Direct answer
A PDF can be perfectly readable to a person and still be unreliable evidence for an AI system. The visible page is only one representation. Retrieval often depends on a hidden text layer, character coordinates, inferred reading order, OCR, table reconstruction, and links between figures, captions, notes, and citations. Any of those layers can fail while the page still looks normal. A defensible GEO workflow therefore audits what survives extraction, not merely what appears on screen. The test is simple: can a machine recover each important claim with its subject, qualifier, unit, date, and source intact?
A PDF is a page description, not a knowledge structure
People often treat PDF as a digital sheet of paper. That is close to its strength and the source of its retrieval problems. The format is designed to preserve presentation across devices. It can store characters at coordinates, images, vector graphics, fonts, tags, and metadata, but the visible arrangement does not automatically state the logical relationship among those objects.
A two-column report illustrates the problem. A reader moves down the left column and then the right. A basic extractor may sort text by internal object order or coordinate and alternate between columns. Every word is present, yet the paragraphs become nonsense. A header repeated on 60 pages may enter the extracted text 60 times. A footnote may appear several pages away from the sentence it limits. A chart may survive only as an image while its caption is detached.
Modern document-conversion systems attempt to rebuild that structure. The Docling technical report describes a pipeline that combines the PDF text layer and geometric coordinates with page images, layout analysis, table-structure recognition, OCR when needed, and reading-order inference. The existence of such a pipeline makes the central point: reliable conversion is an interpretation task, not a copy-and-paste operation.
The six layers where meaning can disappear
PDF retrieval quality is easier to diagnose when the document is separated into layers.
| Layer | What must survive | Typical silent failure |
|---|---|---|
| Text | Characters, words, punctuation, symbols | Ligatures split; minus signs or superscripts disappear |
| Order | Paragraph, column, list, and section sequence | Two columns interleave or a sidebar enters the main argument |
| Layout | Heading levels, blocks, captions, callouts | A caption becomes an unrelated paragraph |
| Tables | Cells, headers, spans, notes, units | Values move under the wrong column or merged headings vanish |
| Visual evidence | Figures, labels, legends, plotted values | The image remains but its meaning is absent from extracted text |
| Provenance | Title, author, date, version, page, source link | An old edition is retrieved without its revision status |
These failures compound. If OCR misreads a decimal and table reconstruction assigns the result to the wrong product, the final statement may be fluent, specific, and false. Human visual inspection of the original page will not reveal what the retrieval system actually received.
Native text, scanned text, and hybrid pages
The first audit question is whether the PDF contains usable text. A native digital PDF normally has selectable characters. A scan may contain only page images. A hybrid document can contain both: some pages have embedded text while appendices, signatures, charts, or historical exhibits are images.
OCR can make image-only pages searchable, but it introduces uncertainty. Small type, skew, compression, uncommon symbols, handwriting, and low contrast all increase error. Domain notation is especially fragile. A mistaken decimal point, inequality, chemical subscript, currency symbol, or model number can alter the claim rather than merely misspell a word.
The Nougat paper approaches scientific documents as page-image-to-markup conversion. Its authors discuss alignment limits in training data and failure modes that include repetition and hallucinated output. That evidence should not be generalized into a universal error rate, but it shows why OCR-derived or model-generated markup needs validation against the rendered page. Searchability is not proof of faithful transcription.
For high-value facts, record the extraction route. “Native text,” “OCR output,” and “manually verified transcription” are different evidence states. A reader may not need to see that label in every paragraph, but the editorial record should preserve it.
Reading order is part of the claim
Order determines which qualifier controls which statement. Suppose a report places this note in a right-hand callout: “Results apply only to the enterprise plan after July 2026.” If extraction inserts the callout before an unrelated paragraph or drops it entirely, a later system may attach the restriction to the wrong result or answer without it.
Reading-order checks should include:
- Multi-column pages and pages with sidebars.
- Headers, footers, page numbers, and repeated navigation.
- Numbered procedures whose steps cross a page boundary.
- Captions placed above, below, or beside visual material.
- Footnotes and endnotes that change scope.
- Hyphenated words and sentences broken across columns.
Do not judge this from a single opening page. Reports often change layout in appendices, comparison sections, and landscape pages. Sample at least one page of every recurring layout type, then inspect every page that contains a fact likely to be quoted.
Tables require reconstruction, not plain-text dumping
A table encodes relationships through rows, columns, spanning headers, indentation, and notes. Flattening it into text can preserve all cell values while destroying those relationships. The PubTables-1M paper treats table extraction as related but distinct tasks: detection, structure recognition, and functional analysis. It also explains that visual structure does not necessarily reveal logical structure.
This is why a table audit must test questions, not screenshots. Ask which entity has a value, what the unit is, which header governs the cell, whether the value is measured or calculated, and which note qualifies it. If answers require guessing from position, the table has not survived extraction.
The neighboring prose should state the most decision-relevant conclusions in complete sentences. This does not replace the table. It supplies a second representation with explicit subjects and conditions. A sentence such as “Model B had the lowest mean power draw in the 22 °C laboratory run” travels better than a detached row containing B | 55 | lowest.
Figures and captions need a textual evidence path
Charts often carry the strongest result and the weakest machine-readable explanation. Alt text may identify that a chart exists without preserving axes, units, sample rules, error bars, or the conclusion. A caption may explain the method but omit the plotted values. OCR may read tick labels but miss color-coded series.
For every important figure, provide a concise text account near it:
- What entities or groups are compared?
- What metric and unit are used?
- What period or experimental condition applies?
- What is the principal result?
- What uncertainty or limitation changes interpretation?
- Where can the underlying data or method be found?
The account should describe the evidence rather than celebrate it. “Conversion increased dramatically” is weak. “In this fictional demonstration, completion rose from 42% to 57% over 28 days; the sample was not randomized” preserves a number, period, and boundary. If the figure is the only place where the numbers exist, publish an accessible data table as well.
A practical PDF information-loss audit
Use the following workflow before treating a PDF as a source for AI-visible content.
- Identify the authoritative edition. Record the canonical URL, document title, publisher, publication or revision date, and version number. Check whether a later edition supersedes it.
- Classify every page. Mark pages as native text, scanned image, or hybrid. Note unusual layouts such as two columns, landscape tables, forms, and foldouts.
- Run at least two extraction views. Compare browser copy, a conventional parser, OCR where needed, or a structure-aware converter. Agreement does not prove correctness, but disagreement locates risk.
- Inspect reading order. Read the extracted output without looking at the page. Check headings, paragraphs, lists, captions, notes, and page transitions.
- Test high-value facts. For each claim likely to be quoted, verify subject, predicate, value, unit, condition, date, and source against the rendered page.
- Audit tables and figures separately. Test cell-to-header relationships and create text summaries for important visuals.
- Record losses and repairs. Keep the original excerpt, extracted version, corrected transcription, page number, repair method, and reviewer.
- Publish a safer derivative when permitted. An HTML article, tagged document, accessible table, or text appendix can expose the same evidence with clearer structure.
The output is not a binary pass or fail. It is a map of which claims are safe to reuse, which need manual verification, and which cannot support automated answers.
A compact claim-level checklist
| Check | Pass condition | Editorial response when it fails |
|---|---|---|
| Identity | The correct document and edition are known | Locate the canonical version before quoting |
| Text fidelity | Critical symbols and numbers match the page | Manually verify and record the correction |
| Reading order | Qualifiers remain attached to their claims | Rewrite as an explicit text excerpt |
| Table context | Row, column, unit, and note are recoverable | Add a prose statement or accessible table |
| Figure context | Metric, axes, conditions, and result are described | Add a data table and evidence-focused caption |
| Citation path | A reader can reach the exact page or section | Add page, section, figure, or table locator |
| Freshness | Revision date and supersession status are known | Mark historical evidence or replace it |
What GEO teams should publish alongside PDFs
PDFs remain valuable for stable reports, downloadable records, print layouts, and long-form evidence. The remedy is not to abandon them. It is to give important facts a second, structured route.
Publish an HTML landing page with a plain definition of the document, its current version, a short findings summary, links to major sections, and a revision history. Put material facts in sentences whose meaning does not depend on page position. Use real HTML tables for key comparisons. Link the PDF as the complete record and state whether the HTML summary is identical, abbreviated, or updated separately.
This approach also improves correction. Replacing a PDF at the same URL can obscure what changed. A visible version log lets publishers state that a number, scope, or methodology changed on a specific date. Retrieval systems may still encounter older copies, so clear version labels inside the file and on the landing page matter.
Common mistakes
“The text is selectable, so extraction is fine.” Selectable text says little about reading order, table structure, or figure meaning.
“OCR confidence is high, so the facts are correct.” An average confidence score can hide one wrong decimal or identifier. Validate consequential tokens directly.
“The citation points to the PDF, so the answer is supported.” Page-level support still depends on whether the cited passage entails the claim and preserves its limitations.
“A newer parser removes the need for review.” Better conversion reduces loss; it does not establish that every recovered relationship is correct.
“The PDF is the official source, so derivative HTML weakens authority.” A well-governed HTML representation can clarify the same official evidence while retaining the PDF as the record copy.
Frequently asked questions
Are PDFs bad for GEO?
- They are risky when important meaning exists only in layout or images. A well-tagged, text-based PDF with clear headings, simple tables, descriptive captions, and version metadata can be useful. It is still wise to publish key facts in HTML.
Is OCR enough for a scanned report?
OCR is a starting point. Critical figures, names, dates, formulas, and qualifiers need comparison with the page image. Preserve the page locator and record manual corrections.
Should every table be repeated in prose?
Repeat the conclusions and conditions that readers are likely to ask about. Large data tables need not become enormous paragraphs, but their units, scope, method, and major findings should have an explicit text path.
How can a team measure PDF retrieval quality?
Build a set of answerable questions tied to exact pages. Score whether extraction preserves the correct fact, context, unit, qualifier, and locator. Track errors by layer—text, order, table, visual, and version—so remediation targets the actual failure.
What is the most important single check?
Take the highest-consequence claim in the document and inspect exactly what a retrieval pipeline sees. If that claim cannot travel with its conditions and provenance, the document is not ready to serve as unattended evidence.
Source and method note
This article synthesizes document-processing research with the Xindar knowledge base's claim-unit and evidence-boundary methods. The main technical sources are the Docling technical report, the Nougat paper, and the PubTables-1M paper. They describe particular systems, datasets, and experiments; they do not establish a universal accuracy rate for every PDF or retrieval product. The audit framework is an editorial procedure derived from those documented failure points. Any numerical example in this article is explicitly fictional and does not describe a commercial product.