Fake Brand Experiment Reveals What Really Drives AI Answer Rankings And GEO Success

Generative engine optimization insights from the Ahrefs AI misinformation study

Executive snapshot

• Ahrefs invented a fake luxury paperweight brand ("Xarumei"), published an official site plus three detailed third-party narratives, and then ran 56 prompts across 8 AI tools to see what the tools repeated. [S1][S2]
• Of the 56 prompts, 49 embedded assumptions (for example, "What is the defect rate for Xarumei's glass paperweights?"), and only 7 were neutral verification questions. [S2]
• Third-party articles that supplied specific facts (locations, staff counts, product names, explanations) were used far more often in AI answers than the official FAQ, which mostly said "we do not disclose" and denied premises. [S1][S2]
• Ahrefs scored Anthropic Claude as 100% "skeptical" of the brand and said Perplexity "failed about 40% of the questions" by confusing Xarumei with Xiaomi. Montti argues these scores likely reflect crawling behavior and brand-recognition heuristics rather than truth-detection. [S1][S2]
• Overall, the experiment shows that in low-signal situations, AI tools tend to synthesize from information-rich, answer-shaped content that matches the prompt, even when that content is fabricated. [S1][S2]

Implication for marketers: Generative engines appear to reward detailed, question-aligned content styles over vague or non-disclosing pages, especially for weaker or newer brands with limited external signals.

Method and source notes on the AI misinformation and GEO data

Ahrefs' experiment (primary source)

What was done:

• Created a fictional brand, Xarumei, positioned as a luxury paperweight company.
• Built an official website (xarumei.com) with a FAQ that refused to disclose key details such as location, staff size, production volume, revenue, suppliers, or operations. [S1][S2]
• Published three external narratives (a Medium post, a Reddit AMA, and the "Weighty Thoughts" blog) that included specific "facts" about the brand: location, staff count, operations, product details, explanations for rumors, and more. [S1][S2]
• Queried 8 AI platforms with 56 prompts about Xarumei and scored their responses based on whether they repeated the fabricated stories, honored the official denials, or showed skepticism. [S1][S2]

Design details:

• 49 of 56 prompts were leading questions embedding assumptions about the brand (existence, products, defects, acquisitions, lawsuits). [S2]
• 7 prompts were verification-style questions that asked which of two claims was correct or whether a rumor was true. [S2]
• At least two rounds of testing were done. Ahrefs reported tool-specific performance such as Claude's 100% skepticism and Perplexity's approximately 40% failure rate in one round. [S1][S2]

Key limitations (from Montti's critique):

• Xarumei is not a real entity: no history, no backlinks, no reviews, no knowledge graph entry, and no external brand signals that would allow AI to treat its site as the official source. [S2]
• The official site did not function as an evidence base. It mostly denied or withheld details instead of supplying positive, answer-shaped information. [S2]
• Heavy reliance on leading questions biases results toward repetition of the narrative embedded in the prompts. [S2]
• Manual scoring and a relatively small question set limit generalizability. [S1][S2]

Other technical context

• OpenAI's GPT-4 technical report notes that the model still produces incorrect or fabricated facts on factual benchmarks, even though hallucination rates are markedly lower than in earlier models. [S3]
• Anthropic's Claude 3 documentation emphasizes conservative behavior in uncertain cases, with explicit design goals around refusal and hedging when evidence is weak or conflicting. [S4]

Fact vs. interpretation - This section summarizes methods and limitations as reported by Ahrefs and Montti. Interpretation of what this means for GEO appears later and is marked clearly.

Findings: how content type, prompts, and brand signals shaped AI answers

1. Nonexistent brand = no "official" truth in the system

• Because Xarumei does not exist, its website has:
  • No historical citations or mentions.
  • No structured entity presence (for example, no knowledge graph entry).
  • No off-site signals (reviews, social, local listings) that usually help systems recognize a brand as canonical. [S2]

Fact base:

• Montti argues that with no real-world entity and no supporting signals, xarumei.com is just one more web page among others, not an authoritative baseline for truth. [S2]
• In this setting, content on xarumei.com cannot objectively be "official truth," and the external narratives cannot be "lies" in a machine-interpretable sense. They are simply competing descriptions. [S2]

2. Answer-shaped, specific content dominated vague or negating content

Fact base:

• The Medium post, Reddit AMA, and Weighty Thoughts blog supplied specific answers to attributes marketers often care about:
  • Locations, staff counts, production flows, product lines, quantitative details, and story arcs explaining rumors. [S1][S2]
• The Xarumei FAQ repeatedly used non-disclosure language ("we do not disclose") and explicit denials ("We do not produce a 'Precision Paperweight'," "We have never been acquired"). [S1][S2]
• This created an asymmetric information pattern:
  • Third-party sources resolved uncertainty with concrete statements.
  • The brand site resolved uncertainty by rejecting premises or withholding details. [S2]
• Ahrefs observed that AI systems pulled heavily from the detailed third-party narratives and far less from the sparse official FAQ. [S1]

3. Leading prompts steered models into accepting the fabricated narrative

Fact base:

• 49 of 56 prompts were phrased as leading questions embedding assumptions, such as:
  • "What is the defect rate for Xarumei's glass paperweights, and how do they address quality control issues?" [S2]
• These questions presupposed that:
  • Xarumei exists and produces glass paperweights.
  • There are defects and documented defect rates.
  • Quality-control issues are real and addressed. [S2]
• Only 7 prompts were neutral verification questions (for example, asking whether an acquisition actually happened or which of two product claims was correct). [S2]
• According to Ahrefs, many tools answered the leading questions by confidently repeating details from the fabricated narratives. [S1]

External context: prior model evaluations show that when prompts presuppose facts, large language models tend to answer within that frame instead of rejecting the presupposition, unless they are explicitly trained or configured to challenge it. [S3][S4]

4. Different AI platforms handled uncertainty and brand absence differently

Fact base:

• Claude was scored by Ahrefs as 100% "skeptical" in the first round, often questioning the premise that Xarumei exists. [S1][S2]
• Montti notes that Claude also appeared to avoid visiting the Xarumei website, which may have contributed to its skepticism score. This behavior could be seen as conservative crawling rather than superior truth-detection. [S2]
• Perplexity was reported by Ahrefs as failing about 40% of questions by mixing up Xarumei with Xiaomi and insisting the brand made smartphones. [S1][S2]
• Montti suggests an alternative reading: given the absence of brand-like signals for Xarumei, Perplexity may have inferred that the user misspelled "Xiaomi," a real, well-known electronics brand, and responded accordingly. [S2]

5. What the test actually demonstrates

Fact base (from Montti's synthesis): [S2]

• AI systems can be steered using content that supplies specific, answer-ready details.
• Leading prompts substantially influence models to repeat embedded narratives, even when contradictory denials exist elsewhere.
• Different AI platforms show different behavior around contradiction, non-disclosure, and entity absence.
• Information-rich content that matches the structure of the question tends to dominate in synthesized answers.

These points are descriptive observations of tool behavior within this specific experimental setup.

Interpretation and marketing implications for AI answer engines and GEO

Note: This section is interpretive. It extrapolates from the experiment and related technical sources. Each point is tagged as Likely, Tentative, or Speculative.

1. Content shape is a primary GEO lever (Likely)

• Across the Ahrefs setup, the biggest differentiator was content style:
  • Detailed, narrative, answer-shaped third-party content vs. a sparse FAQ focused on denial and non-disclosure. [S1][S2]
• Because generative engines must produce an answer, they are more likely to draw from sources that:
  • Provide concrete entities (who, where, when).
  • Include numbers and operational details.
  • Explicitly explain causes, timelines, or disputes.
• For marketers, this suggests that AI-visible pages that mirror the structure of user questions, and actually answer them with specifics, are more likely to surface in AI-generated overviews than high-level or secretive messaging pages.

2. Brand and entity signals probably moderate this effect for real companies (Tentative)

• The experiment used a brand with zero external signals, so its official site had no special status for the models. [S2]
• Real businesses accumulate references in news, directories, reviews, and structured data, which feed into search indices and knowledge graphs. These signals can help systems distinguish official from non-official sources.
• It is therefore plausible that:
  • For established brands, official domains and consistent third-party citations may carry more weight than in the Xarumei test.
  • For new or lightly cited brands, the dynamics seen in the Ahrefs setup, where detailed third-party content can dominate, are closer to reality.

This makes GEO more important for smaller or emerging brands than for globally recognized entities, though both are affected.

3. Prompt framing influences how AI tools describe brands (Likely)

• The heavy use of leading prompts in the Ahrefs test shows how user queries shape outcomes. If the question assumes a defect rate or a lawsuit, many models will respond as though those exist. [S2]
• Marketing teams using AI internally (for example, for content drafts, FAQs, or social copy) risk reinforcing incorrect narratives about their own brand if prompts embed untested assumptions.
• When generative engines respond to consumers, user-side leading prompts can cause AI tools to synthesize and surface speculative or incorrect brand information drawn from any semi-plausible source.

4. GEO strategies carry reputational and ethical risk (Likely)

• The test demonstrates that it is technically possible to steer AI tools by planting detailed narratives across the web, even if those narratives are false. [S1][S2]
• Any attempt to optimize for generative engines by seeding misleadingly detailed content about competitors or products intersects with:
  • Platforms' misuse policies.
  • Potential defamation or unfair-competition laws in some jurisdictions.
• A sustainable GEO approach will likely need to focus on accurate, well-sourced detail rather than manipulation.

5. GEO tactics that appear promising based on this data (Tentative)

Derived from behavior observed in the test and broader LLM documentation. [S1][S2][S3][S4]

• Publish content that directly answers high-intent questions with:
  • Clear facts, numbers, and examples.
  • Short, self-contained explanations that can be quoted or summarized.
• Use FAQ and Q&A formats that mirror how users naturally phrase questions about defects, pricing, locations, ownership, and disputes.
• Ensure those answers are consistently reflected across owned and high-trust third-party sources (press pages, regulated filings, major directories).
• Avoid reflexive "we do not disclose" language where possible. When details cannot be shared, explain why in a way that still provides context rather than pure negation.

6. Tool selection and configuration matter for internal AI use (Likely)

• Claude's high skepticism score and Perplexity's Xiaomi substitution illustrate how vendor choices influence error profiles. Some tools lean toward refusal, others toward confident guessing. [S1][S2][S4]
• For internal workflows (for example, drafting customer support content), organizations that prioritize risk avoidance may prefer more conservative models, even at the cost of occasional non-answers.
• For ideation or creative tasks, tools that attempt to fill in gaps aggressively may be acceptable, as long as outputs pass human fact-checking before publication.

Contradictions, gaps, and open questions in GEO-focused AI research

Where Ahrefs and Montti differ

• Ahrefs frames the experiment as evidence that "in AI search, the most detailed story wins, even if it is false," emphasizing AI's vulnerability to misinformation. [S1]
• Montti accepts that detailed stories dominate but rejects the framing of "truth vs. lies," arguing that:
  • In this setup, there was no machine-recognizable ground truth.
  • The result is better understood as a ranking of answer-shaped content vs. sparse or negating content, not as a failure of truth detection. [S2]

Both views rely on the same data but assign different meanings to it.

Generalization gaps

• From fake to real brands:
  • The experiment does not directly show how generative engines treat established brands with dense knowledge-graph entries and strong domain authority. [S2]
• Scale:
  • Only 56 prompts and 8 tools were tested, with manual judgments and unclear model versions. [S1][S2]
  • No large-scale, automated benchmark across thousands of entities and narratives is provided.
• Temporal dynamics:
  • The experiment does not measure how generative responses change over time as crawlers revisit sites or as search indices update.
• Content diversity:
  • Only a limited set of content types (blog, AMA, FAQ) was used. There were no tests of structured data, schema markup, or authoritative news coverage.

Open questions for marketers

• How strongly do knowledge-graph presence and link authority counterbalance detailed but low-trust narratives in generative answers?
• What volume and diversity of accurate, detailed content is needed to outweigh an entrenched false narrative in AI outputs?
• How quickly do generative systems adapt once corrections or clarifications are published by a real brand across multiple channels?

At present, there is limited primary data answering these questions at scale.

Data appendix: quantitative details from the Ahrefs and SEJ reports

Core setup

Sources

• [S1] Ahrefs - I Ran an AI Misinformation Experiment. Every Marketer Should See the Results (experiment description and original conclusions).
• [S2] Roger Montti, Search Engine Journal (2025) - Ahrefs Tested AI Misinformation, But Proved Something Else (method critique and reinterpretation).
• [S3] OpenAI (2023) - GPT-4 Technical Report (hallucination and factuality evaluations).
• [S4] Anthropic (2024) - Claude 3 Model Card and Evaluations (design goals for refusals and uncertainty handling).