Search for 'AI visibility tool' or 'GEO tool' and you'll find dozens of vendors, from browser plug-ins that screenshot one ChatGPT answer to enterprise platforms with pricing on request. For an agency the wrong choice is expensive: you put clients on it, you build reports around it, and after three months it turns out the numbers aren't reproducible. These ten criteria prevent that.
Why the choice matters
AI answers aren't search results. They vary per run, per model, per country and per wording of the question. A tool that asks one question to one model once a week delivers a number that's different next week without anything having changed. If you report that to a client, you're reporting noise. So the core of a good tool isn't the pretty chart but the tracking method underneath. Everything below follows from that.
Ten criteria
1. Tracking via official APIs, not scraping or a browser. Scraping the ChatGPT interface is unstable, against the terms of service and gives answers that depend on a logged-in account. Via the APIs (OpenAI, Anthropic, Google, Perplexity) tracking is reproducible and in the client's country and language.
2. Several AI models, and the list is configuration. At least ChatGPT, Perplexity and Gemini; Claude as an option. A client can be strong in one model and invisible in another; with one model you miss that. Also ask how new models get added.
3. Repeated tracking on fixed days. Three times a week on fixed days is the minimum to tell a trend from noise; daily is better for large markets. One run a month is a snapshot.
4. Every answer stored verbatim. You must be able to click from a number to the literal answer: which businesses, in which order, with which sources. Without that you can't convince a client and can't correct a classification error.
5. Sources and citations per AI model. The most important lever in GEO is a listing on the sources AI cites. A tool that doesn't show which domains and URLs are cited per model, and whether the client is listed there, leaves you doing the most important work blind.
6. Competitors, including automatic detection. Share of voice and win rate next to the competitors you specify, plus the names AI mentions that you don't track yet. The latter are often the surprise in the first conversation.
7. Sentiment with evidence. Not just 'positive/negative' but the literal sentence, the theme and the source the claim comes from. Otherwise you can't do anything about it.
8. Actions with steps, code and measured effect. A list of 'recommendations' is easy; an action with a step-by-step plan, ready-to-use code or a draft email, and the measured effect in prompts and points thirty days later, is rare. This decides whether the tool creates work or saves it.
9. Transparent metrics. How is visibility calculated: across all prompts or only non-branded ones? What exactly is share of voice? How is the overall score weighted? If that isn't explained, you can't explain it to your client either. See as an example how GeoReady tracks.
10. A price per client, not per seat or per 'credits' you don't understand. For an agency the cost per client must be predictable so you can build your fee on it. Ask what happens when a client wants more prompts or an extra model.
Three questions specifically for agencies
Is there an agency portal? You want all your clients on one screen: score, trend, open actions, who needs attention. Logging into separate accounts per client doesn't scale beyond five clients.
Is white-label complete? Logo, colors and font in the portal, the client app, reports, emails and live views, with the option to hide the tool's name entirely and use your own domain. Half white-label (only the report) means your client still sees which tool you use.
Can the client look without being able to do anything? A read-only login or a live view with a password, without the client being able to change settings or enter your agency portal. Read how GeoReady does this on GEO for agencies.
What to test yourself in a trial
Put one real client in the trial and do four things. Read ten answers and check whether the businesses and sources mentioned match what you see in the AI apps yourself (they don't need to be identical; they should be recognizable). Compare visibility across three tracking days: does it fluctuate within a few points, or does it jump from 10 to 40? Open one action and ask yourself whether a junior could execute it without explanation. And export the data: can you take the answers and numbers with you if you ever leave?
Red flags
- 'Real-time' visibility without fixed tracking days: that's a snapshot you can't compare.
- One overall number without a breakdown per AI model.
- A number that doesn't click through to a literal answer.
- Recommendations without steps, without a source and without measurement afterwards.
- No explanation of the formulas behind the metrics.
- A price you only hear after a demo.
Whichever tool you choose: start with the tracking method, then the sources, then the actions. The charts come by themselves. And never put a client on it before you've watched a trial of your own for a week; the GeoReady trial runs seven days and tracks three AI models from the start.
Look for: tracking via official APIs on several models, repeated on fixed days, every answer stored verbatim, sources and citations per model, competitors, sentiment with evidence, actions with measured effect, white-label and a price per client.