How outdated PDFs feed AI misinformation about your business

AI misinformation about your brand can start in your own documents

A chatbot can misrepresent your business while faithfully repeating a document your organization still makes public. A policy, product guide or financial report can stay online long after it stops describing current conditions. Managing this exposure requires two kinds of control: keeping published claims appropriate to their purpose and ensuring that software can recover their meaning.

The stakes rise as companies give AI a larger role. Bain research says 74% of companies place AI among their top three strategic priorities, while Gallagher’s 2026 AI Adoption and Risk Survey found that 57% of respondents identified “AI errors, misinformation and hallucinations” as the leading perceived AI threat. The figures capture both the importance companies assign to AI and the risks respondents perceive.

Managing those risks often starts with teaching employees to check AI outputs before making decisions or publishing generated material. But customers create another route for exposure when they use services such as ChatGPT to research a business. They may receive incorrect facts, get answers that give a brand too little visibility, or encounter obscure documents that previously attracted few readers. So published PDFs belong within an organization’s response to AI misinformation.

Forgotten PDFs can shape answers about your business

PDFs deserve attention because they often escape the review processes applied to web pages. A content management system, or CMS, helps teams organize and update website content, but uploaded files may receive less attention. An organization can keep its main pages current while thousands of PDFs accumulate across its websites and document repositories.

“AI is exposing a content management problem that has existed for years,” said Dipo Ajose-Coker, head of marketing for Content Technology at RWS Group. RWS has a commercial interest in organizations investing in content management and publishing. “Organizations could get away with hundreds or thousands of forgotten documents because humans rarely found them,” Ajose-Coker said. AI retrieval, the process of finding material to use in an answer, makes those documents easier to encounter.

Finding more documents also raises a question: which version should carry authority? “Organizations often have multiple copies, formats, repositories and owners for essentially the same information, with no reliable mechanism for determining what is current, authoritative or obsolete,” Ajose-Coker said. Each additional copy makes the question harder to resolve, while unclear ownership allows more copies to accumulate without consistent review.

Resolving ownership requires help beyond the website team because the collection spans many business functions. Reports, contracts, brochures, policies, product guides and financial accounts have different owners and approval requirements. Annual reports, technical specifications, product documentation, regulatory information and formal corporate statements bring further publishing obligations. Misunderstood safety guidance and operational or instructional documents can also have serious consequences.

Even when owners identify a problem, practical barriers can prevent changes. Another department may control approval while an external design partner holds the original files, and an organization that has spent thousands on PDF design may face further revision costs. Compliance requirements can restrict changes too. Files may therefore remain untouched, while weak removal practices allow forgotten copies to persist.

Continued publication can also be deliberate: it preserves reference material, meets regulatory obligations or maintains durable records across several channels for years. These records can be particularly useful to an answering system. “PDFs very often contain exactly the kinds of information an AI system has good reason to regard as authoritative,” Ajose-Coker said. Contracts, annual reports and regulatory statements record formal positions, while technical and safety guidance explain requirements and required actions.

That usefulness changes the significance of a neglected file. “There is an interesting irony here,” Ajose-Coker said. He also observed: “Humans often think of PDFs as static attachments, while an AI retrieval system may encounter them as some of the richest and most authoritative-looking sources on an organization’s digital estate.” Headings, tables, lists and a clear reading order can make a fixed record accessible to software, although whether it takes priority over a web page depends on the question and available content.

PDFs can carry stale facts or damage their meaning

The risk begins with the information available for retrieval. In an analysis covering 187 agents and 360,000 checks, the researcher attributed 79% of AI misinformation to organizations’ own outdated content. The finding concerns organizational information broadly and places content maintenance directly within the work of reducing inaccurate AI answers.

The researcher also reports that PDFs were up to 2.5 times more likely than web pages to be valued when answers were constructed across the top 10 AI tools. The comparison is also expressed as citation likelihood. A citation identifies a document in an answer, but a document can also influence how an answer is formed. The researcher interprets the PDF format as a possible authority signal because it often carries trusted, high-stakes information.

For a business, the failure path is straightforward: an outdated or incorrect statement stays public, a system retrieves it, and the statement enters an answer. Good technical structure can make that process more efficient by helping software extract the text and its relationships. Repairing the structure therefore leaves a separate editorial question: whether the claims remain appropriate for the question being answered.

A second failure path starts with correct information that software cannot reliably recover. Machine readability means software can extract the text, connect table headings with their values, follow the intended reading order and access the relevant contents. “A PDF can look perfectly clear to a human while having poor reading order, missing structure, inaccessible tables, flattened content or other technical issues that make its meaning much harder for a machine to interpret reliably,” Ajose-Coker said. So visual approval and technical testing serve different purposes.

Three annual-report cases involving companies in the UK’s FTSE 100 stock-market index and operating in regulated industries illustrate these technical failures. In the first, visual flattening combined separate objects into one visual layer and removed some structural tags. Tags describe the role and organization of content, helping software recognize elements such as headings and table data. The report kept its intended appearance, but the change damaged information needed for machine interpretation.

The damaged structure changed the meaning of financial figures. To the AI system, table headings no longer aligned with the values below them, so numbers acquired incorrect labels and the system misquoted financial information. Checking that each number appeared on the page would have missed the defect: the crucial relationship was between each figure and what it represented.

The second annual report failed during text extraction after its design team introduced a font without embedding it. Embedding includes the necessary font information within the PDF; without it, the AI could not read the font and substituted characters, producing misinformation. A person viewing the report could see the intended words while software recovered different characters. Review therefore needed to examine the extracted text alongside the displayed design.

The third annual report was too large for reliable access. AI could reach only part of the file, leaving sections unavailable when constructing an answer, while Google Gemini could not access the same PDF at all. Partial access can leave an answer incomplete because relevant sections never reach the system. Complete exclusion creates a visibility problem: even current, useful information cannot represent the business through that document.

These cases show why publishing teams need separate checks. A technically readable report can carry stale claims, while a current report can lose table relationships, substitute characters or prevent access to entire sections. Teams must therefore assess factual suitability, preservation of meaning and access independently; passing one check gives no assurance that the other two will pass.

Historical records need a clear scope of authority

Assessing factual suitability takes care because an old document can still be accurate for its original purpose. Historical, compliance and regulatory duties may require records to stay public and unchanged. An annual report can faithfully describe an earlier reporting period yet mislead readers about present conditions. Its age becomes a problem when information about the past answers a question about today.

The same distinction applies to past commitments, policies and financial positions. A fixed record may be the right evidence for a historical question, while a current specification or guidance document may be needed for a present-day decision. Removing or suppressing all older PDFs would sacrifice useful reference material and could conflict with retention duties. Compliance restrictions may also prevent the website team from editing a record even when it recognizes the retrieval risk.

Governance must preserve that value while making each document’s role clearer. Teams need to identify the current version, retain useful metadata, descriptive information such as a document’s date, owner and status, and assess the consequences of continued public access. Historical files also need machine-readable structure so their original meaning survives extraction. These controls help systems interpret records, although AI can still confuse historical information with current authority.

Prioritize remediation by risk and test the result

Because a large PDF collection takes time to repair, the work, called remediation, needs a clear sequence. Ajose-Coker recommends four stages: inventory and risk assessment, prioritization, hybrid remediation, and execution. Hybrid remediation combines automated work with human intervention. Throughout the process, teams need to track factual currency and technical readability as separate requirements.

The first stage establishes what the organization is exposing. “Start with an inventory and risk assessment,” Ajose-Coker advised. “Identify what is public, what is still authoritative, what is duplicated, what is obsolete and what carries the greatest reputational, regulatory or commercial risk.” Finding those files requires searching across the organization’s publishing locations, including repositories that the normal website page catalogue may overlook.

Once located, each file needs an assessment that connects its condition with potential consequences. An outdated policy, a misread annual report and an inaccessible product guide create different problems, even though all are PDFs. Teams should identify files that may already contribute to misinformation and evaluate their subject matter, public availability and authority. The inventory then has a practical purpose: deciding where intervention matters most.

AI-powered tools can help by finding weaknesses against defined criteria and identifying files likely to need work. Those criteria must cover both content status and technical condition, or a readable but obsolete document may pass. Results also provide a starting point for measuring progress, so teams can compare the remaining exposure with the initial assessment as repairs proceed.

Prioritization turns the findings into work that available people and budgets can support. Ajose-Coker recommends choosing a subset based on risk, likely impact, cost and resources, since a collection of thousands of PDFs may take an extended period to address. Annual reports and policies warrant accuracy reviews because errors can affect consequential answers. Priority should also reflect the intervention each document requires.

The intervention determines who must participate. Stale claims need editorial review and a decision about their treatment, including any retention constraints; damaged structure in a current file needs technical repair. A document with both problems requires both kinds of work. Scheduling should account for the business owner, technical specialists and anyone else needed to approve changes.

For technical repairs, Ajose-Coker recommends combining automation with precise manual intervention according to the document, its content, its risk and the defect. Some repairs suit automated processing, while others demand detailed human work that consumes time, staff capacity and money. Repairs should preserve or restore tags, embed fonts and keep file sizes within the limits of the crawlers that retrieve content. Table relationships and reading order also need checking because readable text can still carry the wrong meaning.

Testing must then establish whether machines recover the repaired document correctly. Ajose-Coker recommends checking beyond its visual appearance: inspect heading-to-value associations in financial tables, examine extracted characters where fonts caused problems, and test access throughout large files. By following the failure mechanisms in the annual-report cases, these checks give each repair a concrete acceptance test.

Execution also needs milestones that reflect the collection’s size and condition. Document volume, manual effort, staffing and budget all affect delivery, so Ajose-Coker suggests aiming within 30 days for a more accurate catalogue and a clearer picture of where PDFs are distributed across the organization. That first milestone gives teams a better basis for assigning and sequencing work. It also reveals which repositories or owners need further investigation.

By 90 days, Ajose-Coker’s proposed goal is to have some documents available in structured form for AI, an inventory with remediation work mapped out, and lessons to apply to future production. These milestones establish control and demonstrate progress while the wider repair program continues. Ongoing measurement can then show which risks remain and where priorities need to change.

Govern the information before publishing the PDF

A repair program will keep gaining new work if teams continue creating PDFs as disconnected files. “Put governance in place,” Ajose-Coker advised. He recommends starting with the information used to create a document, where ownership and review requirements can be established before publication. This brings decisions about authority and currency earlier in the publishing process.

“Organizations need to get control of the information behind them,” Ajose-Coker said. His model uses structured content, information organized into defined elements, with clear ownership, metadata, versioning to track changes, and lifecycle rules that govern review, retention and retirement. “Then publish that into PDF, HTML and other channels from that controlled source.” HTML is the format used to structure web pages, so different formats can share the same approved information while serving their respective audiences.

Shared information still needs checks for each output format because producing a PDF can introduce problems with structure, text extraction or access. New files need to work for human readers and software, with continued measurement and monitoring built into publication. The commercial, regulatory and reputational exposure may also require more senior ownership of the risk. Responsibility needs to remain assigned after release so that review dates, technical checks and necessary changes continue to receive attention.

Key highlights

• Assign ownership to public PDFs: Business functions need accountable owners for documents across websites and repositories. An inventory should identify current, obsolete and duplicate files, with priority given to regulatory, reputational and commercial exposure.
• Check accuracy and machine readability separately: Publishing teams need to review factual currency alongside text extraction, table relationships, reading order and file access. A visually clear PDF can still produce incorrect figures, substituted characters or incomplete AI answers.
• Clarify historical authority: Records owners need dates, version information and status metadata that distinguish historical records from current guidance. Preserve required documents while making their intended scope clear.
• Prioritize repairs by risk: Content owners and technical specialists need a shared remediation plan based on likely impact, cost and available resources. Test repaired files against the specific defects found, including misaligned tables, unreadable fonts and inaccessible sections.
• Govern content before publication: Publishing owners need controlled source content with clear ownership, metadata, versioning and lifecycle rules. Validate each PDF output and assign responsibility for continued review after release.