GEO KPIs: the seven numbers worth reporting, and the ones to ignore

What makes a GEO number worth reporting

It has to measure the answer, not the effort. Pages published, markup added and hours spent are inputs, and a report made of inputs cannot say whether anything changed for a buyer. A useful number is read from what the assistants actually said, on questions a buyer actually asks, taken the same way each month so that two months compare.

That last condition decides everything. Answers vary from run to run, so every number below has to come from a fixed panel read the same way each time. Our methodology fixes twelve questions, puts each once per reading to one assistant, Claude, and dates the reading, so the weight sits on the pattern across the twelve and, with monitoring, across months. Change a question and the series starts again. Read one answer as a verdict and you are reporting luck.

The seven, and how each is counted

In our reports the visibility score is the first of them alone, put as a number from 0 to 100: the share of the twelve questions in which the assistant named the business. Being named first and who else was named are reported beside it, not folded into it, and our reports do not record whether an answer recommended the business. The score is useful for a trend line. The rates beside it are useful for deciding what to do, because they move for different reasons.

Read the rates by question type, not only in total

A panel holds three kinds of question: the problem question a buyer asks before knowing any names, the comparison question, and the question that names you. A named rate that comes entirely from the questions that name you means the assistants know who you are and never suggest you to a stranger. The same rate spread across problem questions means the opposite. Report the rates per question type and the total, or the total will mislead.

Read them per assistant too. Each keeps its own index, and a business can be named by one and missing from another for reasons of freshness alone. The probe in our audit of 9,483 small business websites named 14.9% of 9,240 businesses on one Claude question each, and we report that as a Claude number for this reason. A single figure across four assistants hides the one that is failing.

Accuracy and access are the two most often left out

Accuracy is the number that protects the business. An assistant that names you with last year's phone number or a service you stopped offering is sending buyers to a wrong fact. Count the statements about you in each month's answers and mark each right or wrong; the wrong ones point straight at the page they came from. When AI gets facts about your business wrong covers what to do with them.

Access is the number that explains a sudden drop. A CDN setting changed, a security plugin updated, a robots.txt rule was added by someone who meant GPTBot and wrote something broader. In the audit, 7.2% of sites served our crawler a challenge page instead of the site. Check access every month with the crawler access check and a fetch of the homepage; a fall in every rate at once, across every assistant, is usually this.

The numbers to leave out of the report

• A single screenshot of one assistant naming you, or not. One answer is a coin flip, and a screenshot is one answer.
• The number of prompts "tracked". A long panel is not worth more for being long: readings compare only when the same questions are read the same way each time, and a panel that changes every month is not a series.
• Traffic labeled as AI referrals in analytics, taken alone. It counts only the buyers who clicked through; a buyer who read your name and phoned never appears in it. Worth watching beside the panel, never instead of it.
• Rankings in a traditional search engine, reported as if they were AI visibility. Google's AI features draw on its ranking, and the other assistants keep their own indexes; a ranking says nothing about whether ChatGPT named you.
• Any number with no baseline. A named rate is good news or bad news only against the month before.

What to set as a target

A direction, not a figure. We do not promise a named rate, and nobody outside the assistants can, because the choice is made inside systems no agency reaches. What a business can commit to is the inputs that remove the reasons it is not named: access in every month, the four site facts in place, the wrong statements corrected. The rates are then the evidence of whether that work reached the answers. A target that says "named on the problem questions by all four assistants" is a direction a panel can confirm; a target that says "named in half the answers by June" is a guess with a date on it.

Report monthly, on one page: the score and its trend, the three rates by question type and assistant, share of voice with the competitors named, accuracy with the wrong statements listed, access, and the four site facts as done or not done. Preparing GEO reporting for management sets out that page.