How the diagnostic works.

What the result does

It compares eight common PDF problems, chooses a first action, recommends one of seven workflows, creates a retry instruction, and sets the verification level. The copy-and-paste question helps distinguish reliable text from image-only, mixed, or garbled text. You can answer all six questions without choosing a file, and only those answers determine the result.

What the optional local check does

If you choose a PDF, your browser checks whether the file can be read, counts its pages, and looks at orientation and available text on at most 12 pages. It may show one short text preview in browser memory. If the checked text contains the replacement character �, it flags that copied text as worth comparing with the visible PDF. That warning applies only to sampled pages. These observations can help you answer the questions, but they do not change an answer automatically or determine the result.

How the page-and-text spot check works

After a successful local PDF check, you can choose up to three pages from the existing sample and compare a local page rendering with text extracted from the same page. You decide whether each page matches closely, has a mismatch, or is hard to judge. The comparison stays in your browser and never answers a diagnostic question or changes the result. A matching sample does not prove that the rest of the PDF is accurate, complete, accessible, or ready for an AI tool.

Want to perform this comparison directly? Open the PDF Text Layer Checker.

What optional local preparation does

After a successful local check, you can extract pages, split the PDF, or rotate pages. Your browser creates new PDF copies on your device without changing the original or sending the source or output to Before You Retry. Preparation may make the PDF easier to work with, but it does not show what an AI platform will receive, count as diagnostic evidence, change the result or score, or guarantee that AI will succeed.

What the local check cannot establish

The local check cannot show that an AI platform received the same text, tell whether a PDF is “AI ready,” or predict how a platform will behave. It also does not detect tables, columns, reading order, form fields, signatures, OCR quality, accessibility quality, language, or encryption status beyond reporting that a protected file could not be read.

When there is not enough evidence

If the clues are too weak to support a likely cause, the result does not force one or assign a confidence label. Instead, it gives you one copy-and-paste check and asks you to try the diagnostic again with the new observation.

How confidence works

The label describes how well your selected clues agree with the result:

• Possible based on your answers — the clues are limited or could fit more than one cause.
• Likely based on your answers — the leading cause has useful supporting clues.
• Strong workflow match based on your answers — the main symptom is reinforced by another relevant answer.

These labels are not probabilities. Choosing “Not sure” does not count as confirming evidence.

What the result cannot prove

The result is a workflow match based on your six answers. It is not a verified diagnosis or proof of why an AI response failed, and the optional local check does not change that. Before You Retry does not publish or claim a measured accuracy rate.

How the learning measurement works

Schema-v2 measurement defines a qualified candidate session as the first diagnostic start, or 10 seconds of focused and visible attention followed by a trusted pointer or keyboard interaction. Qualification does not certify a person: bots, shared devices, blocked scripts, duplicate tabs, and missing events remain measurement limitations.

For schema v2, a qualified candidate session has a utility start when it contains checker_started or diagnostic_started. This prospective definition is not backcast into schema-v1 records, and historical stored events are not rewritten.

Funnel denominators use distinct qualified tab sessions within one exact product release, analytics schema, and experiment. Result and outcome measures use distinct diagnostic runs. Staging, owner, operator, audit, production-smoke, and synthetic-monitoring traffic are excluded from governed public-learning cohorts.

How outcome evidence is interpreted

The optional result control is a voluntary, fixed-choice self-report. A response does not prove that the recommendation caused the outcome. Reports must show the response rate alongside solved, partly solved, and not-solved shares so response bias remains visible. “I haven’t tried it yet” is counted among respondents but is neither success nor failure and is excluded from those resolution shares.

Learning-cycle and data limits

Workers Analytics Engine can sample at write and query time, so exact event sequences and counts are not guaranteed. Product learning uses a 30–45-day review cycle, subject to minimum sample and response-rate requirements, and must state its data-through time, exclusions, and sampling limits.

Cloudflare’s provider-managed Analytics Engine retention is three months. Governed reports enforce a maximum 90-day lookback. The lookback is a query boundary, not evidence of deletion at exactly 90 days. There is no raw-event archive or export and no stable cross-session identity.

Tool guidance is separate

The diagnosis does not rank AI tools by brand. Tool guidance explains what kind of document support may matter after the first fix. Check the vendor’s current documentation before paying for a tool or using it with a sensitive document. Guidance was last reviewed August 14, 2026.