The answer to "Do we show up in AI?" is usually a screenshot: someone asks ChatGPT, sees the brand, takes a screenshot and sends it to the team. The next day, the same question brings a different answer. That's not a measurement, it's a moment.
This post describes a regular measurement method you can run without buying a paid tool. It has three layers: a prompt test you run with a fixed question list, the official data Google provides, and traffic in your analytics. All three give incomplete data; read together, they give you a real picture.
The same question may not get the same answer every time; the response of the model building the answer can vary with the phrasing of the question, the conditions on a given day and the platform. That's why measurement should rest not on a single answer but on the rate across repeated attempts. This isn't an official standard; it's our working method: ask the same question again and again and look at the rate.
A moment: a screenshot
A measurement: a fixed list
Twenty questions are enough. Split them into three groups:
| Group | Example pattern | What it shows |
|---|---|---|
| Category questions (unbranded) | "Who offers reliable [service] in [city]?" | What someone who doesn't know you sees |
| Comparison questions | "[Brand A] or [Brand B]?" or "What should I look for when choosing a [service] company?" | Whether you're mentioned at the decision stage |
| Brand questions | "What is [brand name] and what does it do?" | Whether the model knows you correctly |
Watch out: Write the questions in the language a real customer would use. A phrase you would never search yourself, like "best GEO agency," will mislead you.
A few companies that set up chatbots in and around Istanbul stand out:
Pick the 3-4 platforms your customers use: ChatGPT, Gemini, Perplexity, Google's AI Mode and so on. Ask each question on each platform at least 3 times. Run each attempt in a new chat; the context of a previous chat can affect the answer. Since accounts or personalization can change the result, use the same account if possible and make a note of it.
20 questions × 3 platforms × 3 attempts = 180 answers. That's not an hour's work, it's a full workday; plan to repeat it once a month.
For each answer, record these fields:
| Date | Platform | Question | Attempt | Brand mentioned? | Position | Sources shown | Note |
|---|---|---|---|---|---|---|---|
| 2026-10-21 | ChatGPT | Question 1 | 1 | Yes | 2 | site-a.com, site-b.com | Wrong price |
The "Note" column looks trivial, but it's the most valuable one: everything the model gets wrong is a to-do list of fixes.
01
Prompt test
Fixed question list, repeated attempts
02
Official data
Search Console generative AI report
03
Traffic
GA4 referral sources
Google opened Search Console's generative AI performance report to all sites on August 31, 2026. The report shows how many times your pages appeared inside AI Overviews and AI Mode, with breakdowns by page, country, device and date. It doesn't give clicks or queries. I covered the details in the Daily Radar post; for the official explanation, see Google's Search Console help page.
Google's AI features page also says that impressions from these features are counted under the Web search type in the Search Console Performance report. So the new tab makes a slice of the existing total visible; it doesn't add a separate total.
How do you use it? Read this report as a complement to the prompt test: the test tells you "am I mentioned for these questions"; the report tells you "which of my pages are shown in AI features on Google."
In GA4's Acquisition report, look for sources such as chatgpt.com, perplexity.ai, gemini.google.com or copilot.microsoft.com. This shows visitors who clicked a link in an AI answer and arrived at your site.
For details, see the related Daily Radar post. Cookie preferences are another limit: GA4 only fully counts visitors who give consent.
OpenAI's crawler documentation defines three bots: OAI-SearchBot (for showing pages in ChatGPT search), GPTBot (for model training) and ChatGPT-User (which visits a page when a user asks). You can search for these names in your server access logs. If you don't see any OAI-SearchBot entries, first check whether the bot is blocked; I walked through the reasons step by step in the post Why doesn't my brand show up in ChatGPT.
Seeing bot entries doesn't mean you'll appear; it only shows that you can be crawled.
| What you see | Likely meaning | What to do |
|---|---|---|
| Never mentioned in category questions, correct in brand questions | They know you but don't recommend you | Work on content and external mentions |
| Wrong information in brand questions | The model is learning from the wrong source | Clarify your own page, find the wrong source |
| Not mentioned anywhere, no bot entries | There may be a technical access problem | Check robots.txt and indexing |
| Mentioned, but no traffic | The answer may not require a click | Put click-worthy information such as price and scope into the answer |
| Direct traffic rose, and so did mentions | It may be AI-referred visits | Don't call it certain based on this data alone |
Measurement setup checklist
0/8Check off the items; the first empty item is where you start.
In our GEO service, we run a monthly prompt test with the same questions; the skeleton of the method is the same as in this post. If you like, we can build your question list together.
Not necessarily. A manual test with 20 questions and a few platforms is enough to get started. Tools save time, but the same limits apply to them: answers vary, and every measurement is a sample.
Twenty questions is reasonable to start with. More makes the measurement more solid, but the effort grows too. What matters is that the list stays fixed; if you ask different questions every month, you can't compare month to month.
Answers aren't pulled from a single fixed result; they're rebuilt every time. That's why measurement should rest not on a single answer but on the rate across repeated attempts.
No. According to Google's announcement, the report gives impressions, pages, country, device and date; it doesn't give clicks or queries.
It may be; there have been reports that tag and referrer information is missing from ChatGPT links. But Direct can also rise from other sources. Attributing it to AI based on this data alone wouldn't be right.
I explain in plain language what GEO is, what information AI tools look at when they recommend a brand, and what you need to change on your site. I've also flagged the parts that get overhyped.
According to Comscore's September 22, 2026 report, ChatGPT's share of prompts across AI assistants fell from 70% to 50% in six months. Being a source and being named in the answer now need to be measured separately.
Why can two GEO quotes differ several times over? We break down the six line items that make up a quote, from the initial analysis and content creation to external mentions and measurement, and show how to compare quotes side by side.
We use cookies to improve the site and measure advertising performance. Analytics and advertising cookies run only if you accept them. Details are in our Privacy Policy and Data Protection Notice.
