GEO Experts in AI Answers: Findings from 110 Responses and Two IndexResearch Studies

One snapshot on October 30, 2026 · 11 interfaces · 10 questions per interface

IndexResearch published two reports based on the same set of responses to questions about GEO specialists in Russia. One counts how often each person's name appears. The other looks at which name appears first. I went through both sets of results to see how the numbers vary across AI interfaces and different types of queries. There are some useful differences.

Who collected the data

I own GAEO.ru and am a co-founder of IndexResearch. GAEO proposed the query topics and paid for response collection through BrandFound. My name is among the 74 people identified in the answers. The original reports disclose these connections.

The published material includes the BMR report on mention frequency and the FRS report on first-name appearances. GitHub contains the full BMR response data and calculation code, together with the FRS dataset and algorithm. The counting methods were finalized after the answers had been collected.

Here on GAEO.ru, I look at my results and at what this exercise can teach us about measuring visibility in AI answers. The original IndexResearch reports contain the complete list of 74 people, exact prompts and evidence excerpts.

How the answers were collected

The measurement took place on October 30, 2026, between 10:34 and 10:51 Moscow time. It covered 11 interfaces: ChatGPT, Perplexity, DeepSeek, Grok, Claude, Google Gemini, Google AI Mode, Google AI Overviews, GigaChat, Yandex Search with Alice, and the standalone Alice AI chat. Each received 10 Russian-language queries about choosing an individual GEO/AEO consultant for a business. The resulting dataset contains 110 responses, one per prompt and interface.

The questions covered several situations: requesting a list of experts, choosing one consultant, finding someone who personally manages client work, commissioning a full-service project, and checking for published case studies. My name did not appear in the prompts.

The researchers identified 74 people across the answers and counted 411 answer–person pairs. One answer could mention several people. All 110 responses were collected successfully.

One detail matters when comparing interfaces. For ChatGPT, Perplexity, Claude, DeepSeek and Grok, six questions included the extra phrase “in Russia.” The corresponding questions sent to the other six interfaces did not contain that phrase. The wording may therefore have contributed to differences in the results.

Mention frequency: BMR

In these reports, Brand Mention Rate (BMR) is the percentage of responses containing a person's recognized full name. Repeating the same name within one response still counts as one appearance.

Formula: BMR = responses containing the name / 110 × 100%.

My name appeared in 75 responses, giving a BMR of 68.2%. The next most frequent names were Vasily Zharkov (45 responses), Vladimir Nazarov (32), Dmitry Sevalnev (27) and Alexander Trigub (25).

The count includes a name appearing in a list of consultants, a comparison, a quotation from another publication or a critical comment. Looking at the passage around each occurrence is therefore useful. BMR tells us how widely the name appeared in this particular sample.

Who appears first: FRS

The second report uses First Recommendation Share (FRS). Its algorithm reads each response from the beginning and finds the first full name from the published dictionary of people. The name may appear in an introduction, in a quoted ranking or in the main list. Each response can have one first recognized person.

Formula: FRS = responses where the name appears first / 110 × 100%.

My name appeared first in 58 responses (52.7%). Zharkov came first in 17 (15.5%), and Nazarov in 5 (4.5%). Nine answers contained no recognized full person name. Of the 74 people, 21 appeared first at least once; the remaining 53 had an FRS of zero.

The name First Recommendation Share comes from an earlier version of the methodology. In this release, the metric measures the order of full names in the text. A name quoted from a third-party ranking can count as first. To determine whether the assistant actually recommends that person, the surrounding answer has to be read.

BMR and FRS for five people

On a phone, scroll the table horizontally →

I calculated the final column for this article by dividing first-name appearances by responses mentioning the same person. These values all come from the original 110 responses.

Two other names make the pattern easy to see. Maxim Melnikov appeared in 19 responses, and another recognized person was named before him in every one. Sergey Sivkov appeared in 9 responses and came first in 2 of them. Melnikov ranks higher on mention frequency; Sivkov ranks higher on first appearances.

In my GEO analyses, I keep both counts. I then read the relevant passages to check how the person was described, whether the facts were correct and whether the answer actually suggested contacting them.

Results by AI interface

Every figure below is a count out of the 10 answers collected from that interface. I compared my results with those of Vasily Zharkov, who ranked second overall on both metrics.

On a phone, scroll the table horizontally →

In ChatGPT, Perplexity and DeepSeek, my name appeared in all 10 answers. It appeared first in 9, 9 and 6 respectively. Google AI Mode mentioned me in 5 answers, while Alice AI chat mentioned me in 3.

Zharkov did particularly well in the Russian interfaces. Yandex Search with Alice mentioned him in 5 answers and named him first in 4. In the separate Alice AI chat, both figures were 4. GigaChat placed him first 3 times, compared with 2 for me.

When planning a client's GEO work, I start by finding out which AI products their customers actually use to look for suppliers. The combined numbers give a useful overview; the individual interfaces and queries show where to investigate more closely.

How query wording changes the figures

The reports also looked at two subsets of the same sample. Q01–Q07 mostly asked about GEO specialists in general. My name appeared in 50 of those 77 responses (64.9%). Q08–Q10 focused on personally managed projects, full-service work and case studies. My name appeared in 25 out of 33 (75.8%). The difference was 10.9 percentage points.

Zharkov's figures moved in the other direction: 37/77 (48.1%) in the first group and 8/33 (24.2%) in the second. This analysis was done after the responses were collected. Because the groups differ in several ways, the numbers cannot isolate the effect of any one word.

For me, the practical implication is clear: save the exact prompts with every set of results. Otherwise, changes in the questions can easily be mistaken for changes in visibility. In client projects, I agree on commercially relevant prompts in advance, preserve their wording and track new questions as a separate series.

Limitations of the study

This is a snapshot from one day, with 10 questions per interface and one response to each prompt–interface pair. GAEO proposed the query topics, and the counting rules were finalized after collection. Some interfaces received additional geographic wording.

Name recognition has its own limitations. The algorithm uses a fixed dictionary of 74 people and shortened patterns derived from given names and surnames. Rare names or similar spellings may produce matching errors. The GitHub repositories contain the original responses and text excerpts needed to inspect individual cases.

The two reports describe how often people were named and where their names appeared. They did not assess all 74 people's professional experience, customer outcomes or the reasons a given AI interface selected its sources. Previously published rankings and articles may also have influenced the answers. These boundaries matter especially because I took part in organizing the collection.

How I apply these findings to client work

I use five rules in my own measurement process.

1. Keep the complete response. I save the prompt, AI interface, date and observation identifier along with the text. That makes every example easy to inspect and every count reproducible.
2. Track distinct measures. A brand mention, the first appearance of a brand, an actual recommendation and a source citation each tell us something different. I examine them separately.
3. Repeat the same questions. A stable prompt panel makes month-to-month comparisons useful. New prompts and interfaces are recorded as additions.
4. Check the underlying facts. I review what the assistant says about a company's specialization, services and case studies. Frequently named companies can still be described inaccurately.
5. Compare visibility with inquiries. When the data is available, I look at leads and sales alongside AI visibility. A separate analysis is needed to understand what contributed to any commercial change.

I have written before about the sources AI systems use when naming GEO specialists and promoting an expert's personal brand in AI answers. My professional biography covers my experience, projects and current specialization.

What I learned from the results

In the October 30, 2026 sample, my name appeared in 75 of 110 answers (BMR 68.2%) and appeared first in 58 (FRS 52.7%). Those were the highest values for both metrics among the 74 recognized people.

I found the interface-by-interface and prompt-group results particularly useful. They help me design better GEO measurement panels and see where a company is already visible and where its answers and sources deserve closer examination. That is how I use this approach at GAEO.ru.