Chapter 7: How to Measure AI Visibility

Last updated: November 1, 2026.

Measurement used to be fairly easy in classic SEO. You checked where you ranked for a keyword, and Search Console gave you impressions, clicks and average position. AI search doesn’t offer that comfort. The same question gets a different answer each time (Chapter 3), each platform pulls from different sources (Chapter 4), and most platforms don’t show you their own data.

What we do have is the list of 30–50 commercial prompts from Chapter 6. This chapter is about measuring how you show up across that list. We’ll use five metrics instead of one score, run each prompt several times, keep the platforms apart and look at how far the tools can be trusted. You’ll end up with a table you can fill in every month.

1. Why a single score misleads

Most AI visibility tools sum up the result in one number: “Your AI visibility score is 42.” It’s an easy number to put in front of a manager. In my view it misleads more than it helps.

In Chapter 1 I described how differently a brand can show up in an AI answer: mentioned, recommended, cited as a source, clicked, and on top of all that described correctly or incorrectly. Chapter 3 showed that these can be independent of each other. Your guide can be cited as a source while the recommendation list in the same answer names your competitors.

Each of these points to a different problem. If you’re never mentioned, the problem is probably in the foundation: either your pages can’t be found for the relevant searches or the brand isn’t well known enough. If you’re mentioned but not recommended, the information that would set you apart may not be written down anywhere, on your site or off it. If you’re recommended but not linked, AI is learning about you from other sources rather than your own site. If you’re recommended but described wrongly, there is outdated or inconsistent information about you somewhere.

A single score mixes these four situations together. When it falls from 42 to 38, it tells you neither what broke nor what to fix. A single score tells you something changed, not what changed. That’s why there are five separate metrics.

2. The five metrics

The metrics build on each other. First, are you in the answer at all? Then, are you recommended, are you linked, how do you do against competitors, and finally, are you described correctly?

3. Prompt Coverage

Prompt Coverage shows how many of the prompts on your list mention your brand. It measures breadth: which questions you show up for and which ones you’re missing from entirely.

There’s a definition problem here. Because of the variability covered in Chapter 3, we run each prompt more than once, and being mentioned in one run out of five isn’t the same as being mentioned in all five. Start by calculating a mention rate for each prompt, meaning the share of that prompt’s runs that mention you. Then set a threshold. Counting a prompt as “covered” when you’re mentioned in at least half of its runs is a reasonable place to start.

Say you have a list of 40 prompts and you’re mentioned at least half the time in 14 of them. Your Prompt Coverage is 35%.

The breakdowns are where this metric earns its keep. Split coverage by the persona clusters and prompt types from Chapter 6 and the gaps show up. If you appear in the purchasing manager’s questions but not at all in the production manager’s technical ones, what’s missing is probably technical content. If you show up in problem-and-solution questions but not in category discovery questions, your guides are being cited but your brand isn’t making the shortlist.

4. Recommendation Rate

Recommendation Rate shows how many answers recommend your brand. Of the five metrics, it’s the one closest to business results. As I said in Chapter 3, it’s usually being recommended that brings the sale.

The hard part is drawing the line between being mentioned and being recommended in the same place every time. If the model presents your brand as an option that fits the user’s need, you’ve been recommended: “For foaming products, X’s models with diving nozzles are a good fit” or “Based on your budget and capacity, take a look at X and Y.” If your name comes up but not as a recommendation, say in a list like “Manufacturers in Turkey include X, Y and Z” or in a negative context, you’ve only been mentioned.

Apply this rule the same way on every platform and every month. Whoever does the measuring should code the same answer the same way. Write the rule down and keep it together with a few example answers.

When you calculate Recommendation Rate, use only unbranded prompts in the denominator. In a prompt like “Is X Machinery reliable?”, your name is expected to come up. The rate tells you something in questions where the user doesn’t know you yet.

Some tools also report what position you hold in the list. In Chapter 3 we saw what Rand Fishkin said about tools that report “ranking position” in AI. The order changes on every run; what you want to know is how often you’re recommended across many runs.

5. Linked Citation Rate

Linked Citation Rate shows how many answers link to your domain. Whether or not your brand name appears, it measures whether your page is a source for the answer.

It helps to calculate it two ways. The overall rate is answers that link to your site divided by total answers; if your guides are cited in problem-and-solution questions, this is where you’ll see it. The second is how many of the answers that recommend you also link to your site. If that second rate is low, AI is recommending you but getting its information somewhere else.

When that happens, look at which sources do get the link. A marketplace, an industry directory, a competitor’s comparison post, a forum? That list shows you who AI is learning about you from. Working with those sources is the subject of Chapter 14.

Allow for platform differences when you count links. As we saw in Chapter 4, ChatGPT and Gemini put some of their sources behind a Sources button instead of in the text. Decide up front which ones you’ll count and apply the same rule every month. It’s better to keep in-text links and source-list links in separate columns, because the chance that a user clicks is very different for the two.

Linked Citation Rate doesn’t measure clicks; a link being shown doesn’t mean anyone clicked it. We get to clicks and business results in Chapter 8, with Analytics and other data sources.

6. Side by side with a competitor: Comparative Win Rate

Comparative Win Rate shows which brand AI puts ahead when you and a competitor appear in the same answer.

For each competitor, pull out the answers where they appear together with you, then count how many of those present you as the better fit. If AI recommends both of you without favoring either, that answer is a tie and doesn’t count.

This metric is most useful for two kinds of prompts. One is direct comparisons such as “X Machinery or Y Machinery?” or “Piston or diving nozzle?” In these, AI usually comes down clearly on one side. The other is shortlist questions: if you’re both on the list in a category discovery question, look at which one is described as the better match for the user’s need.

The reasons matter more than the number. When AI puts a competitor ahead, it usually gives a reason: “Y has a service network across Turkey,” “Y delivers faster,” “Y has a model built for foaming products.” Collect these and you’ll see which of your competitor’s information is reaching AI and which of yours isn’t. Maybe you have a nationwide service network too, but you never state it clearly anywhere.

Stick to three to five competitors. The ones your sales team runs into most often at the quote stage are usually the right choice.

7. Representation Accuracy and the fact sheet

Representation Accuracy shows how much of what AI says about your brand is correct. In Chapter 1 I said I consider being described wrongly the most dangerous situation; this metric puts a number on that risk.

To calculate it, first prepare a fact sheet: a list of the things known to be true about your brand. For a filling machine manufacturer, it might include:

• Machine types it makes, and the ones it no longer makes
• Capacity by model
• Product types it does and doesn’t suit
• Price range, or the factors that affect price
• Service regions and spare parts lead times
• Certifications
• Address, contact details and founding year

Then compare every statement about your brand in the answers against this sheet. Representation Accuracy is the number of correct statements divided by the total number of statements about your brand.

Don’t just count the errors; sort them by type, because each type comes from a different place. Outdated information (an old price, a discontinued model, the address before you moved) usually comes from a page nobody updated or an old directory listing. Wrong information, such as a feature that doesn’t exist or an incorrect capacity, can come from inconsistent information or from something the model made up. Missing information means an entire product group or service region never gets mentioned. Then there’s confusion: your brand being mixed up with a similarly named company or a competitor.

This metric matters most in branded prompts, which is why we put brand verification prompts on the list in Chapter 6. In the Tow Center study from Chapter 3, the tools stated wrong information with full confidence. So don’t leave Representation Accuracy to automated tools; at least for the basic facts, have a person read the answers. Chapter 13 covers how to fix a brand that’s being described wrongly.

8. Running the same prompt several times

All five metrics are rates, and a rate only means something with enough observations. In Chapter 3 we looked at the SparkToro and Gumshoe study. The researchers ran each prompt 60 to 100 times on each platform and found that, across many queries, visibility percentage was far more consistent than ranking position. How many runs you need for reliable visibility data was left open (Search Engine Journal).

Running every prompt 60–100 times a month isn’t realistic for most businesses. A workable routine looks like this:

• Run each prompt at least 5 times on each platform. Even that gives you a far sturdier picture than a single screenshot.
• For your 10 most important prompts, raise that to 10 or more.
• Make decisions based on prompt groups, not single prompts. Running 40 prompts 5 times each gives you 200 answers per platform, which produces a much more reliable rate than the 5 answers for one prompt.

With 5 runs, a prompt’s mention rate can only be 0, 20, 40, 60, 80 or 100%. Going from 2 out of 5 one month to 3 out of 5 the next looks like a jump from 40% to 60%. Given the variability from Chapter 3, that difference is most likely chance. Don’t mistake small swings in a single prompt for a signal; look for trends across prompt groups over several months.

Keep the setup fixed while you run them. As we saw in Chapter 4, memory, account, language and location all affect the answer:

• Open a new chat for every run.
• Use accounts with memory turned off, or accounts created just for measurement.
• Measure in your target audience’s language and country.
• Record the date, platform and, if possible, the model for each answer.
• Measure at the same time each month.

9. Each platform on its own

The platforms behave quite differently. In Ahrefs’ data, even Google’s AI Mode and AI Overviews cited the same URLs for the same question only 13.7% of the time (Search Engine Journal). And only about 12% of the links cited by ChatGPT, Gemini and Copilot were in Google’s top 10 results (Ahrefs).

So calculate the five metrics separately for each platform. If your Recommendation Rate is 60% in ChatGPT and 10% in AI Mode, the average gives you 35%, a number that matches neither platform. Two separate problems disappear behind one figure.

If you need to give a manager a single summary, weight the platforms by their value to you rather than equally. As when setting platform priorities in Chapter 4, you can set those weights by looking at how much traffic, and what quality of traffic, each platform sends you in Analytics. Always include the platform breakdown in an appendix to the report.

For where to start, the order from Chapter 4 still applies: Google’s AI Overviews and AI Mode first, then ChatGPT, then Gemini, Copilot and Perplexity depending on your audience. AI Overviews don’t appear for every search. If one doesn’t show up for a prompt, record that answer separately as “feature didn’t appear” rather than “not visible.”

10. The measurement table

You can start measuring with a simple table, no expensive tool required. Each row is one run, meaning one answer to one prompt on one platform:

The values in the table are examples, there only to show how to fill it in.

At the end of the month you calculate the five metrics from these rows, platform by platform. The monthly report then becomes a table like this:

Build the report with a separate column for each platform and a breakdown by persona cluster and prompt type, and you’ll see right away where to look.

11. How far to trust the tools

Measuring by hand takes time, and past a certain point a tool makes sense. Semrush, Ahrefs and a long list of newer tools built just for this run prompts on a schedule and report visibility. Use them, but know their limits.

The first is the gap between the API and the interface. Some tools run prompts through the model’s developer API rather than the interface users actually see. Genezio, which sells an AI visibility tool, compared GPT-5.2’s API with the ChatGPT.com interface in February 2026, across 3,645 conversations in the UK banking sector. The background searches the model ran had zero queries in common between the two paths, and only 12.3% of cited URLs were shared (Genezio). The study covered one sector and one model and was done by a company that sells the tool, so read the results with care. It’s still reason enough to ask your tool where it collects its results.

Personalization and location matter too. Tools usually run prompts from clean accounts in a specific country. That’s good for comparable measurement, but it doesn’t match what a real user with memory, history and a location sees. If you’re measuring for users in Turkey, make sure the tool can measure from Turkey and in Turkish.

Some tools also generate the prompt set for you. The warning from Chapter 6 about prompts generated by AI applies here: they look reasonable but may not reflect what your actual customers ask. Check the list yourself.

Tools that show a single “visibility score” don’t always explain how they calculate it. Rand Fishkin’s advice from the SparkToro study works here as well: ask your vendor to explain its method and show its math (Search Engine Journal).

Tools can count mentions and links automatically, but to know whether a statement is correct they need your fact sheet. Most tools either don’t measure Representation Accuracy at all or only offer sentiment analysis (positive, negative). A positive sentence can still be wrong.

Finally, models change. Platforms update their models and interfaces often. When your visibility shifts from one month to the next, keep a log of platform updates so you can tell whether the change came from something you did or something the platform did.

12. Search Console and Bing reports

Alongside your own measurement, use the data the platforms give you. As we saw in Chapter 4, there are two sources for this right now.

In June 2026 Google added separate performance reports for generative AI features to Search Console, and as of August 31, 2026 it opened them to all sites. For now the report shows impressions only, broken down by page, country, date and, for Search, device. AI Overviews and AI Mode aren’t shown separately, and there’s no query or click data (PPC Land, Search Engine Journal). Google also notes that sites with few impressions may not see the report at all.

The AI performance report in Bing Webmaster Tools, mentioned in Chapter 2, shows how many times your pages were cited in Copilot and Bing’s AI answers.

Both reports tell you which of your pages appear in AI answers. They don’t tell you which questions you appear for, whether you’re recommended, how you compare with competitors or whether you’re described correctly. They complement your own measurement rather than replacing it. We look at these reports in more detail in Chapter 8.

13. What a low rate points to

The benefit of keeping the five metrics separate is that a low rate leads you to the right problem. Here’s a rough mapping:

This table is a starting point, not a rule; the same symptom can have different causes. Still, a meeting that opens with “our visibility score dropped” doesn’t end up in the same place as one that opens with “our Recommendation Rate in AI Mode fell for category discovery questions, and competitor Y’s service network is being put forward.” In the second one, what to do next is more or less clear.

In Chapter 17 we turn the cycle of measuring, diagnosing, fixing and measuring again into a system that runs continuously.

14. By hand first, then with a tool

Don’t treat manual measurement and tools as alternatives. Use them in sequence.

Start by hand in the first month. Run the 30–50 prompts yourself a few times on each platform and read the answers. It takes time, but you learn a lot. Reading the answers is the only way to learn where to draw the line between mentioned and recommended, which reasons AI gives for putting competitors ahead, and what it gets wrong about your brand.

Once your coding rules have settled, you can hand the regular runs over to a tool. When you choose one, put the limits above to the vendor as questions.

Even after the tool takes over, keep reading part of the output by hand each month. Having a person read the answers to your 10 most important prompts is the most reliable way both to check that the tool is counting correctly and to track Representation Accuracy.

15. From visibility to business results

All of these measurements are an intermediate goal. What really matters is whether demand, sales and revenue go up when visibility does. In Chapter 1 I said that across our clients, AI traffic usually stays under 10% of total traffic but converts at around 6–7%. Connecting the visibility table to that kind of business data is the job of the next chapter.

If you take one habit away from this chapter, make it asking about visibility as five separate questions rather than one number. Answer each platform by platform, averaged over several runs; the rest is filling in the same table with the same rules every month.

Checklist

Sources

• Search Engine Journal: AI recommendations change with nearly every query (SparkToro) (January 30, 2026)
• Search Engine Journal: Google AI Mode and AI Overviews cite different URLs, per Ahrefs report (December 15, 2025)
• Ahrefs: Only 12% of AI cited URLs rank in Google’s top 10 for the original prompt (August 11, 2025)
• Genezio: Zero query overlaps between GPT-5.2 (API) and ChatGPT.com (February 4, 2026)
• PPC Land: Google finally gives Search Console its own generative AI visibility reports (June 2026; Google’s June 3, 2026 announcement)
• Search Engine Journal: Google Search Console AI reports rolled out worldwide (reports Google’s August 31, 2026 global rollout note)
• Koozai: Bing Webmaster Tools launches AI Performance reporting (February 13, 2026)

Previous: Chapter 6, From Keywords to Prompts · Table of contents · Next: Chapter 8, AI Traffic and Business Results