How to measure GEO results

Success in generative engine optimization is not one number. A buyer who asks an AI tool a question meets your brand in four steps, and each step is counted in a different place. Measure all four, because each can move without the others.

Mentions and citations together make up your AI search visibility. An AI citation is its stronger form, because it names you and gives the reader a way to reach you. Visits and enquiries are what the business feels. A report that stops at visibility describes the shelf, not the sale; a report that counts only AI referral visits misses every buyer who read an answer and came back another way.

A result needs a starting point, so the prompt set is written, and run once, before any GEO work begins. It is the unit of everything that follows, and the one thing that must not drift.

1. 1Collect the questions buyers already ask: sales calls, support tickets and chat logs first, then the questions in your Search Console queries. In Bing Webmaster Tools, grounding queries show the phrases its AI used when it retrieved pages that were then cited^[1].
2. 2Word each prompt the way a buyer would, as a full question, and keep the wording fixed. A prompt whose words change is a new prompt, with no history.
3. 3Split the set by stage (questions that learn, compare and buy) and, for a business that sells to both, consumer prompts apart from business ones.
4. 4Keep brand prompts apart from the rest. A prompt that names your brand tells you how you are described; one that does not tells you whether you are found at all.
5. 5Decide, before the work starts, what counts as success: which enquiries and sales count, and how each will be traced back to search. A rule chosen after the results is not a rule.
6. 6Run the whole set once and store every answer, dated, with its sources. That is the baseline.

Size follows the decision the numbers have to support. A small business can learn a great deal from 10 to 20 questions checked by hand each month in a spreadsheet. A brand that sells in several markets needs a set per market and language, and a tool to ask it.

Answers vary from one run to the next and from one tool to another. Google notes that AI Overviews and AI Mode may use different models and techniques, so the responses and links they show vary^[2]. Measurement has to hold everything else still:

• The same prompts, in the same wording, on the same day of each month.
• A new chat for every prompt, signed out or in a clean profile, so that earlier questions do not steer the answer.
• Two or three runs of each prompt, because one answer is an anecdote.
• The same location and language each time, set to the market you sell in.
• The full answer saved with its sources, not only a score.

Read the matrix by row as well as by total. In the sample, the prompts about cost and security are where the brand is weakest, and both are questions a pricing page and a security page should answer in their first lines. A total hides that; a row points at the page to fix.

Then turn the rows into rates: the mention rate, the share of answers that name you; the citation rate, the share that link your pages; and, where you track competitors, your share of all the brands named. Keep the counts beside the rates. Three answers out of eight and thirty out of eighty give the same rate, and they are not the same evidence.

Your prompt set is a sample you control. The engines’ reports are counts that you cannot bend. Use both, and read each for what it is.

Two details change how you read them. Bing defines citation share as your site’s share of all the citations shown for a grounding query, and calls it an observational metric, not a ranking system or a competitive scoreboard^[5]. And ChatGPT search can cite only a site that OAI-SearchBot may crawl: OpenAI says that sites opted out of it will not be shown in ChatGPT search answers^[7]. Before you read a zero as a result, check that the crawler was let in; GPTBot, OAI-SearchBot and robots.txt shows how.

Visits are where GEO meets the numbers a finance team already trusts. In GA4, open the AI Assistant channel^[6] and read it by landing page: the pages that receive visits from AI assistants are pages being cited, and their conversions are the first revenue you can tie to AI answers by a stated rule. Google’s AI Overviews and AI Mode sit outside that channel: a click on a link in either counts as a click in Search Console, with the rest of Google Search^[4].

Google says that when people click from results pages with AI Overviews, the clicks are higher quality, meaning visitors are more likely to spend more time on the site^[2]. Treat that as a claim to check on your own data: compare the engagement and conversions of AI Assistant visits with those of organic search visits, page by page, over a full quarter.

The largest gap is the mention with no click. A buyer reads an answer that names you, then searches your brand or types your address days later, and the visit is counted as brand search or direct traffic, not as AI. Two habits recover part of it: ask new customers where they first heard of you, with AI tools as one of the answers they can pick, and watch brand impressions in Search Console beside your mention rate.

Label revenue tied to AI answers as the rest of the report does: known, influenced, modeled or incremental, and GA4-reported where the figure comes only from analytics, each beside its figure and never added together. The guide to SEO reporting in revenue defines each label.

Citation counts move for reasons that have nothing to do with your work. Bing lists changes in user behaviour, evolving models, freshness signals, partner refresh cycles and broader changes across the web^[5]. A change in the numbers is not a result until you can name its likely cause.

• Compare like with like: the same prompt set, over periods of the same length. Bing’s Compare view overlays an earlier period, such as the prior 30 days, for this reason^[5].
• Check whether the prompt set changed. Adding or removing prompts moves every rate, whatever happened to the answers.
• Check the season. A dip in a slow month is read against the same month a year earlier.
• Flag a small base. When mentions move from 4 to 3, say that the numbers are small instead of reporting a fall.
• Read the measures together. Fewer mentions with more citations is a different story from fewer of both.

Report GEO next to classic search, not in a separate deck: the same buyers, the same pages and the same months. One page is enough if it answers five questions, in this order.

1. 1What changed: one headline sentence, such as which prompts you gained or lost, and on which tools.
2. 2The numbers: mention rate, citation rate, AI referral visits and the enquiries or revenue from them, each for this period, the last period and the baseline, with the count beside every rate and the source of every figure named.
3. 3Why: the likely cause of each change, and the changes that have no cause you can name yet, said as such.
4. 4What it cannot show: mentions without clicks, engines you do not track, and small bases.
5. 5What happens next: three or four actions, each tied to the prompts it should move, and the date the set will be run again.

Keep the full prompt-level sheet linked, not pasted. A reader who doubts a number should be able to open the answer it came from.

• Changing the prompt set mid-period, then comparing the new rate with the old one.
• Mixing brand prompts with discovery prompts, which flatters the mention rate.
• Reading one run of one tool as a result.
• Trusting a score you cannot trace back to answers. Google warns that no third-party tool has access to its internal ranking or AI systems^[8].
• Calling a citation a sale. A citation that sends no visits and no enquiries is visibility, and is reported as visibility.

For the work that moves these numbers, start with the practical guide to generative engine optimization; for the checks to run before the first baseline, use the GEO audit checklist.