Why GEO tracking should start from intents, and how to build them

In AI search, nobody outside the platforms can see what consumers ask. You can only draw a sample of what generative engines say about a topic, and that sample decides everything you see, from the analysis to the actions that follow. In GEO, the right unit for that sample is strategic intent. An intent is a buying need: who is asking, what they want and where they are in the decision. Here's why, and how I build them.

1. The data black box: GEO has no Search Console

In traditional SEO, professionals work on a solid foundation of data. Google Search Console shows real impressions, clicks and queries. Keyword research tools and trend data estimate demand by keyword, geography and time.

Like it or not, in AI search the major AI platforms do not share their query logs. Any “prompt volume” figure you see today is an indirect estimate, and different methods produce very different numbers for the same topic.

And not because of a policy choice. It all comes down to seven reasons:

1. Prompts are long and conversational: users describe situations, constraints and preferences in full sentences instead of typing short keywords. No two prompts are identical, so there is no standard “query” to count.
2. One prompt, many topics: a single request can combine several needs, such as planning, comparing and asking for advice. Assigning a prompt to a single topic is arbitrary, and summing topics inflates volumes.
3. Conversations, not queries: many prompts are follow-ups inside an ongoing conversation and only make sense in context. The same words can express different needs depending on what came before.
4. Personalized answers: responses adapt to the user’s history, memory and preferences. Even the same prompt does not produce a single, observable “result” to track.
5. Fragmented access points: AI is used through browsers, desktop and mobile apps, APIs, third-party products and messaging chatbots. Most of these channels are invisible to any external observer.
6. Use goes far beyond search: people also use AI to write, plan, analyze and create. Separating “search-like” demand from everything else is itself an estimate.
7. Demand spread across platforms: users split their activity across several assistants (mainly OpenAI and Gemini, but also Perplexity, Claude, Grok etc), unlike traditional search where Google dominates.

2. So visibility has to be sampled

If you can't observe real conversations, you have to simulate them. You write the questions your customer would ask, then observe which brands get mentioned, in what position and with what story. That comes with three rules:

• Every visibility metric is conditional on the set of simulated questions. A 30% Share of Voice is only true “for these questions”, never in absolute terms.
• The quality of the simulation depends on how realistic the questions are. Someone might search "best vegan restaurants Milan". But on Chatgpt, the same person would ask: "I'm taking my parents to Milan next weekend. One of them is vegan and the other doesn't really like 'healthy' food. Where can we have a relaxed dinner near Brera without spending more than €60 a head?" That framing changes which brands get recommended.
• The simulation must be repeated. Generative answers are not deterministic and answers change over time. Plus, the engines disagree with each other. A brand can be named often on one engine and not at all on another, at the same moment.

So choosing the questions is a design decision. It defines which portion of the market you see, which competitors you're compared against and which gaps you act on. It even decides whether a rise in share of voice is real progress or just a change in the prompt mix.

In statistics, a sample is useful when it represents the population. In GEO the population can't be observed, so representativeness has to be built by design. That's what intents are for.

3. The limits of a bottom-up, prompt-by-prompt approach

The most common approach of many of the tools I analyzed copies the logic of SEO rank tracking. You define a list of prompts, often derived from SEO keywords, and then the tool runs them periodically. It is a familiar model, and that is exactly its weakness. Here is why:

• Author bias. The prompt set might reflect the internal language of the brand or its own convictions, not the customer's behaviour, and tends to overrepresent the search territories where the brand is already strong.
• Small lexical changes can alter the answer significantly. Without a structure that aggregates variants, the mentions fluctuates for linguistic reasons rather than competitive ones.
• Partial market coverage. Without an upstream map, there is no way to know when the set of prompt is omni-comprehensive or which parts of the market are missing.
• Weak comparisons over time. Adding or removing prompts changes the baseline, making trends hard to interpret.

4. What a GEO intent is

A GEO intent is a broad purchase intention of the consumer, which also resonates with the specific goals of the the brand (commercial goals, marketing goals etc.). Prompts become semantic variations of it: different phrasings of the same question or different angles. Together, they estimate how likely the brand is to appear when that need arises.

A well-designed intent answers three questions.

Who is asking? The same need changes with the profile. A first-time luxury buyer and a seasoned collector get different answers, citing different brands.

What are they looking for? The need, problem or category defines the search territory, and so the competitors that matter.

Where are they in the journey? Visibility has a different value at each stage:

• Exploration: understanding the problem. "My skin has become much drier since I turned 40. What kind of skincare routine should I be looking at?"
• Evaluation: comparing approaches and providers. "I'm a freelance designer with irregular income. Which types of bank accounts or digital banks are best for managing taxes and savings?"
• Decision: verifying and choosing. "I want to buy my first investment handbag for around €5,000. Which models hold their resale value best over ten years?"

Being cited in an exploration question builds awareness. Being cited in a final comparison influences the choice and drives convertion.

Once the intent is defined, then must be deployed into several variants. Phrased as a problem, as a request for advice, with a budget constraint, with reference to products or services already in use, and as a request for alternatives to a market leader. Market, language and stated constraints matter too. The same question in Italian or English, or with a budget or dietary need, can bring up different competitors.

5. Choosing intents: strategy first, tracking second

So, my recommendation is: stop asking “which prompts should we monitor?” before answering “where do we want to win?”. Choosing intents is the operational translation of the brand strategy. If the strategy is not clear, the tracking will return a picture which is misleading.

No single source is enough. The value lies in combining different data:

• Surveys and qualitative research: motivations, choice criteria, alternatives considered, the role of AI in the process (customer surveys, interviews, focus groups).
• Transactional data: which segments and products generate real value, and which are growing (e-commerce, retail sales data).
• Digital analytics: topics and landing pages that generate interest, traffic from AI platforms, on-site search queries (web analytics, internal search, Search Console).
• Observed behavior: how customers describe their problems and needs in their own words (customer care logs, chat transcripts, reviews, forums, social listening).
• Competitive intelligence: where competitors are investing and where they are vulnerable (competitor analysis, existing AI answers).
• Business objectives: markets, categories and positioning the brand wants to grow in (strategic plan, marketing plan).

6. Covering both the present and the future

A well-built intent portfolio does not just capture the current state. It also reflects the brand’s ambition. One way I personally think about it is to distinguish three types of strategy:

• Defend: search territories where the brand is already strong today. Track it to catch any loss of share early.
• Expand: adjacent territories with partial presence. Track it to measure growth before faster competitors fill the gap.
• Conquer: where the brand wants to be in the long term. Track it to set a baseline and follow your investment.

A portfolio of only defend territories gives reassuring numbers of little use. One of only conquer territories gives discouraging numbers that can kill investments still maturing. The mix has to be decided explicitly and reviewed over time.

7. Conclusions

GEO inherits its goal from SEO: being present when the customer is searching. It does not inherit SEO’s measurement tools. How you draw the sample decides what you see, and what you do.

GEO Intents, built from strategy and structured by who, what and when, turn tracking from spot checks into a map of the market. On that map, a brand can decide where to defend, where to grow and where to invest next.