AI-slop content cleanup: spotting hallucinated facts before they cost you

We ground the method in a controlled proof, not anecdote. Patrick Stox at Ahrefs built a synthetic luxury brand called Xarumei, published contradictory narratives, and tested how eight major AI tools handled truth. We documented that experiment in AI experiment: the fake brand Xarumei and LLM hallucinations. The lesson for WordPress publishers is direct: models prefer specific fiction over vague truth, and third-party articles that look like journalism can override your official FAQ. If your site reads like Xarumei’s fake Medium investigation, you are feeding the same failure mode.

When slop compounds with broken code and plugin sprawl, the remediation path sits inside AI-built website rescue. When you need to know whether models already echo your fabrications, pair cleanup with first-party measurement from how we tracked wppoland.com’s own AI citations for 90 days.

#Why YMYL WordPress sites fail this audit first

YMYL publishers ship faster with AI because drafting is the bottleneck, not layout. A clinic network, a fund administrator, or a compliance consultancy can generate forty service area pages in an afternoon. WordPress makes publishing trivial: duplicate a template, swap {city}, hit publish.

The failure mode is not “low quality writing.” It is ungrounded specificity. A model will invent:

• a 47.3% improvement where no study exists,
• a Harvard Business Review citation that points nowhere,
• a Dr. Elena Kowalska, MD, PhD who does not appear in the medical register,
• a case study dated 2024 for a client signed in 2026.

Google’s quality raters and procurement lawyers ask the same question: can you show your work? AI-slop fails that test even when the prose is fluent. EEAT does not mean “sounds expert.” It means demonstrable experience, verifiable expertise, recognizable authority, and trustworthy presentation. Hallucinated facts attack all four at once.

#What the Xarumei experiment proves about your content stack

Stox’s experiment ran in two phases. Phase one asked models about a brand that should not exist. GPT-4 and GPT-5 refused most fabrications. Perplexity confused Xarumei with Xiaomi. Copilot invented elaborate praise to match user framing.

Phase two is the warning for WordPress operators. Stox published:

1. An official FAQ denying rumors in plain language.
2. A glossy blog with 23 master craftsmen at 2847 Meridian Blvd, Nova City, CA, plus a fake Emma Stone endorsement.
3. A Reddit AMA, chosen because Reddit is heavily cited in AI answers.
4. A Medium “investigation” that debunked obvious lies, then inserted new fabrications: founder Jennifer Lawson, location Portland.

Gemini, Grok, AI Overview, Perplexity, and Copilot cited Medium over the FAQ. They preferred the investigative tone. When the FAQ said “we do not publish unit numbers” and fake sources claimed 634 units in 2023, models chose the number.

Translate that to WordPress:

Your site is the FAQ. Your AI-drafted blog cluster is the Medium article. Cleanup means making the FAQ more specific with sources, not more adjectives.

#Red flags and verification steps

Use this table as a first-pass triage before any rewrite budget is approved.

Print the table. Run it against your top twenty revenue URLs first.

#Fake stats: the Xarumei preference for numbers

Models treat numbers as authority signals. Stox showed that 634 units beat “we do not disclose production figures.” On WordPress, the damage shows up in hero sections, comparison tables, and FAQ blocks exported from AI chat sessions.

#Symptoms

• Percentages with one decimal place and no source (41.6% reduction in ticket volume).
• Before/after metrics with no measurement window.
• “Industry average” lines without naming the survey.
• Identical stats repeated across six city pages (copy-paste sprawl).

#Verification workflow

1. Source log: Create a spreadsheet column: claim, primary URL, author, date accessed, owner. No URL, no publish.
2. Round-number rule: Integers like 50%, 10x, 99.9% are hallucination priors unless tied to an instrument.
3. Instrument match: If you claim Core Web Vitals improvement, link CrUX or Lighthouse export with date range.
4. Downgrade path: Replace unsourced stats with qualitative outcomes you can defend in a client call.

For regulated copy, keep lawyer-reviewed PDFs as the canonical source and paraphrase lightly in HTML.

#Fake citations and phantom authorities

AI assistants love brand-name drops. HBR, McKinsey, Forrester, NIST, ISO: the names lend tone without work.

#Symptoms

• Parenthetical citations with no link.
• Links to homepages instead of specific reports.
• DOI-shaped strings that 404.
• “As noted in the 2024 WordPress security whitepaper” with no author.

#Verification workflow

1. Click every outbound reference. Dead link = delete claim.
2. For standards (ISO 27001, WCAG 2.2), link the official spec page, not a SEO blog summary.
3. Store license numbers externally (solicitor register, FCA, KNF) and mirror exactly on site.
4. Add citation microcopy in footnotes for long guides; models scrape footnotes surprisingly often.

If a citation cannot be verified in ten minutes, it was never a citation. It was decoration.

#Duplicate near-identical pages

Plugin sprawl is code duplication. AI-slop sprawl is semantic duplication: twenty URLs competing for one intent.

#Symptoms

• {service} + {city} grids with identical H2 order.
• Blog posts titled “Ultimate guide to X” differing by two paragraphs.
• /v2/ or /new/ paths abandoned after a redesign.
• hreflang pairs where EN and PL say different facts.

#Verification workflow

1. Crawl with Screaming Frog or Sitebulb; export word count and hash near-duplicate titles.
2. Cluster by primary keyword intent, not URL folder.
3. Pick a survivor URL per cluster; 301 the rest.
4. Update XML sitemap and internal links in the same deploy.
5. Re-run a fixed AI prompt set after two weeks to see if contradictions persist in answers.

Merging duplicates often lifts crawl efficiency faster than adding another “fresh” AI article.

#Wrong dates and time-travel case studies

Dates signal freshness. AI models backfill plausible years. YMYL auditors compare dates to contracts.

#Symptoms

• updatedDate in frontmatter newer than any real edit.
• Testimonials from “2022” for a product launched 2025.
• Blog pubDate batches on the same afternoon (mass AI import).
• Copyright footer stuck at 2024 while claims say “current 2026 data.”

#Verification workflow

1. Align visible “Last updated” with Git or revision history where possible.
2. Match case study dates to SOW signatures (redact client name in internal log).
3. For medical or financial content, add lastVerified in frontmatter only when a human re-read sources.
4. Remove future dates entirely; they destroy trust in AI Overviews.

WordPress tip: restrict bulk post_date changes to roles that also own legal review.

#Fabricated team bios and authority pages

Xarumei invented Jennifer Lawson because a founder story is more cite-worthy than “unknown.” Your /team/ page is the same attack surface.

#Symptoms

• Bios with parallel structure (“15 years experience in…”) across every headshot.
• Credentials that do not map to registries (bar number, GMC, CPA license).
• Advisors listed with no corresponding press or filings.
• Stock photography with inconsistent EXIF or reverse-image hits.

#Verification workflow

1. One verified identifier per bio: license URL, ORCID, GitHub for engineers, WordPress.org profile for plugin authors.
2. Match Person schema sameAs to profiles you control.
3. Remove “team” members who are really contractors unless disclosure rules allow.
4. For agencies, link to named contributors on real commits or WordCamp talks instead of invented titles.

EEAT is not headcount. It is traceable humans.

#WordPress-specific audit workflow

Run this sequence on a staging clone before production deletes.

#1. Export and inventory

wp post list --post_type=page,post --format=csv --fields=ID,post_title,post_name,post_date,post_modified > content-inventory.csv

Add custom post types if your theme registers them.

#2. Automated similarity pass

Use a near-duplicate detector (Sitebulb, OnCrawl, or shingles scripts) on HTML stripped of nav. Flag clusters above 0.85 cosine similarity.

#3. Human spot check queue

Sort by organic traffic and revenue attribution. Audit top URLs with the red-flag table. Junior editors check; senior signs YMYL pages.

#4. Schema and GEO pass

• Remove aggregateRating without visible reviews.
• Align FAQPage answers with body copy (no extra claims in schema alone).
• Ensure canonicalUrl matches the survivor URL after merges.
• Run npm run geo:qa if your build includes GEO shape checks.

#5. AI echo check

Pick five branded prompts and five category prompts. Log answers weekly. If an assistant cites a stat you deleted, request recrawl and check external copies (PDFs, aggregators).

We document that cadence in the 90-day AI citation series. Cleanup without measurement is guesswork.

#Cleanup checklist you can hand to a client

Week 1 - Triage

• Red-flag table on top 20 URLs
• Source log template shared with content owner
• List of duplicate clusters with survivor URL picked
• Legal/compliance pages flagged for manual review

Week 2 - Remove and merge

• 301 map for merged pages
• Unsourced stats removed or replaced
• Fake citations deleted
• Team bios trimmed to verifiable identifiers

Week 3 - Reinforce truth

• FAQ updated with specific, sourced answers (Xarumei lesson)
• lastVerified or visible review date on YMYL guides
• Internal links point to survivors only
• Schema matches visible content

Week 4 - Measure

• Fixed prompt AI check logged
• Search Console validation for dropped URLs
• Stakeholder sign-off on remaining claims

#When to escalate beyond editorial cleanup

Editorial cleanup fixes copy. Some AI builds also ship unsafe plugins, broken checkout, and forty-plugin stacks. Signals for AI-built website rescue instead of copy edits alone:

• Custom AI plugins in wp-content/plugins/ without security review.
• WooCommerce flows that fail when unsourced marketing claims are removed (because features never worked).
• Multiple contradictory NAP blocks across theme options and page builders.
• Legal threat letter citing published figures.

Rescue engagements use individual pricing because scope spans code, content, and measurement. Copy-only cleanup is smaller; full rescue is not.

#Practitioner notes from rescue work

On a recent regulated B2B intake, the marketing lead exported 112 URLs from a Divi + AI workflow. 34 pages shared one introductory paragraph. 19 statistics had no source when challenged. Three team bios used the same sentence skeleton with different names. None of the bios matched LinkedIn URLs the client provided after we asked.

After merge and source pass, indexed pages dropped by 28, but organic clicks on survivor URLs rose over eight weeks because contradictions disappeared. The client could answer procurement due diligence without contradicting their own site. That is the business case for cleanup, not a vanity word count.

Another pattern: AI drafts love “independent study” language. We replace it with named instruments (CrUX, CitationOne Geoboard snapshot, client-approved Lighthouse exports) or we delete it. Middle ground is where Xarumei wins.

#Connecting cleanup to GEO and citations

Generative engines scrape what you publish. If your WordPress site says Portland on one URL and Nova City on another, you recreated Xarumei’s phase two on your own domain. Models will pick the most specific paragraph.

GEO work without fact cleanup repeats errors into llmCard.facts, speakable selectors, and FAQ schema. You amplify slop structurally. Fix copy first, then enrich schema.

Our measuring our own AI citations Q2 2026 snapshot showed how instrument choice changes the story. Cleanup is the on-page half; off-page corroboration is the other. Both fail if facts are invented.

#Summary table: from detection to ship criteria

AI-slop cleanup is not anti-AI. It is pro-evidence. Models already hallucinate. Your job is to ensure WordPress is not the training data they quote.

Last updated: 2026-07-10. Methodology references the Xarumei synthetic brand experiment and wppoland.com’s 90-day AI citation tracking baseline. For combined code and content remediation, see AI-built website rescue (individual quote).