An editor needs an evidence-first evaluation that distinguishes a plausible factual answer from a verifiable page or paragraph locator. Here is a scenario-based comparison using documentation checked on 2026-10-25, not an independently tested vendor ranking. The region, plan, exact model version and price are unspecified unless a first-party document establishes them; unmatched performance claims stay unknown.
Audience: fact-checkers and librarians. Evidence checked: 2026-10-25. Editorial status: ready. Comparison scope: named scenario; region, plan and model/version parity require separate confirmation.
Define a source locator before measuring accuracy
A locator identifies the publisher, exact document edition, page or section and the supporting text. A model citing only a publication homepage has not located a claim. For a scanned PDF, page index may differ from the printed page number. For a website, a heading and retrieval timestamp are often more reliable than a line number that shifts after dynamic rendering. Build a labelled answer key manually without AI-generated answer locations.
Use a deliberately adversarial source set
Collect ten born-digital PDFs, ten scanned reports, ten pages with tables, and ten frequently revised web documents. Include two near-duplicate editions and one source that explicitly contradicts the question. Record hashes and retrieval dates. Ask systems for supporting locator plus short quoted evidence; limit quoting to rights-permitted excerpts. Mark wrong publication, wrong edition, fabricated citation and correct locator as distinct outcomes rather than compressing them into overall confidence.
Capability availability is not the same as quality
OpenAI’s file-input guidance differentiates how PDFs and other files are processed; the path and model matter. Gemini Apps documents support for uploaded material but also platform limits. Claude project knowledge is a storage/retrieval context rather than a guarantee of verbatim page references. No cited document provides a comparable, independently controlled multi-vendor page accuracy score. This guide proposes a test; it does not claim that its outcomes were observed.
Publish a reproducible error report
For every example, capture query, input artifact ID, provider surface, model version when disclosed, output and reviewer notes. Show a confusion matrix for supported, unsupported and mislocated claims. Re-run only after intentionally changing a version or corpus. Prefer the workflow with fewer untraceable statements and lower human verification effort; a visually attractive bibliography is not sufficient.
What an actual buyer pilot should measure
Freeze the relevant source files, release identifiers, evaluation prompts, operator permissions and reviewer identities. Use the same rights-cleared materials for every candidate; separate tasks that cannot be configured equivalently instead of forcing them into a single ranking. Review the resulting artifacts against objective acceptance criteria specific to this decision: source traceability, error handling, rework, permission boundaries and whether a competent human approves the final deliverable. Record rejected attempts as well as successful examples. This is a proposed method, not a claim that these vendors have been tested by this publication.
Costs and unresolved evidence
Do not equate an interactive assistant seat, media credit, billable character, hosted token or self-hosted GPU hour. To estimate actual spending, count accepted results, failed attempts, revision cycles, human review, security operations and export/storage costs. AI 10 must supply dated provider-specific pricing records with tier, unit, currency and region; in their absence, monetary figures and claims of price parity remain unknown. Account-level availability and contract addenda require verification. No universal winner score or fictitious hands-on rating is assigned.
Decision for this scenario
Buy for locator reproducibility, not the number of links produced; publish error categories and independent verification effort alongside any benchmark.
The article has documented feature and policy evidence but no independent hands-on performance test. Reconfirm plan, region and terms before publication; editorial ready is not WordPress publication approval.
Sources and Further Reading
Original provider and policy sources are listed here. Source-to-claim identifiers are maintained separately in the editorial CSV rather than inserting numbered references into the main text. Where documents are mutable, the access date describes the verification point only.
- OpenAI — File inputs for the OpenAI API (checked 2026-10-25)
- Google — Upload and analyze files in Gemini Apps (checked 2026-10-25)
- Anthropic — What are projects? (checked 2026-10-25)