What the AI assistants saw
A = B
ChatGPT, Gemini and Claude scored both versions the same. They read the page image, so the chart was found, or missed, in both.
A Report, Read Two Ways.
Same pages, two different files.
We laid out the same four-page brief twice, with identical pages, fonts and chart. Version A was exported with InDesign’s default settings. Version B was exported with semantic structure: tagged headings, a real table, a described chart, a set title and language, and bookmarks.
Ask for the same check on one of your reports ↗Results
What a screen reader gets
8 vs 10 of 10
Read through its structure, Version B answered all ten, in both runs. Version A has no structure to read.
What a validator sees
3,720 vs 0
Content errors in the PAC 2024 accessibility checker. Version B meets every machine-checkable requirement.
What the AI assistants saw
We asked ChatGPT, Gemini and Claude the same ten factual questions about each version, in fresh chats, twice. On the eight questions answered in the text, every assistant got every answer right in both versions. The two that only the chart answers were where they slipped, and they slipped the same way on A and B, because these assistants read the page image.
What a screen reader gets
Read through its structure, the way assistive technology and many retrieval tools read a PDF, Version B answered all ten, in both runs. Version A has no structure to read: eight of ten, and nothing on the chart.
What a validator sees
Version B meets every machine-checkable requirement in the PAC 2024 accessibility checker. Version A fails it, with 3,720 content errors and no tags, title or bookmarks.
| Version A · default export | Version B · tagged | |
|---|---|---|
| Tags | none | 176 |
| Headings | none | H1 6 · H2 6 · H3 3 |
| Table | no structure | column headers marked |
| Bookmarks | 0 | 11 |
| Title shown | none | The Last Seventeen Points |
| Language | pt-BR (wrong) | en-GB |
| Chart description | none | present |
| PAC 2024 | fails (3,720 content errors) | every machine-checkable requirement met |
Same pages, same design. The difference is in the file, and it is set once, in the template.
Method
Ask for the same check on one of your reports ↗Next: the 14-report study →Ten questions with an answer key, scored correct / partial / incorrect / not found. ChatGPT and Gemini, two runs per version; Claude, one run per version. Structure read with the PAC 2024 checker and a tag-tree extraction. Data: Eurostat. A test of extraction and retrieval, not an accessibility certification.