How we fixed “AI slop” after Google’s August spam update

Executive summary

The affected site had published a large group of AI-assisted articles quickly. The pages were technically indexable, but many had the same introduction pattern, repeated advice already available everywhere, weak or missing authorship, shallow intent coverage and claims with no first-hand evidence. Several pages competed for the same topic.

Recovery did not come from making AI text sound more human. It came from changing the publishing system so every indexable page needed a real audience, a distinct job and evidence that justified its existence.

What went wrong: the AI-slop footprint

We compared page groups rather than judging isolated sentences. The failure pattern was visible at template, topic and site level.

Repair 1: map the real user intent

Intent is more than labelling a keyword “informational” or “commercial.” We documented the reader’s situation, decision, constraints, required evidence and successful next step.

For example, “how to fix AI content” can hide several intents: diagnose a traffic loss, decide which pages to remove, improve one article, build a safe editorial workflow or understand Google’s policy. A useful page separates these jobs and lets the reader reach the relevant one quickly.

Repair 2: make entities and relationships explicit

Entity optimization is not inserting a long list of related words. It is making the people, organizations, products, concepts and events in the page unambiguous—and explaining how they relate to the user’s question.

On this page, the update is named precisely as the August 2025 spam update; Google Search is the organization/product context; scaled content abuse is the policy concept; ChatGPT and Claude are tools—not authors or proof of expertise.

Repair 3: show who is responsible

An author box is not a ranking switch. It helps readers evaluate responsibility and expertise, especially when the page gives consequential advice. We replaced generic ownership with accurate bylines connected to a useful author page.

Google’s people-first guidance frames this as Who, How and Why: who created the content, how it was produced, and why it exists. The “why” should be to help an intended audience—not merely to capture search visits.

The recovery workflow we used

1. Freeze uncontrolled publishing

   Disable automatic publication. Keep AI-assisted drafts private until a named reviewer approves them.

2. Segment the loss

   Compare Search Console by landing page, query type, directory, template, country and device. Separate sitewide decline from a small page group.

3. Inventory every indexable URL

   Record purpose, intended audience, organic clicks, impressions, conversions, backlinks, author, last review, word overlap and canonical status.

4. Choose keep, improve, consolidate or remove

   Do not rewrite everything. Keep genuinely useful pages, merge overlapping pages into a stronger destination, and return 404/410 when no relevant replacement exists.

5. Rebuild priority pages from evidence

   Start with customer questions, support logs, product data, interviews, experiments, screenshots, examples and expert decisions. AI may organize this material but must not invent it.

6. Repair intent and entity coverage

   Answer the main task early, cover necessary sub-decisions, define entities and connect them to trusted sources and relevant internal pages.

7. Add accountable review

   Use a real byline, reviewer where needed, disclosures, references and a visible update policy. Remove fake personas and inflated credentials.

8. Validate before requesting reprocessing

   Check status codes, canonicals, noindex rules, renderability, internal links, structured data and sitemap inclusion. Submit important URLs after the page is materially better.

9. Measure cohorts, not anecdotes

   Track improved pages against unchanged controls. Watch query diversity, qualified clicks, engagement and conversions—not only average position.

Prompts for ChatGPT and Claude

These prompts are designed to assist an expert editor. They explicitly prevent the model from fabricating experience, entities or recovery claims. Provide your own page text, Search Console export, customer evidence and source documents.

You are a senior content-quality auditor. Evaluate the page below for usefulness after a Google traffic decline.

Do not rewrite the article. Do not guess Google penalties or claim an update caused the decline.

Return a table with:
1. Intended user and job-to-be-done
2. Primary and secondary search intents
3. Sections that directly complete those intents
4. Commodity statements that add no original value
5. Claims requiring a source, date, test or expert review
6. Missing first-hand evidence or examples
7. Entities that are undefined, confused or unsupported
8. Overlap with other URLs I provide
9. Recommended action: keep, improve, consolidate, redirect, noindex or remove
10. Evidence still needed before editing

If evidence is absent, write “NOT PROVIDED.” Never invent data, experience, credentials or quotations.

PAGE: [paste page]
OTHER URL SUMMARIES: [paste]
SEARCH CONSOLE DATA: [paste]

Create an editorial brief from the evidence supplied below.

The brief must serve one defined audience and one primary task. Include:
- reader situation, knowledge level, constraints and desired outcome
- the answer that should appear in the first screen
- decision steps in the order the reader needs them
- questions and objections that must be resolved
- original evidence to include (tests, examples, screenshots, interviews or data)
- a source plan separating primary, secondary and internal sources
- entities with type, definition, relevant attributes and relationships
- sections to omit because they are generic or outside intent
- conversion step that logically follows task completion

Do not produce headings merely to include keywords. Do not add an FAQ unless it resolves genuine leftover questions.

EVIDENCE PACK: [paste]
AUDIENCE RESEARCH: [paste]
PRODUCT/SERVICE FACTS: [paste]

Extract an entity-and-evidence ledger from these materials.

For each material entity return:
- canonical name
- entity type
- aliases used in the documents
- claims made about it
- relationship to other entities
- best supporting source URL or document
- source date
- confidence: verified / ambiguous / unsupported
- exact correction needed

Only include entities necessary to answer the user's task. Flag name collisions and stale facts. Do not infer a Knowledge Graph ID. Do not fabricate schema properties.

MATERIALS: [paste]

Act as a skeptical human editor. Compare the draft with the evidence pack sentence by sentence.

Label each factual claim:
- SUPPORTED: cite the supplied source
- NEEDS QUALIFICATION: explain the missing context
- UNSUPPORTED: recommend deletion or a question for the author
- EXPERIENCE CLAIM: identify the named person who must verify it

Then check whether the draft:
1. answers the primary intent early,
2. adds something unavailable in generic summaries,
3. distinguishes facts, opinion and experience,
4. uses accurate authorship,
5. discloses material AI assistance when appropriate,
6. avoids fake statistics, quotes, clients and outcomes,
7. contains a useful next step.

Do not silently repair unsupported claims. Preserve an audit trail.

DRAFT: [paste]
EVIDENCE PACK: [paste]

Rewrite the approved section for clarity and task completion using only facts marked SUPPORTED in the evidence ledger.

Requirements:
- lead with the answer, then explain
- conversational, precise language for [AUDIENCE]
- preserve necessary nuance and limitations
- attribute external claims near the claim
- include the author's real decision process where supplied
- define unfamiliar entities at first mention
- never add numbers, examples, credentials or outcomes not in the ledger
- if a required fact is missing, insert [AUTHOR INPUT REQUIRED]

Avoid generic openings, rhetorical filler, keyword repetition, exaggerated certainty and “in today's digital landscape.”

APPROVED BRIEF: [paste]
SUPPORTED EVIDENCE LEDGER: [paste]
SECTION TO REWRITE: [paste]

How to measure a real recovery

Do not call a recovery after one good day or because a few keywords moved. Document the date each page changed and evaluate comparable periods after search systems have recrawled and reprocessed the work.

Google’s core-update guidance notes that improvements can be recognized without waiting for another major core update, but there is no fixed recovery timeline. Competition, crawling, site history and the scale of the change all matter.

Frequently asked questions