LLM Industry Benchmark 2026: Which Sources Shape AI Education Recommendations?

AI assistants like ChatGPT, Gemini and Google AI Overviews increasingly answer study and funding questions before anyone visits a provider’s website. Which sources those answers draw on is therefore a relevant question for education providers. Peak Ace analysed 1,000 standardised prompts across ten industries to investigate this. This article looks at the education lens of the benchmark: which source types AI systems access for education questions, how this compares with other industries, and what it may mean for content and Digital PR.

Note: This article is an excerpt from the LLM Industry Benchmark. The complete analyses, including results for all ten industries, are available in the benchmark itself.

In short:

• In our benchmark, comparison formats lead for comparison and pricing questions in education.
• For recommendations, product pages lead in education, while comparison pages lead across all ten industries.
• For criteria and guidance questions, how-to guides come first in both views.
• The education sources mix differs from the cross-industry picture. This is an observation for this prompt set, not a general rule.
• All findings refer to this prompt set, this time period, and this setup.

Methodology and data corpus

The benchmark is based on a clearly defined data corpus. The following sections describe its structure and the key figures derived from it.

Prompt set and structure

The prompt set comprises 1,000 standardised, brand-neutral bottom-of-funnel prompts. There are 100 commercially relevant prompts per industry, distributed across ten topic clusters. The prompts are geared towards decisions people make shortly before enrolling, applying for funding, or booking.

Search intent

Each industry covers four search intents, with 25 prompts per intent:

• Best/Top/Test Winner: Recommendation questions, such as “Which programme is best for…?”
• Comparisons: Contrasts, such as “Which is better, A or B?”
• Prices/Where to buy: Questions about costs and sources, such as “Where can I apply for…?”
• Criteria/Advice: Questions about selection criteria, such as “What should I check before choosing…?”

Market, language and time period

The prompts are in German and geared towards the German market. The survey was conducted from 1 to 31 August 2026.

Systems and repetitions

Gemini, ChatGPT and Google AI Overviews were observed. Each prompt was executed 15 times, five times per system. Individual responses are treated as samples, which allows for a better classification of fluctuations.

Data source

The data comes from Peec AI exports. Fifty brands per industry were monitored. The report focuses on the top 20 brands by visibility.

Key figures

• Visibility: How frequently an entity appears in the responses to the prompt set.
• Share of Voice (SOV): The proportion of an entity among all mentions in the observed dataset.
• Sentiment: The tone of the responses, rated on a scale of 0 to 100. It is not a quality rating of the entity.
• Retrieval: A source that has been accessed. It does not automatically mean that the source was cited in the final answer.

Search intents and source patterns

Search intent appears to influence the type of source that shapes a response. The patterns differ depending on the industry, so the industry serves primarily as a starting point.

The education lens

Education reaches the following averages in the benchmark: visibility of 4.60 percent, SOV of 3.50 percent, and sentiment of 53.9. Across all ten industries, the averages are 6.94 percent, 3.72 percent, and 54.58.

Source types in education and across industries

The table below compares the leading source type per intent in education with the aggregated picture across all ten industries. The cross-industry figures do not refer to a single industry.

Three patterns stand out:

• Comparisons lead in several intents. Across all industries, comparison formats lead three out of four intents. In education, they lead comparison and pricing questions and appear among the leading formats for recommendations.
• Advice questions follow a different pattern. For criteria and guidance questions, how-to guides lead in both views. Their share of retrieval is 22.33 percent across all industries.
• Product pages support recommendations. In education, product pages lead recommendation questions and appear among the leading formats for all four intents.

The pattern is not universal. There is no single content type that dominates everywhere, and this may indicate that the mix of sources needs to be checked for each industry and intent.

Retrieval means that a source has been accessed. This does not automatically imply that it will be cited or that the provider will appear in the final answer. The format types describe content formats, not ownership or authority.

AI Visibility focus

An average value does not show whether mentions are spread across many entities or concentrated on a few. For education, the benchmark therefore reports two concentration metrics for the top 20, both based on SOV shares normalised to 100 percent:

• HHI: The concentration of SOV shares. Higher values indicate a more uneven distribution. For education, the HHI is 833.
• Top 3 share: The combined share of the three largest SOV contributions. For education, this is 39 percent.

Around two fifths of the SOV within the top 20 falls on three entities. This is a moderate concentration signal for this setup and this observed group. It describes the distribution of mentions, not market share.

In this prompt set, the three entities with the highest visibility are BAföG (13 percent), KfW (9 percent), and Erasmus (6 percent). This is a snapshot of the observation period, not a shortlist of universities or course providers, and not a ranking of institutions. All three are funding and programme-related entities. This may indicate that for financing questions, funding schemes are often directly relevant to the answer.

Concentration scores help to understand the distribution of mentions. However, they do not explain why a particular entity appears. For that, it is worth looking at the questions behind the results and the pages that were accessed.

What this means for education providers and Digital PR

The data is not proof of impact. It offers initial observations that may indicate where SEO and GEO teams should start with baseline work. The following practical implications are based on this snapshot:

• Handle answer types separately. A comparison should explain alternatives, as well as advantages and disadvantages. A guide should clarify requirements and next steps. A general article rarely fulfils both tasks equally well.
• Keep costs, deadlines, and conditions consistent. Comparisons lead pricing questions. Check whether costs, eligibility, deadlines, and conditions are described identically on your own pages and in relevant comparison sources.
• Explain the criteria and link them to programmes. How-to guides and articles dominate criteria questions. Selection criteria can be explained clearly and linked to suitable programmes, so the advice and the offer do not diverge.
• Prioritise your Digital PR. Identify the most important third-party sources per intent and check what they say about your organisation. Not every publication is equally relevant. Use the source mix as a starting point for a prioritised list of domains and formats.
• Keep information consistent. Programme names, eligibility criteria, fees, and deadlines should match across your website and external sources. Outdated or conflicting information increases the risk of inaccurate statements in generated answers.
• Do not forget the basics. Key pages should be accessible, clearly structured, and easy to navigate. This is a prerequisite for information access, but not a guarantee of recommendations in AI-generated answers.
• Treat monitoring as an ongoing task. Brand mentions and retrieved sources can be observed separately, ideally with a consistent prompt set. Analysis by category, intent, and system complements traditional search and enrolment metrics.

Limits of expressive power

The results are valid only for this specific prompt set, time period, and setup. They do not represent a complete measurement of an industry’s or provider’s visibility in AI systems. In particular:

• The prompt set shapes the outcome. Different questions or wording can lead to different answers.
• Individual responses are samples. Small differences should not be overinterpreted.
• Three systems are not one system. Gemini, ChatGPT, and Google AI Overviews cannot simply be equated.
• Retrieval is not citation. The source analysis shows accessed pages, not the sources that were explicitly cited in an answer.
• Knowledge and search are not separate. No causal mechanisms can be derived from the source mix alone.
• Dialogue and personalisation are missing. Typical enrolment and funding decisions cannot be reliably described on this basis.
• Sentiment is not a quality assessment. It describes the tone of the responses, not the quality of an institution or programme.

Read the full benchmark

This article provides an excerpt. The complete LLM Industry Benchmark includes, among other things:

• The source mix for each of the ten industries and all four intent types
• Visibility, SOV, and sentiment within the benchmark
• HHI and top-3 share within the top 20 of each industry
• The leading entities observed per industry
• The cross-industry shares of the source types
• The industry scorecards and the complete methodology

Download the full LLM Industry Benchmark

Had a look at the benchmark and want to strengthen your education provider’s presence in AI answers? Get in touch with Peak Ace’s LLMO and GEO team.

FAQ: AI Visibility, Benchmarks and GEO in Education

What is a benchmark?

A benchmark evaluates defined results against a fixed test set. The LLM Industry Benchmark examines brand mentions, responses, and retrieved source types for commercial questions across ten industries.

What does the LLM Industry Benchmark measure?

It captures visibility, SOV, sentiment, concentration, and retrieval patterns for a defined prompt set, market, and time period. The values are aggregated across Gemini, ChatGPT, and Google AI Overviews.

Which sources do AI systems access for education questions?

In our benchmark, product pages lead recommendation questions, comparisons lead comparison and pricing questions, and how-to guides lead criteria and guidance questions. This applies only to the prompt set and period examined, and retrieval is not the same as citation.

Does AI visibility depend solely on one’s own website?

Comparisons, articles, guides, and other third-party formats play a significant role. It is therefore worth checking which specific URLs shape the answers and how they differ from your own content.

Is AI visibility relevant for education providers?

A provider or programme named in an AI answer may become part of a learner’s shortlist. The benchmark helps to examine source patterns, but it does not measure enrolments or the effect of a specific investment.

What is a GEO strategy?

Generative Engine Optimisation (GEO) combines SEO, content, Digital PR, and monitoring to improve visibility in AI-generated answers. Measures are best prioritised according to the questions the target audience asks and the sources it accesses.

How can providers test whether measures are effective?

In three steps: a source and content audit, a check of the timeliness and consistency of the information, and a remeasurement with the same test questions under the same conditions.