The Paradox of AI Search & Efficiency: How Efficient Is Search Today?

In classic search, we know a phenomenon that any SEO consultant can explain in their sleep: SERP overlap. Hundreds of different keywords lead to the same ten results. The search results pages overlap, rankings are stable, the world is predictable.

In AI search, the opposite is true, and that is exactly what makes it as fascinating as it is frustrating.

The SERP Overlap Phenomenon in Classic Search

The concept of SERP overlap is foundational to how the traditional search ecosystem was built. When a user searches for any of hundreds of related keywords, Google returns largely the same set of ten blue links. This stability is not accidental. It is the direct result of a deterministic index, a fixed, crawled and ranked database of URLs that produces consistent, reproducible results for any given query.

This predictability became the bedrock of an entire industry. Rank trackers, keyword research tools, visibility indices and click-through-rate benchmarks all rest on a single assumption: that the same query, entered at roughly the same time, will return the same result. Because that assumption held, it became possible to measure, compare and optimise with precision.

The numbers from 2025 and 2026 show how dramatically this foundation is shifting. In February 2024, Gartner predicted a 25 percent decline in traditional search volume by 2026 (Gartner, Feb 2024). By early 2026, the actual data painted a picture that was, in several metrics, more severe than the forecast. Google traffic to publishers dropped by 33 percent according to Chartbeat and Press Gazette (Chartbeat/Press Gazette, Jan 2026), compared to the predicted 25 percent. The organic click-through rate in the presence of AI Overviews fell by 61 percent, and the paid click-through rate fell by 68 percent (Seer Interactive, Sep 2025). Zero-click searches reached 69 percent of all queries (Similarweb, Mai 2025).

The paradox is striking: Google is not necessarily losing users at the rate predicted, with searches per user in the United States declining by approximately 20 percent, slightly below Gartner's forecast (SparkToro/Datos, Jan 2026). But users are clicking dramatically less. The deterministic index still exists. The SERP overlap still exists. What has changed is that users increasingly never leave the results page at all.

Why AI Answers Are Non-Deterministic

When you ask the same question in ChatGPT, Perplexity or Claude, you receive a different answer every time. Not just worded differently, but with different sources, different structure and different recommendations. There is no fixed index, no stable ranking, no SERP position that could be tracked in any conventional sense.

In classic search, the rule is: many keywords lead to similar results. In AI search, the rule is: one prompt leads to many results. That is not a bug. That is the architecture.

While Google relies on a deterministic index, large language models work probabilistically. Every answer is a probability distribution over possible token sequences. Even minimal changes in context, one extra word, a different point in time, a different model, shift the entire output.

The Eight Invisible Factors Behind Every AI Answer

Even with an identical query, from the same user, on the same platform, there are dozens of invisible paths that influence the result. SparkToro founder Rand Fishkin summarises the implication: to validly represent the recommendations of an AI system, the same question must be asked at least 60 to 100 times and the results averaged (SparkToro Research, Jan 2026). The eight decisive factors are:

1. Context: Whether an active conversation history exists or a fresh chat has been started. Previous messages fundamentally alter the weighting of the answer.
2. Tools: Whether the model has access to web search, code execution or file uploads. Each activated tool changes the output, often in unexpected ways.
3. Policy filters: Content guidelines can shorten, rephrase or block answers. Different filters apply depending on the topic.
4. Streaming: Whether the answer is streamed or generated as a block. The output method can subtly influence token selection.
5. Temperature: The randomness factor in token selection. Critically, even at temperature zero, no LLM provider guarantees determinism. OpenAI, Anthropic and Google explicitly point this out (Šubonis, MLOps Shenanigans, Apr 2025; Thinking Machines Lab, Sep 2025). The causes lie in floating-point arithmetic, sparse mixture-of-experts architectures and batch processing.
6. Cache: Whether a cached response is served or a new calculation is performed. Cached answers are identical, new ones vary, but the user cannot see which case applies.
7. Language and region: German versus English, Europe versus the United States. Sources and recommendations shift completely, often with different sentiment.
8. Answer length: A short answer versus a detailed explanation. Length influences which details, and critically, which brands are mentioned.

What Non-Determinism Means for Measurement

Many organisations are beginning to measure their AI visibility. This is understandable: if users are increasingly searching in ChatGPT rather than Google, brands want to know whether they appear there. But anyone who engages seriously with the tracking challenge quickly realises that the complexity is orders of magnitude higher than anything known from classic SEO.

The Mathematics of Scenario Explosion

To even approximately represent the result variance of an AI system, the following dimensions must be considered:

Scenarios = Login status × Plan × Platform × Version × Model × Feature flags × Prompt variants

Calculated concretely: login status has three states, plan or tier has four, platform has five, version has three, model has six, feature flags have eight, and prompt variants number fifty or more. The result is more than 4.3 million scenarios per prompt (Ziptie Research, Jun 2025).

To put this in perspective, classic Google rank tracking involves roughly ten scenarios per keyword. SERP features such as snippets and People Also Ask expand that to around fifty. Local SEO with geo-grids reaches approximately five hundred. AI search tracking, by contrast, involves more than 4.3 million scenarios, a complexity factor of 432,000 times that of basic rank tracking (Ziptie Research, Jun 2025).

This means that if an AI tracking tool tests 100 or even 1,000 prompts, it is covering the absolute tip of the iceberg, a coverage of approximately 0.02 percent of the scenario space. It is a sampling attempt in a space that is mathematically almost impossible to map completely.

URL Consistency as Evidence

The empirical data confirms this theoretical picture. In Google's AI Mode, across 10,000 keywords, only 9.2 percent of URLs repeated themselves (Ziptie Research, Jun 2025). This figure alone illustrates why the concept of SERP overlap has no equivalent in AI search. There is no overlap. There is not even a SERP.

The business context makes the measurement challenge more urgent. According to McKinsey, 50 percent of consumers actively use AI search engines for purchasing decisions (McKinsey AI Discovery Survey, Aug 2025). Revenue flowing through AI-assisted search is projected to reach significant scale by 2028 (McKinsey, Okt 2025). Yet only 16 percent of brands systematically track their AI search performance (McKinsey CMO Survey, Sep 2025). Substantial budgets are already being spent on AI visibility analytics, often without validation of the underlying methodology (SparkToro Research, Jan 2026).

What AI Tracking Can and Cannot Do

The consequence is not that AI tracking is meaningless. But it requires honesty about what it can and cannot deliver.

AI tracking can measure directional visibility, identifying rough trends in whether a brand is mentioned more or less frequently in AI answers. It can assess topic field presence, establishing whether a brand appears at all in relevant subject areas. It can detect outliers, identifying sudden changes in mention frequency. It can reveal recurring source preferences of a model, which can be used strategically. And it can provide a competitive comparison, positioning relative visibility against competitors as a share of voice metric.

What is not possible: a stable AI position equivalent to rank three on Google does not exist. Deterministic rankings cannot be produced by probabilistic systems. Complete coverage of the scenario space is not achievable. There is no guarantee that specific sources will be cited consistently. And AI visibility cannot be directly compared to search visibility as a one-to-one metric. It is a different measurement, from a different system.

What Brands Should Do

The actual strategy lies not in tracking alone but in optimising the fundamentals. Edelman emphasises that earned media is the most important driver of brand visibility in AI answers (Edelman, 2025). McKinsey adds that a brand's own website often accounts for only 5 to 10 percent of the sources that AI search references (McKinsey, Okt 2025). Six levers are available.

1. Strengthen E-E-A-T

Experience, expertise, authority and trust form the basis for visibility in both deterministic and probabilistic systems. Gartner states directly: "Companies will need to focus on producing unique content that demonstrates expertise, experience, authoritativeness and trustworthiness." (Gartner, Feb 2024). This is not a new principle, but its importance is amplified when AI systems must decide which sources to draw on without a fixed index to consult.

2. Maintain Consistent Entities

Brand information should be kept consistent across Wikipedia, Wikidata, Schema.org and Knowledge Panels. AI systems weight consistency across multiple sources as a trust signal. Inconsistent or contradictory information about a brand across these reference points reduces the likelihood of reliable citation.

3. Implement Structured Data

JSON-LD, FAQ schema and How-To schema increase the probability of being used as a source by AI systems. Search Engine Land describes the shift from keyword-based to entity-based measurement as central to the new optimisation paradigm (searchengineland.com, Nov 2025).

4. Track Samples, Not Positions

AI visibility should be measured as a trend, not as an absolute position. SparkToro recommends a minimum of 60 to 100 samples per prompt, averaged across different points in time, to achieve statistical validity (SparkToro Research, Jan 2026). Any tracking approach that tests fewer prompts than this should be understood as indicative rather than conclusive.

5. Establish Digital Distinctiveness

A brand must be defined clearly enough that it appears reliably even in probabilistic systems. This means a distinct positioning, consistent core messages and clear differentiation from competitors. Ambiguity at the brand level translates directly into inconsistency at the AI output level.

6. Build Multi-Platform Presence

Optimising only for Google is no longer sufficient. Content should be placed deliberately on sources that AI systems actively draw from: Reddit, Wikipedia, specialist publications, industry directories and YouTube transcripts. Because the brand's own website accounts for only a small fraction of AI-referenced sources (McKinsey, Okt 2025), the distribution of authoritative content across third-party platforms becomes a primary strategic lever.

A Paradigm Shift, Not a Feature Update

AI search is not a new channel alongside Google. It is a fundamentally different system with fundamentally different rules. Anyone who attempts to measure it with the tools and mental models of classic search will struggle, not because the tools are poor, but because the underlying assumptions no longer hold.

SERP overlap was the foundation on which an entire ecosystem of rank trackers, keyword tools and visibility indices was built. In AI search, that foundation does not exist. There is no ranking. There is no fixed position. There are only probabilities.

The scale of the shift is visible in the data: publisher traffic down by a third (Chartbeat/Press Gazette, Jan 2026), click-through rates collapsing in the presence of AI Overviews (Seer Interactive, Sep 2025), and the majority of brands still without systematic measurement in place (McKinsey CMO Survey, Sep 2025). The question for brands is not whether to engage with AI search visibility, but how to do so with an accurate understanding of what can and cannot be known.

Somewhere, one must begin. But it helps to know how large the space is into which that first step is taken.

Sources

• Gartner (Feb 2024): "Predicts Search Engine Volume Will Drop 25% by 2026"
• SparkToro/Datos (Jan 2026): "Google Searches Per User Down Nearly 20%"
• Chartbeat/Press Gazette (Jan 2026): "Global Publisher Google Traffic Dropped by a Third in 2025"
• Seer Interactive (Sep 2025): "AIO Impact: Organic CTR -61%, Paid CTR -68%"
• Similarweb (Mai 2025): Zero-Click Searches at 69%
• McKinsey (Okt 2025): "New Front Door to the Internet: Winning in the Age of AI Search"
• McKinsey AI Discovery Survey (Aug 2025): Consumer AI search usage for purchasing decisions
• McKinsey CMO Survey (Sep 2025): Brand AI search performance tracking
• SparkToro Research (Jan 2026): "AIs Are Highly Inconsistent When Recommending Brands"
• Edelman (2025): "How Brands Can Stay Visible in an AI-Driven Search World"
• Search Engine Land (Nov 2025): "Measuring Zero-Click Search: Visibility-First SEO"
• Ziptie Research (Jun 2025): "AI SEO Tracking Tools 2026: Comparative Analysis"
• Šubonis (Apr 2025): "Zero Temperature Randomness in LLMs"
• Thinking Machines Lab (Sep 2025): "Defeating Nondeterminism in LLM Inference"