How to Tell If Your AI Visibility Efforts Are Actually Working

The problem with "I think it's working"

If you've added schema markup, rewritten your meta description, or added an FAQ section to your site because someone told you it would help with AI visibility, you're probably wondering how you'd even know if it worked.

This is harder than checking Google rankings. There's no dashboard built into ChatGPT that shows you where you rank. There's no "position 3" the way there is in search results. AI models generate an answer in the moment, and that answer can change based on how the question is worded, which model you're using, and sometimes just by chance.

That uncertainty is exactly why it's easy to fool yourself. You make a change, you ask ChatGPT about your business a couple of times, it mentions you, and you conclude the change worked. But you don't actually know if it would have mentioned you anyway.

Ask the same questions, the same way, over time

The only real way to tell if something is working is to test consistently, not randomly.

That means picking a small set of realistic questions a customer might actually ask — "best plumber in [your town]," "who should I hire for [your service] near [your area]" — and asking those same questions across the same set of AI models on a regular basis. Not once. Not just after you make a change. Regularly, so you can see a pattern instead of a single data point.

A single response where you're mentioned or not mentioned doesn't tell you much. What matters is whether your visibility is trending in one direction over several checks, across several questions, across more than one AI model.

This is also why it's worth being skeptical of anyone who shows you one screenshot of ChatGPT mentioning your business and calls it proof. One good answer isn't a trend. It might just be luck, or a slightly different phrasing of the question that happened to favor you that day.

Separate the signals you can control from the results you can't

It helps to think about AI visibility in two layers.

The first layer is your own site: does it have clear schema markup, a meta description that actually explains what you do and where, headings that describe your services in plain language, and reasonable page speed. These are things you can check directly, fix directly, and verify are done. This is the layer a tool like Belcher can actually inspect for you — reading your site and telling you what's missing or broken, and generating the schema, FAQ, or meta content you can paste in yourself.

The second layer is the actual output: whether AI models mention you when someone asks a recommendation-style question. This is the layer that matters to your business, but it's also the layer you don't fully control. AI models are trained on a mix of sources, and no one outside the companies that build them can say exactly how any single business's website affects a given answer. Fixing your site's signals makes you more likely to be understood and referenced correctly, but it isn't a guarantee, and no one honestly telling you about this space should promise otherwise.

Belcher's role is to check the first layer for problems and to test the second layer by asking real recommendation-style questions across ChatGPT, Gemini, Claude, and Perplexity, so you can see where you currently stand and whether that changes over time. It's not able to tell you what's being said about you on Reddit or review sites, and it won't pretend to.

What actual progress looks like

Real progress usually looks unremarkable at first. Your AI Recommendation Score moves a little. You start showing up in answers to some questions but not others. One model mentions you consistently while another still doesn't.

That's normal, and it's more honest than a story where everything suddenly works after one fix. If someone tells you they can guarantee you'll be the first result in every AI answer, be skeptical. Nobody controls that fully, including the platforms doing the testing.

What you should expect is a clear before-and-after picture: what your site's signals looked like, what got fixed, and whether your recommendation rate across models moved in response. If a tool can't show you that basic before-and-after comparison, it's hard to know if you're actually improving or just paying for reassurance.

Keep it simple

You don't need a complicated tracking system to know if this is working. You need the same questions, asked consistently, across the same models, with a clear record of what you changed and when. That's the difference between guessing and actually knowing.