The first question most SEO managers (or AEO or GEO managers) ask when they set up their prompt tracking tool usually comes from a very good place:

"Why these prompts?"

"Where did they come from?"

"Which should I prioritize so that my credit spend doesn't get out of hand?"

Because they know by now that a single answer from any AI model is one possibility from a system that gives a different answer next time a local customer asks.

The questions are valid. When our GEO/AEO analyst Katya Shishchenko measured how 5 AI models recommend local businesses — 9 to 11 prompts per vertical, 50 to 100 runs each, more than 120,000 mentions across 3,700 businesses — she confirmed that only 5.5% of businesses mentioned by any model appear in all 5 models' answers to the same prompt. And nearly 2 in 3 businesses cited by any AI tool show up in only one model's recommendations, which means the other 4 models don't mention them at all.

So how do you, as the person running search for your brand, know which prompts to track, how many you need, and which to prioritize — especially when tracking thousands would give you way better data but costs more credits?

The prompt set you build gives you the AI visibility numbers you report to leadership; it informs your content strategy across locations. But if it's not done efficiently, it's going to cost money and deliver metrics you'll be nervous to share with your team.

If you're still at the stage of finding prompts, we've covered how prompt research works in GEO Studio separately. This is about how to track AI visibility properly — turning your list of prompt ideas into a solid set that gives you and leadership the insights they need per location. And it's about knowing what that set can't tell you even when you've built it well.

4 Rules on The Prompts You Should Track

This is where you start in GEO Studio. The Prompt Universe pulls candidate prompts from sources like Google Search Console, Reddit, YouTube, and competitor domains, and organizes them so you can see what people are asking AI models about your business category — whether that's healthcare, retail, hospitality, or services.

It will admittedly hand you more prompt candidates than you should run at once — our team suggests these 4 rules when you build out your prompt set.

1. Keep Branded and Unbranded Prompts Separate

"Is [brand] good for X?" tells you what AI says about you once a customer already knows you. "Best X near [city]" tells you whether AI brings you up at all.

"If unbranded prompts measure competitive visibility, branded ones help us understand brand recognition, business-entity accuracy, reputation, and general sentiment, and whether AI recommends other brands when yours is already mentioned in the prompt. It's like a temperature check for your brand." – Katya Shishchenko

Measure both branded and unbranded prompts, but remember that branded prompts will always give you a high mention rate, so blending them into one average dashboard or report will hide some of your biggest unbranded opportunities.

It's tempting to keep the prompts that get you cited consistently and signal brand authority, but don't let them skew any overall impression of your AI visibility. You'll need to know when your multi-location brand starts losing visibility for unbranded queries — volatility is more common there.

2. Cover the Intents That Lead to a Visit, Not Just Discovery

Intent coverage shows where your gaps are and where to focus content.

Based on her research — a separate study from the one I mentioned above — Katya Shishchenko, our AEO/GEO Analyst at Uberall recommends tracking:

• 7 to 11 types of intent, with
• 5 to 12 phrasing variations each – 5 phrasings per intent is enough to see whether your visibility is moving, and 12 shows you the full set of businesses AI recommends
• across at least 2 AI models (ChatGPT and Gemini)

That works out to roughly 35 to 130 prompts per market.

She ran 182,430 prompts through 3 AI models: GPT-4o, Claude 3.5 Sonnet, and Gemini 1.5 Pro. These prompts covered 16 types of local search intent across 5 industries — dining, auto repair, legal, veterinary, and hair salons — and 4 US cities. Each prompt ran 10 to 30 times to account for how AI answers vary from run to run.

The intents you track decide which competitors you see. Urgent, specialist, and comparison prompts, such as "affordable dentist in Chicago" and "emergency dentist open Sunday near Wicker Park," will usually favor national chains. These are closer to driving on-site visits. Meanwhile, "best," "highly rated" and "where do locals go" favor smaller, more independent local businesses and are the kind of prompts that will help your brand be discovered.

In Katya's monitoring framework study, urgent prompts surfaced national chains 20% of the time, while "highly rated" and "best" prompts surfaced chains less than 1.5% of the time.

You can also save credits by merging overlapping queries like "best," "most trusted," and "near me." These, from Katya's research, return largely the same businesses in AI responses. These were the only 2 of the 16 intents we tested that overlapped heavily.

3. Put Neighborhoods and Landmarks Inside the Prompt Text

GEO Studio has two separate fields when you set up tracking:

1. a location (where your business sits)
2. the prompt text (what the customer asks).

Our Technical Product Manager Luma Ghazaleh explains: "Neighborhoods, districts, and local terms like 'SoHo' or 'Downtown Austin' represent search intent modifiers rather than physical business entities — treat them as prompt variations, not standalone locations."

In practice, that means entering the following in GEO Studio:

• Location field: Chicago (the city-level market your store operates in)
• Prompt text: "dentist near Wicker Park" (not "dentist in Chicago")

Tracking only a city-level prompt like "dentist in Chicago" is too broad to show you where your brand wins or loses within that city. When queries are more neighborhood-specific (Wicker Park, not Chicago), AI models will vary their responses as they weigh local reviews, proximity, and relevance at the neighborhood level.

Track those neighborhood variations as separate prompts, all with "Chicago" as the location, and you can see how AI models recommend your multi-location brand in each part of the city — and where a competitor is winning a neighborhood you thought you had.

4. Test Wording Before You Commit Credits

Query fan-out shows you the sub-questions a model asks itself before it compiles an answer or local recommendation. You can use it in two ways to build a better prompt set in your prompt tracking tool.

Cut redundant prompts. Before you commit credits to two similar-sounding prompts, run fan-out on both by asking prompt variations in your own AI model of choice. If "affordable dentist in Chicago" and "cheap dentist in Chicago" seem to break down into the same sub-questions and recommend or mention the same brands, drop one and save your credits.

If they fan out differently — maybe "affordable" pulls in insurance-related sub-questions while "cheap" pulls in walk-in pricing — they're testing different parts of your visibility, and you should track both if you want to be recommended in both scenarios.

A specific, multi-detail prompt will often fan out into more sub-questions than a generic one — test both to see. Even if there may not be much search volume for specific, detailed prompts, you're more likely to cover the sub-questions local customers ask before choosing which business to visit.

For example, "Where can I get running shoes for flat feet near me that's open Sunday?" might break into:

• "best running shoes for flat feet"
• "running shoe stores near [user's location]"
• "stores offering gait analysis near [location]"
• "[store name] Sunday opening hours"
• "[store name] reviews flat feet fitting"
• "stability vs motion control shoes"

That's 6 sub-questions from one detailed "parent" prompt. A generic prompt like "running shoe stores near me" might only fan out into two. When you're choosing between similar prompts, run fan-out on both and pick the one that gives you richer search and visibility insights.

How Many Prompts, How Often, and on Which Models?

A small set of prompts gives you false positives and false negatives.

For example, say you track 10 prompts across 3 models (ChatGPT, Gemini, Perplexity), once a week. That's 30 answers. Your brand appears in 9, so your Mention Rate reads 30%.

Next week you appear in 6, because of how AI models vary their answers, making your Mention Rate 20% – even though your content, prompts, and competitors have stayed the same. That's a false negative because your visibility hasn't necessarily fallen; 30 answers is just too few to account for that variation.

Or you risk introducing false positives into your prompt tracking reports. Say you publish 3 new store location pages, appear in 12 of 30 answers the following week, and report a 40% Mention Rate to your CMO as proof the content update worked.

Actually, the jump from 30% to 40% is just 6 more mentions out of 30 answers — well within the range AI models vary on their own between runs. That's a false positive: The content may or may not have helped your AI visibility, but this sample was too small to confirm that finding.

{{quote}}

You should make sure you're hitting the Goldilocks amount of data to make robust analyses about your brand's AI visibility and take the right actions from them.

This is our advice for GEO Studio.

Prompts

Track more than feels necessary. If your budget means you need to be cautious with your credit spend, fewer prompts run more often beats more prompts run rarely.

Track intent coverage alongside mentions. A brand mentioned 100 times across 2 intents is visible to fewer customers than one mentioned 50 times across 8. For example, if an Italian franchise establishment shows up in "affordable" and "where do locals go" prompts but not in "best" or attribute-specific ones ("vegan options"), AI models might perceive that restaurant brand as a known local option, but not as a quality leader or a specialist.

That usually means the restaurant needs more high-quality reviews and more detailed service attributes on its Google Business Profile so that AI models have the signals to recommend it for those higher-intent prompts.

Frequency

You can set up daily, weekly, or monthly streaming in GEO Studio. Pick one and keep it. Rationing runs to save credits might cause gaps in your data, where a drop in visibility is actually due to a drop in data. Not what you want to be reporting.

In a separate monitoring study, Katya found that weekly and daily tracking surfaced the same brands with very similar mention rates and rankings. Weekly is therefore enough for most setups.

Models

Each engine builds answers from different sources. GEO Studio tracks ChatGPT, Gemini, Perplexity, Grok, Copilot, DeepSeek, and Google AI Overviews, and a visibility decline on one model with no decline on the others is more likely a model update than a loss of visibility. Track all of them and compare how your brand does.

In Katya's research, ChatGPT and Claude agreed strongly on which businesses to recommend, but both differed a lot from Gemini — so if you're only tracking 2 models, make sure Gemini is one of them.

Before You Report That GEO Metric

Only 5.5% of businesses show up across all 5 AI models for the same prompt, and in verticals like dental it's even less than that.

You need to track local visibility, not just brand visibility with the right prompt-tracking setup. A national average can look healthy while individual locations are invisible.

If your prompt set is small, your frequency is inconsistent, or you're only tracking 2 or 3 models, the Mention Rate you hand to leadership is measuring how the models happened to answer that week — that's not robust.

Measure branded and unbranded prompts separately, intents that lead to a visit rather than just discovery, neighborhoods and landmarks inside the prompt text, and keep to a streaming schedule.

GEO Studio gives you the prompts, the models, and the response-level detail behind every rate — how much you can trust the number depends on the set you choose to run. Our experts can walk you through your setup.