Where AI gets its answers. The anatomy of 15 million citations

At the end of 2025, Russia's non-resource, non-energy export reached $156.8 billion. Development institutions, export centers, and large-scale government support programs work toward this goal. Massive capital is invested in logistics and promotion infrastructure. But market competition is shifting to artificial intelligence interfaces. If an algorithm does not see a company, the company does not exist for the global market. This is a new type of economic filter.

When a language model answers a user, it builds the response from a narrow set of sources. Most large corporations do not participate in shaping this set. It operates by rules that direct advertising budgets cannot change.

The GolOps Lab research center measured this field. We analyzed 15 million citations from actual responses, 1,174 tracked brands, 265,000 domains, and over a million unique URLs. The data comes from live outputs of commercial systems that form algorithmic choice for international business right now.

1. Problem statement and algorithmic invisibility

Modern systems do not have to explicitly exclude Russian enterprises. It is enough for their data to be fragmented and poorly indexed on authoritative platforms in third countries. Algorithmic invisibility acts as a hidden mechanism for long-term exclusion of an object from the global field of future decisions.

Previously, the process of choosing a contractor began with traditional search. Now, 49% of procurement teams are already piloting generative models to automatically narrow down options. AI systems compile a preliminary shortlist. If a corporation falls out of this layer, subsequent marketing becomes completely useless. The decision of whom not to invite to a tender is made by a machine before direct negotiations begin.

We link this process to the mathematics of the economics of trust. Researchers Boryana Dimitrova, Daniel Korschun, and Yoto Yotov proved a direct connection between country reputation and export volume. Every 1-point improvement in a country's position in the global trust ranking leads to a real 2% increase in exports to the target jurisdiction.

GolOps transfers this strict logic to machine-mediated choice. The Choice Control Index (CCI) is a measurable proxy for digital environment advantage. We measure how often an object appears in recommendations, in what specific position, with what degree of argumentation confidence, and against which actual market alternatives.

The CCI shift ties directly to the economic contour. We factor in market size, the share of algorithmic influence on decision-making, the probability of entering the consideration set, and final conversion into deals. For the government sector, this is a matter of protecting national projects and preserving export influence in BRICS+ countries. For large private business, it is a matter of contract acquisition cost, which multiplies when digital visibility is lost.

The main risk is that the foreign market stops considering you altogether. Every missed deal goes to competitors who built an infrastructure of influence in time.

2. Observation infrastructure. How we calculate indices

Our approach relies on a strict separation of raw facts and the calculation layer. We do not rely on analysts' subjective feelings or outdated advertising metrics.

Measurements occur at several levels of detail. The Choice Control Index is calculated for a single specific business. The Sector Index shows the aggregate position of an industry. The RU Export Index aggregates data across the country's entire non-resource export basket. The Cultural Index and Labor Market Index complement the macro perspective.

We use a rolling 90-day window. Every day, the Command Center captures 1.2 million signals from seven key systems. The list includes ChatGPT, Perplexity, Gemini, Claude, DeepSeek, YandexGPT, and Google AI Search. Data comes exclusively from live outputs that real users receive in production interfaces. We do not use simulations or detached laboratory environments.

Two key metrics form the foundation of the calculation layer. The Participation Index captures the sheer fact of making the final shortlist. The Win Index measures argumentation strength and position relative to direct competitors. The intersection of these values produces the final CCI.

For example, PhosAgro's CCI is 12 out of 100 possible points. This means that among 100 target choice scenarios in international systems, the object appears in recommendations in about every eighth answer. The presentation remains neutral, lacking strong argumentation in favor of closing a deal. Such numbers translate vague talk about image into a strict managerial metric understandable to a board of directors.

3. Map of key citation sources

Fifteen domains with the highest number of citations over 30 days:

This is not a list of traditional media or opinion leaders. The top positions belong to structured knowledge bases, video platforms, review aggregators, and highly specialized directories. AI turns to places where data is marked up and verifiable. Brand loudness without this strict structure does not make the cut.

4. The power law of link distribution

The full database of 15 million links aligns into a classic power-law function. The facts completely refute the popular hypotheses of traditional marketing agencies.

The Wikipedia platform takes 17% of the entire visible output field, and the en.wikipedia.org domain alone accounts for over 4% of all URL citations. One resource holds a sixth of all mentions. It is the primary directory for forming basic factual answers.

The next source is half the size. By the twentieth position, the share drops below 0.15%, and by the hundredth it falls past 0.06%. Beyond that lie tens of thousands of small sites, each gathering fractions of a per mille. The remaining 83% of the flow is shared among more than 60,000 domains. Real competition for visibility unfolds in the 0.01% to 0.1% band. Large industry resources, niche directories, corporate blogs, and specialized aggregators live in this zone. This is the controllable field of choice.

Most resources at the top of the distribution hold their ground firmly for six months. The window to change the situation is much shorter than strategy directors assume. In a few years, the field will solidify completely, and the cost of market entry will multiply.

5. Debunking the social platform phantom

We categorized the 15 million links into clear groups. The result dismantles the conventional promotion picture pushed by consultants.

The industry resources category takes 86.5%. Specialized sites, independent directories, and B2B platforms form the actual database for answer generation. This is the citation infrastructure that companies must bring under systemic control.

Meanwhile, an industry myth persists that forums and user-generated content are the key to success. The data shows a completely different picture. All social platforms combined yield only 4.7% of mentions. Reddit accounts for a meager 0.96%. Directories bring in 4.5%, but they do fundamental work. A single call to a structured encyclopedia in a factual query completely shapes the structure of the final text.

Global language models use forums exclusively for subjective queries from individuals. In strict corporate scenarios, machines rely on databases with verified structure and predictable markup. A 95% field operates under laws where emotional user discussions play no role. Building a corporate presence architecture on the social layer is pointless.

6. Structure of intents and algorithmic levers of influence

We analyzed 23,093 unique choice scenarios. The distribution clearly demonstrates actual user intent in a commercial environment.

Every third query requires the system to compile a structured list of candidates. The probability of your presence there depends strictly on proper formatting of source materials. Data reveals four mathematical amplification metrics working right now.

Brand presence. A page mentioning a brand is cited 1.5 times more often than a page without it. This rule requires tight integration of your name into authoritative third-party contexts. The system must see confirmation of your expertise from several independent sources.

List headings. A ten-point format is cited 20% more often than standard product pages. Machines find it exceptionally convenient to extract ready-made fragments from them to form the final answer for the client.

Comparative formats. Match-up pages receive an algorithmic boost of 1.1x. Nearly a tenth of queries are directly comparative. Systems purposefully look for materials formatted to contrast features.

Instructions and data freshness. Practical guides get a 1.1x amplification factor. Including the current year in the heading provides a similar increase in material usage frequency. In the technology and financial sectors, 25.3% of links point to materials no older than 60 days. For a demanding corporate audience, older data is automatically treated by the machine as a dead asset.

7. The mechanics of economic loss

When an object falls out of the consideration set, four levels of damage arise for a business or the state. This is a direct financial loss, not an image problem.

The first level of loss lies in lost revenue. If a company does not appear in the answers of systems used for market analysis and preliminary comparison of alternatives, it stops receiving inbound inquiries. The deal goes to a competitor who took a spot in the machine answer in time. The displacement map clearly shows exactly who takes your contracts. Regional manufacturers often yield to international giants simply because the latter structured their data better.

The second level brings an increase in acquisition cost. If the client did not see you in artificial intelligence recommendations, the sales department has to break through a wall of distrust with cold calls and expensive advertising. The cost of acquiring a single contract increases several times over.

The third level creates a price discount. A supplier unknown to algorithms is forced to dump prices to compensate for the trust deficit. You lose profit margin on every new contract.

The fourth level hits strategic subjectivity. In an extreme scenario, systematic absence in algorithms leads to the status of a non-existent company in the global market. You voluntarily give up your place in the future architecture of decision-making.

8. Counter-infrastructure of influence. How to turn risk into an asset

Attempts to solve this problem through standard PR agency work are guaranteed to fail. It requires an architecture of digital presence.

GolOps turns this opaque zone into a controllable system. The process is built in a strict sequence, eliminating chaotic marketing actions. We take the position under full control.

The first step is Index Diagnostics. The initial research identifies specific points of algorithmic invisibility. We provide a detailed displacement map showing all competitors occupying your rightful place in the output. Index Diagnostics takes about 10 business days and forms the basis for board-level decision-making. For the government sector, we apply the Strategic Diagnostics format, which accounts for the specifics of national interests.

The second step opens the Strategic Pilot. We execute a full work cycle in 10 weeks. The mechanics include recording the initial position, developing an adjustment plan, making changes to the trust infrastructure, re-measuring, and preparing an executive document. Management receives confirmed figures proving the position shift. You scale only a strong and verified position, not a mistake.

The third step transitions the client to the Command Center. This is a continuously operating monitoring infrastructure. The system collects new signals daily, conducts deep attribution of every model answer, and ranks actions by expected market impact. The platform includes quality control modules and a built-in company knowledge base. All data is stored on servers in strict compliance with the federal security standards of large corporations.

Summary. The cost of inaction

Inaction today turns into direct financial losses tomorrow. A company outside the zone of systemic control disappears from the consideration set at the very first selection stage.

According to the Gartner research agency, 90% of corporate purchases will be fully delegated to autonomous agents by 2028. Conversion from new channels already exceeds classic organic search figures by 4.4 times, according to Semrush platform estimates. We estimate current losses from weak algorithmic representation in the export sector at $47.3 billion annually.

The shortlist is formed without you. The procurement scenario runs without you. The recommendation landing in the foreign buyer's chat window is assembled from sources where your brand name is not even mentioned. Every quarter without a properly built index infrastructure results in decisions made without your presence in the meeting room.

Discuss pilot →