QA & fact-checking: Claims checked before anything ships

Most AI writing tools stop at a confident draft. Jeenie runs a fact-check between write and publish — extract claims, gather evidence, verdict, optional rewrite — so shaky statements get caught while the job is still in flight.

Extract checkable claims from the draft

After the writer finishes, Jeenie pulls statistics, dates, comparative statements, and named facts — not opinions, brand voice, or vague advice.

Each claim gets a short search query. The job stays in flight; nothing publishes until this pass completes.

Gather live SERP evidence

Claims are weighed against organic search results — the same kind of public sources a careful editor would open — not against the model’s memory.

Research from earlier in the job is reused as fallback evidence. Live lookups run when search is configured, so verdicts aren’t a closed-loop LLM self-check.

Verdict each claim: supported, unsupported, or uncertain

Every claim gets a verdict, a confidence score, and a rationale tied to the evidence snippets. Thin evidence prefers uncertain over a fake “unsupported.”

The report is attached to the job. You can see what was checked, what backed it, and what didn’t — not a chatbot disclaimer at the bottom of the post.

Revise unsupported claims before publish

Unsupported statements can be removed or softened in the HTML while structure and brand tone stay intact. No new statistics are invented to paper over a miss.

Then the pipeline continues: cover image, persist, publish. QA is a step in the loop, not a Slack thread after the URL is live.

A report on the job — not a disclaimer on the post

After the pass, you get verdicts, rationales, and source domains attached to the job. If something was rewritten, that’s on the record too.

Fact-checking sits between draft and cover image in every content job. See the full loop from brand analysis to live CMS post.

Fact-check report

serp+llm · revised

Checked 3 claims: 1 supported, 1 uncertain, 1 unsupported. Softened the unsupported statistic before publish.

  • supported

    Ecommerce email open rates typically land in the high teens to low twenties for mid-market brands.

    Industry benchmarks from major ESPs report average open rates in this range.

    Evidence e.g. mailchimp.com

  • uncertain

    Most Shopify stores see a lift within two weeks of publishing weekly category content.

    Evidence describes content cadence benefits, but not a two-week lift for most stores.

    Evidence e.g. shopify.com

  • unsupported

    Customers typically cut content costs by 73% in the first quarter.

    No source backs a 73% first-quarter reduction. Specific figure looks invented.

    Evidence e.g. No matching source

What actually gets checked

Opinions and brand voice stay. Checkable facts — the ones that can be wrong in public — go through the pass.

  • Statistics and numbers

    Percentages, dollar figures, and “typical” rates get pulled out and checked — the claims most likely to embarrass you after go-live.

  • Dates and named facts

    Specific years, product names, and absolute assertions are treated as checkable, not decorative color.

  • Comparative statements

    “Faster than,” “most teams,” and similar comparisons need evidence. If the SERP doesn’t back it, the verdict is uncertain or unsupported.

  • Invented specifics

    Hallucinated stats and dead references get flagged. Unsupported claims can be softened before the cover image even runs.

Why the check comes before publish

Speed without checks just moves embarrassment earlier. QA is the constraint that keeps autopilot responsible when nobody is in the doc.

  • Shaky stats don’t ship

    Numbers and absolute claims are checked while the job is still running — not after a customer screenshots the post.

  • A report you can audit

    Verdicts, rationales, and evidence stay on the job. Your team can see what the model was told to stand behind.

  • Rewrites happen in-flight

    Unsupported claims can be softened before cover and publish. Editorial review starts from a cleaner draft, not a hallucination hunt.

  • Editorial time goes to judgment

    This is editorial QA, not legal or medical review. Humans still decide what to publish — they just spend less time catching invented numbers.

Put Your Content Publishing on Autopilot.

Connect your CMS, set your cadence, and let Jeenie handle the rest from research to draft to post.