How to choose an AI visibility tool

An AI visibility tool measures which answers from ChatGPT, Gemini, Claude, Copilot or Perplexity a business appears in when the user asks about the category. In the market they're also called GEO tools, for Generative Engine Optimization; if the term is new to you, it's defined in what is GEO?.

They all look alike on the pricing page. They differ in eight things, and only three of them show up in the sales material.

1. Does it measure in the language and from the country where you sell?

It's the filter that rules out the most candidates at once, and the one most often overlooked.

When an engine answers in Spanish, it consults a different set of pages than it does in English. It isn't the same content translated: it's other sources, other outlets, other threads. A measurement run with prompts in the wrong language gives you an accurate portrait — of another market. In South Florida, where many businesses sell to both English and Spanish speakers, that usually means measuring in both.

And within a language, the location the query is run from also moves the answer: the same question can name different businesses from Miami than from Orlando, or from Quito than from Bogotá. Ask whether the tool lets you set the country — and ideally the city — of the run, not just the language.

What to ask: what language are the prompts run in, from what country, and can I choose?

2. How many times does it run each prompt?

This is the technical question that most quickly separates the serious from the cosmetic.

A generative model is probabilistic: the same prompt, run twice in a row, can give two different answers. A tool that runs each prompt once a day and shows you a number is reporting noise to two decimal places. You'll see rises and drops every week that correspond to nothing you did.

The right approach is to repeat each prompt several times per run and report the range, not a point. If a tool doesn't say how many runs are behind each figure, that figure can't be interpreted.

What to ask: how many repetitions are behind each number, and do you show me the range?

3. Does it tell apart when the engine searched the web and when it answered from memory?

Almost none report it, and it completely changes how the results read.

An engine can answer by consulting the web in that moment, or it can answer from what stuck in its training. In the second case, no change to your site could have affected that answer. If you mix both kinds of run into one average, you'll credit your work with movements that aren't its doing, and dismiss work that did pay off.

What to ask: can I see, run by run, whether there was a web search?

4. Can I read the full answer, or just the number?

A label you can't audit is worth nothing, and this applies above all to sentiment.

“Sentiment: positive” is one model's claim about another model's output. Without the exact sentence that justifies it, you have no way of knowing whether the engine praised you, named you as the example of what to avoid, or confused you with another company. All three cases can be labeled the same.

The same goes for mentions: you need the exact prompt, the engine, the date, the time and the full text of the answer, saved. It's how you audit the tool, and it's what you need in the uncomfortable case where an engine says something false about your business and you need proof.

What to ask: do you keep the full text of every answer, and can I export it?

5. Does it show which sources each engine is citing?

Measuring mentions tells you that you don't show up. The cited sources tell you why.

When an engine answers a discovery question, it cites third parties: comparisons, threads, directories, media coverage. If you don't know which pages those are for your category, you don't have an action plan — you have a thermometer. And almost all the work that moves the needle happens off your site, on those pages.

What to ask: do you tell me which URLs each engine cites for my prompts, and how often?

6. Who builds the prompts?

There are two models and neither is bad, but it's worth knowing which one you're buying.

In some tools you write the list of questions. It's flexible and cheaper, and it works if you already know how your customer asks. In others the set is built from your Search Console, your category and your competitors. It costs more and it's what you need if you don't have that research done.

The expensive mistake is buying the first while thinking it's the second, writing twenty questions that name your business, and concluding six months later that AI visibility doesn't work.

What to ask: do I set the prompts or do you build them, and from what inputs?

7. Who executes the changes?

This is the question that decides the return, and the one no feature comparison answers.

Most tools in this category deliver data. The work that comes out of that data — rewriting the page that doesn't answer, earning the mention in the source the engine cites, fixing the outdated listing that describes you wrongly — someone else has to do. If nobody in your company has time assigned to do it, the pattern is predictable: the tool gets bought, someone opens it for the first few weeks, and the report becomes an email nobody reads.

Before comparing prices, answer this: who, by name, is going to take this month's report and turn it into changes? If the answer is “nobody yet,” what you're missing isn't data.

There are three ways to solve it: an in-house team that already executes, an agency that receives the findings, or a provider that measures and executes. Apoda is the third case, and it isn't the best in the abstract: if you already have someone to execute, a measurement-only platform costs you less and serves you just as well.

What to ask (yourself, not the salesperson): do I already have hands, or am I just missing data?

8. What's the real price, not the one on the homepage?

Prices in this category are published per domain and billed annually, and the three things you need most are usually add-ons: extra prompts, extra engines and extra users.

To really compare, build the price around your specific case, not around the entry plan:

What to ask: with my number of prompts, my engines and my users, what's the monthly bill?

When don't you need any tool?

It's worth saying, because it saves money.

• If you've never checked your robots.txt. Before buying anything, make sure AI crawlers can read your site. It's free, takes ten minutes and is the most common reason for not showing up. It's explained in my business doesn't show up in ChatGPT.
• If you sell to a closed list of known accounts. AI visibility matters in discovery. If your customers already know you and deals come in through relationships or tenders, this isn't your bottleneck.
• If you're not going to execute anything in the next six months. Measuring without intervening produces a nice data series and no change.

A checklist to bring to the sales call

1. What language and what country are the prompts run in?
2. How many repetitions are behind each number?
3. Does it tell apart when the engine searched the web?
4. Do you keep the full answer and the sentence that justifies the sentiment?
5. Do you show the sources each engine cites?
6. Who builds the prompts?
7. Who executes the changes that come out of the report?
8. What's the real bill with my number of prompts, engines and users?

If you want to see how Apoda, Profound, Peec AI and Otterly answer these side by side, it's in comparisons. And if you're already evaluating the category leader, the in-depth comparison is in a Profound alternative for Spanish-speaking markets.