GEO metrics: the KPIs that prove results and the ones that mislead

The GEO KPIs that prove results to a client, in three layers: presence, traffic and business. What each one measures, where it comes from, how often to read it, its margin of error and which metrics mislead.
Written with AI, reviewed by humans

In this article

See how your brand shows up in AI answers

Book a demo and find out what to fix first, with a GEO specialist.

Book a demo

Reading time: 12 minutes

The GEO performance metrics that prove results to a client are presence, citation, recommendation, share of voice and AI traffic. Presence is measured by the same prompts every month. By contrast, the GEO KPIs that mislead are AI position, prompts run once, a single screenshot and AI traffic read alone.

In short

  • GEO metrics that prove results form three layers: presence in answers, traffic from AI and business.
  • GEO KPIs for presence come from fixed prompts, run 5 times a month in each engine. With 1,000 answers a month, the margin runs from ± 3 to ± 7 points.
  • AI position, a single screenshot and prompts run once mislead, because the list of brands changes almost every run.
  • AI traffic is not enough as the only KPI. After all, GA4 puts AI Overview clicks in organic search, and some visits arrive with no source.

Which GEO metrics show results?

The GEO performance metrics that show results answer three questions. Does the brand appear in the answers? Does the answer send visits? And do those visits become business? Each layer has its own source. So the report carries few numbers, always with the same source, the same frequency and the baseline beside them. By our criteria, the frequency is monthly for all of them.

KPIWhat it measuresSource and frequency
Presence rateAnswers that name the brand ÷ totalFixed prompts, monthly
Citation rateAnswers that cite the site as a sourceFixed prompts, monthly
Recommendation rateAnswers that suggest the brand as an optionFixed prompts, monthly
Share of voiceBrand mentions ÷ mentions of the monitored brandsFixed prompts, monthly
Sentiment and accuracyTone and errors in answers about the brandPrompts with the brand, monthly
AI impressionsSite links in AI Overviews and AI ModeSearch Console, monthly
Citations in Copilot and BingTimes the site became a source in AI answersBing Webmaster Tools, monthly
AI sessions and leadsVisits and key events from assistantsGA4, monthly
Brand searchesClicks on queries with the brand nameSearch Console, monthly
Sales that cite AIDeals with a stated AI originForm and CRM, monthly

Where do the presence numbers come from?

For the presence KPIs, the prompts are the sample, so the list defines the result. The first five metrics come from the collection in the protocol to measure brand mentions and citations in AI. That means 40 to 60 prompts without the brand, frozen for a quarter, 5 runs per engine over the month and 3 to 5 competitors.

Share of voice has its own choices, as the guide on how to calculate share of voice in AI shows. The slice of the answers that belongs to the brand and its competitors also has its own name. See how to measure share of model. By contrast, the ready-made score in each tool uses a different formula, as the guide on how to read the AI visibility score explains.

What is the margin of error of GEO metrics?

Among GEO performance metrics, only the ones that come from prompts have a sampling margin of error, because each month measures part of the possible answers. GA4, Search Console and Bing count events, but they also have limits, such as delay and visits with no source. The table is our own calculation, using the method from the measurement protocol, at 95% confidence.

Answers per monthCollectionMargin of error
3010 prompts, 3 runs, 1 engine± 18 to ± 31 points
25050 prompts, 5 runs, 1 engine± 6 to ± 14 points
1,00050 prompts, 5 runs, 4 engines± 3 to ± 7 points

The real margin falls inside the range, because repeated runs of the same prompt resemble each other. Since both months carry error, the gap between them must exceed about 1.4 times the margin. So, with 1,000 answers a month, treat only a change above 4 to 10 points as real, and confirm it the next month.

According to Google Analytics Help, GA4 processing can take 24 to 48 hours. Also, the attribution credit of key events can change up to 12 days after registration. So, by our criteria, close a month’s GA4 numbers only after the 12th day of the following month.

GEO performance metrics: illustration of a KPI dashboard with donut, bar and line charts

How do you link GEO metrics to traffic, leads and sales?

Link the layers in sequence: presence, visit, lead and sale, always in the same period. Presence responds first, while traffic and sales show only the part that left a trace. So the report places the GEO performance metrics side by side, without adding the numbers, and explains where the trace gets lost.

  • Visits: since May 2026, GA4 has the AI Assistant channel, with sources such as ChatGPT, Gemini and Copilot, according to the channel definitions. Clicks from AI Overviews and AI Mode, however, stay in Organic Search. To combine the slices, create a custom channel group.
  • Impressions on Google: the generative AI performance report in Search Console shows the site’s impressions in AI Overviews and AI Mode. That data already sits in the Web type of the performance report, so don’t add the two. See what Search Console shows about AI Mode.
  • Leads: mark the form or the purchase as a key event in GA4. Also, add the question how did you hear about us to the form, with an “AI recommendation” option.
  • Sales: match each lead with the CRM, because the closed revenue lives there.

Where does a sale from AI disappear in the report?

A sale disappears at two points on the path. First, Cloudflare noted in 2025 that traffic from Claude’s native app arrived without the referrer header. Without it, analytics tends to count the visit as direct. Second, no GA4 attribution model gives credit to direct visits, unless the whole path is direct, according to Google Analytics Help.

So someone who met the brand in ChatGPT and came back through Google becomes an organic search sale. How to credit that path is covered in our guide to the attribution model in GA4. This blind spot isn’t only in your account. In the US edition of a study by PYMNTS Intelligence, only 24% of 584 US merchants could tell when a sale came from an AI platform. The survey ran in March 2026. Visa Acceptance Solutions, which commissioned the study, sells payment solutions.

For that reason, read brand searches and the origin stated on the form as complementary signals. The Search Console brand filter uses AI and can be wrong. Neither signal proves cause, but both show what the click doesn’t show.

Which GEO metrics mislead?

GEO performance metrics mislead when they change while visibility doesn’t, such as position in the answer, prompts run once and the screenshot. Averages that mix prompts with and without the brand also mislead. Likewise, numbers from a single run look precise but measure chance. Finally, AI traffic misleads when it appears alone.

  • AI position: the order of the brands changes almost every run. In the SparkToro research, 600 volunteers ran 12 prompts 2,961 times in ChatGPT, Claude and Google’s AI. The chance of ChatGPT or Google’s AI repeating the same list of brands was below 1 in 100. Also, the chance of repeating the same order was below 1 in 1,000.
  • Prompts without repetition: “the brand appeared in 30 of 50 prompts”, with one run each, measures a day, not a trend. So the GEO KPIs use 5 runs per prompt in each engine.
  • Single screenshot: a saved answer shows only that run. Likewise, the sources rotate. According to SISTRIX, over 17 weeks and six countries (Germany, the US, the UK, Italy, Spain and France), 56% of the domains cited by AI Mode were new each week.
  • Prompts with the brand name: they tend to produce answers with the brand and, so, inflate presence. Read that group separately.
  • Average position in Search Console: an AI Overview takes a single position, and all the links in it get that position, per Search Console Help.

Rand Fishkin, co-founder of SparkToro, changed his mind about what counts as a metric after the study:

“I’ve been swayed from my initial position and now believe visibility % across dozens to hundreds of prompts run multiple times is a reasonable metric.”

Rand Fishkin, SparkToro (2026)

Why is AI traffic not enough as the only KPI?

Because traffic shows only the part of presence that became a click and left a trace. An answer can settle the question with no visit, while AI Overview clicks fall into organic search and apps arrive with no source. Likewise, according to Bing Help, a citation in AI Performance doesn’t represent traffic, clicks or engagement.

So flat AI traffic doesn’t prove that presence stopped. Use it as one of the GEO performance metrics, next to presence and sales, and never as the sole proof.

What are the baseline and a reasonable timeline for GEO KPIs?

By our criteria, the baseline comes from one to three months of collection before the first actions, with the same prompts, engines and competitors. Then compare month by month. Also by our criteria, the first review of the GEO KPIs comes at the end of the first quarter, because a rise counts only after it repeats.

Not all data goes back in time. So connect the sources on day one of the contract:

  • GA4: marking an event as a key event affects reports only from that moment and doesn’t change historical data, according to Google Analytics Help. The custom channel group, however, also works for past data.
  • Bing Webmaster Tools: AI Performance shows only the dashboard’s time ranges, from 7 days to 3 months. So export the data every month.
  • Prompts: GEO collection doesn’t go back in time either. If the project started with no baseline, use the first measured month and say so in the report.

As for timing, citations rotate within weeks. Scrunch, which sells AI monitoring, analyzed 3.5 million citation events across AI platforms with Stacker. On average, citation activity from sources that appeared in the same week fell by half in about 4.5 weeks.

So presence on the GEO prompts can swing from one month to the next. Therefore, track the GEO performance metrics every month and confirm any rise the next month.

How do you build the monthly GEO metrics report?

By our criteria, the monthly report fits on one page of numbers and one of commentary. Each of the GEO KPIs appears with the month’s value, the baseline, the margin of error and the source. Then come the likely causes and next month’s actions, each with an owner and a deadline.

  1. Summary: what changed, whether it passed the margin and what changes next month, in three sentences.
  2. Business: leads and sales that cite AI, brand searches and sessions from the AI channel.
  3. Presence: presence, citation, recommendation and share of voice, by engine and with the margin.
  4. Google and Microsoft: AI impressions in Search Console and citations in Bing, not added to traffic.
  5. Reputation: sentiment and errors found, with the source the AI cited.
  6. Method: tool, prompt list, engines, country, period and any change in collection.

The AI block fits inside the SEO report template for clients. To bring GA4, Search Console and the prompt spreadsheet onto one screen, build an SEO dashboard in Looker Studio, now called Data Studio.

The method block looks like a detail, but it supports the GEO KPIs and the prompts that produce them. After all, without it nobody can redo the math next month, and a competitor swap looks like a change in the AI. Likewise, keep the list of GEO prompts and the full answers, to audit any number the client questions.

If the client asks for a single score, show the tool’s score with its formula beside it, not in place of the rates. Finally, open the report with the business, because the board decides by revenue, not by presence.

Frequently asked questions

How many GEO KPIs should you show a client?

By our criteria, five to eight GEO KPIs in the summary. For example, presence, citation, share of voice, AI sessions, brand searches and leads or sales that cite AI. The other numbers, such as presence by prompt and sentiment by engine, go on the detail pages, because they explain the main ones.

Can you measure GEO metrics without a paid tool?

Yes, while the volume is small. Manual prompt collection, GA4, Search Console and Bing Webmaster Tools cost nothing. Above a few hundred answers a month, however, the collection of GEO performance metrics usually moves to a tool or an in-house script.

Is the AI visibility score the same as the presence rate?

Not always, because each tool calculates its own score. According to each tool’s documentation, Peec AI divides the answers with the brand by the total answers. Profound, by contrast, divides by the answers that mention at least one brand. So with 10 answers, the brand in 4 and 2 with no brand at all, the math gives 40% in one tool and 50% in the other.

Want to know how your brand shows up in AI answers today?

Book a demo and see how your brand shows up in AI answers, with support from a GEO specialist.

Main sources

About this articleWritten with AI and reviewed by humans, with verified sources. Found something out of date? Write to the NotHuman team.

Get new NotHuman articles by email

By subscribing, you agree to receive blog articles by email and to the Privacy Policy. Unsubscribe any time.

Keep reading

Demo

Want to know how your brand shows up in AI answers today?

Book a demo and see how your brand shows up in AI answers, with support from a GEO specialist.

Book a demo

Your privacy

We use necessary cookies to make the site work and Google Analytics and Microsoft Clarity statistics cookies to improve our content. You can decline the statistics ones; in the EU, UK and Switzerland they are only turned on after you accept. Cookie Policy.

Cookie preferences

Cookie Policy