How does AI describe your business? What GEO is and how to measure it

Your customers no longer ask Google “what’s the best … near me?”. They ask ChatGPT, Gemini or Claude. If your name isn’t in the answer, you don’t exist for that customer. GEO (Generative Engine Optimization) is the work of strengthening the information AI assistants need to describe your business accurately and to mention you for the right questions.

Where does AI get its information?

Assistants don’t describe you from memory. They collect information from maps, business directories, review platforms, your website and structured data (schema.org), then compare those sources with each other. If the sources disagree (a different phone number here, an old address there), the assistant avoids recommending you or describes you wrongly.

That’s why GEO is not a “ranking trick”. The point is that what AI finds about you is accurate, current and trustworthy. The first academic paper on GEO reached the same conclusion: citing sources and adding quotations and statistics raised visibility in generative search answers by up to 40%, while keyword stuffing made almost no difference.¹

How is GEO measured? Three questions

We ask ChatGPT, Claude and Gemini real customer questions and record the answers together with their sources. We look at three things:

• Recognition: “Do you know this brand?” Does the assistant remember your founding year, number of locations and address correctly? Do assistants contradict each other?
• Discoverability: “What’s the best … in my city?” Does it recommend you for a category question? Which competitor is recommended instead?
• Reputation: “Is … trustworthy?” Is the answer positive or mixed? Does the assistant describe you through your best-rated locations or through unanswered complaints?

These results change slightly every time you run them, so a single measurement is a snapshot, not a score. The same questions are asked again every month and the direction is tracked.

The evidence behind the answer: four dimensions

Before recommending a brand, an assistant weighs its sources. We look at the evidence behind every answer from four angles:

1. Source coverage: Maps, directories, review platforms, your website: can the assistant find enough sources?
2. Source consistency: Are name, address, phone and category the same everywhere?
3. Review strength: Are review count, rating and response rate enough?
4. Freshness: How current are your photos, posts and reviews?

The picture we see most often: plenty of sources, but they don’t agree.

Where to start: five steps

1. Open the door to crawlers. Check whether your robots.txt blocks AI crawlers. Many sites do this without knowing. The companies use separate crawlers for training and for search: at OpenAI, a site that blocks OAI-SearchBot isn’t shown in ChatGPT search, while GPTBot is used for model training.³ Anthropic’s ClaudeBot, Claude-User and Claude-SearchBot⁴ and Perplexity’s PerplexityBot⁵ can each be managed separately. Decide one by one which to allow and which to block.
2. Tie your details to a single truth. Name, address, phone, opening hours and category should be identical on your website, Google Business Profile, Apple Maps and directories.
3. Add the language machines read. Schema.org structured data and an llms.txt file⁶ make your information read, not guessed. On Shopify stores, /llms.txt shows a shopping-focused text by default; since May 2026 it can be written for your brand with the agents.md.liquid theme template.⁷
4. Reply to reviews. Unanswered complaints become the raw material of AI summaries.
5. Measure and repeat. Ask the same questions once a month and record the change.

How we run the measurement

Every audit follows the same protocol, so this month’s result can be compared with last month’s.

1. Question set. Together with the brand we prepare a set of 12–20 questions in three groups: brand questions (“Who is X, when was it founded?”), category questions (“Which agencies build Shopify stores in Istanbul?”) and reputation questions (“Is X trustworthy?”). Questions use the customer’s own words and stay exactly the same every month.
2. Assistants. The set is asked in ChatGPT, Gemini, Claude and Perplexity. Each question is asked in a new session; date, model version and location are recorded.
3. Recording. The answer is saved as it is, without editing, together with the source links the assistant shows.
4. Reading. For each answer we check four things: Was the brand mentioned? Is the information correct? Which source did it come from? Who was recommended instead?
5. Comparison. We don’t produce a single score. We report how the same questions develop over the months and which sources wrong information comes from.

Before this measurement we look at five layers of the site: access (robots.txt, sitemap, llms.txt), readability (content without JavaScript), identity (schema matching the page text), confirmation (sources outside the site) and testing. Details are on the GEO consulting page.

We audited our own site. While writing this note we applied the same method to brandmentor.com.tr. Technically the site was sound: open to crawlers, structured data in place. But our name didn’t come up in AI searches, because the outside sources confirming us were weak. We’ve started fixing that.

Our own case showed best that GEO is also won outside your website.

Nobody can guarantee it

No service can guarantee “the top spot in ChatGPT”; if someone does, stay away. What can be done is to make sure the assistant can reach you, read you and pass on your information correctly. That is exactly what we do, as a Shopify Partner and Obenan’s Premium Partner for Türkiye. Our method, steps and deliverables are on the GEO consulting page.