Rand Fishkin, the founder of SparkToro and a titan in the world of search engine optimization, recently published what many are calling the most critical piece of primary research the AI visibility industry has seen to date. In collaboration with Patrick O’Donnell, Fishkin’s study meticulously dismantles the long-held assumption that AI tools function like traditional search engines with stable, predictable rankings.
His core conclusion is striking: AI models produce wildly inconsistent brand recommendation lists. This variability is so high that the very concept of a “ranking position” in an AI world is effectively meaningless. While many in the marketing world were stunned by these findings, the research highlights a deeper, more structural reality about how Large Language Models (LLMs) operate. They are not deterministic lookup tables; they are probability engines. Fishkin’s data proves the problem, but to solve it, we must look deeper into the mechanics of “confidence” and how AI systems build trust in a brand.
The Death of the AI Ranking Position Myth
For decades, SEO professionals have obsessed over “Rank #1.” Whether it was on Google or Bing, the goal was to secure a specific spot on a page. When ChatGPT, Claude, and Gemini emerged, marketers naturally tried to apply this same logic. They wanted to know: “How do I rank #1 in ChatGPT?”
Fishkin and O’Donnell’s research proves that this question is fundamentally flawed. They ran 2,961 prompts across the leading AI platforms, focusing on brand recommendations across 12 distinct categories. The results were chaotic. Fewer than 1 in 100 runs produced the same list of brands, and fewer than 1 in 1,000 produced the same brands in the same order. As Fishkin puts it, treating these platforms as deterministic ranking systems is “provably nonsensical.”
However, Fishkin also discovered a pattern within the chaos. While the specific “rank” was inconsistent, some brands appeared much more frequently than others. This led to a shift in focus from “rank position” to “visibility percentage.” If a brand shows up in 95% of queries for a specific category, it is a dominant player, regardless of whether it appears first or third in a specific session. This variance is where the real story of AI optimization begins.
Why AI Recommendations Are Inconsistent
To understand why Fishkin’s lists changed every time he hit “enter,” we have to understand that AI platforms are confidence engines, not recommendation engines. When you ask ChatGPT for the “best cancer care hospitals,” it doesn’t search a database. Instead, it generates a response based on a probability distribution shaped by three key factors:
- What the model “knows” from its massive training corpus.
- How confidently it knows that information based on the weight of the data.
- What specific information it retrieved or “grounded” itself with at the exact moment of the query.
When a model is highly confident about an entity’s relevance, that entity appears consistently. For example, in Fishkin’s data, “City of Hope” appeared in 97% of cancer care responses. This isn’t luck; it’s the result of deep, corroborated, multi-source presence in the data the AI consumes. Conversely, brands that appear only 5% to 10% of the time reside in a “low-confidence zone.” The AI knows they exist, but it hasn’t found enough corroborating evidence to commit to them consistently.
The Framework of Cascading Confidence
To fix the inconsistency problem, brands must move from the “inconsistent pile” to the “consistent pile.” This requires navigating what is known as the “Cascading Confidence” framework. This is a multi-stage pipeline—formalized as DSCRI-ARGDW—that every piece of content must pass through before it can influence an AI recommendation.
The pipeline consists of ten distinct gates: Discovered, Selected, Crawled, Rendered, Indexed, Annotated, Recruited, Grounded, Displayed, and Won. At every single stage, the AI system asks: “How confident am I in this content?”
The Multiplicative Nature of AI Trust
Confidence in an AI system is not additive; it is multiplicative. This is a crucial distinction that many marketers miss. If a brand has 90% confidence at each of the ten stages, the final end-to-end confidence is not 90%—it is 0.9 raised to the tenth power, which equals roughly 35%. If confidence drops to 80% per stage, the total confidence plummets to 11%.
One single failure point—such as a website that is slow to render or has inconsistent information—can destroy the entire “bid” for an AI recommendation. This principle was echoed years ago by Google’s Gary Illyes, who noted that a zero on any single ranking factor kills the entire ranking bid. In the age of AI, this “cascading confidence” is what determines whether your brand is a 97% “City of Hope” or a 5% “also-ran.”
The Three Graphs Model: How AI Sees the World
AI systems do not rely on a single source of truth. Instead, they pull from three different knowledge representations simultaneously. Understanding how your brand lives within these three “graphs” is the key to achieving universal visibility.
1. The Entity Graph (Knowledge Graph)
This is a database of explicit entities and their relationships. It contains binary, verified facts. Either a brand is in the knowledge graph, or it isn’t. This graph has low “fuzziness.” It is the foundation of identity.
2. The Document Graph (Search Engine Index)
This is the traditional territory of SEO. It consists of annotated URLs and ranked pages. It has medium fuzziness. AI models use this graph to “ground” their answers in real-time web data to prevent hallucinations.
3. The Concept Graph (LLM Parametric Knowledge)
This is the learned association within the model itself. It is where “fuzziness” is highest and where Fishkin’s documented inconsistency originates. This graph is built during the training phase and represents the AI’s internal “understanding” of a topic.
Brands that achieve near-universal visibility are present across all three graphs. They have a strong presence in the Knowledge Graph, high-ranking authoritative pages in the Document Graph, and deep encoding in the Concept Graph. If a brand is missing from one, the AI hedges its bets, leading to the inconsistency Fishkin observed.
Crossing the Corroboration Threshold
There is a specific point where AI behavior shifts from hesitant to assertive. This is the “corroboration threshold.” Below this threshold, the AI uses “hedging” language—it might say a brand “claims to be” a leader rather than stating it “is” the leader. It will include the brand in some outputs but omit it in others.
Research across 73 million brand profiles suggests that the threshold for consistency is approximately 2 to 3 independent, high-confidence sources corroborating the same claim made by the brand’s “Entity Home” (its canonical website). These sources must be high-authority, such as Wikipedia, major industry databases, or top-tier media outlets.
Without this corroboration, the AI lacks the “certainty” required to include the brand every time. This explains why broad categories like “science fiction novels” produce diverse lists—there are thousands of options with thin corroboration. In contrast, narrow categories with dense corroboration produce much more stable lists.
Why You Can’t “Fake It” in the Age of AI
Some marketers have attempted to “game” the system by flooding the web with AI-generated mentions or fake expertise. A study by Authoritas in December 2025 titled “Can you fake it till you make it in the age of AI search?” put this to the test. They followed a case where a company created 11 entirely fictional experts, complete with AI headshots and faked credentials, and seeded them into 600 press articles.
The result? Across nine different AI models and 55 topic-based questions, not a single fake expert appeared in an AI recommendation. Despite the high volume of mentions, the fake personas lacked “cascading confidence.” They existed only in the Document Graph (the press articles) but had no presence in the Entity Graph or Knowledge Graph. The AI models were smart enough to see through the surface-level noise. AI evaluates confidence; it doesn’t just count mentions.
The Accelerating Gap: The Rich Get Richer
One of the most alarming findings from the Authoritas data is how quickly the “top” of the market is pulling away. They used a metric called the Weighted Citability Score (WCS) to measure how much AI engines trust and cite specific entities. Between December 2025 and February 2026, the concentration of AI citability among the top 10 experts in digital marketing increased by 92%.
Even though the total number of experts being cited grew, the leaders saw their dominance compound. This is a “flywheel” effect. Confident entities generate confident AI outputs, which lead to higher user trust and more positive engagement signals, which then further reinforce the AI’s confidence in those entities. For brands and experts who aren’t actively managing their digital footprint, the gap is widening at an exponential rate.
How to Fix Inconsistent AI Visibility
Knowing that AI recommendations are inconsistent is the first step. The second step is taking practitioner-level action to stabilize your brand’s presence. Here is how to move from being an “inconsistent sample” to a “reliable recommendation.”
1. Optimize the Entity Home
Your “Entity Home” is the canonical web property—usually your main website—that anchors your brand in the digital world. It must be semantically clean, use structured data, and state clearly who you are and what you do. If your own site is ambiguous, you are training the AI to be uncertain about your brand.
2. Build Independent Corroboration
Identify your most critical claims and ensure they are confirmed by at least three independent, high-authority sources. This isn’t about general PR; it’s about strategic entity corroboration. You need the AI to see the same “fact” about your brand in multiple places it already trusts.
3. Focus on Pipeline Integrity
Traditional technical SEO matters more than ever in the age of AI. If an AI’s crawler cannot render your JavaScript correctly or if your site structure is a maze, your “confidence score” will drop at the very beginning of the pipeline. You cannot optimize the “output” (the recommendation) if you haven’t optimized the “input” (the discovery and rendering).
4. Manage the Algorithmic Trinity
Don’t just focus on “links” for the Document Graph. Work on your Knowledge Graph presence through Schema.org and database entries. Focus on your Concept Graph presence by ensuring your brand’s narrative is consistent across the entire corpus of data the AI is likely to train on. If these three graphs are in alignment, inconsistency vanishes.
The Future of Algorithmic Education
Rand Fishkin has provided the industry with a necessary reality check. The era of “ranking” is being replaced by the era of “visibility and confidence.” Measuring your visibility percentage is a good start, but it is merely a symptom. The underlying “condition” is the confidence the AI has in your brand’s data.
As AI systems become more integrated into our daily search habits, the brands that win won’t necessarily be the ones with the most content. They will be the ones that have systematically built “cascading confidence” across the entire algorithmic pipeline. The gap between the trusted and the ignored is widening. Fishkin proved the problem; now it is up to marketers to implement the solution.
The work of building this confidence is an ongoing discipline, not a one-time project. In a world where AI recommendations can change with every click, the only true defense is a foundation of unshakeable, corroborated, and technically sound brand authority.