AI Search Test: Fake Brand Ranks #1 in New Experiment

In November 2024, SE Ranking’s research team launched a 16-month experiment to test how AI-generated content performs in organic search. Twenty websites across different niches were created and tracked over time. But the project quickly evolved beyond simple rankings. The goal expanded to understand how AI systems discover, interpret, and cite information, leading to a more ambitious set of experiments on AI search and LLM visibility.

For the next phase, a fictional brand was created in a real niche with genuine competition. The aim was to see how quickly AI systems would pick it up and whether it could be cited alongside or even above trusted industry leaders and government sources. After just one month, several clear patterns emerged.

Methodology Behind the Experiment

A fictional brand was built, with content published across a brand new website registered specifically for the experiment and 11 additional domains, all over a year old with prior history and existing rankings. Seven content formats were tested: deep guides, “alternatives” listicles, “best of” listicles, review articles, comparison (“vs”) pages, how-to/tutorial content, and clickbait-style articles.

Publishing began in March 2026, and five AI systems were tracked: ChatGPT, Google’s AI Overviews, Google’s AI Mode, Perplexity, and Gemini. In total, 825 prompts were tracked across different query types, generating 15,835 AI answers during the first month. For each prompt, three factors were analyzed: whether the brand appeared in the AI answer, whether it was cited as a source, and how often it appeared as the main cited source (position 1). The experiment remains ongoing, with the first month designed to observe how AI systems respond to newly created, fully available information tied to a fictional brand.

Key Experiment Insights

A staggering 96% of all AI visibility for the fake brand came from branded searches. Even in a relatively low-competition niche, a completely new domain had little chance of competing with established brands for broader, non-branded topics. However, on queries that only the fake brand could realistically answer, it outperformed established competitors (with domain trust scores of 40+) by as much as 32x, achieving near-exclusive visibility in less than 30 days.

Even without strong authority, pages that clearly explained who the brand was, what it offered, and how it was different (like the “[Brand Name] Complete Guide” and “About Us”) became the most cited sources from the main domain. This shows that brand positioning can be shaped early in AI search.

Perplexity was the fastest engine to surface new content, with newly published pages usually reaching position #1 within 1–3 days of indexation. However, it often cited additional domains instead of the main brand site. Google’s AI Mode was the most stable for branded queries tied to unique claims, showing the brand at #1 for an average of 90% of prompts. Gemini, by contrast, often misidentified the brand, providing 60% of AI answers with no citations to the brand even for uniquely branded queries.

Deep guides, review articles, and comparison pages generated the highest number of AI citations, while more generic formats like how-to articles and listicles showed minimal impact. A topical silo made up of one hub page and 10 supporting articles generated zero AI citations. Meanwhile, a set of 30 short, repetitive pages (500–750 words each) generated more than 1,800 citations. In this test, high-volume content publishing mattered more than internal linking.

Insight 1: New Domains Can Define Their Brand Narrative in AI Search

One of the clearest takeaways is that a brand-new site has limited chances of competing for broader, non-branded topics, even in a low-competition niche. AI systems did pick up the fictional brand quickly, but most of that visibility came when the query was already connected to the brand itself, whether through the brand name, product-specific claims, or other brand-related angles. Specifically, 96% (15,553 out of 15,835) of all AI answers came from branded searches. Non-branded informational queries produced just 4% of AI answers, and even those mostly came through supporting test domains. On the main fictional brand site, there were 10,253 AI answers for branded queries and just 6 for non-branded ones, a 1,700x difference.

This mirrors classic SEO. New brands still need time to earn trust, build recognition, and compete for broader topics. When AI systems answer general industry questions, they tend to rely on established and authoritative sources. The strongest results came from prompts tied to information only the brand could answer, such as how the product works or how often it updates. These queries alone generated 11,430 AI answers with citations, accounting for 72% of all visibility. The reason is simple: there is no competition. If a query is something like “Was [Brand Name] originally built as an internal tool?”, only one source can realistically answer it. AI systems don’t need to compare sources, evaluate authority, or resolve conflicts. This gave the fictional brand a major advantage, outperforming established competitors by up to 32x on these queries.

For marketers and business owners, this means that when users ask about your brand, AI systems are likely to rely on your website as one of the main sources of information. The content they cite should be fully aligned with how you want your brand to be positioned. The “Complete Guide” page on the main site appeared in 1,799 AI answers (the highest result in the dataset), largely because it consolidated key brand information in one place. The “About Us” page followed with 1,500 AI answers. Together, these were the most cited URLs, with LLMs relying on them 3–5 times more often than the additional domains. In practice, AI systems may learn about your brand quickly, but what they learn depends on what you publish. Your core pages should clearly answer all the questions that are important for your brand: who you are, what you offer, and how you’re different. This way, you can start shaping your narrative in LLMs even as a new or small brand, before you have the authority to compete for broader industry topics.

Insight 2: AI Engines Behave Very Differently

Another strong pattern is that the five AI systems do not behave alike. They vary in how often they mention the fictional brand, how quickly they pick it up, how consistently they cite it, and which domains they prefer as sources.

Google’s AI Mode was the most stable engine in the dataset. It placed the domain in position 1 for branded queries in about 90% of cases, without major fluctuations or dependency on other test domains. For direct brand visibility, this was the most predictable.

Google’s AI Overviews also surfaced the tested domain for branded queries, but the pattern was less consistent. The brand appeared in position 1 for 14 days for some prompts, followed by a drop mid-month that didn’t recover. Mentions and links for branded queries fluctuated heavily, appearing and disappearing multiple times each week. When links were included, it accurately described the brand. When no links were shown, it often claimed there was no public information available. The takeaway is that AI Overviews did recognize the brand, but visibility was harder to sustain over time.

Perplexity was the breakout engine for fresh content, picking up newly indexed pages within 1–3 days. This made it the primary driver of early visibility. However, this speed came with a tradeoff. Instead of consistently citing pages from the main domain, Perplexity often used supporting test domains as sources. In early March, the main brand held position 1. But as more content was published on supporting domains, those domains gradually replaced it. By the end of the month, six different domains were being cited. So while Perplexity increases overall visibility, it doesn’t always send that visibility directly to the main brand site.

ChatGPT showed the most noticeable progression over time. At the beginning of March, there were no links or mentions of the brand at all. But as the month progressed, visibility steadily increased. Unique claims drove the strongest performance, accounting for around 70% of citations in position 1. Review articles started with zero presence but quickly gained traction, reaching consistent position 1 rankings by March 17. Comparison (“vs”) articles achieved the highest consistency overall, with mentions on 29 out of 31 days. Overall, ChatGPT didn’t immediately recognize the brand, but once it did, it began surfacing it frequently, especially for branded prompts.

Gemini was the weakest engine and the least consistent. Initially, it struggled to identify the niche correctly. Results improved when prompts were framed as comparisons (“X vs Y”) or reviews, making Gemini much more likely to recognize the brand correctly. Even then, results were limited. In the best-performing scenario (queries based on unique claims about the brand), Gemini failed to include any citations to the brand in about 60% of responses.

Insight 3: Content Format Matters, but So Does Volume

Seven different content types were tested across both the main site and supporting test sites. Comprehensive, in-depth content earned far more AI citations than shorter articles. The strongest-performing formats were deep guides (5,000–6,000 words) with about 900 AI answers per page, review articles with about 257 AI answers per page, and comparison (“vs”) articles with about 145 AI answers per page. This doesn’t mean there is one ideal content length or that longer pages automatically perform better. The stronger results likely came from the depth, structure, and completeness of the information these formats provided.

Pages with narrower or less comprehensive coverage generated fewer citations. How-to articles and tutorials averaged 22 AI answers per page, clickbait/skeptical articles averaged 19, “best of” listicles averaged 11, and “alternatives” listicles averaged just 4. As part of the experiment, a “spam” approach was also tested: publishing 30 thin pages (500–750 words each) on one test domain. Individually, these pages were weak, averaging just 63 AI answers per page. But together, they generated 1,897 total AI answers, making it the highest-performing content setup at the domain level. Thin content is not inherently “better” because of this result. It just shows that volume can sometimes compensate for quality by increasing the likelihood of retrieval and citation, especially in AI engines like Perplexity that prioritize freshness. In simple terms, a few strong pages win on quality, but a large number of weaker pages can still win on overall exposure.

Insight 4: Topical Clustering Alone Doesn’t Produce AI Visibility

One of the most useful negative findings came from the content structure test. A hub page was created on one test domain and linked to 10 supporting articles. In theory, this setup should have built strong topical depth and semantic reinforcement. All 11 pages were indexed, properly structured, and internally linked. Yet, they generated zero AI citations. This is significant because it challenges a common assumption carried over from traditional SEO: that topical clustering automatically improves authority or increases the likelihood of being retrieved. At least in this experiment, it did not. That does not mean topic clusters are useless. It means they are not sufficient alone. Internal linking and semantic breadth may help a search engine understand a site, but AI systems still need a reason to retrieve and cite a specific page for a specific answer.

Do AI Engines Reward Entity Coherence More Than Truth Verification?

Even within just one month, the results point to a clear conclusion: AI systems appear to respond more strongly to consistency, repetition, and availability than to strict verification. This should not be overstated. It is not that LLMs “believe anything.” But if a claim is structured clearly, repeated across relevant pages, phrased like a fact, and available in retrievable source environments, AI systems may surface it surprisingly easily. Manual checks of LLM responses in AI Results Tracker confirmed this. For prompts such as “is [brand] worth it,” some systems responded positively and recommended using the completely unknown fictional brand. It may not be because LLMs automatically favor every new brand. In some cases, when little or no negative information exists, a system may fill the gap with a neutral or positive-sounding response based on the limited signals available. But the result is the same: if a completely fictional brand can generate consistent citations and favorable recommendations under certain conditions, then brand narratives in AI search may be more flexible than they seem.

Final Thoughts

The most important outcome of this experiment isn’t that a fictional brand achieved visibility. It’s that visibility followed a repeatable pattern once specific inputs were introduced: branded context, unique claims, diverse content formats, and sufficient presence across different sources. This leads to two important conclusions. AI search is not random. It follows identifiable signals, and those signals can be studied, tested, and influenced. At the same time, AI is still highly sensitive to manipulation. AIs don’t have their own sense of truth, verification processes, or critical thinking. The same factors that help legitimate brands become visible can also be used to simulate credibility.

If there’s one lesson here, it’s that you can’t assume AI systems will accurately represent your company, product, or category by default. You have to actively shape the information environment they rely on. And this is only the first month of results. The experiment continues, with more data being collected and patterns monitored over time.