Why Traditional SEO No Longer Provides the Full Picture
Internet user behavior is rapidly transforming: instead of short keywords, people are increasingly asking chatbots detailed questions, requesting product comparisons, or asking for suitable service recommendations. According to Gartner analysts, traditional search traffic volumes could drop by a quarter by the end of 2026, while traffic to Russian websites from neural networks grew sixfold in 2025.
Generative search lacks the familiar rankings from positions one through ten. It operates on binary logic: a language model either includes a specific company in its list of recommendations or ignores it entirely. Understanding a business's true visibility to AI assistants requires a fundamentally different approach to data collection.
Prompt Scenarios and Generation Randomness
The basic unit of analysis is no longer an individual keyword, but a prompt scenario reflecting real user experience. To get an objective picture, different types of queries must be tested:
- Overview queries: collecting options and roundups within a specific niche.
- Comparative queries: direct comparison of the advantages of several solutions.
- Transactional queries: questions about purchasing, ordering, or signing up for a service.
A single one-off query to a chatbot carries no statistical value. Due to the built-in generation temperature parameter, neural networks respond with a degree of variability: the list of brands can differ even when submitting the exact same question twice in a row. To obtain reliable conclusions, each scenario must be run dozens or hundreds of times.
Separating Modes: With and Without Web Search
It is crucial to account for the mode in which the language model operates:
- Without web search: the answer is generated solely based on accumulated training weights. If a brand emerged recently or had limited coverage in the training dataset, the model will not know about it.
- With web search enabled: the system queries up-to-date web pages in real time, and the result depends on current website indexing.
Mixing the metrics of these two modes into a single overall score is methodologically incorrect: they must be analyzed as independent data streams.
Key Brand Visibility Metrics in AI
Specialized metrics are used for regular monitoring:
| Metric | Meaning | Common Evaluation Mistake |
|---|---|---|
| Brand Mention Rate (BMR) | Percentage of responses mentioning the company out of the total number of tests | Analyzing an overly small sample (5–10 prompts), which leads to random spikes |
| Share of Voice (SoV) | Share of mentions for a specific brand among all competing brands in the niche | Ignoring small niche players that the neural network might cite more often than market leaders |
| Citation Rate | Frequency with which links to brand resources appear in the model's footnotes and citations | Equating a simple text mention with a clickable website link |
| Context Sentiment | Tone of the mention (positive, neutral, or critical) | Automatically treating any inclusion in an answer as positive |
Data Distortions and Platform Specifics
Different providers demonstrate different preferences: some services prioritize well-known corporations, while others more actively recommend highly specialized alternatives. Therefore, analysis should encompass multiple platforms simultaneously (including popular international and domestic solutions such as GigaChat and Alice).
Model hallucinations present an additional challenge: a neural network may mention a brand but attribute non-existent features, incorrect pricing, or third-party products to it. Such distortions require separate qualitative assessment, as formal presence in an answer in such cases carries reputational risks.
Given frequent algorithm updates, one-off measurements quickly lose relevance, and tracking trends requires continuous, cyclical monitoring.
Comments
to leave a comment.
No comments yet.