← All resources Guide 8 min read

How to start tracking GEO: a no-nonsense beginner's guide

Summary

Most teams either don't measure their AI search visibility or drown in dashboards that measure the wrong thing. Here's how to start in a week.

More and more buyers ask an AI first, then type your URL into the browser a few days later. ChatGPT now has more than 1 billion weekly active users. In B2B, 94% of buyers use AI somewhere in the process, and twice as many as a year earlier rated it more important than vendor websites or sales.

Your analytics won’t show most of it. The clicks are rare, and many land in “Direct.” Very few teams track this properly yet. The method isn’t complicated.

What you are measuring

GEO (generative engine optimization), also called AEO (answer engine optimization), is getting your brand named, cited and described correctly in AI answers. ChatGPT, Perplexity, Gemini, Google’s AI Overviews and AI Mode, Claude, Copilot, Grok.

GEO tracking is the measurement side: which questions you show up for, what the answer says about you, which sources it uses, and whether it turns into pipeline.

It overlaps with SEO, but it isn’t the same job. SEO reports rankings and clicks. An AI answer is a shortlist, often built from several searches, and it sends few clicks. Ranking for a keyword doesn’t tell you what that shortlist says.

One run is not a measurement

Three findings change the setup. Then we build it.

  • The same answer almost never comes back. SparkToro ran 12 prompts through ChatGPT, Claude, and Google’s AI 2,961 times. There was less than a 1-in-100 chance of the same brand list twice. One hospital appeared in 69 of 71 ChatGPT answers and came first in only 25. How often you appear is a real metric. A rank is not.
  • Other people’s sites do the talking. A University of Toronto study found that for U.S. software queries, 72.7% of AI citations were reviews, publishers, and other third parties, against 45.4% on Google. A passing mention at the end is not the same as being the answer.
  • Analytics undercount it. Pew found that when Google showed an AI summary, people clicked a result on 8% of visits, against 15% without one, and a link inside the summary on 1%. They ended the session on 26% of those pages, against 16%. Claude’s app sends no referrer, and Google counts AI Overviews inside ordinary search data, not on their own line.

So you need two measurements: what the AI says, and a way to catch the deals it sends you.

Fifty prompts, in clusters

Romana Kuts, founder of SaaStorm, laid out a practical framework at the Demand Day Summit in September 2026, and much of what follows builds on it. Her own numbers make the case for measuring at all: SaaStorm’s CRM showed 0 AI-sourced deals before March 2025, and 20 in a single month by July 2026. Her comment: “Not the product. Not the team. We started measuring it, so we could grow it.”

Where the 50 prompts come from

A prompt is the GEO version of a keyword: a question a real buyer might ask an AI. If your prompt list is wrong, every number built on it is noise.

Hundreds of prompts sound thorough, but half are ones no buyer would type. With around 50, every prompt maps to a real sales question and you can actually read the answers.

Where to find them:

  • Sales calls. The exact questions prospects ask in discovery. Your call recordings are the best prompt database you own.
  • Google’s People Also Ask. Already phrased as questions, already backed by search volume.
  • Reddit and communities. How people describe the problem when no vendor is listening.
  • Support tickets. Post-sale questions about integrations, limits and pricing make good pre-sale prompts.
  • Comparisons. “X vs Y,” “alternatives to X,” “is X worth it.” This is where the narrative gets decided.

Don’t just turn your SEO keyword list into questions. Exporting terms from an SEO keyword tool and asking an LLM to rewrite them is, in Romana’s words, “the biggest mistake you can make.”

Write prompts the way a buyer types. Not “partner relationship management software,” but “best PRM for a 10-person partner team on HubSpot.” Long questions are also the ones that trigger AI answers. In Pew’s data, only 8% of one- or two-word Google searches produced an AI summary, against 53% of searches with 10 or more words.

A good shortcut: the Google Search Console queries that already bring you traffic show real demand, and in ChatBakers you can turn them into tracked prompts directly.

Group them into clusters

Put the 50 prompts into 5 to 10 clusters, usually one per use case, feature area or buyer problem. Each cluster gets an owner, a target and a content plan.

Clusters show what a long list hides. In an example Romana shared, an anonymised seed-stage sales-tech company had a 44% mention rate on HubSpot-partner prompts and 2% on partnership-strategy prompts. As she put it: “That is a content brief, not a dashboard.” Start narrow, then add clusters.

What to write down

A tracker that repeats each prompt

You can’t check six AI engines by hand every week. We build a tracker, so we’re biased, but the prompt list matters more than the tool. Spend a week on your prompts, not a quarter comparing vendors. Check that your tracker:

  • Covers the engines your buyers use: ChatGPT, Perplexity, Gemini, Copilot, Claude, Grok, and Google AI Overviews and AI Mode. Google says the two can use different models, so their answers and links differ.
  • Runs each prompt several times and reports how often you appear, not a single “rank.”
  • Shows the full answer text and every cited URL.
  • Exports cleanly, so you can tie clusters to pipeline.

Run them on a schedule, several times per prompt. The same list of brands came back less than once in 100 runs, so one check tells you almost nothing.

Three numbers, not a dashboard

Keep three:

  • Mention rate per cluster. How often you appear across repeated runs of your 50 prompts.
  • Citation share. Which sources the AI cites: you, competitors, or third parties.
  • Narrative and position. What the answer says you are, and who you’re listed next to.

Treat the rest as context. Romana calls them “the new pageviews.” Total mentions rise when you add prompts. Sentiment rarely moves in B2B. AI referral sessions undercount by design. Share of voice across all prompts averages away the clusters that matter. And prompts tracked is “a vanity metric with a price tag.”

Narrative matters most. “An attribution dashboard for small SaaS teams” and “a revenue attribution platform for mid-market SaaS” both count as a mention. Only the second gets the right buyer to book a call. As Romana said: “It doesn’t matter how much you are cited if you are cited wrong.”

Read the answers yourself

Once a month, read about 10 answers per cluster and score each cluster red, amber or green on three questions:

  • Right category? Does the model call you what you call yourself?
  • Right comparison set? Are you next to the competitors you want, or positioned as the budget option?
  • Right use case? What does it say you’re best for, and not for? The “not for” line is the one that costs pipeline, and often the easiest to fix.

A quick first check: ask each engine “What is [your company]?” and “What are its main competitors?” If the competitors are wrong, you probably don’t show up when buyers ask for alternatives to them either.

Pull the deals out of Direct

A buyer asks ChatGPT, checks in Claude, types your name into Chrome days later and books a demo. Your CRM calls it Direct. Four non-technical fixes:

  • Referrer sub-sources. Where a referrer exists (chatgpt.com, claude.ai, perplexity.ai, Gemini, Copilot), map it to an Inbound sub-source on the contact and deal.
  • Ask on every form. Add a required “How did you hear about us?” field that lists AI tools by name, above Google.
  • Sales confirms on the call. Ask early in discovery: “What did you search or ask before you found us?” The rep sets the sub-source.
  • Mine call recordings. Search your transcripts for “ChatGPT,” “Claude” and “Perplexity.”

Some AI referrers do show up. Native apps often pass none, and Google folds AI Overviews and AI Mode into Search Console’s “Web” data. The form and the question on the call fill that gap.

Change what the answers are citing

  • Third-party citations. The Toronto study’s earned-media bias showed up in Romana’s client data too: 73% of citations came from sites the company doesn’t control. Brief those sources, because “citations carry narrative.”
  • Deeper product pages. Who it’s for and not for, integrations, limits, pricing logic, honest comparisons. Don’t expect magic from markup: Google says no special schema is needed for AI Overviews or AI Mode.
  • Evidence and real people. Named authors and original data help. Romana’s line: “Generated content raised the floor. Human content is now the ceiling.”

The first week

  • 50 prompts from sales calls, People Also Ask, Reddit, support tickets and comparisons
  • 5 to 10 clusters, each with an owner
  • A tracker that runs prompts repeatedly and shows full answers and cited URLs
  • 20 to 30 minutes a month to read answers and fill in the red/amber/green sheet
  • AI referrer sub-sources in your CRM, and an AI-tool option on every form
  • “What did you search or ask before you found us?” in discovery calls

The tracking itself isn’t hard: fifty good prompts, three metrics, twenty minutes of reading a month, and four changes to your CRM and sales process. Most teams skip the reading and the attribution, and that’s where the pipeline shows up.

We built ChatBakers to watch how a brand shows up across ChatGPT, Perplexity, Gemini, Google AI Overviews, Claude, Copilot and Grok: which prompts name you, what the answer says, and which sources it cited. Pick your 50 prompts before you change anything. Measure it.

Sources

Monitor your AI visibility

See how AI models talk about your brand. Start your free 7-day trial.

Start Free Trial →