Ahrefs, Semrush, Webflow and Yext measure different parts of AI visibility. Their findings point beyond backlinks, content volume or any single platform…
The strongest observed relationship with AI visibility in a 75,000-brand Ahrefs study was not backlink volume, Domain Rating or the number of pages on a company’s website. It was YouTube mentions, followed closely by the extent to which the brand was discussed across the wider web.
That is a counterintuitive finding for marketers trained to explain search performance through links, authority scores and content production. It is also easy to misuse.
A correlation of 0.737 between YouTube mentions and AI visibility does not mean that publishing more videos will cause ChatGPT to recommend a brand. Large, successful companies tend to have greater awareness, search demand, media coverage, reviews, creator content, advertising and AI visibility at the same time. YouTube mentions may partly measure something broader: how extensively the brand exists in public discourse.
Five vendor studies now provide enough evidence to ask a more disciplined question. Not ‘Which AEO tactic works?’ but: <strong>what observable conditions travel with brands that appear frequently in AI answers, and what underlying explanation best fits those patterns?</strong>
The current answer is still a hypothesis. AI visibility appears to favour brands that are widely known, specifically relevant, clearly associated with the subject, supported by credible evidence and represented consistently enough for an AI system to connect the brand with the question. Accessibility helps make that evidence usable. It does not guarantee preference.
<h2>Ahrefs found public brand presence, not content volume, at the top</h2>
Ahrefs screened 75,000 domains with a Domain Rating above 40 and a highest-volume keyword of at least 800 monthly searches. It then used Brand Radar to examine millions of responses in ChatGPT, Google AI Mode and Google AI Overviews and calculated Spearman rank correlations between brand visibility and a set of search, web and YouTube variables.
YouTube mentions showed the strongest correlation across the variables tested at approximately 0.737. YouTube mention impressions were close behind at approximately 0.717. Branded web mentions correlated between 0.66 and 0.71, depending on the AI surface. For ChatGPT, branded search volume correlated at 0.352 and Domain Rating at 0.266. Website page count was approximately 0.194.
The pattern is important, but the boundary is equally important. The sample was deliberately screened towards established, searched-for domains, and the analysis was observational. Ahrefs explicitly states that improving a correlated variable will not automatically improve AI visibility.
YouTube is therefore not ‘the secret to AEO’. A mention in a video title, description or transcript can be a form of public evidence. It can also be a proxy for brand scale, cultural presence or marketing investment. The study cannot separate those mechanisms.
What it can support is narrower: <strong>within Ahrefs’ screened sample, broad contextual brand presence travelled with AI visibility more strongly than page count, backlink volume or Domain Rating.</strong>
<h2>The web-mention pattern appeared in Ahrefs’ earlier AI Overview study</h2>
An earlier Ahrefs analysis used the same broad 75,000-domain screening logic for Google AI Overviews. Branded web mentions had the strongest measured correlation with AI Overview visibility at 0.664. Branded anchors followed at 0.527 and branded search volume at 0.392. Domain Rating was 0.326, referring domains 0.295, backlinks 0.218 and website page count 0.17.
Brands in the highest quartile for web mentions averaged 169 AI Overview mentions, compared with 14 in the next quartile. That is more than a tenfold difference in the observed averages, not proof that moving a brand between quartiles would cause a tenfold increase.
Two related analyses from the same commercial provider do not create independent causal proof. They do make the association difficult to dismiss. Brands discussed extensively across the web tended to be substantially more visible in the AI surfaces Ahrefs measured.
Contextual mentions may matter conceptually because they contain more than a connection between two pages. They can express a relationship among an entity, an attribute and a context: <strong>Brand X → Attribute Y → Situation Z.</strong>
Volvo and safety provide a familiar illustration. An AI system answering ‘Which SUV brands are associated with safety?’ needs more than evidence that Volvo has a website or attracts links. It needs accessible information connecting Volvo with safety through product specifications, crash testing, regulation, journalism, reviews, consumer discussion and the company’s own historical communications.
The Volvo example is not a finding from these studies. It illustrates the difference between general awareness and a decision-relevant brand association.
<h2>Semrush found that broad SEO strength did not consistently identify the topic owner</h2>
Semrush and Kevin Indig tracked 1,094 US ChatGPT categories monthly from January through June 2026. Each category contained five representative prompts covering definition, comparison, alternatives, use case and buying questions. The dataset included more than 50,000 brands, 220,000 domains, 600,000 citations and 220,000 URLs.
A ‘category owner’ had to appear in at least four of five prompts and lead the runner-up by at least five percentage points. Only 15.2% of categories met that threshold. Another 31.2% had an emerging leader and 53.7% were unsettled. In the higher-demand half of the categories, only 11.3% had a clear owner, compared with 19% in the lower-demand half.
This does not prove that 85% of topics are easy to win. It shows that roughly 85% did not have a leader meeting Semrush’s five-prompt consistency and margin test during the study period.
Semrush then compared each topic owner with its runner-up. The owner had higher branded search volume in 55.7% of matched pairs, higher organic traffic in 48.4% and a higher Authority Score in 52.5%. Only branded search volume was statistically significant, and the source described that edge as modest.
The finding does not show that SEO is irrelevant. Traditional SEO remains important for access, retrieval and first-party information quality. It shows that broad domain-level strength did not consistently explain which brand owned a specific ChatGPT topic.
That creates a strategically important distinction. <strong>General brand strength can increase the probability of entering the consideration set. Specific relevance may help determine whether the brand belongs in a particular decision.</strong> The second sentence remains an AEO Updates interpretation because Semrush did not experimentally isolate relevance as the missing cause.
<h2>Brand strength still creates a substantial inherited advantage</h2>
The same studies also resist an overly convenient narrative that AI levels the playing field.
Ahrefs found that the same major brands tended to appear across AI surfaces. Correlations between brand mentions were 0.821 for AI Overviews and AI Mode, 0.749 for AI Overviews and ChatGPT, and 0.769 for AI Mode and ChatGPT.
Webflow’s July 2026 AEO Maturity Index assessed 2,000 live US company websites on 123 gates, queried ChatGPT, Gemini, Claude and Perplexity with buyer questions, and used deterministic code plus LLM-based graders to score the responses.
Large brands in Webflow’s sample had a 23% mention rate, compared with 11% for smaller brands. Their citation rate was 8% versus 5%, and their share of voice was 20% versus 11%. Larger brands were also scored somewhat higher on sentiment, accuracy and message pull-through.
Established brands do not start from zero in AI. They arrive with accumulated awareness, usage, media, reviews, distribution and public conversation. That human brand equity can create a large public evidence footprint that an AI system may encounter through training, retrieval or both.
This should not be interpreted as a closed market. It means challengers should be realistic about the territory they can own.
<h2>Smaller brands may compete through narrower decision relevance</h2>
A smaller company is unlikely to reproduce Nike’s entire information footprint. It may not need to.
‘Best running shoe’ is a broad territory where established brands carry enormous accumulated evidence. ‘Best trail-running shoe for wet technical terrain’ is narrower. A brand that genuinely excels in that use case can build concentrated first-party information, independent testing, expert reviews, demonstrations and customer evidence around a more specific conclusion.
This motivates an <strong>AI Challenger</strong> hypothesis: a brand that receives materially more AI consideration or recommendation than conventional human brand strength would predict. It is a proposed research construct, not a label that can be applied from a few favourable prompts. Testing it would require a declared measure of human brand strength, a stable prompt universe, repeated AI observations and a clear definition of consideration or recommendation.
The opportunity in Semrush’s data is therefore not ‘85% of topics are available’. It is that many tracked categories lacked a consistent leader under the study’s definition. A smaller brand may be able to build a stronger connection to a narrow decision territory, but that causal proposition still requires testing.
<h2>Webflow connects visibility with authority and a wider maturity system</h2>
The median company in Webflow’s study appeared in 16% of answer-engine responses and was cited with a link in 6%. Webflow grouped its 123 gates into Content, Technical, Authority and Measurement maturity.
Companies scored at Level 3 had 2.3 times the mention rate and 3.7 times the citation rate of Level 1 companies. This was a cross-sectional comparison, not evidence that the same companies moved between levels and produced those gains.
Webflow reported on-site authority as the strongest correlate of AEO mention rate. Its authority definition included claims supported with data, sources, credentials and clear authorship. Third-party recognition as an industry leader, thought leader or authority was the next strongest relationship. Seventy-three per cent of companies appeared in fewer than one-quarter of the places cited for their category.
This is a commercial study from a website platform that sells AEO capabilities and offers an AEO assessment. The rubric is proprietary, several qualitative outcomes were LLM-graded, and correlation does not prove that raising one maturity score will create a specific visibility lift.
Even with those caveats, the Webflow findings fit the Ahrefs pattern. Visibility travelled with a broader authority and evidence environment, not simply the number of pages a company published.
<h2>First-party information matters, but volume is the wrong objective</h2>
The Ahrefs page-count correlations do not mean the corporate website has become unimportant. A first-party site remains the most authoritative place for product capabilities, specifications, pricing, availability, terms, documentation and company facts.
The more defensible conclusion is that <strong>useful evidence quality and information architecture matter more than publishing volume for its own sake</strong>. First-party information should be clear, specific, current, credible, extractable and machine accessible. It should also be consistent with what credible independent sources can verify.
<div class="my-8 overflow-x-auto"><table class="w-full min-w-[720px] border-collapse text-sm"><thead><tr class="border-b-2 border-[#20160E]"><th class="px-3 py-3 text-left font-semibold">Evidence function</th><th class="px-3 py-3 text-left font-semibold">What it helps establish</th><th class="px-3 py-3 text-left font-semibold">Likely source types</th></tr></thead><tbody><tr class="border-b border-[#D7C8B5]"><td class="px-3 py-3 font-semibold">Product truth</td><td class="px-3 py-3">Features, specifications and capabilities</td><td class="px-3 py-3">Brand website, documentation, official databases</td></tr><tr class="border-b border-[#D7C8B5]"><td class="px-3 py-3 font-semibold">Commercial truth</td><td class="px-3 py-3">Price, availability and terms</td><td class="px-3 py-3">Brand website, retailers, marketplaces, listings</td></tr><tr class="border-b border-[#D7C8B5]"><td class="px-3 py-3 font-semibold">Independent validation</td><td class="px-3 py-3">Quality, relative performance and credibility</td><td class="px-3 py-3">Journalism, analysts, trade sources, testing, reviews</td></tr><tr class="border-b border-[#D7C8B5]"><td class="px-3 py-3 font-semibold">Expert interpretation</td><td class="px-3 py-3">Professional judgement and technical explanation</td><td class="px-3 py-3">Named experts, professional publications, academics</td></tr><tr class="border-b border-[#D7C8B5]"><td class="px-3 py-3 font-semibold">Lived experience</td><td class="px-3 py-3">Real-world usability, common problems and owner experience</td><td class="px-3 py-3">Reviews, forums, communities, Reddit, video</td></tr><tr><td class="px-3 py-3 font-semibold">Demonstrated evidence</td><td class="px-3 py-3">Whether a product or claim holds up in practice</td><td class="px-3 py-3">Demonstrations, independent tests, video</td></tr></tbody></table></div>
The practical principle is not ‘get on Reddit’, ‘publish on LinkedIn’ or ‘make YouTube videos’. It is: <strong>build the evidence required to support the conclusion, then identify where that evidence can credibly and naturally exist.</strong>
<h2>Yext shows why there is no universal source recipe</h2>
Yext analysed 17.2 million distinct AI citations collected globally through Yext Scout in the fourth quarter of 2025. The study covered Gemini, Claude, Perplexity and SearchGPT, four intent quadrants at location level and seven sectors. The majority of queries were US-based.
Citation-source patterns varied by model, sector and industry. Listings represented 54.53% of distinct URLs, while websites produced 4.31 citation occurrences per URL versus 2.46 for listings. Claude’s reliance on Yext’s ‘Limited Control’ category ran at two to four times the rate of competing models across the seven sectors. SearchGPT cited official hotel websites at 38.1%.
These are citation patterns, not brand-mention rates, recommendations or consumer outcomes. Yext’s categories involve judgement, several industries had small samples, and a Q4 2025 snapshot may not survive later model changes. The study cannot establish why one source type appeared more often.
It does support an operational conclusion: there is no stable universal source mix. The evidence required to answer ‘What time does this Starbucks close?’ differs from the evidence required to answer ‘Which national coffee chain has the best-quality coffee?’ Model, industry, query type and claim all change the relevant source environment.
<h2>Accessibility is a gate, not the entire competitive advantage</h2>
AI systems need to access, parse and sometimes retrieve information before they can use it. Crawlability, indexability, site structure, structured data, entity clarity and machine-readable content therefore remain important foundations.
Accessibility and preference are different questions. <strong>Accessibility determines whether an AI system can use the evidence. It does not necessarily determine whether that evidence is strong enough to make the brand the answer.</strong>
Schema should be treated as part of machine comprehension, not as a proven recommendation lever. The current studies do not show that adding structured data causes ChatGPT to recommend a brand. The same caution applies to freshness. Webflow found stronger-performing companies had fresher information environments, but those companies may also have larger teams, more coverage and more public discussion.
Freshness is best classified as a probable, context-dependent contributor. Backlinks remain important to conventional search and the wider web ecosystem, but Ahrefs found backlink volume much less correlated with AI visibility than branded mentions in its screened samples.
<h2>A plausible architecture, not a proven ranking system</h2>
The studies can be organised into a causal hypothesis without pretending they prove every arrow:
<div class="my-8 border-y-2 border-[#20160E] bg-[#F4EBDD] px-5 py-6 text-center font-semibold text-[#20160E]">Human Brand Equity → Public Evidence Footprint → Machine-Readable Brand Representation → Decision Fit → AI Consideration → Recommendation</div>
Human Brand Equity includes awareness, familiarity, usage and reputation. The public evidence footprint includes brand mentions, journalism, reviews, expert commentary, user-generated content, video and independent validation. Machine-readable brand representation covers entity recognition, claims, associations and category relevance. Decision fit depends on the specific purpose, audience, criteria, location and context.
This is an AEO Updates synthesis of an architecture consistent with the observed associations. The studies do not establish that the sequence is complete, prove the size of any link or show that AI recommendation causes final human choice.
The model does, however, fit the evidence better than a simple ‘publish an article, earn a ChatGPT recommendation’ theory.
<h2>Decision-Relevant Evidence Density offers a testable construct</h2>
A useful working construct is <strong>Decision-Relevant Evidence Density</strong>: the strength, breadth and independence of retrievable evidence connecting a brand to a specific claim within a defined decision territory.
Its dimensions are straightforward. Strength asks how convincing the evidence is. Breadth asks how widely it exists. Independence asks whether it comes from genuinely separate sources. Relevance asks whether it supports the specific decision criterion. Retrievability asks whether an AI system can access and understand it.
The Prompt Group and AEO Updates can use this construct to form hypotheses, but it is not a validated industry metric. The current studies do not identify optimal weights or an ideal first-party/third-party ratio.
Consider two hypothetical safety claims. Brand A publishes 100 pages stating that it is the safest. Brand B has 20 relevant pieces of evidence spread across product specifications, independent testing, journalism, expert reviews, customer experience, regulatory data and demonstrations. Brand B may provide a more credible evidence environment despite having less total content. That is theoretically plausible and consistent with the distributed-mention and authority findings. Direct causal research is still required.
<h2>What the studies support today</h2>
<div class="my-8 overflow-x-auto"><table class="w-full min-w-[760px] border-collapse text-sm"><thead><tr class="border-b-2 border-[#20160E]"><th class="px-3 py-3 text-left font-semibold">Potential condition</th><th class="px-3 py-3 text-left font-semibold">Current evidence</th><th class="px-3 py-3 text-left font-semibold">Confidence in association</th></tr></thead><tbody><tr class="border-b border-[#D7C8B5]"><td class="px-3 py-3 font-semibold">Overall brand strength</td><td class="px-3 py-3">Large-brand advantage across Ahrefs and Webflow samples</td><td class="px-3 py-3">Strong within the studied samples</td></tr><tr class="border-b border-[#D7C8B5]"><td class="px-3 py-3 font-semibold">Broad contextual web mentions</td><td class="px-3 py-3">Highest or near-highest Ahrefs correlations</td><td class="px-3 py-3">Strong within Ahrefs’ screened samples</td></tr><tr class="border-b border-[#D7C8B5]"><td class="px-3 py-3 font-semibold">Topic-specific relevance</td><td class="px-3 py-3">Consistent topic ownership is not explained by broad SEO metrics alone</td><td class="px-3 py-3">Moderate; relevance was not isolated</td></tr><tr class="border-b border-[#D7C8B5]"><td class="px-3 py-3 font-semibold">Third-party recognition</td><td class="px-3 py-3">Strong Webflow-reported relationship</td><td class="px-3 py-3">Moderate; one vendor rubric</td></tr><tr class="border-b border-[#D7C8B5]"><td class="px-3 py-3 font-semibold">Technical accessibility</td><td class="px-3 py-3">Important eligibility foundation</td><td class="px-3 py-3">Moderate; not a proven preference mechanism</td></tr><tr class="border-b border-[#D7C8B5]"><td class="px-3 py-3 font-semibold">Backlink volume alone</td><td class="px-3 py-3">Weak Ahrefs correlations</td><td class="px-3 py-3">Strong confidence that volume alone is insufficient within those samples</td></tr><tr class="border-b border-[#D7C8B5]"><td class="px-3 py-3 font-semibold">Website page count</td><td class="px-3 py-3">Very weak Ahrefs correlations</td><td class="px-3 py-3">Strong confidence that volume alone is insufficient within those samples</td></tr><tr><td class="px-3 py-3 font-semibold">One universal platform tactic</td><td class="px-3 py-3">Contradicted by Yext’s model and context variation</td><td class="px-3 py-3">Strong confidence that no single recipe fits every question</td></tr></tbody></table></div>
The table does not assign causal coefficients. It separates replicated observations from working interpretations and unproven mechanisms.
<h2>Machine Positioning changes the marketer’s question</h2>
Machine Positioning can be defined as <strong>the degree to which intended brand positioning is faithfully and advantageously represented in AI-mediated decisions</strong>. It should not become a separate positioning strategy for machines. Human and machine positioning should remain consistent.
The most useful diagnostic is not whether an AI system knows the brand. It is whether the system connects the brand with the attribute, audience, use case and decision that matter. Volvo → safety is different from Volvo → largest automotive company. A specialist CRM may not dominate ‘best CRM’ but could become strongly associated with ‘best CRM for a 15-person independent financial advisory firm’.
Marketers can begin with five actions. Define the narrow decision territories the brand has a legitimate right to own. Specify the true claims and associations required to support that position. Audit first-party and independent evidence by function, not by channel. Measure Access, Visibility, Brand Representation, Consideration and Recommendation separately across the relevant prompts and platforms. Test changes over time without treating correlation as proof of causation.
<h2>AEO Updates Takeaway</h2>
What makes a brand more likely to appear in AI? The current evidence cannot provide a universal causal formula. It supports a more defensible working answer.
AI systems appear more likely to surface brands they can connect to the specific subject or decision being asked about. That connection travels with some combination of existing brand strength, broad contextual mentions, topic relevance, authoritative first-party information, third-party recognition, machine accessibility and fit with the decision context.
The objective is not to manipulate an answer engine into mentioning a brand. It is to create an evidence environment in which mentioning the brand becomes a reasonable conclusion.
If the evidence is strong, distributed and relevant, the particular source an AI system retrieves can change while the underlying conclusion survives. That may be the more durable form of AI visibility.
<h3>References</h3>
[1] <a href="https://ahrefs.com/blog/ai-brand-visibility-correlations/" target="_blank" rel="noopener noreferrer">Ahrefs, ‘Top Brand Visibility Factors in ChatGPT, AI Mode, and AI Overviews’</a>, 12 December 2025.
[2] <a href="https://ahrefs.com/blog/ai-overview-brand-correlation/" target="_blank" rel="noopener noreferrer">Ahrefs, ‘An Analysis of AI Overview Brand Visibility Factors’</a>, 26 May 2025.
[3] <a href="https://www.semrush.com/blog/chatgpt-topic-authority-study/" target="_blank" rel="noopener noreferrer">Semrush, ‘AI visibility is a topic-level game: A study of 50,000 brands in ChatGPT’</a>, 20 July 2026.
[4] <a href="https://webflow.com/blog/aeo-maturity-index" target="_blank" rel="noopener noreferrer">Webflow, ‘The AI discovery gap: we analyzed 2,000 websites, and almost nobody is ready for answer engines’</a>, 16 July 2026.
[5] <a href="https://www.yext.com/research/articles/ai-citation-behavior-across-models/" target="_blank" rel="noopener noreferrer">Yext, ‘AI Citation Behavior Across Models: Evidence from 17.2 Million Citations’</a>, 28 April 2026.