AI Visibility in Russian Property Search 2026: Which Real Estate Platforms Do AI Systems Recommend?

100 API responses, four AI systems, ten apartment-buying scenarios and 775 citations. I studied which Russian real estate platforms most often appear in AI recommendations, which one takes the first position, how much the recommendation output changes when the same query is repeated, and which sources accompany the responses

When a user asks an AI system «где лучше искать квартиру» (“where is the best place to look for an apartment”) instead of running a regular search, the result looks to them like a ready-made list of recommendations. For a brand, the same answer is far more complex. It is not enough to know whether a platform was mentioned at least once: what matters is whether it made it into the recommendation list itself, what position it took, whether the same set will reappear on the next run, and in what information context the system formed its answer

That is why in this study I deliberately did not try to build yet another ranking of the “best platforms for AI.” Its purpose is different: to break AI visibility down into measurable components and see how well they agree with one another

The main conclusion turned out to be fairly simple: AI visibility is not a single number. A brand may appear often in responses but rarely open the list; a system may consistently reproduce the overall set of platforms but change the top three; a brand’s own website may be cited often but remain a small part of the overall body of sources surrounding its recommendations

In short: the findings

A distinct core of four platforms emerged in the sample. Cian (cian.ru) was recommended in 87 of 100 API responses, Domclick (domclick.ru) in 81, Avito Nedvizhimost (avito.ru) in 75 and Yandex Nedvizhimost (realty.yandex.ru) in 69. The fifth platform, Etagi, appeared in recommendations 28 times, and every other platform in no more than 13 responses

First position was concentrated even more strongly. Of the 91 responses in which the first recommendation could be determined unambiguously, Cian held it 62 times — 68.1%. Avito Nedvizhimost came first 9 times, Domclick 8

At the same time, the lists themselves cannot be called fully deterministic. The overall repeatability of the recommendation set between runs of the same query was macro Jaccard 0.694, while the overlap in the composition of the top three was 0.561. The spread across scenarios was even wider: from a Jaccard of 0.471 to 0.917

Finally, 93 of 100 responses contained links. In total, the collection held 775 individual citations, which after normalization came down to 191 domains and 628 unique “response × domain” pairs. Real estate portals and marketplaces accounted for 41.7% of these pairs. At the same time, the recommended brands’ own domains remained a minority of the accompanying sources in every case: the maximum Owned Source Share across the entire sample did not exceed 15%

These numbers describe a specific sample of API responses, not the quality of the platforms and not the behavior of the consumer apps ChatGPT, Claude, Perplexity or Grok

---

1. Research question and methodology

The working research question was formulated as follows:

Which online platforms do generative AI systems recommend to users for finding and buying an apartment in Russia, how stable are these recommendations, and which sources accompany the responses?

The study covered four specific search-enabled API systems:

• anthropic/claude-sonnet-5;
• openai/gpt-5.6-terra;
• perplexity/sonar-pro-search;
• x-ai/grok-4.5

For brevity, I refer to these four API systems as Claude, OpenAI, Perplexity and Grok throughout the rest of the text. These short names refer only to the API configurations listed above and do not mean the corresponding companies’ consumer apps

This distinction is essential. I did not test the ChatGPT, Claude, Perplexity or Grok user interfaces and do not extend the results to them. The unit of analysis here is a specific API response from a specific system, in a fixed configuration, during the collection period

Ten user scenarios

To avoid reducing the experiment to a single wording such as «какие сайты посоветуешь» (“which websites would you recommend”), I used ten scenarios, each with a different user task. The queries were submitted in Russian; the table gives the original wording followed by an English translation

Five scenarios — RU_001, RU_002, RU_004, RU_006 and RU_010 — were run three times on each system. The other five were run twice. This yielded 25 responses per system and 100 API observations in total

The repeated runs were not there simply to add “bulk” to the data. They made it possible to examine a second layer of AI visibility: how well a system reproduces its own answer when a user repeats the same question

How responses were turned into data

First, all possible mentions of services were extracted from the responses, and the different spellings were then mapped to unified brand names. For example, M2 had to be distinguished from a separate service called «Квадратный метр» (Kvadratny Metr, literally “square meter”), and the ordinary Russian phrase «цена за квадратный метр» (“price per square meter”) must not be mistaken for the name of a platform. After the mentions had been extracted and the spellings consolidated into unified brands, the data went through a separate review procedure. Formal rules were used only for preliminary classification; further review was carried out in an AI-assisted process, and I approved the final set of decisions myself — before the metrics were calculated. After that, the dataset was frozen and contained 476 unique “response × brand” combinations: 452 recommendations and 24 neutral mentions

Of these, 118 were coded as explicit recommendations, 334 as brands included in a recommended list, and 24 as neutral mentions. No negative or undetermined units remained in the final coding. In other words, the main recommendation dataset contained 452 units, while another 24 mentions were kept separately as neutral. This matters: the mere presence of a brand name in the text was not automatically counted as a recommendation

For the rank-based metrics, the order could be verifiably reconstructed for 446 of 452 recommendations — 98.67%. Six units sat in structurally parallel, unordered groups. I did not assign them artificial positions: they were counted as recommendations but excluded from the metrics that require an exact rank

How I guarded against fitting the results

Before calculating the results, I fixed the counting rules for the main metrics in advance. This was necessary so that the methodology would not shift after the leaders had become visible. Later, two specific cases needed clarification: how to treat recommendations without an unambiguous order, and exactly how to calculate the source metrics. These rules were also written down before the corresponding metrics were calculated, rather than adjusted to fit the resulting numbers

It is important to separate the protocol of the experiment itself from the subsequent processing of the texts. The research question, the set of AI systems, the ten user scenarios, the repetition scheme and the main metrics, together with the rules for interpreting them, were fixed before the results were calculated. The rules for extracting and classifying mentions, however, were refined after the API responses had been collected, during data preparation. Before the final coding and the calculation of the metrics, the extraction and the set of formal rules were frozen and were not readjusted to the results afterwards

I deliberately did not build a single composite AI Visibility index and did not assign weights to frequency, first position, stability or sources. Statistical significance and confidence intervals were not calculated either. The aim of the study is to provide a transparent descriptive profile, not to artificially collapse different properties into a single place in a ranking

---

2. Four platforms formed a separate top tier

The first metric is Recommendation Share: the share of the 100 API responses in which a platform was explicitly recommended or included in a recommended list

The result split the sample into clearly distinct tiers. Cian received 87 recommendations out of 100, Domclick 81, Avito Nedvizhimost 75 and Yandex Nedvizhimost 69. After fourth place there is a sharp break: Etagi — 28, N1 — 13, Restate and Youla — 12 each, M2 and Mir Kvartir — 11 each

In other words, the gap between the fourth and fifth results is 41 percentage points: 69% versus 28%. For this sample, it is more accurate to speak not of a smooth ladder of a dozen competitors but of a top core of four brands and a long tail of other platforms

These four also share another property: each brand was recommended at least once in all ten scenarios and by all four systems. This does not make them the “best” services, but it shows how broad their observed AI visibility was within this experiment

Notably, recommendation frequency on its own says nothing about position within the list. Domclick almost catches up with Cian on Recommendation Share — 81% versus 87% — but the next metric shows a very different degree of concentration

---

3. First position is more concentrated than simple presence

For First Recommendation Share, the denominator is not 100. Nine responses contained no recommendations at all, so they are excluded from the denominator. In the final coding, there were no further responses that had to be excluded because of an ambiguous first position. As a result, the first recommendation could be determined unambiguously in 91 responses

The six recommendations without a verified ordinal rank were not turned into artificial “first places”: they were counted in the overall recommendation frequency but excluded from the metrics that required a specific position. In this sample, they did not cause any additional exclusions from the First Recommendation Share denominator

Cian came first in 62 of 91 cases — 68.1%. Avito Nedvizhimost opened the list 9 times, or 9.9%, and Domclick 8 times, or 8.8%. Yandex Nedvizhimost and Youla each took 3 first positions, Novostroy-M and BN.ru 2 each, and Restate and M2 one each

The gap between the first and second brand is much wider on this metric than on Recommendation Share. Cian was recommended in 87% of responses versus 81% for Domclick, but on first position the difference is already 68.1% versus 8.8%

Average rank confirms the same picture, though from a different angle. Among ranked recommendations, Cian’s average position is 1.55 with n=87, Avito Nedvizhimost’s is 2.83 with n=75, Yandex Nedvizhimost’s is 3.16 with 67 ranked cases out of 69 recommendations, and Domclick’s is 3.23 with n=81

Domclick is therefore a good example of why AI visibility cannot be described by a single presence share. It is second in overall recommendation frequency, but only third by number of first positions, and it trails the other three leaders on average rank. A brand can be an almost obligatory element of the list and still rarely open it

---

4. Different AI systems produce a similar core but different recommendation density

If Recommendation Share is calculated separately within each system, the four leading brands remain, but their frequency varies noticeably

In Claude, Cian appears in recommendations 22 times out of 25, Domclick 20, Avito Nedvizhimost 18 and Yandex Nedvizhimost 11. For OpenAI, the values are 20, 19, 15 and 18 respectively. For Perplexity — 21, 18, 18 and 17. In Grok, the first three brands appear 24 times out of 25 each, and Yandex Nedvizhimost 23 times

At the level of the chart, it is easy to draw too strong a conclusion — for example, that Grok “prefers” these brands. The data do not show this. The systems differ not only in their choice of specific platforms but also in the length of the lists they produce. Grok’s 25 responses contained 163 coded recommendation units, compared with 76 in OpenAI’s 25 responses. Grok’s high share for several brands at once therefore partly reflects a broader set of recommendations, not some hidden attitude of the system toward a particular platform

That is why Figure 3 is more useful when read as a map of visibility rather than as a contest between models. It shows that the shared top tier is reproduced by all four systems, but the density and composition of the list around this core differ

---

5. The same question does not guarantee the same list

Up to this point, the study has mainly answered the question “who gets recommended?” Repeated runs make it possible to test something more complex: will we get roughly the same answer if we ask the system exactly the same question again?

For this, I used two metrics. The first is the Jaccard coefficient, which compares the entire set of brands. The second is Top-3 overlap, which looks only at the composition of the first three recommendations

What exactly the Jaccard coefficient shows

Imagine two responses from the same system to the same query. In the first, the system recommends Cian, Domclick, Avito Nedvizhimost and Yandex Nedvizhimost; in the second, Cian, Domclick, Yandex Nedvizhimost and Etagi

The intersection of the two sets is three brands: Cian, Domclick and Yandex Nedvizhimost. The union is five brands: Avito Nedvizhimost and Etagi are added to those three. The Jaccard coefficient equals the size of the intersection divided by the size of the union: 3 / 5 = 0.60

If the two sets match completely, Jaccard equals 1. If they have no brands in common, it equals 0. The coefficient ignores order within the list: all that matters is which platforms are in the set

Each question was run either two or three times. Two runs give one comparison of responses; three runs give three: the first with the second, the first with the third, and the second with the third. In total, this produced 80 pairwise comparisons

Simply averaging all 80 values, however, would be wrong: questions with three repetitions would carry three times the weight of questions with two. So I first calculated one average value for each “AI system × question” combination. There were 40 such combinations — four systems × ten questions. Only then did I average these values. As a result, each scenario for each system had an equal influence on the final metric

At the same time, these 80 pairwise comparisons were not treated as 80 independent statistical observations: several pairs may include the same original response. They served as intermediate elements of a descriptive repeatability metric, and the main unit of averaging was the “AI system × user scenario” combination. This averaging equalizes the weight of the scenarios in the final metric, but it does not make the estimates for scenarios with two and three repetitions equally precise

How Top-3 overlap differs from Jaccard

Jaccard looks at the entire recommended set. The beginning of a response, however, may behave differently, so Top-3 overlap — the match in the composition of the top three between repetitions — was measured separately

If both top threes consist of the same three brands, the result is 3/3 = 1, even if the brands swap places within the top three. If two brands match, it is 2/3 ≈ 0.67; one, 1/3 ≈ 0.33; none, 0

This is fundamentally not an exact-rank metric. It answers how well the composition of the top three is preserved. If two responses contain Cian, Domclick and Avito in the Top-3 but in a different order, the overlap remains 1

Overall result: the set is more stable than its top

Averaged over the 40 cells, macro Jaccard was 0.694. The pair-weighted value, given only as a supplementary figure, is 0.712. In the pre-specified three-repetition subset of 20 cells, macro Jaccard was 0.731

For Top-3, the result is lower: macro Top-3 overlap = 0.561

In other words, when the same query was repeated, the systems often returned a large part of the same set of brands, but the composition of the top three positions changed more. This is an important distinction: an AI system can keep the core of a list while reshuffling its top

There is also a technical edge case. In three pairwise comparisons, neither response contained any recommendations. For Jaccard, these are two identical empty sets, so the value is 1; for Top-3, there was nothing to preserve, so the overlap is 0. This is not a calculation error: the two metrics answer different questions

---

6. Repeatability depends on the AI system — but not in the way a single metric would suggest

OpenAI GPT-5.6 Terra has the highest macro Jaccard — 0.822. It is followed by Grok — 0.723, Claude — 0.686 and Perplexity — 0.545. Looking at this metric alone, one could say that OpenAI’s full set of recommendations repeated most often in this sample

Top-3, however, changes the picture. Claude’s top-three overlap is 0.700, the highest of the four systems. OpenAI and Grok score 0.544 each, and Perplexity 0.456

So OpenAI reproduced the whole set better than the others, but not the top of the list. Claude, conversely, produced a less stable set but more often kept the same brands in its Top-3. The two metrics do not contradict each other: they measure different properties

This is exactly where it is especially important not to mistake repeatability for quality. A higher Jaccard does not make a response more correct, more useful or more reliable for a buyer. One system may consistently repeat the same poor set; another may vary its list while remaining genuinely useful. This experiment assesses only how reproducible the composition of the set is

---

7. Query wording affects repeatability more than it might seem

If the same metrics are grouped by scenario rather than by system, the spread becomes even wider

The most stable scenario was RU_010 — a request to name two or three platforms that are best used at the same time. Its macro Jaccard is 0.917, and its Top-3 overlap 0.944. Almost as stable was RU_009, about searching by map, filters, price and neighborhood: Jaccard 0.908, Top-3 1.000

At the opposite end are scenarios with noticeably lower repeatability. RU_007 — choosing an apartment remotely when relocating — had a Jaccard of 0.471 and a Top-3 overlap of 0.417. RU_003, in which the user is buying an apartment for the first time, had a Jaccard of 0.500 and a Top-3 overlap of 0.167. For new-build apartments in Moscow, RU_004, the values were 0.540 and 0.361

The Jaccard range across queries is 0.471–0.917, compared with 0.545–0.822 across systems. This does not prove that the scenario is “more important” than the model in a causal sense, but it does show that, in this sample, the differences between task formulations were at least as large as the differences between systems

One could propose an intuitive hypothesis that queries with a narrower expected format — for example, «назови 2–3 платформы» (“name 2–3 platforms”) — leave the system less room for variation. The study data, however, do not establish this mechanism on their own, so this causal interpretation is not part of the conclusions

---

8. Sources: 93 of 100 responses contained links

The third part of the study concerns not the brands as such but the information context of the responses

93 of 100 API responses contained at least one link. In total, the systems provided 775 individual source citations. All of them had a parseable URL. After normalization, they resolved to 191 canonical domains

For the analysis of domain frequency, however, I did not use the raw number of citations. If the same domain appears three times in one response, it should not automatically carry three times the weight. So an additional unit was introduced — the “response × domain” pair. There were 628 such unique pairs

This difference matters. 775 is the number of individual citations, while 628 is the number of unique occurrences of a domain in a specific response. The second level is better suited to analyzing the structure of sources because it does not reward a domain for several links repeated within a single text

---

9. What types of sources accompany the responses

Each of the 191 domains was assigned to one category. The category describes the type of source, not its owner. These are two independent axes: for example, cian.ru is a real estate portal by type and, at the same time, Cian’s own domain by ownership

The largest category is real estate portals and marketplaces: 262 pairs out of 628, or 41.7%. In second place is the broad other category — 108 pairs, 17.2%. Next come ratings and reviews — 81 pairs, 12.9%, industry media — 65, 10.4%, and business and news media — 62, 9.9%

The “other” category brings together heterogeneous commercial, reference and corporate resources that did not fit any of the main categories. It should therefore not be interpreted as a single type of source

Smaller categories: forums and communities — 18 pairs, 2.9%; government and regulatory resources — 14, 2.2%; social platforms — also 14, 2.2%; video — 4, 0.6%

At the structural level, it is clear that AI responses are not confined to the real estate aggregators themselves. The information context includes rating projects, media outlets, government sources, app stores, communities and other types of resources. These data, however, cannot establish which of them caused a particular platform to be chosen

---

10. The most frequent domains: the leaders’ own websites are prominent, but not the only ones

Looking at individual domains rather than categories, first place goes to cian.ru, which was cited in 46 of 100 responses. Next come domclick.ru — 41, rskrf.ru — 38, realty.yandex.ru — 33 and всеостройке.рф — 24. They are followed by realty.rbc.ru — 13, and avito.ru and etagi.com — 12 each

The mere fact that cian.ru and domclick.ru are at the top does not support the conclusion that “models recommend brands because of their own websites.” A single response may contain several sources at once: the platform’s own website, an industry media outlet, a rating, another portal and a government resource. To understand what share of this context belongs specifically to the recommended brand’s own domains, a separate metric is needed

---

11. Owned Source Share: what the share of owned domains actually shows

Owned Source Share is one of the metrics that is especially easy to misinterpret. It does not measure the “contribution of a brand’s website to the recommendation.” It describes only the composition of sources around the responses in which the brand had already been recommended

The calculation works as follows. Take Cian: it was recommended in 87 responses. These responses contain 577 unique “response × domain” pairs. Of these, 47 pairs belong to domains owned by Cian. Owned Source Share is therefore 47 / 577 = 8.1%. In other words, across the whole set of domains in the responses where Cian was recommended, roughly 8% of unique domain occurrences come from the brand’s own resources, and the rest from sources external to it

Among the brands shown in the chart, the highest share belongs to Yandex Nedvizhimost — 8.4%: 39 owned pairs out of 463. Next come Cian — 8.1%: 47/577, Domclick — 7.5%: 40/534, N1 — 7.4%: 7/95, Restate — 5.9%: 7/118, Mir Kvartir — 5.3%: 6/113, Etagi — 3.5%: 8/230, Avito Nedvizhimost — 2.2%: 11/511, M2 — 1.2%: 1/80 and Youla — 0 of 83

Figure 7 shows only brands with at least ten recommendations, so that small samples do not take center stage in the visualization. Across the full set, the highest Owned Source Share belonged to Novostroy-M — 15.0%, but that is 6 owned pairs out of 40, with only seven recommendations of the brand in total. Comparing this percentage with the large platforms without stating the denominator would therefore be misleading

The main observed result here is a different one: for no brand did its own domains make up the majority of the sources around its recommendations; the maximum across the full sample is 15%. This allows us to say that recommendations exist within a broad external information context. But it still does not prove that external websites influence recommendations more strongly than a brand’s own

---

12. Why cian.ru can be the most frequent domain while Cian’s Owned Source Share is only 8.1%

At first glance, these two results look contradictory: cian.ru appears in 46 of 100 responses — more often than any other domain — yet the share of Cian’s own sources among responses recommending the brand is only 8.1%

There is no contradiction, because the denominators are different. Domain occurrence frequency answers the question: “in how many of the 100 responses does a particular domain appear?” That is why a single frequent domain can score highly. Owned Source Share asks: “among all unique domains in all responses where a particular brand is recommended, what share do the brand’s own domains make up?” A single response recommending Cian may simultaneously include cian.ru, Roskachestvo, a specialist media outlet, another portal and several other resources. All of them count toward the denominator

So a frequently cited owned domain can be prominent overall and at the same time account for a small share of the brand’s wider source context. For AI visibility analysis, this is an important distinction: the frequency of an individual source and the structure of sources around a brand are not the same thing

---

13. The key limitation of source analysis: a citation is not a trace of the model’s reasoning

Source analysis invites many appealing claims that the data do not support. For example: “AI recommends Cian because it often cites cian.ru” or “the model trusts third-party sources more.” Neither of these statements follows from the study

A citation shows that a source was present in the visible API response. A link does not reveal the internal mechanism by which the response was formed, and it does not prove that this particular source was the reason the brand was chosen

A technical limitation reinforces this caution. For 537 of 775 citations, the provider did not report a usable local position of the link within the text. For Anthropic and Perplexity links, the provider did not indicate which specific fragment of the response a source referred to. OpenAI provided such anchoring for all 76 of its links, and Grok for 162 of 259

For a large part of the material, therefore, only co-occurrence within the same response can be established with confidence. Locally linking a specific citation to a specific recommendation is possible only on a limited subset, mainly OpenAI and Grok. That is why the main source metrics were calculated at the level of the whole response, rather than on the assumption that a particular link supports a particular recommendation

---

14. What the study says about AI visibility — and why a single ranking is not enough

Looking only at Figure 1, one could end this piece with the statement “Cian is first, followed by Domclick, Avito and Yandex.” The other layers, however, show how little information such a statement conveys

Recommendation frequency. Here there is a top tier of four platforms: 87%, 81%, 75% and 69%. This is the most obvious layer of visibility

First position. The picture is much more concentrated: Cian takes 62 of 91 first places, while the nearest brands take 9 and 8. Overall frequency and dominance at the start of a response therefore diverge substantially

Coverage of systems and scenarios. All four leaders appear in all ten scenarios and in all four systems, but the intensity of their visibility varies from system to system

Repeatability. The same question does not guarantee an identical list. Overall macro Jaccard is 0.694, and Top-3 overlap is 0.561. Repeatability also varies between systems, and even more between scenarios

Source context. Links are present in 93% of responses, but the recommended brands’ own domains remain a minority among all accompanying domains

None of these axes is “the main one” by definition. To produce a single ranking, I would have to decide on my own how many points to give to frequency, how many to first position, how many to stability and how many to the source profile. Such weights would reflect not the data but my subjective judgment

That is why the study has no composite AI Visibility Score. I consider it more accurate to describe a brand as a profile across several dimensions

---

15. What can be used in practice

The study does not yield a universal instruction of the “do X and AI will start recommending your brand more often” kind. It shows which sources accompanied the responses, but it does not establish whether they influenced the choice of brand. Testing the causal influence of sources would require a separate controlled experiment — for example, comparing results before and after changing the availability of specific sources

The study does, however, support a more practical conclusion about the measurement of AI visibility itself. If a company monitors its brand’s presence in generative search, a single “how many times were we mentioned” metric is not enough. A minimal profile should distinguish:

• the share of responses in which the brand is recommended;
• first position and average rank;
• presence across different systems and user scenarios;
• repeatability of responses;
• the structure of accompanying sources

Such monitoring makes it possible to describe which sources appear alongside recommendations, but it does not turn their presence into proof of influence on the choice of brand. A brand may be consistently present everywhere yet rarely come first; it may be first in one scenario and disappear in another; it may show high frequency at one point in time but low reproducibility between repetitions

For GEO/AEO (generative engine optimization and answer engine optimization) practice, this is, in my view, the more useful unit of analysis: not “a brand’s position in AI” but the profile of its observed visibility

---

16. Limitations of the study

The results refer to 100 API observations from a single period — August 2026. AI systems, search indexes, the ways search is invoked and providers’ internal strategies can change. A rerun of the same protocol a month or a year later is therefore not guaranteed to reproduce the same values

The study covered four specific API systems in a fixed configuration, not consumer apps. It is not valid to replace openai/gpt-5.6-terra with “ChatGPT” in the conclusions or to extend the results to the Claude, Perplexity or Grok interfaces

Nor were the system configurations identical at the level of their internal search and generation mechanisms. Generation parameters such as temperature and top_p were not forcibly standardized: each provider’s default values were used. Differences in repeatability should therefore be read as a characteristic of a specific AI system in the configuration used, taken as a whole — the model, the provider’s default settings and the search pipeline — rather than as a measurement of the “pure randomness” of the language model itself under identical generation parameters

The formal recall of brand detection was not evaluated against an independent, manually annotated corpus. Extraction relied on a prepared dictionary of name variants, entity normalization, contextual checks and a subsequent audit. Rare platforms or unanticipated spellings could therefore, in theory, have been missed

The study does not measure the objective quality of real estate platforms. Frequent recommendation does not mean that a service is better, safer, more convenient or more cost-effective for a buyer. Nor does repeatability equal correctness: a high Jaccard shows the stability of the set, not the quality of the response

The rank-based metrics cover 446 of 452 recommendation units. The six structurally unordered units were not given an artificial rank. First Recommendation Share was calculated on the 91 responses with an unambiguously determinable first position

For 537 of 775 citations, there is no usable local position in the text, so local analysis of “which link refers to which recommendation” is limited. The co-occurrence of a source and a brand is not interpreted causally

Statistical significance and confidence intervals were not calculated. The study is descriptive and does not claim statistical generalization to the AI search market as a whole

---

Technical write-up

I have also prepared a separate technical write-up of the study: how the data collection process was organized, how brands and sources were normalized, how mentions were reviewed in an AI-assisted process and contested decisions approved, how the final dataset was frozen and how reproducibility was checked. A link will appear here once it is published on Habr

---

Methodological summary

• Collection period: August 2026
• Observations: 100
• AI systems: 4
• User scenarios: 10
• Final “response × brand” units: 476
• Recommendation units: 452
• Neutral mentions: 24
• Recommendations with a verified ordinal rank: 446 of 452
• Responses with an unambiguously determinable first recommendation: 91
• Pairwise comparisons of repetitions: 80
• “AI system × query” combinations in the repeatability analysis: 40
• Individual source citations: 775
• Canonical domains: 191
• Unique “response × domain” pairs: 628
• Citations with a known position in the text: 238 of 775
• Composite AI Visibility index: not built
• Causal influence of sources: not assessed

---

Rules for republication and adaptation

The Russian original of this text is the primary source of the study; this is its full English translation. When adapting it for other platforms, it is acceptable to shorten it and simplify the methodological explanations, but it is not acceptable to change numerical values or denominators, to present repeatability as “quality,” to transfer API results to consumer apps, to call citation frequency a source’s “influence,” or to present the Recommendation Share order as a ranking of platform quality