News publisher AI readiness checklist: 24 verifiable checks

Why check this

A tool purchase is not an editorial policy. Reporters and editors need a shared boundary before confidential sources, copyrighted material, personal data or unverified output enter a workflow.

Example

The policy permits reviewed transcription and document triage, restricts translation of sensitive reporting, and prohibits AI-authored allegations or fabricated source material.

How to verify it

1. List current and proposed AI uses by newsroom task.
2. Classify each use as permitted, restricted, pilot-only or prohibited.
3. Name the approving editor, legal/privacy reviewer and review date.
4. Publish the reader-facing principles and keep the operational controls internally.

Evidence to record

• Policy URL
• Approval date
• Named owners
• Review date

Common mistakes

• Publishing a vague statement that everything is human reviewed.
• Allowing individual staff to select tools with no data boundary.

Useful tools

• Paris Charter on AI and Journalism
• Newsroom policy register

Why check this

A useful output can still create source-protection, privacy, copyright, security or vendor-lock-in risk.

Example

The register names the contract, data location, retention, training preference, access roles, export method and material that must never be uploaded.

How to verify it

1. Reopen the current product terms and privacy documentation.
2. Test access controls, deletion and export on a non-sensitive sample.
3. Classify source, unpublished, personal and licensed data.
4. Record the approved configuration and an exit path.

Evidence to record

• Approved tools
• Restricted data classes
• Retention
• Exit path

Common mistakes

• Relying on a marketing claim such as enterprise secure.
• Uploading unpublished source documents before checking terms.

Useful tools

• Vendor documentation
• Data-protection assessment
• Access log

Why check this

Human oversight is not testable when no person, evidence requirement or rejection condition is recorded.

Example

The reporter checks every quote against time-coded audio; the desk editor checks names and numbers; any missing source link stops publication.

How to verify it

1. Select one workflow and map input, transformation and output.
2. Assign an owner at every irreversible editorial decision.
3. Define evidence that must accompany the output.
4. Write the stop and escalation conditions.

Evidence to record

• Named owner
• Required evidence
• Stop rule
• Escalation route

Common mistakes

• Treating a final glance as verification.
• Allowing an AI confidence score to replace a source check.

Useful tools

• Workflow card
• Editorial sign-off log

Why check this

A commodity rewrite is easy to substitute and gives a reader little reason to visit the publication.

Example

The brief names one municipal dataset, two local interviews and a reproducible calculation absent from the original announcement.

How to verify it

1. State what is already public.
2. Name the reporting action that will add information.
3. Record the source owner and verification method.
4. Remove or reframe the commission when it adds only new wording.

Evidence to record

• Original input
• Verification method
• Reporter
• Reader value

Common mistakes

• Calling a rewritten press release analysis.
• Treating novel wording as information gain.

Useful tools

• Reporting brief
• Source ledger

Why check this

Fluent text is not evidence, and a source that exists may not support the sentence placed beside it.

Example

The 38% claim is labeled as a U.S. aggregate referral observation rather than a causal France-wide estimate.

How to verify it

1. Search the draft for numbers, dates, names, superlatives and causal verbs.
2. Open the primary source for each claim.
3. Record verified, qualified or remove.
4. Keep attribution and limits close to the claim.

Evidence to record

• Verified claims
• Qualified claims
• Removed claims
• Broken sources

Common mistakes

• Citing a secondary article for a dataset it summarizes.
• Dropping geography or denominator from a number.

Useful tools

• Claim ledger
• Primary documentation

Why check this

Search and answer systems may retrieve a passage without the paragraphs that originally qualified it.

Example

A percentage remains attached to its sample, period and country, and the paragraph states what the evidence does not prove.

How to verify it

1. Copy the paragraph into a blank document.
2. Underline the direct finding and boundary.
3. Confirm the source and qualification remain visible.
4. Rewrite the paragraph if a reader could overgeneralize it.

Evidence to record

• Finding
• Boundary
• Visible source
• Limitation

Common mistakes

• Turning the article into disconnected FAQs.
• Moving every caveat to the conclusion.

Useful tools

• Citation-ready passage test
• Skeptical review

Why check this

Operators document different bots or control tokens for search discovery, model training and a user-requested fetch. Blocking one job should not accidentally block another.

Example

The matrix treats Googlebot, Google-Extended, OAI-SearchBot and GPTBot as separate policy entries with owners and review dates.

How to verify it

1. Inventory official operator documentation.
2. Record purpose, control mechanism, business decision and owner.
3. Implement matching robots, CDN and WAF rules.
4. Schedule a policy review after platform changes.

Evidence to record

• Operator
• Purpose
• Decision
• Implementation
• Review date

Common mistakes

• Blocking every bot containing AI.
• Treating a control token as a request user agent.

Useful tools

• AI crawler guide
• robots.txt
• CDN/WAF policy

Why check this

An allow rule in robots.txt does not override a firewall denial, rate limit, interstitial or challenge page.

Example

Tests cover a fresh story, updated story, evergreen explainer, paywall preview, lead image, news sitemap and main sitemap.

How to verify it

1. Choose representative URLs and declare expected behavior.
2. Verify bot identity using the operator's method.
3. Record response, redirect, bytes, latency and body signature.
4. Fix challenge, 403, 429, 5xx or false-200 behavior and retest.

Evidence to record

• Final status
• Challenge state
• Body match
• Image access
• Latency

Common mistakes

• Testing only the homepage.
• Returning a challenge page with 200 OK.

Useful tools

• Server/CDN logs
• HTTP test
• Bot verification

Why check this

Conflicting dates and URLs make it harder for readers and systems to understand which version is authoritative and what changed.

Example

A developing story keeps one canonical; a material update changes dateModified and adds a note; the news sitemap contains appropriate recent URLs.

How to verify it

1. Compare visible and machine-readable dates.
2. Inspect canonical, redirects, sitemap and internal links.
3. Define what counts as a material update.
4. Remove artificial freshening and repair contradictory versions.

Evidence to record

• Canonical
• datePublished
• dateModified
• Update note
• Sitemap state

Common mistakes

• Resetting the publication date after a spelling fix.
• Creating a new URL for every minor development.

Useful tools

• News sitemap
• URL Inspection
• Structured data validator

Why check this

Readers need to know who is accountable and when the reporting was produced or materially changed.

Example

The H1, title and article link identify the same story; the byline names the journalist; published and updated labels are unambiguous.

How to verify it

1. Review the logged-out article template on mobile and desktop.
2. Compare headline sources and remove contradictions.
3. Open the author page and verify credentials and recent work.
4. Check date labels and timezone output.

Evidence to record

• Headline match
• Human author
• Author URL
• Published date
• Updated date

Common mistakes

• Using AI as the author name.
• Displaying By followed by an empty author.

Useful tools

• Rendered page
• Accessibility tree
• Article validator

Why check this

A reader should not have to infer whether a claim is reported fact, interpretation, personal opinion or paid material.

Example

The article shows Analysis above the headline and explains the method; sponsored work is labeled before the reader reaches the claim.

How to verify it

1. Inventory content types and paid relationships.
2. Choose consistent visible labels and definitions.
3. Check feeds, cards, article pages and structured data.
4. Correct unlabeled or misleading examples.

Evidence to record

• Content type
• Disclosure visibility
• Feed consistency
• Owner

Common mistakes

• Hiding a sponsorship note in the footer.
• Presenting a columnist's inference as reported fact.

Useful tools

• Editorial taxonomy
• Commercial disclosure policy

Why check this

Structured data can reinforce visible facts but cannot repair missing authorship, publisher information or accountability.

Example

NewsArticle names the visible headline, dates, author, publisher and representative image; policy and contact links resolve.

How to verify it

1. Open the About, contact, editorial, corrections and ownership pages.
2. Validate Article or NewsArticle data.
3. Compare each property with the visible story.
4. Remove duplicate generators and invented values.

Evidence to record

• Policy links
• Contact route
• Schema errors
• Visible-data match

Common mistakes

• Assuming valid schema guarantees Google News inclusion.
• Publishing dates in markup that readers cannot see.

Useful tools

• Rich Results Test
• Schema validator
• Publisher transparency audit

Why check this

Discover is driven by interests rather than only typed queries. A keyword does not provide the story, audience or editorial turn.

Example

Interest: electric mobility in Île-de-France; turn: a new local dataset shows a charging-access gap by commune.

How to verify it

1. Name the durable interest or community.
2. Write what changed and who is affected.
3. Identify the original evidence.
4. Cancel packaging when the story lacks a real turn.

Evidence to record

• Interest
• What changed
• Affected reader
• Original evidence

Common mistakes

• Treating a broad category as a story.
• Choosing a trend before verifying a reportable development.

Useful tools

• Story brief
• Audience questions
• Google Trends as a lead only

Why check this

Google recommends non-misleading headlines and large, relevant images; logos, generic stock art and text-heavy images weaken the reader proposition.

Example

A 1200×675 first-hand photograph or original explanatory visual is crawlable, captioned where needed and allowed through max-image-preview:large.

How to verify it

1. Compare headline, deck, image and article conclusion.
2. Verify image rights, subject relevance and source.
3. Confirm width, pixels, responsive variants and large-preview permission.
4. Inspect feed-like, 4:3 and square crops.

Evidence to record

• Promise match
• Image width
• Rights
• Crop result
• Large preview

Common mistakes

• Withholding the central fact as clickbait.
• Using the publication logo as the story image.

Useful tools

• Image inspector
• Search Console Discover guidance
• Crop QA

Why check this

A volatile feed visit has little durable value when the article is a dead end and the publisher offers no reason to return.

Example

A local housing story links to the maintained rent tracker and a weekly local-policy briefing with a clear editorial promise.

How to verify it

1. Identify the next reader question.
2. Choose one maintained destination or direct relationship.
3. Add a descriptive contextual link after delivering value.
4. Measure qualified continuation and return behavior.

Evidence to record

• Next destination
• CTA visibility
• Continuation rate
• Return or subscription

Common mistakes

• Adding an unrelated newsletter popup before the article.
• Linking every story to the homepage.

Useful tools

• Topic hub
• Newsletter
• Preferred Sources deep link

Why check this

AI answer systems may extract material outside its original paragraph sequence.

Example

A traffic figure includes sample, geography, period and the explicit statement that the observation does not prove AI caused the decline.

How to verify it

1. Find the three most quotable claims.
2. Check source, scope, distinction and limit in place.
3. Add an as-of date for changing product facts.
4. Remove unsupported causal or universal wording.

Evidence to record

• Attribution
• Scope
• As-of date
• Limitation

Common mistakes

• Adding an AI-specific schema that does not describe the page.
• Writing fragments only for citation at the cost of a coherent article.

Useful tools

• Citation-ready passage test
• Claim ledger

Why check this

First-hand media can differentiate a publisher, but hidden, inaccessible or poorly attributed media cannot reliably support the story.

Example

An interview video has a reviewed transcript, named speakers, time codes, a descriptive poster and a canonical article linking the claims to source context.

How to verify it

1. Inventory the evidence carried by each media asset.
2. Check rights, URLs, captions, alt text and transcripts.
3. Verify playback and access logged out on mobile.
4. Keep critical facts in accessible article text.

Evidence to record

• Stable URL
• Rights
• Transcript
• Alt text
• Playback

Common mistakes

• Placing the only important number inside an infographic.
• Publishing an unreviewed automatic transcript.

Useful tools

• Media inventory
• Transcript editor
• Accessibility review

Why check this

Multi-format distribution expands reach only when every version preserves the source, caveat, rights and useful destination.

Example

The short video shows one finding and links to the canonical investigation; the newsletter adds context; every version carries the corrected number.

How to verify it

1. Lock the approved facts, media and correction state.
2. Define the reader job for each format.
3. Use AI only for bounded drafting or adaptation.
4. Run a cross-format factual and destination review.

Evidence to record

• Formats
• Conclusion match
• Source link
• Correction propagation

Common mistakes

• Copying the same text to every platform.
• Correcting the article but leaving the false social card live.

Useful tools

• Distribution brief
• Platform properties
• Correction log

Why check this

One organic total cannot show whether a loss came from access, demand, visibility, answer substitution, weak landing pages or poor conversion.

Example

The baseline stores bot status, indexed URLs, News/Discover/Search metrics, AI reports where available, landing sessions and new subscribers for the same period.

How to verify it

1. Define the property, timezone, period and content taxonomy.
2. Export each available report separately.
3. Record absent or thresholded data as unavailable, not zero.
4. Save crawler and analytics evidence beside editorial changes.

Evidence to record

• Access
• Index
• Search
• News
• Discover
• AI visibility
• Referrals
• Outcomes

Common mistakes

• Merging impressions and visits.
• Treating missing report rows as zero exposure.

Useful tools

• Search Console
• Bing Webmaster Tools
• Analytics
• Server logs

Why check this

AI answer exposure, Discover volatility and reader intent are not uniform across every article or audience.

Example

The review compares original local reporting with commodity explainers in France on mobile rather than assigning one sitewide decline to AI.

How to verify it

1. Create a fixed editorial taxonomy.
2. Tag original versus derivative work.
3. Record geography, device, language and dates.
4. Compare matched groups and list confounders.

Evidence to record

• Content type
• Originality
• Country
• Device
• Exposure group

Common mistakes

• Using one viral story as the benchmark.
• Comparing different countries and periods without qualification.

Useful tools

• Content inventory
• Search Console filters
• Analytics exploration

Why check this

A visit or citation is fragile when the publication does not earn consent to continue the relationship.

Example

A reader of the transport investigation can follow the topic, add the domain as a preferred source or join a briefing that promises verified local updates.

How to verify it

1. Define the value of the next relationship.
2. Place the CTA after a relevant story payoff.
3. Preserve consent and attribution.
4. Measure signup quality, return behavior and cancellation—not clicks alone.

Evidence to record

• Qualified CTA
• Signup
• Return
• Subscription
• Cancellation

Common mistakes

• Showing an unrelated modal before the headline.
• Optimizing email captures without measuring reader value.

Useful tools

• Newsletter
• App or alert
• Preferred Sources
• First-party analytics

Why check this

A workflow can appear efficient while silently moving errors or risks downstream.

Example

A monthly sample records hallucinated names, transcript errors, missed caveats, false positives and the corrective action by tool and task.

How to verify it

1. Choose a representative random and high-risk sample.
2. Use a stable rubric and two-person review for severe risks.
3. Record error type, severity and source.
4. Pause or repair workflows that cross the stop threshold.

Evidence to record

• Sample
• Error rate
• Severity
• Correction
• Workflow status

Common mistakes

• Reviewing only published successes.
• Changing the rubric after seeing results.

Useful tools

• Quality audit
• Corrections log
• Incident register

Why check this

Changing headlines, images, templates, crawler rules and formats together makes the result impossible to interpret.

Example

A fixed story group receives improved first-hand images while headline policy, template and distribution cadence remain stable for the window.

How to verify it

1. Write the hypothesis and non-claim.
2. Save the pre-change baseline.
3. Apply one controlled change to a defined group.
4. Review primary and guardrail metrics before deciding.

Evidence to record

• Baseline
• Change
• Primary metric
• Guardrail
• Decision rule

Common mistakes

• Calling a before/after observation causal.
• Stopping when the first positive result appears.

Useful tools

• Experiment card
• Release log
• Performance reports

Why check this

Product names, crawler purposes, report fields, policy controls and site configuration can change after the workflow is approved.

Example

The monthly record notes documentation changes, expired policy decisions, new errors, unavailable reports and the owner of each action.

How to verify it

1. Review the official source registry and platform updates.
2. Retest one normal and one edge-case article.
3. Inspect errors and documentation drift.
4. Assign actions, deadlines and the next review date.

Evidence to record

• Documentation changes
• Failed checks
• Expired decisions
• Action owners

Common mistakes

• Using a year-old crawler list.
• Assuming a plugin setting proves the live output.

Useful tools

• Primary-source registry
• Template audit
• Monthly review log