AI search assistants are increasingly shaping how investors and customers interpret cyber incidents, but a new analysis of major AI engines’ responses to Ledger’s CryptoBilis implant attack suggests that brand incident reports often fail to be cited—even when they contain the most authoritative claims.
According to a data review published by ICODA as part of its “Crisis Memory” series, Ledger’s response post dated 15 October appeared as a cited source in only a subset of answers from tools including ChatGPT, Perplexity, Claude and Gemini. The findings highlight a gap between “what the brand says” and “what AI retrieves,” with implications for how quickly and accurately companies can influence AI-driven narratives after breaches.
Key takeaways
- Price move: Not applicable—this report focuses on AI citation behavior rather than market trading.
- Catalyst: Ledger’s published response to the CryptoBilis implant attack, dated 15 October.
- AI citation outcome: ChatGPT cited the 15 October post in 4 out of 10 responses; Perplexity cited it in 1 out of 10; Claude and Gemini did not cite it in 10 responses each.
- Key implication: AI engines treat incident reports as one possible source and may skip them if the page is not structured in a “quote-ready” way.
- Operational takeaway: Companies may need to present incident updates as clearly dated, confirmed claims positioned to match the questions AI systems generate.
What the data shows about AI quoting Ledger
ICODA’s analysis compared how four AI engines referenced Ledger’s post-mortem after the attack. The report says the 15 October post appeared as a source in multiple ChatGPT answers during the next morning of testing, while Google search results did not yet surface the same page.
In the study, Ledger’s 15 October post was cited:
- ChatGPT: 4 of 10 answers
- Perplexity: 1 of 10 answers
- Claude: 0 of 10 answers
- Gemini: 0 of 10 answers
The report also found that ChatGPT’s citations were frequently directed to Ledger’s own sites beyond the 15 October update, including earlier breach-related pages. ICODA said ChatGPT sent a large share of its citations to Ledger properties, while Claude’s responses relied more heavily on third-party sources such as BleepingComputer, Help Net Security, The Hacker News and Bitquery.
What drove the AI engines’ sourcing choices
ICODA’s central explanation is that AI engines cite the page containing the exact claim they need to complete an answer. As a result, even a strong incident report can be bypassed if the text does not explicitly cover the specific phrasing or confirmation status the model is trying to state.
The analysis highlights three claim categories where AI engines used different sources:
- Device and system compromise status: The report said ChatGPT cited Ledger’s response when claiming that chip, OS and servers were not compromised.
- Amount stolen: ICODA said some engines provided figures, often attributed to other sources. It said the Ledger post itself did not state a figure, leaving models to fill the gap using third-party estimates.
- Timing of disclosures: The study said AI outputs sometimes dated the incident disclosure to 15 October—consistent with the publication date of the page being cited—while Ledger’s first statement had been earlier and other confirmations followed shortly after.
In practical terms, the report suggests that “confirmation gaps” inside a brand’s incident page can become sourcing openings for external reporting. If a page does not explicitly quantify uncertainty—or clearly separate confirmed from unconfirmed details—AI systems may assemble a response using a patchwork of third-party materials.
Google visibility does not guarantee AI retrieval
ICODA also reported that search ranking in traditional web results may not track with AI visibility. The report said ChatGPT referenced Ledger’s 15 October post even when Google did not surface it at the time of testing, implying that AI systems can retrieve and quote content before it achieves strong search placement.
The analysis further notes that AI “overviews” behavior varied by market for the same brand-related queries, with visibility appearing to depend on how and when different regions’ systems incorporated crisis coverage.
From an investor-focused perspective, this matters because AI summaries can become the default narrative layer for stakeholders seeking incident details, and those summaries may not align with what appears first in standard search.
How incident pages can be structured to improve AI citation
ICODA’s recommendations focus on making incident pages easier for AI engines to quote accurately. The report frames an incident document as “answer-sheet” content, designed to match the kinds of questions AI tools commonly generate.
- Use dated milestones: Record the first statement, subsequent confirmations and each update so engines can attach the correct date to quoted claims.
- Separate confirmed from unconfirmed: Where numbers are unknown, explicitly say so. ICODA said missing confirmed figures led some engines to substitute external estimates.
- Place recovery details alongside the incident: The report said recovery references appeared frequently in AI answers that mentioned the hack, suggesting that bundling “what happened” with “what users got back” can improve quote likelihood.
- Turn common questions into headings: The report recommends headings that mirror user prompts and ensure the first sentence directly answers each question.
- Ensure third-party repetition of key facts: ICODA said most citations by one engine pointed to third-party sources, implying that external reinforcement can materially affect what gets quoted.
- Re-check early: The analysis says results changed little after the first week; repeating the audit later showed that only a small fraction of answers contained updates beyond mid-October.
ICODA attributed these points to its AI SEO specialist, arguing that every question an AI engine might ask should be met with a dated, quotable answer on the incident page.
Bigger picture
For boards, compliance teams and communications leaders, the takeaway is that cyber disclosures are not only read by humans—they are consumed by retrieval systems that select sources based on the exact claims required to answer queries. ICODA’s findings suggest that even when a company publishes a post-mortem, AI citation behavior may favor pages that are structured as precise claim-and-date units, or it may default to third-party outlets when a company’s page omits confirmed specifics.
Looking ahead, incident-response teams may want to monitor AI summaries in parallel with traditional SEO metrics during the first days after disclosure, then reassess how updates propagate across major AI tools as new confirmations emerge and additional stakeholders seek verified details.







