A user asks ChatGPT "what is the best bookkeeping software in Türkiye". The answer names three or four brands and sometimes adds source links. Is your brand on that list? In which sentence does your competitor appear? The name for this set of questions is GEO, generative engine optimization.
The term comes from an academic paper, "GEO: Generative Engine Optimization", published on arXiv in 2023 and presented at the KDD conference in 2024. Researchers from Princeton and other institutions measured ways to raise content visibility in generative engine answers on a benchmark they built. This post is about the step before improvement: measuring where you stand today.
How it differs from search results
On Google, your rank for a keyword is a measurable position. An AI answer has no such position. Your brand is either mentioned or not, your site is either cited or not. And if you are mentioned, the tone of the sentence matters: "the most recommended" and "some users complain" don't count as the same mention.
The second difference is repeatability. Ask the same question twice and you may not get the same answer. The model changes, the sources it searches change, the length of the answer changes. So a single screenshot is not evidence; at best it is an example.
AI summaries on Google are a separate surface. We covered how to measure them in another post: Getting cited in AI Overviews
Build a question set
Measurement starts with a fixed question set. Write the questions the way a customer would actually ask them, not in your own marketing language. Mixing four types works well:
- Category question: "what is the best ... app".
- Comparison question: "is X or Y better for ...".
- Problem question: "how do I ...", "what should I use for ...".
- Brand question: "is X reliable", "what does X do".
A few dozen questions are enough to start. Don't change the set often. If you add and remove questions every week, week-to-week comparison loses its meaning. If you need new questions, add them as a separate group and leave the old group's series intact.
What to record for every answer
- Question, model and time of the read.
- Whether your brand was mentioned, and in which sentence.
- Which competitors were mentioned.
- The domains the answer cites as sources.
- The answer text itself, or at least its opening, as evidence if a dispute comes up later.
One answer is not a sample
Write the report as a ratio, not as individual answers: "how many of the 40 questions mentioned us this week". Compute the same ratio for competitors. A swing of a point or two in a week is most likely noise. A shift that keeps going in the same direction for several weeks is worth discussing.
Don't count a question that got no answer as not mentioned. If the model returned an error that day or the request timed out, that cell should stay unmeasured. Otherwise you will report an infrastructure problem as a drop in brand visibility.
Keep models separate too. ChatGPT, Claude, Gemini and Perplexity can name different brands for the same question. Folding them all into one ratio hides which assistant you are falling behind in.
Which bot ChatGPT uses to read your site
To be cited you first have to be readable. According to OpenAI's documentation on its crawlers, OAI-SearchBot is used to surface websites in ChatGPT's search features, and sites that block it in robots.txt are not shown in ChatGPT search answers. GPTBot does a different job: it signals whether content may be used to train foundation models.
The practical consequence: a site that blocks GPTBot to opt out of training, and blocks OAI-SearchBot on the same line, has also closed off its chance of being cited in ChatGPT search. Read your own robots.txt once with this in mind. The same document says a robots.txt change can take about 24 hours to take effect on the search side.
The source list is a to-do list
When you collect the domains cited in answers over several weeks, a pattern appears. Certain comparison sites, forums or trade publications may keep coming up as sources. For a brand that isn't named on any of them, the job is clear: get onto those pages, and get missing or outdated information corrected.
The limits of the measurement
The answer your measurement tool sees may not be the answer a signed-in user sees with their own history and settings. What you are measuring is a trend under fixed conditions, not individual users' screens. Say so plainly in the report, or someone will challenge the whole series with a single question they asked on their own phone.
With Adversee
In Adversee's SEO module, the questions you track are put to ChatGPT, Claude, Gemini and Perplexity on a regular schedule. For each model it records whether your domain or brand was mentioned in the latest answer, which tracked competitors were mentioned, the sources and the opening of the answer. A model that returned no answer shows as not measured. The sources view also lists the sites AI answers cite most and the sources where you are not mentioned.
Start tracking competitor moves automatically.