A source-aware AI writing workflow: citations are not enough

A cited draft is still a draft

Current research tools make source discovery much better. OpenAI documents that ChatGPT deep research can work across the public web, uploaded files, connected apps, and user-selected sites, then return a report with citations or source links. That is a useful research surface. It is not a guarantee that every adjacent sentence is supported by the cited page.

OpenAI's own accuracy guidance tells users to verify important information, including quotes, data, technical details, and references. A 2026 preprint, Cited but Not Verified, evaluated 14 closed and open models by checking whether citation links worked, contained relevant material, and actually supported the claim. In that study, the strongest models had link validity above 94% and relevance above 80%, but factual accuracy across evaluated systems ranged from 39% to 77%.

Those numbers belong to that paper's models, prompts, retrieval depth, and evaluation method. They are not a universal accuracy rate for every cited answer. The durable lesson is narrower: a working, relevant link and a supported claim are three different checks.

A citation gives you somewhere to look. Verification decides whether the sentence survives.

THE FIRST SEPARATION

Use saved sources to find the question, not to settle it

João's source index is deliberately broad. X likes capture fast commentary and demonstrations. Arc bookmarks retain pages. YouTube queues hold talks and implementation walkthroughs. This run found recurring interest in research agents, citation integrity, agent skills, MCP authorization, tool-use evaluation, and MCP Apps interfaces.

That recurrence is a topic signal, not audience demand and not proof. A saved post from Perplexity framed research as two jobs: search wide enough to find the relevant set, then investigate deeply enough to support every claim. It led to Perplexity's WANDR benchmark. The post was useful discovery context; the research page and its disclosed method are the citable source.

The same boundary applies to product behavior. A practitioner post can reveal a sharp failure mode or useful vocabulary. Current documentation from the product owner decides what the product claims to do. Primary research or a dated firsthand test supports measured outcomes. Commentary can then help explain why those outcomes matter.

THE SOURCE MAP

Map claims before writing paragraphs

The source map is the most important artifact in the workflow. It forces the research agent and editor to agree on what is being claimed, what kind of source could support it, when the source was checked, and where its authority ends.

This format prevents source laundering. A social post cannot quietly become evidence for a protocol requirement. A vendor benchmark cannot become a universal model ranking. A primary study cannot escape its tested dataset. A firsthand result cannot turn one run into a market claim.

WIDE, THEN DEEP

Retrieval coverage is part of factual quality

A claim can be well supported by the wrong slice of the literature. Source-aware writing therefore needs both coverage and claim-level verification.

AutoResearchBench tests scientific literature discovery as two separate tasks: finding a specific target paper through multi-step search and collecting an unknown set of qualifying papers. Its strongest evaluated agents reached 9.39% accuracy on the deep task and 9.31% intersection-over-union on the wide task. That is a deliberately difficult scientific benchmark, not a score for general web research. It demonstrates why "the agent searched" is not evidence that the right evidence set was found.

DeepResearch Bench makes another useful separation. It evaluates report quality with one framework and information retrieval plus citation accuracy with another. Perplexity's DRACO benchmark likewise grades factual accuracy, breadth and depth, presentation, and primary-source citation as distinct dimensions. The benchmark comes from a product vendor and uses an LLM-as-judge protocol, so its claims should be read with those disclosed limits.

The practical editorial rule is simple: generate multiple candidate searches, record what each source contributes, and name the missing evidence. Do not reward the agent for source count alone.

THE REAL CONTROL

Let an evidence gap stop the article

The failure state of most writing automations is not an obviously fabricated citation. It is a polished draft that makes the editor feel almost done. Once prose becomes coherent, narrowing a claim or publishing nothing feels like throwing away work.

A source-aware workflow makes gaps explicit before that pressure arrives:

• Authority gap: only commentary supports a product or protocol claim.
• Coverage gap: the research found examples but not the relevant set or counterexample.
• Freshness gap: the source is authoritative but may predate a fast-moving product change.
• Provenance gap: a useful assertion cannot be mapped to the exact source passage or firsthand observation.
• Visual gap: a screenshot, chart, or diagram would appear evidentiary but cannot be reproduced or attributed responsibly.

PaperTrail, a CHI 2026 study, explored a claim-evidence interface for scholarly question answering. Its interface exposed supported assertions, unsupported claims, and omissions. In a study with 26 researchers, that visibility lowered trust compared with the baseline, but the added caution did not consistently change reliance behavior. The result is a useful warning: showing provenance is valuable, but it does not remove the cognitive cost of verification or guarantee that a person will act on the warning.

That is why the gate must be procedural. If a material claim still has an authority, coverage, freshness, or provenance gap, the workflow narrows it, researches again, or publishes nothing.

A reusable checklist

1. Define one reader question. Record the direct answer and check the current site for intent collisions.
2. Search personal and public sources. Use saved material for vocabulary, examples, and leads; use the current web to refresh and challenge them.
3. Classify provenance. Separate official documentation, primary research, firsthand testing, practitioner analysis, community commentary, and inference.
4. Build the claim map. Record the source, type, checked date, and limitation before drafting material claims.
5. Search wide, then verify deep. Look for missing entities and counterexamples, then open the exact sources that support the final answer.
6. Draft within the evidence boundary. Attribute commentary, scope measurements, and label every inference that could otherwise read as fact.
7. Verify the rendered page. Check source links, structured data, visual provenance, internal links, mobile layout, and a prominent last-verified warning.
8. Keep a no-publish branch. If a material gap remains, return the research batch instead of converting schedule pressure into a weak article.