fix this today: replace the vague “AI crawlability” task with a page URL, the target service, and a specific access check. Keep decisions about training permissions with the people authorized to make them.

A recommendation that says “improve authority” can sit untouched for months.

Turn a finding into a change someone can finish. If the wrong service area appears on your website, identify the page and the person who can correct it. If buyers keep asking a pricing question you never answer, give the product owner an exact question to resolve.

Each task needs a page or system, an owner, a concrete change, and a way to check the result. Include the date it actually shipped.

Then repeat the relevant measurements under comparable conditions. Keep the original answers. If visibility changes after an edit, describe the sequence accurately; a before-and-after observation alone doesn't prove that the edit caused the change. Other sites and the answer engines may have changed too.

Fix this today: choose one finding from the latest report, assign the smallest useful correction, and set the next observation window. Stop expanding the backlog until that loop works.