GEO is new enough that everyone is still learning, and established enough that the mistakes are already repeating. These eight are the ones we see most often at agencies that are just starting. They have one common cause: doing something before you've tracked. That's understandable (doing feels productive) and avoidable.
Mistake 1: one check as a measurement
A staff member types 'best HVAC contractor in Boston' into ChatGPT, doesn't see the client in it and notes: 'not visible'. A week later the client is in it. Nobody knows why. AI answers vary per run, per wording and per day; one check is a snapshot with no comparison value. The fix: a fixed set of 35 to 50 questions, on fixed days, with the answer stored verbatim. Then you talk about a win rate ('mentioned in 6 of 9 runs') instead of yes or no. See how to track AI visibility.
Mistake 2: only one AI model
Most agencies start with ChatGPT because it's the best known. But Perplexity cites other sources, Gemini leans on the Google ecosystem, and Claude chooses differently again. A client can be in 40% of answers on Perplexity and 8% on ChatGPT. Tracking one model misses not only half the picture but also the clue to where the work is: the difference between models almost always points to a source one model reads and the other doesn't.
Mistake 3: writing content without knowing the questions
'Let's write a blog about heat pumps' is an SEO reflex. For GEO the question is: which sub-questions does an AI model ask itself when someone asks for an installer? ('What does a heat pump cost in a 1970s home?', 'Which rebates apply in 2026?', 'Is a hybrid heat pump suitable without underfloor heating?') Those AI search queries are visible in the tracking. If the client has no page that concretely answers such a sub-question, that is the content gap. A generic article doesn't fill it.
Mistake 4: skipping sources
This is the most expensive mistake. Agencies put months into the client's website while the AI model builds its answer from six comparison sites and directories the client simply isn't listed on. AI doesn't invent businesses; it repeats what the sources say. So look at the citations first: which domains come back in almost every answer, on which is the competitor listed and the client not? Getting a listing on such a source is a few hours of PR work and often the fastest win there is. Read why sources and citations are the key.
Mistake 5: forgetting the technical side
A robots.txt that blocks GPTBot, PerplexityBot or ClaudeBot (often unintentionally, through a generic 'block all bots' rule from a previous agency) makes all other work useless: the model can't read the site. The same goes for a missing sitemap, business details only in an image, or inconsistent name/address/phone across the web. Do the site check before anything else; the technical actions are a precondition, not an option. See the technical GEO audit.
Mistake 6: promising results
'You'll be in ChatGPT within three months' is a promise you don't control, and every client who got it and didn't see it delivered cancels. Promise what you do control: tracking runs, actions get executed, the effect gets measured, and the client sees every month exactly what changed. That's a promise you keep every month. That's why you have the first conversation with numbers, not promises: see the first conversation about GEO.
Mistake 7: reporting noise
Celebrating a 3-point rise one week and explaining a 3-point drop the next costs you credibility. Within a few points a week, AI visibility is noise. Report trends over four weeks or longer, put the change versus last month next to every number, and say when something falls within the noise. Then, when there's a real jump (a listing on a new source that shows up in the answers), the client believes it too.
Mistake 8: stopping after two months
Source actions need time: the AI model has to read the source again before the listing shows up in answers. Content actions take one to three months. Whoever concludes after eight weeks that 'GEO doesn't work' is measuring exactly at the moment the first actions start to land. Agree a first period of three months with the client, and measure the effect thirty days after every action: in how many prompts the client is mentioned since, compared with before. That number, per action, is the proof that it works, and the reason to continue.
The principle behind all eight: track first, then act, then track again. Whoever keeps that rhythm doesn't make the mistakes, and after a year has something no tool delivers: knowledge of which sources and actions work in which industry. The plan for the first thirty days is in starting GEO at your agency.
Most GEO mistakes come from acting before tracking: one check as a baseline, one model, content without the questions, ignoring sources, forgetting the technical side, promising results, reporting noise and stopping too early.