Skip to content

Checks

AI search risk check

Flags AI visibility tactics that Google's spam policies target, such as self-ranking lists, scaled pages and paid or fake mentions, before they cost your site.

Other tools

Add this skill to Claude

Claude takes skills as a ZIP file. Your download has started; if it didn't, use the button below.

  1. 1In Claude, open Customize, then Skills.
  2. 2Choose +, then Upload a skill, and pick ai-search-risk.zip.
  3. 3Ask for it by name, or let Claude use it when a task matches.

Skills need code execution turned on in Claude's settings. On a Team or Enterprise plan, an admin can share it with everyone.

Download the folder

Try it now, without installing

See the skill work on a made-up input before you add it. Nothing is installed, and you can swap in your own input.

Open in the Claude app

Opens a new chat in the Claude desktop app with the prompt filled in. Nothing is sent until you press enter.

Includes the whole SKILL.md and the sample input. Paste it into Claude, ChatGPT or any chat.

Influenced by
Spam policies for Google web search, Optimizing for generative AI features, Self-promotional lists in AI Overviews
Needs
The tactics, planned or live
Works with
Claude, ChatGPT, Codex, Cursor and any agent that reads SKILL.md
Version
1.0.0 · Last checked · License MIT
Maintained by
Sabih AhmedOrganic visibility
SkillSEO and AI search

AI search risk check

Example output

Check a GEO plan
TacticSeveritySafer move
14 self-ranking listsCritical1 fair comparison
Monthly date bumpsWarningDate real edits
Review programFineKeep it

Flags AI visibility tactics that Google's spam policies target, such as self-ranking lists, scaled pages and paid or fake mentions, before they cost your site.

The skill

The whole skill, in plain text.

Nothing is hidden. Read it, change it to fit your team, or take the whole folder.

Scripts
Optional Python scripts for the counting. It works without them and says which way it ran.
Network
Opens web pages or your tools if your agent can. Works from pasted data if not.

3 test cases

ai-search-risk/SKILL.md

In the folder

  • ai-search-risk/SKILL.md
  • ai-search-risk/references/tactics.md
  • ai-search-risk/scripts/risk_scan.py
---name: ai-search-riskdescription: >-  Checks a B2B SaaS company's planned or live AI visibility tactics against Google's spam policies and its guidance  on generative AI features, before they cost the site: self-ranking best-of lists, pages scaled from a template, AI  text at volume, hidden instructions to AI, cloaking, paid or fake mentions, fake reviews and sock-puppet posts.  Scans the site's page list for the patterns, maps each tactic to the policy it meets, gives the safer move and  returns PASS, PASS WITH NOTES, FAIL or BLOCKED. Use it before launching a GEO or AEO tactic, or when visibility  drops after one. Not for crawler access (use agent-readiness) or measuring visibility (use ai-answer-panel).license: MITcompatibility: >-  Works in Claude, ChatGPT, Codex, Cursor and any agent that reads SKILL.md. Reads the sitemap and pages if the agent  can browse, otherwise a page list and the tactics the reader describes. Python 3 is optional, for the pattern scan.metadata:  author: SHAKEUP  version: "1.0.0"  stage: check  plugin: seo-ai-search  homepage: https://www.shakeuphq.com/skills/ai-search-risk--- # AI search risk check Flags the AI visibility tactics that Google treats as spam, or that published data ties to lost visibility, before they ship or while they can still be undone. ## Purpose - One job: check planned or live tactics and the site's pages against the spam policies and Google's AI guidance, and give each finding a severity and a safer move.- Stage: check. It ends in PASS or FAIL with severities.- It never helps carry out a critical tactic: no hidden instructions to AI, no fake reviews, no sock-puppet posts, no undisclosed paid mentions. It says why and offers the move that works.- It never claims a penalty it can't see. A pattern is a risk; a manual action is shown in Search Console. ## Use when / not | Situation | Use || --- | --- || An agency, tool or teammate proposes a GEO, AEO or "LLM optimization" tactic | this skill, before it ships || Visibility or traffic dropped after a content push | this skill, on the live pages || Planning outreach to cited pages | `cited-pages` (it routes risky plays here) || AI crawlers can't reach the site | `agent-readiness` || A single page's claims need checking | `claim-check` | ## Inputs | Input | Required | Ways to provide, in order | If missing || --- | --- | --- | --- || The tactics: planned or live, in the reader's words | yes, or pages | the reader's request; an agency proposal; a content plan | check the pages only || The site's pages: URL, title, published date; for lists, the first item; text where available | yes, or tactics | a crawl or sitemap if the agent can browse; a CMS or crawl export; a pasted list | judge the described tactics only; coverage `partial` || The same page fetched as a browser and as an AI crawler | no | 2 fetches if the agent can make them; 2 HTML files the reader saves | the cloaking check is skipped || Search Console messages and traffic around a drop | no | a connected Search Console; a screenshot of Manual actions; a Pages export | say whether a manual action is known or not | ## Profile read From `company-profile.md`: `## Company` (website and brand terms, to spot self-ranking lists), `## Competitors` (mention policy, for comparison pages), `## Content` (site sections, authors, to check author claims), `## Claims and proof` (what comparison pages may claim), `## Channels` (who runs outreach and community work), and every `## Learnings` entry for `ai-search-risk` with status `noted`, `fact` or `accepted`. No profile: ask for the website and brand terms, in one message. Say that `company-profile` saves this next time. ## Steps 1. **Apply learnings.** Read them first: tactics already reviewed and the decision, and any accepted threshold change. 2. **List what is being checked.** Each tactic in one line (what, where, how many pages, live or planned), and the page list if there is one. 3. **Map each tactic** to the table in `references/tactics.md`: the policy or guidance it meets, the default severity, the safer move. A tactic not in the table: judge it by the 2 questions the spam policies turn on, and say you did. Is it made mainly to manipulate rankings or AI answers rather than to help a user? Would a user, seeing how it was made, feel misled? 4. **Scan the pages.** With Python, write the page list to JSON and run:    ```   python3 scripts/risk_scan.py site --json pages.json   python3 scripts/risk_scan.py compare --browser as_browser.html --bot as_bot.html   ```    Without Python, check by eye: list titles that start "best", "top" or end "alternatives" and see which put the brand first; sort titles and look for runs that differ by a word or 2; count pages published per month; search page text for lines addressed to AI and for hidden text. Say which way you ran it.    | Pattern | Threshold | Why |   | --- | --- | --- |   | Self-ranking lists | any is a warning; 10 or more, or 10% of pages, is critical | at scale it reads as a scheme, and sites that leaned on them lost visibility across the domain |   | Template groups | 10 or more titles that differ by 2 words or fewer | the shape of programmatic pages; each must add something the others don't |   | Publishing spike | 3x the usual 30-day output | a scaled push; check what each page adds |   | Near duplicates | 5-word overlap of 0.6 or more between 2 pages | thin variants of one page |   | Instructions to AI, hidden text | any | an attempt to steer AI answers with content users don't see |   | Crawler gets different content | word overlap under 0.7 between browser and crawler fetches | cloaking, if on purpose | 5. **Judge value where the scan only finds a pattern.** For each template group or spike, open 3 pages: does each carry data, an example or advice the others don't? Yes, the group drops to a suggestion. No, it stays a warning, and becomes critical when the pages were made in bulk with generative AI and nobody added anything. 6. **Write each finding:** the tactic or pages, the severity, the policy or guidance it meets (named, linked), the evidence (counts, URLs), and the safer move. For a critical tactic the reader asked to carry out, decline that part plainly and give the safer move in its place. 7. **Order the fixes** for anything live: critical first, then by the number of pages. Removal, noindex or redirects are the reader's call. ## Check and verdict | Verdict | Condition (first match wins) || --- | --- || `BLOCKED` | No tactic described and no page list after asking || `FAIL` | 1 or more critical findings || `PASS WITH NOTES` | Warnings or suggestions only || `PASS` | No findings | No named stops: the reader decides what to remove. A request to carry out a critical tactic is declined in the output, not paused. ## Learn Append a learning to the profile when: - the reader decides on a flagged tactic (a `fact`: what was decided, so the next run doesn't relitigate it);- a drop or a recovery is measured after a change (a `fact` with dates and the source);- the reader marks a template group as valuable after review (a `proposal` to exempt that page type, applied once accepted). Example: `- 2026-10-24 | ai-search-risk | fact | 14 self-ranking "best" pages set to noindex and replaced with 1 fair comparison page | Why: FAIL on 2026-10-24; reader approved | Status: noted` ## Output ```Verdict: <PASS | PASS WITH NOTES | FAIL | BLOCKED> · Coverage: <complete | partial (what's missing)> · Scan: <script | by eye>Counts: <n> critical, <n> warnings, <n> suggestions · Pages scanned: <n> | Tactic or pages | Severity | Meets | Evidence | Safer move |(critical first) Declined: <any part of the request the skill won't carry out, and why>Fix order (live tactics): <list>Manual action known: <yes, no, or not checked>Learnings applied: <list, or none>``` Then the handoff block. ## Hand off ```handoffskill: ai-search-riskversion: 1.0.0verdict: FAILcoverage: partial (no page text, so near duplicates not checked)for: <site or proposal>outputs:  critical: 2  warnings: 1  self_ranking_pages: 14  declined: [undisclosed paid placement]next: [cited-pages, refresh-audit]open: [decide noindex or rewrite for the 14 list pages]``` ## Failure modes | Problem | Do this || --- | --- || The reader cites an agency that says a tactic "works" | Judge the tactic, not the claim; a short-term lift is how most of these patterns start || A comparison page names competitors | Fine when fair, sourced and disclosed, and the profile's mention policy allows it; check claims with `claim-check` || Programmatic pages carry real data per page (pricing by plan, integrations by tool) | Not scaled abuse; a suggestion at most || Traffic dropped and the reader wants the cause | Check for a manual action first; then compare the drop date with the tactic's launch and Google's update dates; say plainly that correlation is not proof || The reader asks for the hidden text "just for AI crawlers" | Decline; it meets hidden text and cloaking policies whatever the audience | ## Influenced by - Google's spam policies for web search, including scaled content abuse, site reputation abuse, cloaking, hidden text and link spam, and its note that attempts to manipulate generative AI responses count: https://developers.google.com/search/docs/essentials/spam-policies- Google's guide to its generative AI features, which calls inauthentic mentions less helpful than they seem and says no special files, markup or rewriting are needed: https://developers.google.com/search/docs/fundamentals/ai-optimization-guide- Lily Ray's study of self-promotional "best" lists in B2B AI Overviews, which found most were cited without the publishing brand being recommended, and tracked sites that leaned on them losing visibility: https://lilyraynyc.substack.com/p/why-calling-yourself-the-best-could

How it thinks

Inputs in, a verdict out, and a stop only where you decide.

What we built

A pattern is a risk, not a verdict, so it opens pages and judges what each one adds before it flags a template. It won't write hidden text, fake reviews or sock-puppet posts, and says why.

It reads your company profile if you have one, or asks for only what it needs. Then it works through its steps, checks its own work and ends in a verdict.

The steps

  1. 01

    List what's being checked

    Each tactic in one line, where it runs and how many pages it touches, live or planned.

  2. 02

    Map each tactic to the policy it meets

    Scaled content abuse, site reputation abuse, cloaking, hidden text, link spam, or Google's note that inauthentic mentions help less than they seem.

  3. 03

    Scan your pages for the patterns

    Its script finds lists that rank your own brand first, runs of pages from one template, publishing spikes, near duplicates and text written to steer AI answers.

  4. 04

    Judge value where a pattern isn't proof

    It opens 3 pages from each template group. If each adds real data or help, the flag drops.

  5. 05

    Give the safer move

    One honest comparison page instead of 14 self-ranking lists, a segment pitch instead of a paid ranking, a review program instead of bought reviews.

What you need

  • Required

    The tactics, planned or live

    Your request → An agency proposal → Your content plan

  • Optional

    Your site's pages, with titles and publish dates

    It reads your sitemap if your agent can browse → A CMS or crawl export → A list you paste

  • Optional

    A page fetched as a browser and as an AI crawler

    2 fetches, if your agent can make them → 2 saved HTML files

  • Optional

    Search Console messages around a traffic drop

    A connected Search Console → A screenshot of the manual actions report

What it can return

  • PASS

    Nothing found.

  • PASS WITH NOTES

    Warnings or suggestions only, such as date bumps or tactics with no effect.

  • FAIL

    At least 1 tactic meets a named spam policy or steers AI answers with content people can't see.

  • BLOCKED

    There's no tactic or page list to check.

Where it stops for you

It runs through without stopping. You review the result.

Example output

What you get back.

A run for a made-up company. Your output depends on your inputs.

Here's a run for a made-up B2B SaaS company that sells attribution software. An agency had shipped part of a plan for AI answers and proposed the rest.

Verdict

FAIL

Coverage: complete

Scan: script

3 critical, 0 warnings

Tactic or pagesSeverityMeetsSafer move
14 best-of posts ranking its own product first, 23% of the siteCriticalGoogle's note on inauthentic mentions, at a scale its spam policies targetNoindex or merge them into 1 fair comparison page that says who wrote it
The same 14 posts, written with an AI tool from one templateCriticalScaled content abuse: only the industry name changesKeep only the industries where it has data or a customer to show
A paid top spot on a review blog's list, with no labelCriticalLink spam: a paid placement without disclosurePitch the author the buyer the list misses, with proof they can check
A sourced comparison with one competitor, saying who wrote itFineNothingKeep it
Declined

recommending the paid placement. Manual action known: not checked, since no Search Console report was shared.

The handoff block the next skill reads
skill: ai-search-risk
version: 1.0.0
verdict: FAIL
coverage: complete
for: agency plan for AI answers
outputs:
  critical: 3
  self_ranking_pages: 14
  declined: [unlabeled paid placement]
next: [cited-pages]
open: [decide noindex or merge for the 14 posts]

What good looks like

  • Every finding names the policy or guidance it meets, with a link
  • Counts and URLs as evidence, not a feeling
  • A safer move for every flagged tactic
  • Template groups weighed on what each page adds
  • No claim of a penalty without the manual actions report

Install

Use it in the tool you already have.

Pick your tool. Each tab has the shortest path we know of.

  1. 01
  2. 02
    In Claude, open Customize, then Skills.
  3. 03
    Choose +, then Upload a skill, and pick the ZIP. Code execution must be on in your settings.
  4. 04
    Ask for it by name, or let Claude use it when a task matches.

Questions

Before you add it.

Are self-promotional best-of lists against Google's rules?

No single policy names them. Google's guide to its AI features says inauthentic mentions help less than they seem, and at scale they look like the patterns its spam policies target.

Does llms.txt count as a risk?

No. Google says Search ignores it, so it neither helps nor hurts there. It's flagged as wasted effort for Google visibility, not as a risk.

Will it tell me if I've been penalized?

Only if you share Search Console's manual actions report. A traffic drop near a tactic's launch is a lead to check, not proof.

What does the script do?

It scans your page list for self-ranking lists, template runs, publishing spikes, near duplicates and hidden instructions to AI. Without Python, the skill checks the same patterns itself and says so.