AI Search Visibility in 2026: How to Track It and What to Measure

What AI search visibility actually means

AI search visibility is the share of AI-generated answers about your category that name you. It is not a ranking, because there is no ranked list to hold a position in — an assistant asked for the best tool in a category typically names two to five products and omits everything else. You are inside that set or you are invisible, and there is no page two to be on.

That makes it a rate, not a position, and rates need a denominator. The denominator is your prompt set: the specific questions a buyer would type. Change the prompt set and the number changes, which is why a visibility figure quoted without the prompts behind it is not comparable to anything, including your own figure from last month.

It also differs from classic search visibility in who is doing the reading. In Google, a human decides whether to click your result. In an AI answer, a model decides whether to mention you at all, using a handful of retrieved documents. Everything in the rest of this guide follows from that one difference.

The four metrics worth tracking

Most AI visibility dashboards report more than this. These four are the ones that change decisions:

• Mention rate — the share of your prompt set whose answers name you at all. This is the headline number and the one to move first. Going from 0 to 2 out of 8 is a bigger event than any refinement that follows.
• Position within the answer — whether you are the first name, one of several, or the afterthought after the recommendation. Being named third in a list of five is much closer to invisible than the mention rate suggests.
• Competitor share on the same prompts — who wins the answers you lose. This is the metric that tells you whether the gap is your content or the category's source set, because a competitor appearing everywhere you do not is almost always a third-party presence gap.
• Cited sources — which URLs the model pulled from. This is the most actionable output of tracking and the most commonly ignored, because it names the specific pages you need to be on.

Why one-off checks mislead in both directions

AI answers are non-deterministic. The same prompt, same model, same day can name you once and skip you the next time. A single check therefore has an error bar wide enough to contain almost any conclusion you want.

The false-positive version: you check, you appear, you conclude you are fine, and you stop. The false-negative version: you check, you are absent, you conclude the channel is hopeless, and you stop. Both are common, and both come from treating one sample as a measurement.

There is also a subtler trap. If your prompt mentions your product by name, the model will almost always discuss it, which feels like visibility and is not. The only reading that means anything is the unprompted one — a question phrased the way a buyer who has never heard of you would phrase it.

Building a prompt set that will not flatter you

Five to ten prompts is enough to start, and the discipline is in how you write them:

• Never include your brand name. If the buyer already knows you, the answer is not measuring discovery.
• Use the words a buyer uses, not the words your marketing uses. If you call it revenue intelligence and they call it a sales dashboard, measure the sales dashboard.
• Cover several intents, not several phrasings of one. Best-in-category, cheapest option, alternative to the market leader, best for a specific use case, and a direct comparison are five different retrieval paths and often five different answers.
• Include at least one prompt you expect to lose. A set you always win is a set that has stopped being informative.
• Freeze the wording. Changing prompts between runs makes the trend meaningless, which is the whole point of tracking.

How often to run it

Daily is mostly noise for a small product. The underlying inputs — your content, the roundups you appear in, the models themselves — do not move daily, so daily readings mainly measure the non-determinism.

Weekly is right if you are actively shipping changes and want to attribute movement to them. Monthly is right for everyone else, and is the honest default: it is fast enough to catch a real decline and slow enough that a single odd run does not dominate the picture.

The cadence that fails is the one that depends on someone remembering. Manual tracking almost always stops within a few weeks, and the reading that matters is the one in month four. Automating it at any cadence beats a better cadence you will not sustain — that is the argument for a scheduled scan, and it is why ours costs $19 a month rather than nothing.

What AI visibility tracking should cost

The floor is free: a one-off scan to find out whether you have a problem. That is worth doing before you spend anything, and it is what our /audit scan does with no card.

Entry-level scheduled monitoring for a single product runs roughly $19 to $35 a month across the category. At that price you are buying a recurring reading and a trend line, not a platform.

Mid-market tools with competitor benchmarking, multi-engine coverage and client-ready reporting run roughly $100 to $300 a month, which makes sense for agencies reporting on several brands. Enterprise answer-engine analytics with crawler-level data are quote-based and generally four figures.

The sizing rule is simple: the tracking should cost a small fraction of what the channel is worth to you, and you cannot know what it is worth until you have taken the baseline. So take the free reading first.