AI search is more cited, more fragmented — and still unreliable as a source of product visibility

The June edition of the Odyssiant AI Search Tracker analysed 4,623 AI-generated answers across four engines, up 23% on the previous period. The evidence layer kept expanding, with 24,870 citations extracted and citations appearing in 80% of answers. But the average visibility score fell to 26/100, and Not Mentioned remained the dominant verdict at 57%. AI search is becoming more observable, but not more favourable.

The June edition of the Odyssiant AI Search Tracker analysed 4,623 AI-generated answers across four engines. That is up from 3,751 in the previous period, a rise of 23%.

On the surface, the evidence layer continued to expand. The tracker extracted 24,870 citations across the full dataset, with citations appearing in 80% of answers. But the overall visibility score fell to 26/100, down from 31/100 in the previous period.

That is the central signal from June.

AI search is becoming more observable, but not necessarily more favourable. More answers are being generated. More citations are being extracted. More domains are appearing in the evidence layer. Yet for most products, the default outcome is still absence.

The most common verdict in June was Not Mentioned, accounting for 57% of all answers. In other words, AI systems were usually able to answer commercially relevant buyer questions without including the product being tested.

That is the visibility gap marketers need to understand.

It is not simply that AI systems describe brands inaccurately. In many cases, they bypass the product entirely.

Key takeaways

• 4,623 AI answers were analysed in June, up 23% from the previous period
• The average visibility score fell to 26/100, down 5 points month on month
• 80% of answers included citations, down from 85%
• The full dataset produced 24,870 citations
• After removing audited company domains, 22,803 non-audited-domain citations remained
• Product audits accounted for 100% of the analysed answers
• 57% of answers did not mention the product being tested
• 12% of answers returned no usable answer, driven largely by Google Gemini 2 Flash
• Early-stage research remained dominant, with Frame & Clarify and Explore Landscape at 57% of prompts
• Compliance, legal, energy, fibre, employment, corporate, security and sportsbook questions were prominent

Product visibility weakened, even as the dataset grew

June’s tracker expanded from 3,751 to 4,623 answers.

By engine, the dataset was weighted towards OpenAI GPT-4o:

The average visibility score was 26/100, down from 31/100 in the previous period.

The engine-level scores show why the overall number moved:

The sharp fall in Gemini’s score reflects a high number of no-answer outcomes in this tracker period. That matters because AI visibility is not stable across engines. A product can appear reasonably well in one AI environment and disappear completely in another.

For marketers, that makes single-engine testing risky. A ChatGPT result is not an AI visibility strategy. It is one view of a fragmented answer environment.

Absence is still the dominant commercial risk

The verdict distribution shows the problem clearly:

The most important figure is not the number of weak mentions. It is the number of missing ones.

In 57% of answers, the product being tested was not mentioned at all. That means the AI system answered the buyer’s question, but did not connect the buyer’s need to the product.

That distinction matters.

A weak recommendation can be improved. A poor citation can be replaced. A comparison gap can be filled with better proof. But a Not Mentioned verdict means the product never made it into the buyer’s AI-assisted research path.

If AI search is becoming a discovery layer, then absence is not a reporting issue. It is a demand problem.

Early-stage prompts still shape the shortlist

The June tracker remained heavily weighted towards the early and middle stages of buyer research.

The first two stages accounted for 57% of the tracker.

This reinforces one of the most important patterns in AI-led discovery: buyers do not begin with a final shortlist. They begin by trying to understand the problem, explore the landscape and work out which kinds of solutions might be relevant.

If a product is absent at that stage, it may never reach the comparison stage.

That is why product pages alone are not enough. Buyers are asking AI systems about problems, risks, trade-offs, costs, compliance requirements, alternatives and implementation considerations. The content and evidence layer has to meet those questions before the buyer is ready to search for a named provider.

The citation layer is wide, but not neutral

Across the full dataset, the tracker extracted 24,870 citations. For this edition, audited company domains were removed from the citation league tables so that the source environment could be viewed more clearly.

After those removals, the tracker contained 22,803 non-audited-domain citations.

The leading cited domains were:

This is not a conventional ranking environment. The evidence layer includes encyclopaedias, community forums, academic sources, publishers, regulatory bodies, analyst firms, professional directories, review sites, social platforms and vendor-owned content.

That makes AI visibility a broader problem than website optimisation.

Owned content still matters. But it is only one part of the evidence system AI engines use when constructing answers.

Source types show the shape of the evidence environment

After removing audited company domains, the June source mix looked like this:

Vendor and other web sources still dominated, but the wider evidence layer remains commercially important.

Government and regulatory sources were prominent in compliance-heavy categories. Analyst and research sources appeared where buyers were trying to understand market direction or validate claims. Community and social sources continued to appear in answer construction, especially where AI systems drew on lived experience, reviews or practical discussion.

This matters because many AI visibility gaps are not caused by a missing blog post. They are caused by a weak evidence environment.

If AI systems cannot find credible proof, third-party validation, comparison material, implementation evidence, customer outcomes or regulatory support, they may choose safer, better-evidenced alternatives.

Engine behaviour remains inconsistent

The top cited domains varied materially by engine.

OpenAI GPT-4o leaned heavily on broad information and research sources, including Wikipedia, Reddit, ScienceDirect, Energy.gov and LegalClarity.

Anthropic Claude 4 Sonnet drew more visibly from institutional, sector and regulatory sources, including the American Bar Association, Ofcom, iGaming Business, SHRM and the American Gaming Association.

Perplexity Sonar showed a stronger tendency towards direct web and social citations, with YouTube, Facebook, Chambers, Legal 500 and Uswitch among its most cited sources.

This is one of the practical lessons from June.

AI visibility is not one channel. It is a set of overlapping answer environments. Each engine has different retrieval behaviour, citation behaviour and tolerance for different kinds of sources.

That means marketers need to measure visibility by engine, prompt type and buyer stage. The useful insight is rarely in the overall score alone. It is in where the product appears, where it disappears, and which sources influence the answer.

Buyer questions are becoming more proof-led

The June prompt set showed a strong presence of compliance, comparison, cost, implementation, security and evidence language.

After removing audited company names from the keyword analysis, the most frequent prompt terms included:

This reflects the way buyers use AI for decision support.

They are not only asking which suppliers exist. They are asking how solutions compare, what evidence supports the claims, which compliance requirements matter, what security proof is available and what trade-offs they should consider.

Examples of the underlying prompt patterns included:

• How do different solutions compare in terms of cost and effectiveness?
• What regulatory requirements should be considered when selecting a solution?
• What security certifications are necessary in this category?
• What independent assessments are used to verify vendor integrity?
• What evidence should buyers ask for before making a decision?
• How should total cost of ownership be compared?
• What proof points support claims about quality, performance or compliance?

This is where AI visibility starts to connect directly with sales enablement.

If the evidence does not exist in a form AI systems can find and use, the answer may still be generated — but the product may not be included.

The Odyssiant view this month

June’s tracker points to five conclusions.

1. More data does not automatically mean more visibility

The dataset grew, but the average score fell. AI search is becoming more measurable, but the measurable result is not always positive.

2. Product absence remains the default

The most common answer is still one that does not mention the product being tested. This is the central commercial issue for marketers.

3. The evidence layer is broader than owned content

AI systems are drawing from thousands of domains. Websites matter, but so do third-party proof, regulatory sources, analyst material, directories, publisher content, reviews and community discussion.

4. Engines behave differently

OpenAI, Claude, Gemini and Perplexity do not produce the same visibility environment. Testing one engine creates false confidence.

5. Proof is becoming part of discovery

Buyers are using AI to validate claims, compare options and reduce risk. That means content strategy, PR, evidence, sales enablement and governance need to work together.

What marketers should do next

The June tracker points to five practical actions.

1. Measure product visibility, not just brand awareness

Test whether your products, services and propositions appear when buyers ask real problem, category, comparison and verification questions.

2. Build for the early journey

Create content that helps buyers frame the problem, understand the landscape and identify credible approaches before they know which vendor to search for.

3. Strengthen the proof layer

Prioritise case studies, quantified outcomes, compliance evidence, security documentation, customer proof and third-party validation.

4. Map the citation environment

Find out which sources AI engines use in your category. Some will be owned. Many will not be.

5. Track by engine and buyer stage

Visibility is fragmented. The important question is not only whether AI mentions you, but where, when, why and with which supporting sources.

Closing view

The June tracker shows an AI search environment that is becoming more measurable, but not more forgiving.

Answers are being generated at scale. Citations are visible. Source patterns can be tracked. Engine differences can be observed.

But most products are still missing from most relevant AI-generated answers.

That is the challenge for marketing teams.

Report generated by Odyssiant AI Search Tracker — June 2026