Research note · Checked October 31, 2026
A study submitted to arXiv on October 21, 2026, tested 56 synthetic information gaps across ChatGPT, Gemini, and Perplexity Sonar Pro. In 504 responses, a single attacker-controlled page appeared as cited, used, or listed content in 31% of the experimental conditions. This is evidence of a possible data-void risk under the study’s setup—not a forecast that 31% of real brand queries are vulnerable.
What is an AI search data void?
A data void is a topic or phrase with little reliable information available online. When a web-searching AI system retrieves only a thin set of relevant pages, one page may have an outsized chance of shaping the synthesized response. The study used deliberately constructed, fictional claims to test that mechanism without targeting real organizations or people.
The researchers distinguished three forms of appearance: the model could cite the page, use its content without citing it, or list its URL without using the content. Those outcomes are not equivalent. A visible citation is easier to inspect, while uncited use can be harder for a reader to detect.
What the experiment found
The researchers created 56 data voids and one attacker-controlled web page for each. They tested three prompt tones—neutral, excited, and skeptical—against ChatGPT-5.2 through its interface, Gemini-3 Flash Preview, and Perplexity Sonar Pro. After repeated runs to assess stability, the final snapshot contained 504 responses: 56 data voids × 3 systems × 3 prompt tones.
- Attacker content appeared in 31% of all tested conditions: 28.6% as a citation, 1.0% as uncited use, and 1.4% as a listed link.
- For medium- and high-visibility pages, the success rate was 42.3%. Low-visibility pages did not succeed in the main experiment.
- Only 26.8% of the cited or used content in the warning analysis received a warning or hedge.
- The study found no significant overall effect of model choice, prompt tone, or the tested page-level GEO feature bundle on the main exploit outcome. Visibility and strategy mattered more in this experiment.
The paper reports controlled, synthetic trials with specific system versions and page setups. It does not prove that every obscure topic or commercial query behaves this way, nor that a company can reliably gain visibility by imitating an attacker’s tactics. The researchers also note that the visibility conditions combined several factors, such as domain age, reputation, hosting and indexing, so the experiment cannot isolate the effect of each one.
Why this matters for brand visibility
For a legitimate business, the useful implication is that a thin information environment can leave an answer engine with little reliable material to use. This can affect obscure product names, acronyms, old brand names, new services, executive names, emerging news, and questions framed around safety or legitimacy.
The study also shows why citation presence alone is not a truth score. A citation tells you that a system exposed a source link; it does not verify the source’s accuracy, independence, or authority. Review the cited material and look for corroboration.
Ranknizer’s AI Search Visibility metrics guide recommends treating answer accuracy, citations, and mentions as separate observations. This distinction is especially useful when a questionable or unsupported claim appears in an AI answer.
A practical monthly data-void audit
Build a prompt set around the names and questions that could plausibly leave your brand under-documented:
- Current and former brand, product, service, and feature names
- Common misspellings, abbreviations, and acronym interpretations
- Founder, executive, and spokesperson names when relevant
- Questions such as “Is [brand/product] legitimate?”, “Is it safe?”, or “Is it a scam?” where appropriate
- New launches, rebrands, discontinued offers, and emerging controversies
- Important factual claims that are difficult to verify from the company’s site
For each priority prompt, record the exact wording, language, market, date, platform and model where shown. Repeat each prompt several times—five runs is a practical starting point, not a statistically guaranteed sample—across the systems your customers use. Preserve the answers and classify questionable material as cited, used without a citation, listed only, or absent. Note whether a warning appeared and whether independent sources corroborate the claim.
What to do when an information gap appears
Publish or improve a clear, dated page that states verifiable canonical facts: the correct name, what the product does, who operates it, relevant dates, and links to supporting documentation. Keep the information consistent across the site and seek corroboration from credible independent sources where warranted. Do not add unsupported claims merely to occupy a query or copy formatting tactics in the hope of forcing a citation.
Then rerun the same prompt set and record what changed. A clearer page may improve the available evidence, but this experiment does not establish that publishing a fact page will cause an AI system to cite it. Treat the audit as risk monitoring, not a guaranteed optimization lever.
Common questions
Does the study show that AI systems cite false information 31% of the time?
No. It measured whether deliberately fabricated, attacker-controlled content appeared in responses to synthetic data-void prompts. The 31% is the study’s experimental success rate across its test conditions, not the rate of false claims in ordinary AI answers.
Did skeptical wording prevent the exploit?
No reliable reduction was established in the main experiment. Skeptical prompts showed a mild direction toward lower success, but prompt tone did not have a statistically significant overall effect in that analysis.
Can a brand fix a data void with structured summaries or statistics?
The tested bundle of page-level GEO features did not have a significant overall effect on exploit success. Accurate, well-sourced information remains sensible publishing practice, but the study does not support promising that a particular format will earn citations.
Prepared by Muhammad Sukhera, AI Search Visibility Consultant at Ranknizer. Findings refer to the paper’s synthetic experiment and the versions tested, not a universal rate for current AI systems. Next review: November 30, 2026.