Auto
HTML vs PDF vs Slides: Which Formats Do AI Systems Prefer to Cite?
By Kiran Sharma, Growth-Oriented Content Strategist
If you publish the same idea as an article, a PDF, and a slide deck, which version will an AI system cite?
The short answer
For most businesses, use HTML as the main, public version of an article. Offer a PDF when readers need a fixed-layout or downloadable copy. Use slides to support a talk, not as the only version of an important written explanation.
Search support varies by platform. Google lists both PDF and PowerPoint among its indexable file types. OpenAI, meanwhile, tells publishers that allowing its OAI-SearchBot helps content appear in ChatGPT search. Make sure your content is accessible before worrying about citations. Google’s indexable file types, OpenAI’s crawler guidance
Document-parsing research tests whether systems recover content and layout. AI-citation research tracks which sources answer engines retrieve and cite. OmniDocBench, a document-parsing benchmark that includes slides, a 2026 preprint on AI citation and retrieval
Comparing HTML, PDF, and slides
A well-made PDF can be easier to extract than a poorly built web page. Slides with clear text are easier to read than slides made of screenshots.
| Format | Crawlability | Rendering across devices | Text extraction and snippet risk | Citation frequency |
|---|---|---|---|---|
| HTML article | Strong when public and linked. Client-side JavaScript can add friction if the main text is missing from the initial page or does not appear after rendering. | Usually the easiest to adapt to phones, tablets, and desktops. Keep the same main content available on mobile and desktop. | Often the lowest-risk choice when headings, paragraphs, lists, and tables use clear structure. | No dependable rate by format. |
| Text-based PDF | Google supports PDF indexing. Access, file settings, and the quality of the PDF still matter. | Reliable for preserving page layout, but small text may be hard to read on phones. Reflow depends on the PDF’s structure and reader support. | Can be good when text is searchable and the reading order is correct. Columns, tables, and untagged layouts can scramble extracted text. | No dependable rate by format. |
| PowerPoint or slide deck | Google lists PowerPoint files as indexable. That does not establish support across every AI search system or presentation viewer. | Works well for presenting a visual story. A fixed slide canvas can require zooming on a small screen. | Text may be split among objects, placed out of visual order, or embedded in images. Accessible titles, reading order, and descriptions reduce some of this risk. | No dependable rate by format. |
Google explains that some JavaScript content can be difficult for search systems to render, and recommends approaches such as server-side or static rendering where needed. For mobile indexing, Google also advises keeping the main content equivalent between mobile and desktop versions. Google’s JavaScript SEO guide, Google’s mobile-first indexing guidance
PDF and slide decks have their own structure problems. A tagged PDF carries information about the intended reading order; without correct tags, a multi-column page may be read in an order that makes little sense. Microsoft’s accessibility guidance warns that a slide’s screen-reader reading order can differ from the order people see on the slide. Visual clarity does not automatically create machine clarity. W3C guidance on PDF reading order, Microsoft guidance on accessible PowerPoint
How to measure citation frequency and snippet accuracy
A useful test must separate three questions:
- Could the system reach the file?
- Could it extract the right content?
- Did it choose to cite the file for the query?
If those questions are blended together, a PDF that was never crawled could look like a citation failure, when the real problem was access.
Build a matched test set
Create a set of short content pieces with the same text, claims, numbers, tables, and image descriptions in each format. Keep a “gold” version of the content so reviewers know exactly what each version should contain.
Publishing identical copies can lead Google to choose one as the main version. Link to the version you want readers to find. Google’s canonicalization guidance
Make sure each URL loads without an account, allows relevant crawlers, and shows the important text. Then read it on desktop, tablet, and phone.
Check your page at a narrow phone width. W3C uses 320 CSS pixels for many text-based pages; presentations may need different handling. W3C guidance on reflow
An AI answer may paraphrase several sources. Google selects search-result snippets based on the query, so one carefully written sentence will not always appear. Google’s site-appearance FAQ
A decision tree for choosing a format
Is this a written article or explanation that should keep attracting readers over time?
- Yes: Publish an HTML article as the main version.
- No: Continue.
Does the reader need a downloadable copy or a stable page layout?
- Yes: Add a searchable, accessible PDF, with an HTML page that explains what it contains.
- No: Continue.
Is the content mainly a visual aid for a live presentation?
- Yes: Publish the slides, but also provide an HTML summary or transcript for people who need the full explanation.
- No: Use the format that best serves the reader, but keep important facts available as readable text.
The simplest default for most product teams is HTML for discovery, PDF for download, slides for presenting.
How to make a PDF easier for machines to read
A PDF is not automatically machine-hostile. It becomes risky when its text, structure, or reading order is missing or wrong.
Start with real text
Create the PDF from a document with selectable text when possible. If you must scan a page, run OCR to add a searchable text layer—and proofread the result. Adobe notes that a scanned PDF may contain only image data until OCR creates searchable text. Adobe’s OCR guidance
Add structure and check the reading order
Use tags for headings, paragraphs, lists, tables, and figures. Check that columns, captions, footnotes, and tables are read in the intended order. W3C warns that even complex PDFs converted to tagged format may need manual correction. W3C’s PDF reading-order technique
Describe visuals and keep key information in text
Give meaningful figures and charts text alternatives. Do not put a critical claim only inside a chart image when it can also be stated in the surrounding text. This helps people who use assistive technology and gives extraction tools a better chance of finding the point of the visual. W3C’s PDF accessibility techniques
Add clear metadata and a useful landing page
Give the file a clear title and author. Use a descriptive filename and link to it with meaningful link text, not just “download.” Google identifies PDF title metadata and links pointing to a PDF as signals it can use for the title shown in search results. Google’s PDF search guidance
If the HTML page should be the main search result and the PDF is a duplicate copy, Google supports a rel="canonical" HTTP header for non-HTML files such as PDFs. Google’s canonical-header guidance
Do not treat archival standards as citation tricks
PDF/UA concerns accessibility; PDF/A concerns long-term preservation. Neither standard guarantees that an AI system will cite a file. The Library of Congress’s current format guidance favors accessible, searchable, tagged PDFs for page-layout works, alongside structured and markup formats. Library of Congress Recommended Formats Statement
Archiving and DOI practices that protect the work
A DOI helps people identify and cite a specific work over time. It is not a ranking boost, and it does not automatically preserve a file.
For a lasting report, research article, or presentation:
- Keep a stable HTML landing page with the title, creator, date, summary, version, and access to the file.
- Make the DOI resolve to that landing page, rather than directly to a PDF. DataCite recommends landing pages because they give both readers and machines context for the item. DataCite DOI landing-page guidance
- Update the DOI’s location metadata when the work moves. Crossref provides for updating resolution URLs as hosting changes. Crossref metadata guidance
- Use versioning when a major update changes the substance of the work. DataCite recommends linking major versions with related identifiers; Zenodo gives published versions separate records and identifiers so a cited version’s files remain fixed. DataCite versioning guidance, Zenodo record guidance
- Preserve the source files as well as the final PDF. Zenodo notes that it guarantees bit-level preservation, but a file format may still become hard to read in the future.
The simplest content system for a busy founder
Choose the format after deciding what the content should say and who it should help. Keep one source of truth, then publish the right version for each use: an HTML article for readers, a PDF for download, and slides for a talk.
AutoPod’s Creator, Audience, and Taste inputs guide article generation so your point of view carries through.
Measure each asset by its business job. Track article readers who reach a product page, PDF downloads that assist later conversions, and leads or visits from slide decks. **AI citations are one useful outcome, not the whole content strategy.
