An industry roundtable hosted by Rating Runet spent two and a half hours reviewing the first draft of a GEO (Generative Engine Optimization) standard. The central finding: step-by-step site-edit procedures for generative search go stale within three months, because classical SEO methods are structurally incompatible with Retrieval-Augmented Generation (RAG) architecture.
Four methods that produce no result
Pumping brand mentions through users accomplishes nothing. Commercial LLMs do not retrain on live dialogue — model weights are frozen after pre-training, session context is wiped when the conversation closes, and likes or dislikes in assistant interfaces feed RLHF datasets rather than shaping answers in real time.
Placing an llms.txt file at the site root is equally ineffective. The RAG layer queries the general search index of actually indexed documents; llms.txt is useful to external parsers and auditors but has no effect on ranking by Google or Yandex algorithms.
Link equity behaves differently in generative search than in traditional SEO. If a donor page does not appear in the thematic cluster of sources the model itself cites for a target query, the passed weight is zero.
Schema.org micromarkup is necessary but not sufficient on its own. When a JSON-LD declaration carries the Organization or Offer type but the page body lacks explicit service parameters, deadlines and numerical ranges, the retriever drops the document for low semantic density.
How a generative assistant selects a supplier
Faced with a complex query, a language model applies Query Fan-Out, splitting the input into a set of parallel search sub-queries. Using a question about packaging equipment for a pharmaceutical production line as the example, the model separately searches for regulatory compliance (such as GMP) and regional service availability. A page that offers only generic phrases is cut at the candidate re-ranking stage: the model finds no answers to its sub-queries and routes its recommendation to a competitor whose parameters are stated explicitly.
Server-side crawling traps
GPTBot, ClaudeBot, PerplexityBot and YandexRenderResourcesBot run into a common infrastructure problem. Cloudflare, DDoS-Guard or ModSecurity rules typically pass the first 3–5 crawler requests, then classify deep catalogue scanning as suspicious and begin returning a JS challenge or blank HTML with HTTP status 200 OK. Server logs show no 403 or 500 errors, so the problem is invisible to the site owner while the crawler receives empty responses. The roundtable cited a real case: a manufacturing platform spent a month and a half eliminating AI hallucinations after the fix turned out to be simply unblocking crawlers in the WAF configuration.
Generative crawlers operate under strict text-extraction timeouts and frequently do not execute heavy client-side JavaScript, so a document must deliver its content in the raw HTML. Document Chunking algorithms process <table> elements significantly more reliably than layouts built on nested <div> or <ul> blocks, because tabular structure preserves the relationship between an attribute and its numerical value even when a page is split into chunks.
Acceptance criteria for results
Experts agreed on three mandatory rules.
First, averaged 'AI visibility' scores are not acceptable. DeepSeek draws on industry directories, forums and review databases; ChatGPT uses encyclopaedic sources and falls back on the English-language web with auto-translation when data is thin; Yandex's Alice is tightly coupled to the current organic search index. Results must be recorded separately for each target system.
Second, measurements must be run in series. Because language models operate with temperature > 0, a single query in a web interface is not a metric. A valid measurement uses waves of 3 to 5 repeated runs across a pool of blind, comparative and branded queries. Switching the test environment during a monitoring campaign — moving from API to web interface, for instance — is prohibited, as interfaces apply hidden system prompts and search plugins that undermine comparability.
Third, the content of AI answers must be verified, not just the fact of mention. The illustrative case: a company raised its delivery price from 2,000 to 3,000 roubles but left old figures on secondary pages, and the assistant continued quoting 2,000 roubles to prospective customers.
To prevent this, a company builds a digital passport — a single verified fact register covering official registration details checked against state databases (EGR/MNS) with an update date, a full service list with exact specifications and conditions, documented projects with confirmed metrics, and answers to commercial questions such as leasing terms, warranty periods and on-site service conditions. AI answers are validated against this reference at every acceptance check; any discrepancy between the model's output and the passport is logged in a discrepancy register and routed back to the source for correction.