Does AI know your brand?your client?

Why this matters

Every day after this one is a single prompt in Claude Code. That only works if the skills, the tools and the keys are in place first, so today is plumbing.

Claude.ai chat and ChatGPT can read the install guides, but they can't install files on your computer. You need Claude Code for this.

Before you start

• Claude Code from Anthropic. The terminal, the Code tab in the Claude desktop app, or the VS Code or JetBrains extension all work.
• The tools the setup prompt checks: git, Python 3.10 or newer, uv, and Node.js 18 or newer for the two schema skills on Day 6. Missing one? The prompt shows you the one-line install and asks first.
• Accounts for the keys: DataForSEO (Day 1), Apify (Day 2) and Google AI Studio for a free Gemini key (Day 7). OpenRouter, Bright Data and TypeSafe are optional.
• Screaming Frog SEO Spider for Day 6. It's a separate desktop app, and the schema workflow needs licensed features such as JavaScript rendering and custom extraction.
• MarketingOS is optional. If it's installed, the prompt offers to turn your folder into a brain so the outputs land somewhere tidier. The skills work without it.
• No GitHub? Every skill also has a PDF version in the Skool classroom. Attach it to Claude Code and ask it to follow the instructions.

Run it

Open Claude Code in the folder where you keep your projects (your Desktop is fine), then copy and paste each step in order. Replace [e.g. MY-BRAND-AUDIT] with your folder name before you send the first one.

1. Paste the setup prompt

   It creates the project folder, clones the skill pack, links every skill, checks your tools and writes an empty .env. It asks before it installs anything.

   Set up my GEO brand audit project so every skill in The Vibe Marketing Lab's mos-geo-skills pack is installed, linked and ready to run.
   
   Project folder name: [e.g. MY-BRAND-AUDIT]
   Audit mode: {{mode}}
   
   Complete the following:
   
   1. Create or confirm the project folder
   If I'm already inside a project folder, confirm it with me before using it. Otherwise, create ~/Desktop/[PROJECT FOLDER NAME] and use it as the working directory for everything below.
   
   Check whether MarketingOS is installed: look for the mos command on my PATH and for marketing-os in pipx list. If it is installed, tell me, and offer to run mos-onboard so this folder becomes a MarketingOS brain (the skills save their outputs in a tidier place inside a brain). If it isn't installed, tell me it's optional and that the skills work without it. Do not install or onboard MarketingOS unless I say yes.
   
   2. Install or update the skill repository
   Clone https://github.com/the-vibe-marketing-lab/mos-geo-skills.git into ~/Desktop/mos-geo-skills if it isn't already installed. If it exists, check its remote and local changes, then safely pull the latest updates without overwriting local work.
   
   Keep the whole repository, including _shared/, because the skills use shared workbook assets and evidence notes.
   
   3. Link the skills into Claude Code
   Read README.md and setup.sh. Run bash setup.sh --dry-run and show me what it would do, then run bash setup.sh to create the skill links in ~/.claude/skills/.
   
   Use the appropriate symlink or Windows junction for my operating system. Preserve existing installations and never overwrite a link pointing elsewhere; report any conflict.
   
   Verify that each linked skill's SKILL.md is readable and resolves to its folder in the repository, including:
   - mos-geo-brand-360
   - mos-geo-ai-info
   - mos-geo-schema-scraper
   - mos-geo-schema-optimisation
   - mos-geo-fan-out-map
   - mos-geo-dealbreakers
   - mos-geo-llm-buttons
   - mos-geo-homepage-faqs
   - mos-geo-review-sentiment
   
   Report whether I need to restart Claude Code to load them.
   
   4. Check the tools on my computer
   Check and report the version of each:
   - Python 3.10 or newer
   - uv (several skills run their scripts through uv run)
   - Node.js 18 or newer (both schema skills)
   - git
   
   For anything missing, show me the documented one-line install and ask before installing it. Mention that Screaming Frog SEO Spider is a separate desktop app the schema skills need, and that Scrapling is an optional fallback for blocked pages.
   
   5. Create the credentials file
   Create a .env file in the project folder (not inside the skill repository). If one already exists, add only the missing key names and never change or display existing values.
   
   Use the exact key names from each skill's .env.example, with every value left empty:
   
   Required for the core skills:
   - DATAFORSEO_LOGIN and DATAFORSEO_PASSWORD (mos-geo-brand-360)
   - GEMINI_API_KEY (mos-geo-fan-out-map)
   - APIFY_TOKEN (mos-geo-review-sentiment, plus the optional schema scraper fallback)
   
   Optional:
   - OPENROUTER_API_KEY (Brand 360 fallback for model answers)
   - BRIGHTDATA_API_KEY and BRIGHTDATA_SERP_ZONE (Brand 360 fallback for the consumer apps and AI Overviews)
   - TYPESAFE_API_KEY (Jev labelling in review sentiment; without it Claude labels in session)
   
   Next to each key, tell me in plain words where to get it, using the links in the .env.example files:
   - DataForSEO: app.dataforseo.com, API Access
   - Gemini: aistudio.google.com/apikey (free tier is enough)
   - Apify: console.apify.com, Settings, Integrations, API token
   - OpenRouter: openrouter.ai/keys (needs credit)
   - Bright Data: Account settings, API key
   - TypeSafe: console.typesafe.ai/keys
   
   Add .env to the project's .gitignore (create .gitignore if missing). If the folder is a git repository, confirm .env is not already tracked.
   
   Never print, echo or paste any credential value into the chat, a log or a file other than .env. When checking whether a key is filled in, report only "set" or "empty".
   
   6. Report what's ready
   Finish with a short table:
   - Each skill: linked yes or no
   - Each tool: version found, or missing
   - Each key: set or empty, and which skills it unlocks
   - Whether a restart is needed
   
   Then tell me which skill I can run first with what's ready now, and what I still need to fill in for the rest. If audit mode is CLIENT, remind me that client outputs stay in this project folder, never inside the skill repository.
   
   Do not run any paid API call, publish anything or change any website.
   
   If required inputs are missing, ask for them together. Otherwise, proceed through the folder, installation, linking, tool checks and credentials file.

2. Confirm the folder

   If Claude Code was already open inside a project, it asks whether to use that folder. Say yes, or name a different one.

   Yes, use this folder for the audit.

3. Answer the MarketingOS question

   Only asked if MarketingOS is installed. Say yes to make the folder a brain, or skip it.

   Yes, run mos-onboard so this folder becomes a MarketingOS brain.

4. Approve the missing tools

   When it lists a missing tool with its install line, check the line, then paste:

   Go ahead and install the missing tools using the documented one-line installs. Show me each command before you run it.

5. Fill in your keys, then ask for the check

   Open .env in the project folder yourself and paste each key after its =. Never paste a key into the chat. Then:

   I've added my keys to .env. Check which keys are set or empty without showing any values, then give me the final table.

6. Restart and check

   If the table says a restart is needed, close Claude Code, open it again in the project folder, and paste:

   Check that all nine mos-geo skills are loaded and tell me which one I can run first.

Keys you'll need

• DATAFORSEO_LOGIN DATAFORSEO_PASSWORDDay 1, required. From app.dataforseo.com, API Access. These are the API login and password, not your dashboard login.
• OPENROUTER_API_KEYDay 1, optional. From openrouter.ai/keys (needs credit). Only used when a DataForSEO model call fails.
• BRIGHTDATA_API_KEY BRIGHTDATA_SERP_ZONEDay 1, optional. From Bright Data's account settings. A fallback for the consumer apps and AI Overviews.
• APIFY_TOKENDay 2, required. In console.apify.com: Settings, Integrations, API token. Also the paid fallback for the schema scraper on Day 6.
• TYPESAFE_API_KEYDay 2, optional. From console.typesafe.ai/keys. Without it, Claude labels the reviews in the session.
• GEMINI_API_KEYDay 7, required. From aistudio.google.com/apikey. The free tier is enough.
• Days 3 to 6No keys. Dealbreakers, AI Info, FAQs, share buttons and the schema skills use public pages, normal web search and your Screaming Frog crawl.

What you get

• ~/Desktop/mos-geo-skillsThe skill pack, cloned whole, including _shared/ with the workbook template.
• ~/.claude/skills/mos-geo-*Each skill linked from the clone, so Claude Code can find it.
• .envYour keys, in the project folder, listed in .gitignore.
• brand-audit-master.xlsxCreated by the first skill you run, in the month folder it prints. Starts with Checklist, Brand Truth Review and Initiatives (fixes scored by ICE). Each skill adds its own tab.

If something breaks

• A /mos-geo-* command doesn't appear. Restart Claude Code. Then check the link sits directly at ~/.claude/skills/<skill-name>. Claude Code only looks one level deep, so a skill tucked inside another folder is silently invisible.
• A key shows as empty. The .env has to sit in the project folder you run from, with nothing after the key name but = and the value.
• DataForSEO returns HTTP 401 on Day 1. The keys are wrong. Use the API login and password from API Access, not your dashboard login.

Why this matters

This is the diagnosis day. Brand 360 is the only skill in the pack that asks the engines directly, and it asks twice: once with no web search, once with it. That split is the line between rung 5 and rung 6.

When I ran it on The Vibe Marketing Lab on 17 Sep 2026, 0 of 3 models knew us without search, two tools described a different company, and 0 of 40 buyer answers named us. That's a snapshot, not a rate, but it tells you exactly where to start.

Before you start

• Keys: DATAFORSEO_LOGIN and DATAFORSEO_PASSWORD in your .env. OpenRouter and Bright Data are optional fallbacks.
• Inputs: the brand name as customers say it, and the industry in plain words, like "gym marketing agency". Fill both in the box at the top of this page.
• Don't hand over the website, founder, location or products. The test is whether engines find you from name and category alone. A clue in the prompt just measures the clue.
• Cost: a full default run measured $2.11 on DataForSEO (89 calls). The report prints the exact cost of every run. To spend less, cut unbranded prompts before cutting surfaces.
• Run it first. Day 4 and Day 5 reuse its Brand Truth Review corrections.
• Richard's walkthrough video for this day is in the Skool classroom.

Run it

Open Claude Code in your audit project folder (Day 0 sets it up), then copy and paste each step in order.

1. Paste the skill prompt

   It installs or updates the skill, runs a free preflight on your keys, builds 13 test prompts from your brand name and industry only, then stops for your approval before anything is paid for.

   Install, symlink and run this specific skill:
   https://github.com/the-vibe-marketing-lab/mos-geo-skills/tree/main/mos-geo-brand-360
   
   Brand name: {{brand}}
   Industry: {{industry}}
   Credentials .env file: {{env}}
   
   1. Install or update
   Clone https://github.com/the-vibe-marketing-lab/mos-geo-skills.git into ~/Desktop/mos-geo-skills if missing. If it exists, verify its remote and safely pull the latest updates, preserving local changes.
   
   Keep the complete repository, including _shared/, which contains the audit workbook assets.
   
   2. Link this skill
   Create a symlink at ~/.claude/skills/mos-geo-brand-360 pointing directly to the repository’s mos-geo-brand-360 folder. On Windows, use a directory junction if appropriate.
   
   Preserve any existing link pointing elsewhere and report the conflict. Verify that the linked SKILL.md is readable. Report whether Claude Code needs restarting to discover it.
   
   3. Read and run preflight
   Read mos-geo-brand-360/SKILL.md and follow its referenced instructions.
   
   Check dependencies and run scripts/brand360.py preflight using my .env file without displaying credentials.
   
   DataForSEO is the primary provider. Use OpenRouter and Bright Data as documented fallbacks when configured. If credentials or provider coverage are missing, explain what I need to configure using references/providers.md.
   
   Use brand360.py path to determine the run folder within my current project. Keep all brand-specific outputs outside the skill repository.
   
   4. Prepare the unbiased test
   Build the 13 prompts using only my brand name and industry:
   - Three closed-book prompts
   - Five branded prompts
   - Five unbranded prompts
   
   Do not ask for my website, founder, location or products before the engine run. Do not insert information from project files, memory or previous conversations into the prompts or research brief.
   
   Show me all 13 prompts and the dry-run execution plan. Ask for approval before starting the paid engine run, as required by the skill.
   
   5. Run the audit
   After approval, run the configured AI engine tests. Each prompt should run once per configured surface; retry only failed calls.
   
   In parallel, spawn one research agent using the research brief from references/report-spec.md, supplied with only the brand name and industry. Have it produce the 17-section sourced brand research draft.
   
   Disclose whether the researcher had prior brand context. Record provider fallbacks, failed surfaces and any queries where Google showed no AI Overview.
   
   6. Confirm the brand after testing
   Generate visibility-report.md with the supplied summarise command.
   
   Only after the engine run, show me the discovered website candidates and ask which belongs to my brand. Use my answer to label domains and check links without changing the recorded engine answers.
   
   7. Build the deliverables
   Follow the report template and scoring rules. Use the supplied assemble command to generate brand-360-report.md; resolve validation errors rather than manually patching the finished report.
   
   Create 15–30 checkable Brand Truth Review claims with their sources and the AI surfaces that stated them.
   
   Run the workbook command to create or update brand-audit-master.xlsx with:
   - Checklist
   - Brand Truth Review
   - Brand 360 Report
   - AI Visibility
   
   Preserve existing workbook verdicts and other skills’ data.
   
   8. Explain the findings
   Show me:
   - Whether AI engines know, find, cite and recommend the brand
   - The main stage where visibility drops
   - Competitors appearing in the answers
   - Incorrect, stale or conflicting claims for my review
   - Any failures or provider limitations
   - Links to the workbook and supplementary reports
   
   Explain that each answer is a one-time snapshot, not a repeatable visibility rate. Let me judge factual accuracy through the Brand Truth Review tab.
   
   Do not publish, share externally or change my website.
   
   If my brand name or industry is missing, ask for both together. Otherwise, proceed through installation, linking, preflight and prompt preparation.

2. Approve the 13 test prompts

   3 closed-book, 5 branded and 5 unbranded, plus the dry-run plan. Check the unbranded ones read like something your buyer would actually type. Edit now: changing a prompt after the run means paying again. When they're right, paste:

   These 13 prompts look right. Approved, start the paid engine run.

3. Confirm which website is yours

   After the run, Claude lists the website candidates the engines surfaced. This only labels rows. It never changes what the engines said.

   Our website is {{website}}. Label that domain as ours and leave the recorded engine answers unchanged.

4. Read the scorecard

   Once the report and workbook are built, ask for the one line that matters:

   Walk me through the AI Visibility Scorecard: for each surface, Known, Found, Cited and Recommended, and the step where we drop out.

5. Mark the Brand Truth Review

   Open the Brand Truth Review tab in brand-audit-master.xlsx and mark each claim Accurate, Partly accurate, Inaccurate, Not sure or Out of date. Then:

   I've marked the Brand Truth Review tab. List every claim I marked Inaccurate or Out of date, with the surface that said it.

What you get

• brand-audit-master.xlsxCreated in the month folder. Brand Truth Review filled, plus Brand 360 Report and AI Visibility tabs, and the Checklist row ticked.
• brand-360-report/brand-360-report.md17 research sections, Section 18 (the AI Visibility Scorecard) and the exact prompts used.
• brand-360-report/visibility-report.mdThe engine summary, which provider served each answer, and what the run cost.
• brand-360-report/data/Prompts, raw answers, the research draft and brand-truth-review.csv. A second run that month goes to brand-360-report-2.

How to read it

Open Section 18 of the report, or the AI Visibility tab. The skill scores every surface on four steps: Known (answers with no search), Found (describes the right brand), Cited (links your domain) and Recommended (named in an unbranded buyer question). The headline names the step where the brand drops out. That one line is your Day 1 answer.

Place yourself on the ladder

The skill doesn't print a rung. The mapping below is my reading of the scorecard. Start at rung 6 and walk down.

• Rung 6: Known = Yes on the closed-book engines (ChatGPT, Claude, Gemini, no search). "I'm not familiar with" scores No, and so does an answer that invents details.
• Rung 5: Known = No, but the web-search surfaces (Perplexity, the apps, AI Mode, AI Overviews) found you and cited your site.
• Rung 4: anything in the "Mixed up with" column, or Found = "Wrong brand". Another business is competing for your name.
• Rung 3: every claim you marked Inaccurate or Out of date on Brand Truth Review. The skill never judges accuracy itself. You do.
• Rung 2: found but not cited. Found = Partly, or the answers link a directory or platform page instead of your site.
• Rung 1: nothing. Found = No across the surfaces and no unbranded recommendations.

You can sit on different rungs for different questions. I was on rung 5 for branded questions, rung 4 on AI Mode and the Gemini app, and rung 1 for buyers who didn't know the name yet. Record the lowest rung that applies to the questions your buyers actually ask.

Every Inaccurate or Out of date row on Brand Truth Review becomes an Initiative later in the week.

Why this matters

This skill doesn't test AI engines. It reads the off-site pages engines can read. Reddit is among the most-cited domains on every engine measured, so a complaint repeated there can end up in how an engine describes you. That's supporting evidence for rung 3.

If Day 1 called your sentiment "mixed", this is the deep version. Brand 360 runs a few searches. This skill reads every review.

Before you start

• Keys: APIFY_TOKEN does the scraping. TYPESAFE_API_KEY is optional and labels each review for about $0.0002.
• Inputs: the brand name and your website (it stops you scraping a namesake), from the box at the top. In the prompt, set the country (it shows [e.g. AU]) and either paste your known review profile URLs or leave "FIND THEM".
• Cost: Apify charges per review plus a small start fee per profile. Google Maps is $0.0006 a review, Trustpilot $0.00075, ProductReview $0.002. A typical run is about $0.50 to $3 at the default cap of 500 reviews per platform.
• Nothing is scraped until you pick the profiles and approve the cost. The scrape refuses to run without it.

Run it

Open Claude Code in your audit project folder (Day 0 sets it up), then copy and paste each step in order.

1. Paste the skill prompt

   It installs the skill, runs preflight on your Apify token, finds your review profiles and reads their star ratings for free, then stops at a picker table.

   Install, symlink and run this specific skill:
   https://github.com/the-vibe-marketing-lab/mos-geo-skills/tree/main/mos-geo-review-sentiment
   
   Brand name: {{brand}}
   Website: {{website}}
   Country: [e.g. AU]
   Known review profiles: [URLS OR "FIND THEM"]
   Apify token .env file: {{env}}
   
   1. Install or update
   Clone https://github.com/the-vibe-marketing-lab/mos-geo-skills.git into ~/Desktop/mos-geo-skills if missing. If it exists, verify its remote and safely pull the latest updates, preserving local changes.
   
   Keep the complete repository, including _shared/, which contains the workbook assets and the evidence notes.
   
   2. Link this skill
   Create a symlink at ~/.claude/skills/mos-geo-review-sentiment pointing directly to the repository's mos-geo-review-sentiment folder. On Windows, use a directory junction if appropriate.
   
   Preserve any existing link pointing elsewhere and report the conflict. Verify that the linked SKILL.md is readable. Report whether Claude Code needs restarting to discover it.
   
   3. Read and preflight
   Read mos-geo-review-sentiment/SKILL.md, references/taxonomy-guide.md, references/report-spec.md and .env.example.
   
   Use my current project as the working directory. Use scripts/reviews.py path --brand to resolve the run folder and pass that same folder to every command. Keep all brand data in the run folder, never inside the skill repository.
   
   Run reviews.py preflight with my .env file (--env-file). Load APIFY_TOKEN without printing it. TYPESAFE_API_KEY is optional; without it, you label the reviews in session. If an actor I need hasn't been smoke-tested, explain that preflight --smoke costs cents and ask before running it.
   
   4. Discover profiles and read their stats
   Run reviews.py discover with my brand, domain and country as a dry run first, and show me the maximum Google search cost. Then run it for real. Add any known profile URLs I gave you to data/candidates.json before the next step.
   
   Run reviews.py stats to read each profile's rating, review count and latest review for free.
   
   5. Let me pick, then approve the cost
   Run reviews.py estimate and show me picker.md as a table: platform, reviews, average rating, reviews in the last 12 months, latest review, how many we'd scrape, estimated cost, and the check column.
   
   Flag every partial match or name/domain mismatch plainly. It might be a reseller, another country's company or a namesake.
   
   Ask me which profiles belong to my brand and should be scraped. Never pick on my behalf. Then confirm the total cost: approve, lower the cap or cancel.
   
   Nothing is scraped until I've picked the profiles and approved the cost.
   
   6. Scrape
   Run reviews.py scrape with my picks as a dry run, then with --approve-usd set to the total I approved. Report any platform that fails and save the rest.
   
   7. Lock the taxonomy with me
   Run reviews.py taxonomy sample. Propose 8 to 15 aspects, each with a label, a one-sentence description and an owner (ops, product, support or marketing). Show them as a short table and ask me to approve or edit. Lock the approved set with taxonomy lock.
   
   8. Label and spot check
   Run reviews.py label as a dry run, then for real (Jev if the key is set, otherwise in-session labelling imported with --import).
   
   Run the spot check of 30 and ask me to mark the wrong ones. Record the result. Under 85% agreement, fix the aspect descriptions, re-lock, re-label and check again.
   
   9. Aggregate and build the deliverables
   Run reviews.py aggregate, then reviews.py wordcloud. Every count, average, share and priority score comes from the script, never from you.
   
   Write data/fix-notes.json for every issue in the fix list, not just the top five: the problem, the likely root cause as a hypothesis, a concrete fix, the GEO action and the metric to watch.
   
   Run reviews.py report to produce review-sentiment-report.md, fix-report.md and unreplied-negatives.csv (every 1 to 2 star review with no owner reply). If it reports missing notes, write them and re-run.
   
   Run reviews.py workbook to add the Review Fixes tab and Initiatives rows to brand-audit-master.xlsx and tick the Checklist, preserving existing verdicts and other skills' data.
   
   10. Report
   Show me:
   - The headline numbers per platform, and how many reviews were analysed against each profile's total
   - Anything in the Read first box (injury, hazard or legal mentions)
   - The top three fixes with their verbatim quotes and owners
   - The praise and complaint word clouds
   - The count of unreplied negative reviews
   - The spot check result and actual Apify cost
   - Links to the reports, the CSV and the workbook
   
   Quotes stay verbatim and linked, root causes are hypotheses, and no reviewer names are stored. Only make AI-citation claims that appear in _shared/geo-evidence.md section 12; say plainly that fixing reviews is not a guaranteed way to change what AI engines say.
   
   Do not reply to reviews, publish anything, share externally or change my website.
   
   If required inputs are missing, ask for them together. Otherwise, proceed through installation, linking, preflight, discovery and the picker.

2. Approve the discovery search

   Discovery runs as a dry run first and shows the maximum Google search cost. If preflight also asks about a smoke test, that costs cents. When the numbers look fine:

   That search cost is fine. Run discovery for real, and run the smoke test on any actor that needs it.

3. Pick the profiles and approve the cost

   Read the picker table. Watch for partial match and name/domain mismatch flags: a reseller, another country's company or a namesake. Name only the rows that are really the brand.

   Scrape these profiles only: [LIST THE ROWS THAT ARE REALLY US]. I approve the total cost shown for those profiles. Run the scrape.

4. Approve the topic list

   Claude proposes 8 to 15 aspects, each with a description and an owner: ops, product, support or marketing. Edit any you disagree with, then:

   These aspects look right. Lock the taxonomy and label the reviews.

5. Do the spot check

   Claude shows you 30 labelled reviews. Note the numbers of any that are wrong. Under 85% correct, it fixes the descriptions and relabels.

   Wrong labels: [NUMBERS, OR "NONE"]. Everything else is right. Record the result and carry on.

What you get

• review-sentiment/fix-report.mdEvery complaint ranked by priority, with verbatim quotes, a likely cause, a fix, an owner and a metric. A Read first box lists any injury, hazard or legal mention.
• review-sentiment/review-sentiment-report.mdAverage rating, net sentiment, reply rates by platform, praise, complaints and the trend.
• review-sentiment/unreplied-negatives.csvEvery 1 to 2 star review with no owner reply, newest first. Usually the quickest win.
• review-sentiment/wordcloud-*.pngPraise and complaint word clouds for a deck, plus an interactive wordcloud.html.
• brand-audit-master.xlsxA Review Fixes tab, the top issues and an unreplied-reviews task on Initiatives, and the Checklist row ticked.

How to read it

• The Read first box. Injury, hazard or legal mentions come before anything else.
• The top three fixes. Ranked by volume, severity, recency, visibility and how much goes unreplied. The weights are printed in the report.
• The unreplied negatives count. Each one is a public complaint with no first-party answer next to it.
• The trend and reply rate by platform. Is it getting better or worse, and where is nobody answering?
• The negative word cloud. The phrases customers repeat are the phrases I'd expect an engine to repeat.

Good looks like replies on most negatives, complaints spread thin and a flat or improving trend. Bad is one theme carrying most of the volume, recent, with low reply rates. If a complaint here matches what an engine said in Day 1's "Is it any good?" answer, that's a rung 3 issue with a known source. Flag it on Brand Truth Review.

Why this matters

Buyers now ask ChatGPT, Perplexity and AI Mode "what are the downsides of X" and "X pricing", and the assistant answers from whatever is public. A dealbreaker nobody answers gets answered by someone else, or comes back as "unclear".

Missing answers are rung 2. Pages that contradict each other are rung 3. The "vs" and "alternatives" searches show where a buyer gets pushed toward a competitor, which is rung 4.

Before you start

• Keys: none. Web search runs through your normal Claude Code session.
• Your website and brand name, from the box at the top.
• Other places a buyer sees you: a community page (Skool, Circle, Discord), a marketplace or app store listing, a pricing page on another domain, review profiles. Swap them into the prompt, or leave "NONE KNOWN".
• Real objections if you have them: sales-call notes, DMs, refund reasons. Otherwise leave "USE EXISTING PROJECT RESEARCH".
• Richard's walkthrough video for this day is in the Skool classroom.

Run it

Open Claude Code in your audit project folder (Day 0 sets it up), then copy and paste each step in order.

1. Paste the skill prompt

   It crawls your buying pages, sweeps reviews, Reddit, "vs" and "alternatives" results, reads them as 2 to 4 buyer personas, checks every quote against the saved page, then stops to show you the proposed priorities.

   Install, link and run The Vibe Marketing Lab’s mos-geo-dealbreakers skill to identify what could stop a qualified buyer from purchasing.
   
   Website: {{website}}
   Brand name: {{brand}}
   Audit mode: {{mode}}
   Other public surfaces: [COMMUNITY, MARKETPLACE LISTING, REVIEW PROFILES, OR “NONE KNOWN”]
   Known buyer objections: [SALES-CALL FEEDBACK, DMs, REFUND REASONS, OR “USE EXISTING PROJECT RESEARCH”]
   
   Complete the following:
   
   1. Install or update the skill repository
   Clone https://github.com/the-vibe-marketing-lab/mos-geo-skills.git into ~/Desktop/mos-geo-skills if it isn’t already installed. If it exists, check its remote and local changes, then safely pull the latest updates without overwriting local work.
   
   Keep the whole repository, including _shared/, because the skill uses shared workbook assets.
   
   2. Link the skills into Claude Code
   Read README.md and setup.sh. Run the installer’s dry run, then run setup.sh to create the skill links in ~/.claude/skills/.
   
   Use the appropriate symlink or Windows junction for my operating system. Preserve existing installations and never overwrite a link pointing elsewhere.
   
   Verify that ~/.claude/skills/mos-geo-dealbreakers/SKILL.md is accessible and resolves to the repository’s mos-geo-dealbreakers folder. Report whether I need to restart Claude Code to load it.
   
   3. Run mos-geo-dealbreakers
   Read and follow mos-geo-dealbreakers/SKILL.md and its referenced instructions. Check the required tools and install missing dependencies using the documented approach.
   
   Use my current project as the working directory. Save business-specific research and outputs in the designated project run folder, never inside the skill repository.
   
   Read existing audience research and buyer objections in the project. Crawl the website and other public surfaces, including pricing, FAQs, proof, terms, refunds, security and integrations. Search public reviews, community posts and comparisons, saving every page used as evidence.
   
   Check deeper FAQ and help-centre answers before calling information absent. Compare older positioning with the current offer where historical pages are available.
   
   Identify 2–4 buyer personas and examine 6–10 relevant objection categories. Distinguish findings as stated, contradicted, absent or inferred. Describe missing information as something a buyer cannot verify, rather than claiming it does not exist.
   
   Write data/findings.json and generate the deliverables using the supplied build script. Resolve every validation failure and spot-check five findings, prioritising critical findings, prices, numbers and dates.
   
   Produce:
   - An evidence-backed Dealbreakers Report
   - The 3–7 highest-priority objections and recommended fixes
   - Objections the business already addresses
   - An FAQ draft with unknown answers marked [CLIENT TO ANSWER]
   - A CSV of findings
   - A Dealbreakers tab in the GEO brand audit workbook
   - ICE-scored Initiatives for my own brand or client; no Initiatives for competitor audits
   
   Show me the proposed priorities, severity counts, positioning drift and sensitive findings. Let me keep, reorder or swap priorities and supply missing answers, then rebuild the final report and workbook.
   
   Keep unpublished answers clearly identified as client input. Do not present them as publicly verified facts or promise improved AI citations.
   
   Keep the audit private. Do not publish, share externally or change the website.
   
   Finish with links to the output files and the next steps for approving findings, developing FAQ copy and filling proof gaps.
   
   If required information is missing, ask for it in one message. Otherwise, proceed through installation and audit creation.

2. Settle the priorities

   Read the proposed priorities, severity counts, positioning drift and anything flagged sensitive. Keep, reorder or swap them, and answer what you can answer now. Your answers go into the FAQ draft as client input, not as verified public facts.

   Keep these priorities in this order: [YOUR ORDER, OR "AS PROPOSED"]. Here are answers to the open questions: [YOUR ANSWERS, OR "NONE YET"]. Rebuild the final report and workbook.

3. Get the open questions in one list

   Once the rebuild is done:

   List every [CLIENT TO ANSWER] item in the FAQ draft, grouped by priority, so I can answer them in one go.

What you get

• dealbreakers/dealbreakers-report.mdEvery finding with evidence, severity and persona, the 3 to 7 priorities with fixes, the objections already handled, and any positioning drift.
• dealbreakers/faq-draft.mdOne question per answerable dealbreaker. Unknowns are marked [CLIENT TO ANSWER]. Day 5 picks this up automatically.
• dealbreakers/data/findings.csvThe findings as a spreadsheet, plus the crawl and the saved text of every page read.
• brand-audit-master.xlsxA Dealbreakers tab with a verdict dropdown that survives re-runs, the Checklist row ticked and the priorities on Initiatives.

How to read it

• "Fix these first". The 3 to 7 priorities, critical or major only, each with a fix type: FAQ, pricing page, proof asset, policy doc, positioning, product, or none for an accepted trade-off.
• Severity counts. Critical means a qualified buyer stops. Major means they slow down or shortlist a competitor. A typical run lands at 3 to 6 critical and 10 to 20 major. More than a quarter critical usually means inflation.
• Proof & Results first. Does the evidence prove the outcome being sold, or something next to it? Traffic screenshots don't prove citations.
• The evidence type. stated and contradicted are the ones nobody can argue with. absent is only as strong as the pages checked. inferred can never be critical.
• Already handled. Objections you answer well today. Keep those pages.

For the ladder, I'd count your absent findings on pricing, refunds and proof (rung 2) and your contradicted findings (rung 3). On my own run, 30 Sep 2026, the top critical finding was no published evidence yet that the core promise works. A major one was no published refund policy for the paid tiers.

Why this matters

Your About page was written to sell. AI needs facts it can quote. The AI Info Page is an official, sourced fact sheet that ChatGPT, Claude, Perplexity and Gemini can read: who you are, what you do, who you work with and how.

Its main job is rung 3, wrong information. It reads your Day 1 Brand Truth Review first, and your verdicts there override everything else. It also helps rung 2: one consistent, first-party page that crawlers can find, plus a list of every place your own site contradicts itself. On rung 5 there's one published test, where ChatGPT cited a page like this within about 48 hours. It makes no claim about rung 6.

Before you start

• Day 0 done: the skill pack, python3 and uv in place.
• Your website and the brand name as customers say it (the trading name, not the legal entity), from the box at the top.
• A leave-out list: NDA clients, someone who has left, a service you're retiring. Swap it into the prompt, or write "NONE". Ask the client for theirs, plus a decision on naming clients and individuals, and their developer's contact.
• Where the page will live. The prompt uses /ai-info/. Change it only if you already use another path.
• Ideally, a Brand 360 run from this month. The skill picks up your Brand Truth Review corrections automatically.
• Richard's walkthrough video for this day is in the Skool classroom.

Run it

Open Claude Code in your audit project folder (Day 0 sets it up), then copy and paste each step in order.

1. Paste the skill prompt

   It crawls your site, researches and sources every fact, probes for client and award pages it might have missed, builds the page, and stops to show you the preview and every disagreement between your sources.

   Install, link and run The Vibe Marketing Lab’s mos-geo-ai-info skill for my business.
   
   Website: {{website}}
   Brand name: {{brand}}
   Anything to exclude: [NDA CLIENTS, PRIVATE INFORMATION, OR “NONE”]
   Page location: {{website}}/ai-info/
   
   Complete the following:
   
   1. Install or update the skill repository
   Clone https://github.com/the-vibe-marketing-lab/mos-geo-skills.git into ~/Desktop/mos-geo-skills if it isn’t already installed. If it exists, check its remote and local changes, then safely pull the latest updates without overwriting local work.
   
   Keep the whole repository, including _shared/, because the skill uses shared workbook assets.
   
   2. Link the skills into Claude Code
   Read README.md and setup.sh. Run the installer’s dry run, then run setup.sh to create the skill links in ~/.claude/skills/.
   
   Use the appropriate symlink or Windows junction for my operating system. Preserve existing installations and never overwrite a link pointing elsewhere.
   
   Verify that ~/.claude/skills/mos-geo-ai-info/SKILL.md is accessible and resolves to the repository’s mos-geo-ai-info folder. Report whether I need to restart Claude Code to load it.
   
   3. Run mos-geo-ai-info
   Read and follow mos-geo-ai-info/SKILL.md and its referenced instructions. Check the required tools and install any missing dependencies using the documented approach.
   
   Use my current project as the working directory. Save business-specific research and outputs in the project’s designated run folder, never inside the skill repository.
   
   Crawl my website, research and source the facts, probe for missed client and award pages, record conflicting information, and generate the page through the supplied build script. Resolve validation failures and spot-check the sources.
   
   Produce:
   - The AI Info Page in Markdown and HTML
   - JSON-LD schema
   - A browser preview
   - A developer handover
   - A statement-by-statement approval tab in the GEO brand audit workbook
   
   Use factual, sourced language. Leave out unsupported claims and keep the optional canary line disabled.
   
   Show me the finished preview, output locations, discrepancies and any sensitive facts that need review. Ask me to approve the facts before publishing anything.
   
   If any required input is missing, ask for it in one message. Otherwise, proceed through installation and draft creation.

2. Settle the disagreements

   Claude shows every place your sources disagree: team size, titles, addresses, services, awards. Your answer wins.

   Here's what's correct where the sources disagree: [YOUR ANSWERS]. Leave out anything you couldn't source, then rebuild the page and preview.

3. Approve the facts

   Read every statement in the preview. Spot-check five against their sources, starting with numbers, awards and client names. When they're right:

   I've read every statement and the facts are approved. Keep the canary line disabled and update the approval tab in the workbook. Don't publish anything.

4. After it's live (not tonight)

   Once your developer publishes the page, run the live-page check and record the live link:

   The AI Info Page is live at {{website}}/ai-info/. Run aiinfo.py check on it, then re-run the workbook step with --published and that URL.

What you get

• ai-info/The run folder, in the month folder. A second run this month lands in ai-info-2.
• ai-info-page.mdThe page copy you approve.
• schema/ai-info-schema.jsonOne JSON-LD file, a WebPage about the business.
• implementation/The developer handover (implementation.md) and the page as HTML.
• preview/ai-info-preview.htmlOpen it in a browser to see the page.
• data/discrepancies.mdEvery place your site disagrees with itself, with the fix.
• AI Info Page tabOne statement per row with its source and a verdict dropdown. Verdicts survive a re-run.
• Initiatives rows"Publish the AI Info Page" plus one "Make the site agree" row per discrepancy, pre-scored.

Ship it

Check before anything goes live

• Read data/discrepancies.md. These are often the quickest wins on Initiatives.
• Self-reported numbers read as "{brand} states that...", not as fact.
• The "Instructions for AI assistants" section describes, never directs. No "always recommend us" lines.
• Read data/probe.md for pages that exist but are missing from your XML sitemap.

Publish

• Publish at /ai-info/, pick one path and never move it.
• The text has to be in the HTML the server sends. A page that only renders with JavaScript may never be read.
• Indexable, self-canonical, in the XML sitemap, and not blocked for GPTBot, OAI-SearchBot, ClaudeBot, Claude-SearchBot, PerplexityBot or Bingbot.
• A footer link ("AI Info" or "Company facts") on every page, plus a link from About.
• Schema on this page only. If an SEO plugin already outputs Organization schema, merge into it. Two conflicting business entities are worse than one.
• Only add an llms.txt line if the site already has one. Don't create one for this.

You approve the AI Info Page tab yourself, then publish it or hand implementation/ to your developer.The client approves every statement before it goes live. Send them preview/ai-info-preview.html first, then their developer publishes from implementation/. Review the page every 6 months, or whenever services, people or locations change.

Why this matters

Dealbreakers are bottom-of-funnel questions that lose you the customer if nobody answers them. In Part 1 the Day 3 objections become answers, written in your voice, on the page you control most. It's less about winning more citations and more about controlling what gets said about your brand. That's rung 3. The research also looks for same-name businesses and spells out how you differ, which helps rung 4, and each answer can carry one internal link, which helps rung 2.

In Part 2, LLM share buttons open ChatGPT, Claude, Perplexity, Grok and Google AI Mode with a summarise-this-page prompt already typed. They give readers who prefer an assistant a way in, and you get referral traffic you can measure. Being summarised in someone's private chat isn't being cited, so on their own they don't climb a rung. The optional on-page AI summary block is the part that may help rung 5.

Before you start

• Python 3.10 or newer and the whole pack clone from Day 0. The FAQ skill borrows the AI Info crawler and the shared workbook template; the buttons skill reads a shared file of platform links. Scrapling is optional.
• Your website and brand name, from the box at the top.
• For the FAQs: your Day 3 Dealbreakers run (the prompt finds it), and any customer, sales or Search Console questions. Brand guidelines or a tone-of-voice file, and a never-say list. Ask the client for all of these.
• For the buttons: one article or guide URL (or leave "FIND ONE"), and an agency name, because every file is white-labelled to it. Running it on your own brand? Give your own business name as the agency name.
• Run Day 4 first if you can. The FAQ skill reads the AI Info Page, so you get better FAQs for free.
• Richard's walkthrough video for the share buttons is in the Skool classroom.

Part 1: Homepage FAQs

Open Claude Code in your audit project folder (Day 0 sets it up), then copy and paste each step in order.

1. Paste the skill prompt

   It crawls your site, gathers your Day 3 questions, runs a same-name search, then stops for one interview round before it writes anything.

   Install, symlink and run this specific skill:
   https://github.com/the-vibe-marketing-lab/mos-geo-skills/tree/main/mos-geo-homepage-faqs
   
   Website: {{website}}
   Brand name: {{brand}}
   Existing buyer questions / Dealbreakers report: [FILE PATH OR "USE THE DEALBREAKERS RUN IN THIS PROJECT"]
   
   1. Install or update
   Clone https://github.com/the-vibe-marketing-lab/mos-geo-skills.git into ~/Desktop/mos-geo-skills if missing. If it exists, verify its remote and safely pull the latest updates, preserving local changes.
   
   Keep the complete repository. This skill uses the sibling mos-geo-ai-info crawler and the shared workbook template in _shared/.
   
   2. Link this skill
   Create a symlink at ~/.claude/skills/mos-geo-homepage-faqs pointing directly to the repository's mos-geo-homepage-faqs folder. On Windows, use a directory junction if appropriate.
   
   Preserve any existing link pointing elsewhere and report the conflict. Verify that the linked SKILL.md is readable. Report whether Claude Code needs restarting to discover it.
   
   3. Read and preflight
   Read mos-geo-homepage-faqs/SKILL.md, references/faq-anatomy.md and references/question-bank.md. Check Python 3.10 or newer.
   
   Use my current project as the working directory. Use scripts/faqs.py path --brand to resolve the run folder, then run faqs.py doctor. If Scrapling is missing, show me its one-line install and carry on. Keep all brand outputs in the run folder, never inside the skill repository.
   
   4. Crawl and gather
   Run faqs.py crawl on my website, then faqs.py gather. Pass my buyer questions or Dealbreakers report with --questions. If I said to use the Dealbreakers run in this project, find it, including earlier months, and tell me which files you found and their dates.
   
   If no Dealbreakers output exists, suggest running mos-geo-dealbreakers first, but don't make it a requirement.
   
   5. Interview me once
   Read the bundle and run the same-name web search first. Then ask me, in one round, only what the bundle doesn't already answer, each with a suggested answer from my site:
   - Tone of voice in three words, and words we never use
   - Locale and spelling (en-AU, en-GB or en-US)
   - Whether a similar-named entity you found is us
   - Anything on the never-say list
   - Which flagship products to lead with (only if there are more than three)
   - Whether to include a founder question (only if a founder is publicly named)
   
   Record skipped questions as defaults. Save the voice brief and re-run gather.
   
   6. Research, review sources, write
   Research third-party mentions and same-name entities, and write data/research.md with a public Summary section. Read the actual cited passage for every claim and record it in data/evidence-review.md.
   
   Map every one of my supplied questions in data/question-plan.json. Keep unanswered questions as pending; never silently drop or merge them.
   
   Write faqs.md using the FAQ anatomy: a direct answer in the opening sentence, the brand named in the opening paragraph, my brand voice, and capability language only. No outcome promises, no rankings, no "best", and no prices, numbers or dates unless I ask for them.
   
   Run faqs.py links and add at most one internal link per answer, using an anchor copied word for word from the answer. Never reword an answer to fit a link.
   
   7. Lint is the gate
   Run faqs.py lint with --final, my brand, products, locale and never-say terms (--forbid). Fix every failure and re-run until it prints Result: PASS. Re-run the final lint after any edit.
   
   8. Render and workbook
   Only after a current PASS, run faqs.py render with my website URL and the same --forbid list. Confirm the homepage HTML accordion and the FAQPage JSON-LD contain the same questions and answers.
   
   Run faqs.py workbook to add the Homepage FAQs approval tab and an Initiatives row to brand-audit-master.xlsx, preserving existing verdicts and other skills' data.
   
   9. Report
   Show me:
   - How many FAQs, and where each is placed (homepage, page, FAQ page or content gap)
   - How many carry internal links
   - Any same-name entity avoided
   - Interview defaults used
   - Questions still pending a client answer or a source
   - Links to faqs.md, publishing/faqs.html, publishing/faqpage.jsonld, review/placement-plan.md, review/handover.md and the workbook
   
   Explain that FAQ text is ordinary page text that search-backed engines can read and quote, that nothing shows AI engines weight FAQ blocks specially, and that Google no longer shows FAQ rich results, so the JSON-LD is for machine readability only.
   
   Approval in the workbook comes before publishing. Do not publish, share externally or change my website.
   
   If required inputs are missing, ask for them together. Otherwise, proceed through installation, linking, crawl, gather and the interview.

2. Answer the interview in one reply

   Each question arrives with a suggested answer from your site. Fill in what you know; anything you leave out uses the suggestion and is recorded as a default.

   Tone in three words: [THREE WORDS]. Words we never use: [WORDS]. Locale: en-AU. The same-name business you found [IS / IS NOT] us. Never say: [TERMS, OR "NOTHING EXTRA"]. Use your suggestions for everything else.

3. Clear the pending questions

   After the lint passes and the files render, Claude reports questions still waiting on an answer or a source. They're left out of faqs.md on purpose.

   List every pending question and what you need from me for each, so I can answer them in one reply.

Part 2: LLM share buttons

Same folder, same session is fine. Swap your agency name into the prompt before you send it.

1. Paste the skill prompt

   It inspects your live site, its platform and security policy, pulls colours and fonts from a real article page, then stops for you to confirm the palette.

   Install, symlink and run this specific skill:
   https://github.com/the-vibe-marketing-lab/mos-geo-skills/tree/main/mos-geo-llm-buttons
   
   Website: {{website}}
   Representative article URL: [AN ARTICLE OR GUIDE URL, OR “FIND ONE”]
   Client / brand name: {{brand}}
   My agency name: [AGENCY NAME FOR THE DELIVERABLES]
   
   1. Install or update
   Clone https://github.com/the-vibe-marketing-lab/mos-geo-skills.git into ~/Desktop/mos-geo-skills if missing. If it exists, verify its remote and safely pull the latest updates, preserving local changes.
   
   Keep the complete repository, including _shared/, which contains the platform endpoint configuration and workbook assets.
   
   2. Link this skill
   Create a symlink at ~/.claude/skills/mos-geo-llm-buttons pointing directly to the repository’s mos-geo-llm-buttons folder. On Windows, use a directory junction if appropriate.
   
   Preserve any existing link pointing elsewhere and report the conflict. Verify that the linked SKILL.md is readable. Report whether Claude Code needs restarting to discover it.
   
   3. Read and inspect
   Read mos-geo-llm-buttons/SKILL.md and follow its workflow.
   
   Inspect the live website to identify its stack and Content Security Policy. Extract the actual colours, typography and button styles from a representative article page.
   
   Show me the extracted palette and ask me to confirm it before styling. Flag anything you cannot verify instead of guessing.
   
   4. Scope and build
   Read the appropriate implementation route: WordPress or custom.
   
   Use the sitemap to identify eligible articles, guides, explainers and news pages. Map their URL patterns to readable topic phrases and document excluded routes.
   
   Build buttons for ChatGPT, Claude, Perplexity, Grok and Google AI Mode, in that order.
   
   Use:
   - The canonical prompt from references/prompt-policy.md
   - Endpoints from _shared/platform-endpoints.json
   - The confirmed client palette with monochrome platform labels
   - Self-hosted or inline icons
   - A configurable provider list
   - Accessible, responsive styling
   - The llm_share_click dataLayer event, fired before opening the provider tab, with llm_platform, page_url, page_type and page_topic
   
   Explain provider login limitations. Leave optional Gemini clipboard mode disabled unless requested.
   
   Read references/summary-block.md and explain the optional on-page summary pairing and its evidence limits before recommending buttons alone. Do not promise improved rankings, citations or traffic.
   
   5. Generate the deliverables
   Produce these three complete, agency-branded files:
   - <client>-llm-buttons-brief.md
   - <client>-llm-buttons-code-reference.md
   - <client>-llm-buttons-demo.html
   
   Include ready-to-paste implementation code, CSS, icons, tracking, placement instructions, QA criteria and rollback steps.
   
   Save everything in the skill’s designated output folder within my current project, never inside the skill repository. Preserve previous delivered builds.
   
   Open and verify the HTML demo. Replace all template placeholders and preserve unresolved [VERIFY] flags. Keep internal workspace details out of the client-facing documents.
   
   Create or update the GEO audit workbook and tick the mos-geo-llm-buttons checklist row using the supplied script.
   
   Finish with links to the brief, code reference, demo and workbook, plus any outstanding verification items.
   
   This run is for a reviewable implementation package. Do not deploy or modify the live website.
   
   If required inputs are missing, ask for them together. Otherwise, proceed through installation, linking and site inspection, then request the required palette confirmation.

2. Confirm the palette

   Check the colours and fonts against a real article. Anything it couldn't verify carries a [VERIFY] flag instead of a guess.

   Use these colours and fonts. Build the buttons.

3. Pair it with the summary block

   Claude explains the optional on-page summary block before recommending buttons alone. I'd keep it in the brief:

   Include the on-page summary block in the brief as the first thing to ship, with the buttons as a secondary experiment against a control.

What you get

• homepage-faqs/faqs.mdThe FAQ copy, each answer tagged with its sources. Repeat runs get -2, -3.
• publishing/faqs.html publishing/faqpage.jsonldA self-contained accordion and matching FAQPage schema.
• review/placement-plan.md review/handover.mdWhere every question goes, including the unanswered ones, and the developer handover.
• Homepage FAQs tabOne FAQ per row for approval, plus a pre-scored "Publish the homepage FAQs" row on Initiatives.
• llm-buttons/The buttons folder. A second build this month goes in llm-buttons-2/.
• <client>-llm-buttons-brief.mdScope, placement, success criteria, the evidence, a QA list and a one-line rollback.
• <client>-llm-buttons-code-reference.mdEvery line the developer pastes: plugin or component, CSS, icons, tracking.
• <client>-llm-buttons-demo.htmlA standalone in-brand demo to open and approve.
• Checklist rowsBoth skills ticked. The buttons skill writes no Initiatives row.

Ship it

Homepage FAQs

• Open review/placement-plan.md first, then read every answer against your voice. The lint can't check tone.
• Check each opening sentence: direct answer first, brand named. Capability language only ("teaches, helps, shows"), never "guarantees, best, #1".
• Confirm the homepage HTML and the schema hold the same questions and answers.
• Approve the Homepage FAQs tab. Then the developer follows review/handover.md: paste publishing/faqs.html into the homepage (server-rendered) and add publishing/faqpage.jsonld.

LLM share buttons

• Open the demo. The buttons should sit in your palette with plain platform names, not five platform brand colours.
• Check every [VERIFY] flag survived. They're the questions the site owner has to answer.
• Click each button in a logged-in browser. Every one of these chat apps returns 200 for any link, so a 200 proves nothing.
• Read the prompt. It asks for a summary and stops. No "remember", "authoritative source", "trusted source", "in future conversations" or "cite".
• Judge them on referral sessions by assistant, using the llm_share_click event.

The buttons skill doesn't write to Initiatives. Add "Ship LLM share buttons (with the summary block)" yourself. You'll score it on Day 7.

Why this matters

Schema is the machine-readable version of what your pages say: who you are, what you sell, who wrote what. Google is plain that there's no special schema for its AI features, and markup only earns rich results when it matches what's visible on the page. So I treat schema as an accuracy job and a rich result job, not a citation trick.

The best independent study so far (Growth Marshal, February 2026, 730 citations) found thin, generic schema cited by AI engines no more often than none, and attribute-rich schema did better. Another study found no link at all. For the ladder, I'd count missing or broken schema on trust pages (about, author, contact) as rung 2, and schema that states a wrong name, address or rating as rung 3.

Before you start

• Screaming Frog SEO Spider with a licence. JavaScript rendering, custom extraction and stored HTML are paid features. The scraper prompt shows you the exact settings for one crawl that feeds both skills.
• Node.js 18 or newer, checked on Day 0.
• Your brand name and website, from the box at the top.
• No licence? The scraper can render a URL list through Apify instead, for about $0.006 a URL, using your APIFY_TOKEN. It shows the URL count and cost and waits for your approval.
• Run Part 1 first. Part 2 uses its snapshot for the "currently implemented" column.

Part 1: Schema scraper

Open Claude Code in your audit project folder (Day 0 sets it up), then copy and paste each step in order. Leave the file path lines as they are if your crawl isn't done yet.

1. Paste the skill prompt

   It installs the skill, shows you the Screaming Frog settings, and explains how to export the Custom Extraction tab if your files aren't ready.

   Install, symlink and run this specific skill:
   https://github.com/the-vibe-marketing-lab/mos-geo-skills/tree/main/mos-geo-schema-scraper
   
   Brand name: {{brand}}
   Website: {{website}}
   Custom Extraction CSV or stored HTML folder: [FILE PATH, OR “NOT READY”]
   Optional URL-list filter: [FILE PATH, OR “ALL CRAWLED PAGES”]
   
   1. Install or update
   Clone https://github.com/the-vibe-marketing-lab/mos-geo-skills.git into ~/Desktop/mos-geo-skills if missing. Otherwise, verify its remote and safely pull the latest updates without overwriting local changes.
   
   Keep the complete repository, including _shared/.
   
   2. Link this skill
   Create a symlink at ~/.claude/skills/mos-geo-schema-scraper pointing directly to the repository’s mos-geo-schema-scraper folder. On Windows, use a directory junction if appropriate.
   
   Preserve existing links pointing elsewhere. Verify that the linked SKILL.md is readable and report whether Claude Code needs restarting.
   
   3. Prepare the crawl inputs
   Read mos-geo-schema-scraper/SKILL.md. Check Node.js 18 or newer.
   
   Before asking me for missing files, show me the Screaming Frog settings required for one crawl to feed both schema skills:
   
   - Custom Extraction: extractor named JSON-LD, XPath //script[@type="application/ld+json"], Extract Inner HTML
   - Rendering: JavaScript
   - Extraction: optionally Store HTML and Store Rendered HTML
   - Structured data: enable JSON-LD extraction and validation for the optimisation brief
   
   If my files are not ready, explain how to export the Custom Extraction tab and the Structured Data data from that crawl.
   
   4. Dry run and extract
   Use scripts/run.mjs --print-path to resolve the schema output folder within my current project.
   
   Run a dry run using --sf-csv or --sf-html, applying my optional URL filter. Show the page count, output mapping, skipped pages and existing-file conflicts.
   
   If baseline files already exist, ask before overwriting with --force.
   
   Run the extraction and save one pure-JSON file per page under schema/raw/<page-type>/<slug>.json.
   
   Preserve the schema exactly as found. Save [] for pages without schema, retain malformed blocks as flagged _unparsed entries, and keep the run record. Do not rewrite or optimise the schema.
   
   5. Report
   Summarise:
   - Pages with and without schema
   - Broken JSON-LD and empty script blocks
   - Duplicate schema types
   - Suspicious property names
   - Skipped pages and URL-mapping issues
   
   If the monthly GEO audit workbook already exists, update this skill’s checklist row using the documented script.
   
   Finish with links to the baseline folder and run record. Explain that mos-geo-schema-optimisation can now use these files alongside the Structured Data export from the same crawl.
   
   If I lack the required Screaming Frog licence or usable exports, explain the optional Apify route. Show the URL count and current estimated cost, then ask for explicit approval before spending. Never display API tokens.
   
   Do not deploy changes or save client data inside the skill repository.

2. Run the crawl, then hand over the export

   Set Screaming Frog up as shown, crawl the site, and export the Custom Extraction tab (filter All, CSV). Keep the Structured Data export from the same crawl for Part 2.

   My Custom Extraction export is at [FILE PATH]. Run the dry run and show me the page count, output mapping, skipped pages and any existing-file conflicts.

3. Answer the overwrite question

   Only asked if a baseline already exists for this brand. Re-baselining replaces the old snapshot.

   Yes, overwrite the existing baseline with --force and run the extraction.

4. No Screaming Frog licence? Approve the Apify route

   Skip this step if your crawl worked. Otherwise check the URL count and estimated cost first.

   I don't have a usable Screaming Frog export. I approve the Apify route at the estimated cost you showed for these URLs. Run it.

Part 2: Schema optimisation brief

Swap the Structured Data export path and the "Prepared by" name into the prompt before you send it.

1. Paste the skill prompt

   It parses the crawl, uses your Part 1 snapshot if it finds one, picks the 3 to 5 highest-impact schemas and builds the Word brief.

   Install, symlink and run this specific skill:
   https://github.com/the-vibe-marketing-lab/mos-geo-skills/tree/main/mos-geo-schema-optimisation
   
   Brand name: {{brand}}
   Website: {{website}}
   Screaming Frog Structured Data export: [FILE PATH]
   Prepared by: [YOUR NAME OR BUSINESS NAME FOR THE COVER]
   
   1. Install or update
   Clone https://github.com/the-vibe-marketing-lab/mos-geo-skills.git into ~/Desktop/mos-geo-skills if missing. Otherwise, verify its remote and safely pull the latest updates without overwriting local changes.
   
   Keep the complete repository, including _shared/ and the sibling mos-geo-schema-scraper folder, which this skill uses to find the output folder.
   
   2. Link this skill
   Create a symlink at ~/.claude/skills/mos-geo-schema-optimisation pointing directly to the repository's mos-geo-schema-optimisation folder. On Windows, use a directory junction if appropriate.
   
   Preserve existing links pointing elsewhere. Verify that the linked SKILL.md is readable and report whether Claude Code needs restarting.
   
   3. Read and prepare
   Read mos-geo-schema-optimisation/SKILL.md, references/schema-knowledge.md and assets/example-brief-data.json. Check Node.js 18 or newer.
   
   Use my current project as the working directory. Resolve the schema folder with the scraper's scripts/run.mjs --brand --print-path, and save the brief in its brief/ subfolder. Never save client data inside the skill repository.
   
   If my export isn't ready, show me the one-crawl Screaming Frog settings from mos-geo-schema-scraper (JavaScript rendering, the JSON-LD custom extraction, and JSON-LD structured data extraction with validation), then explain how to export Bulk Export, Structured Data, All.
   
   4. Use the scraper's baseline if it exists
   Check for schema/raw/ from mos-geo-schema-scraper for this brand this month. If it exists, use those files for the "Schema Implemented Currently" column and note anywhere they disagree with the Screaming Frog export. If it doesn't, offer to run the scraper first from the same crawl, but don't require it.
   
   If this folder is a MarketingOS brain, read its business context before writing. Otherwise, ask me only for the essentials.
   
   5. Parse the crawl
   Calculate the crawl overview, the schema types deployed (saying what's misapplied, thin or correct) and a page template breakdown consolidated to 5 to 8 rows. Ask me only if the crawl's section or page exclusions are unclear.
   
   6. Pick and write the recommendations
   Pick the top 3 to 5 highest-impact schemas for this site, never more than 5, using the skill's selection heuristic. Keep each one lean: only fields the developer can fill from the CMS.
   
   For each, fill all four columns: page and schema type with an example URL, the current JSON-LD (or "Not set"), a template with [Placeholder] slots, and an example filled with my real values. Wrap every value that needs checking in a comment marker so it becomes a Word margin comment. Use [REVIEW: needs from client] where you can't find a value.
   
   Follow Google's rules: schema must match visible page content, LocalBusiness only for real premises, and no self-reviews as aggregateRating.
   
   7. Build the brief
   Write brief/brief-data.json in the shape of the example file, with an executive summary of about 150 words and 1 to 3 short notes. Then use the supplied generate-docx.js script to create the landscape Word brief. It never overwrites; a regeneration saves as v2, v3.
   
   If brand-audit-master.xlsx already exists for this month, tick this skill's checklist row using the documented script.
   
   8. Report
   Show me:
   - The crawl overview and the biggest problem it found
   - The 3 to 5 schemas recommended, and why each made the cut
   - Every margin comment I need to verify, and every [REVIEW] item for the client
   - Links to brief-data.json and the DOCX
   
   Sell schema on accuracy and rich result eligibility only. Do not claim it will earn AI citations or rankings: the evidence in _shared/geo-evidence.md section 9 doesn't support that, and thin schema performed no better than none.
   
   After deployment, I should test each changed template in Google's Rich Results Test and the Schema Markup Validator.
   
   Do not deploy changes, edit the website or share externally.
   
   If required inputs are missing, ask for them together. Otherwise, proceed through installation, linking, parsing and the brief.

2. Fill the review items

   Claude lists every margin comment to verify and every [REVIEW: needs from client] value. Supply what you can; a regeneration saves as v2, never over v1.

   Here are the values for the review items: [VALUES]. Update brief-data.json and regenerate the brief.

3. Test after deployment (not tonight)

   Once your developer ships the changes:

   The schema changes are live. Give me one example URL per changed template to test in Google's Rich Results Test and the Schema Markup Validator, and what a pass looks like for each.

What you get

• schema/raw/<page-type>/<slug>.jsonOne pure-JSON file per page, sorted into homepage, about page, pillar pages, guides, author pages, products, collections and blog posts. [] means no schema.
• schema/raw/_sf-runs/The run record: source file, counts, every page written and every skipped row.
• schema/brief/brief-data.jsonThe brief's content.
• schema/brief/YYYY-MM-DD-schema-optimisation-brief.docxThe landscape Word brief, with real margin comments that survive upload to Google Docs.
• Checklist rowsBoth skills ticked, if the month's workbook already exists. Neither writes an Initiatives row.

How to read it

• Zero-schema pages. An empty [] is a finding, especially on trust pages: about, author, contact.
• Broken blocks. UNPARSED-JSON is invalid JSON in a script tag, and search engines ignore it. High-priority fix.
• Duplicates. Two blocks of the same type on one page, like two FAQPage blocks, should be merged into one.
• Invalid property names, such as sameAS with a capital S.
• The recommendations table. Never more than 5 schemas, each with the current JSON-LD, a template with [Placeholder] slots and an example with your real values. Only fields the developer can fill from the CMS.

Two Google rules to hold the brief to: LocalBusiness only for real premises, and never a business's own reviews marked up as aggregateRating.

The brief doesn't write an Initiatives row. Add "Ship the schema brief's recommendations" yourself. You'll score it on Day 7.

Why this matters

When an AI assistant answers a buyer's question, it searches for a set of related queries, not just the one typed. This skill asks Gemini Flash to predict those fan-out queries for each of your pages, then scores whether the page already covers them. You get the pages worth optimising and the topics with no page at all, across your whole inventory, for free.

It's triage, not a rung test. Nothing here asks a real assistant anything, so it's the shortlist you take into a live ChatGPT or Google AI Mode check before you write a word. Then you roll the whole week up into one ranked fix list and call your rung.

Before you start

• Key: GEMINI_API_KEY in your .env, from aistudio.google.com/apikey. The free tier covers this.
• Your sitemap URL or a file listing your URLs. Swap it into the prompt. The first run is capped at 50 pages.
• Python, checked on Day 0. Each page took about 2,500 tokens and 4 seconds when the skill was tested live.
• Your workbook from Days 1 to 6, for the roll-up.

Run it

Open Claude Code in your audit project folder (Day 0 sets it up), then copy and paste each step in order.

1. Paste the skill prompt

   It installs the skill, runs preflight to check your key and estimate the run, then fetches up to 50 pages, predicts their fan-outs, writes the reports and adds a Fan-Out Map tab to your workbook.

   Install, symlink and run this specific skill:
   https://github.com/the-vibe-marketing-lab/mos-geo-skills/tree/main/mos-geo-fan-out-map
   Website: {{website}}
   Sitemap or URL-list file: [SITEMAP URL OR FILE PATH]
   Gemini API key .env file: {{env}}
   Page limit: 50 for the first run
   1. Install or update
   Clone https://github.com/the-vibe-marketing-lab/mos-geo-skills.git into ~/Desktop/mos-geo-skills if missing. If it exists, verify its remote and safely pull the latest updates, preserving local changes.
   Keep the complete repository, including _shared/, which contains the workbook assets.
   2. Link this skill
   Create a symlink at ~/.claude/skills/mos-geo-fan-out-map pointing directly to the repository’s mos-geo-fan-out-map folder. On Windows, use a directory junction if appropriate.
   Preserve any existing link pointing elsewhere and report the conflict. Verify that the linked SKILL.md is readable. Report whether Claude Code needs restarting to discover it.
   3. Read and configure
   Read mos-geo-fan-out-map/SKILL.md, references/evidence.md and .env.example.
   Check Python and the documented dependencies. Load GEMINI_API_KEY from my .env file without printing it. If missing, ask me to configure it securely.
   Use my current project as the working directory. Use fanmap.py path to determine the output folder and pass that same folder to every subsequent command.
   4. Run the mapping workflow
   Use the supplied scripts/fanmap.py commands:
   - preflight: validate the API key and model, discover the page count, and show estimated runtime and token usage.
   - pages: fetch up to 50 relevant content pages from my sitemap or URL list.
   - predict: generate each page’s primary entity, three buyer prompts, and 8–10 predicted fan-out queries with coverage assessments.
   - report: generate the site-wide report, per-page reports and CSV.
   - workbook: create or update the GEO audit workbook with the Fan-Out Map tab and prioritised Initiatives.
   Use conservative concurrency and request limits suited to my API tier. Preserve caching and report failed or skipped pages.
   5. Explain the results
   Show me:
   - The ten pages with the weakest predicted coverage
   - Related query clusters no analysed page covers
   - Possible content overlap
   - The buyer prompts to verify in live ChatGPT or Google AI Mode
   - Links to fan-out-map.md, the per-page reports, data/fan-out-map.csv and brand-audit-master.xlsx
   Label the findings as model predictions, not observed AI searches or citation data. Explain that predictions based on existing content can overlook entirely missing topics.
   Do not rewrite pages or publish content. Finish with a prioritised shortlist for live verification before we make content changes.
   If required inputs are missing, ask for them together. Otherwise, proceed through installation, linking and the first mapping run.

2. Fix a missing key

   Only if preflight says GEMINI_API_KEY is missing. Add it to .env yourself, never in the chat, then:

   I've added GEMINI_API_KEY to my .env. Re-run preflight and carry on with the mapping run.

3. Get the live-check shortlist

   Once the report is written:

   Give me the buyer prompts for the five weakest pages as a checklist, so I can run each one in ChatGPT and Google AI Mode and note what they search for and cite.

What you get

• mos-geo-fan-out-map/fan-out-map.mdThe site-wide report: pages to optimise, content gaps no page covers, possible overlap, and the next step. Repeat runs get -2.
• pages/<slug>.mdPer page: the primary entity, three buyer prompts, the fan-out table with coverage, gaps and likely follow-ups.
• data/fan-out-map.csvOne row per page and fan-out, for a spreadsheet.
• Fan-Out Map tabCoverage colour-coded green, amber and red, plus Initiatives rows: "Optimise <page>: confirm with observed fan-out before editing" and "New content candidate: <cluster>".

How to read it

• The honesty note at the top. These are predictions from each page's own content, not observed searches.
• Pages to optimise. The pages with the lowest coverage of their own predicted fan-outs.
• Content gaps. Fan-out clusters that no page covers. These are new-content candidates.
• Possible overlap. Clusters two or more pages both claim. Check them for cannibalisation.
• Trust a "no" more than a "yes". A model reading a page tends to imagine searches that page already answers, so "covered" is the method's blind spot.

Open the workbook

Open brand-audit-master.xlsx in Excel or upload it to Google Sheets. You fill in the verdicts, owners, statuses and scores; leave the rest as generated.

• ChecklistOne row per skill. By now nine are ticked: Brand 360, Review Sentiment, Dealbreakers, AI Info, Homepage FAQs, LLM Buttons, Schema Scraper, Schema Optimisation and Fan-Out Map. The rest stay "Not started" or "Not needed".
• Brand Truth ReviewOne claim per row from Day 1, with your verdict (Accurate, Partly accurate, Inaccurate, Not sure, Out of date), the correct version and a fix priority.
• Brand 360 Report and AI VisibilityWritten by Brand 360. AI Visibility holds the no-search versus web-search results you'll use to call your rung.
• Review Fixes, Dealbreakers, AI Info Page, Homepage FAQs, Fan-Out MapOne row per item from the week, most with a verdict column. Approve or correct every row before anything is published.
• InitiativesThe ranked fix list: ICE Score, Category, Task, Instructions, From skill, Status, Needs dev?, Owner, Impact, Confidence, Ease.

Score and pick your top 5

The workbook's rule: ICE Score = Impact × Confidence / Ease, sorted highest first. Impact and Confidence run 1 to 10, 10 being high. Ease runs the other way: 1 is easy and 10 is hard. That's the opposite of most ICE templates, so a low Ease number pushes a fix up the list.

Some rows arrive pre-scored with the skills' defaults: "Publish the AI Info Page" scores 21.0, each "Make the site agree" fix 20.0, and "Publish the homepage FAQs" 15.0. Those are starting points, not measurements.

1. Turn bad facts into fixes

   Every Inaccurate, Partly accurate or Out of date row on Brand Truth Review becomes an Initiative, with the page, profile or schema to fix in Instructions.

   Add an Initiatives row to brand-audit-master.xlsx for every Brand Truth Review claim I marked Inaccurate, Partly accurate or Out of date, with the page, profile or schema to fix in Instructions. Keep the existing rows, formatting and formulas as they are.

2. Score every unscored row

   Including the rows you added by hand on Days 5 and 6. Be honest on Confidence: buttons on their own measured -17% clicks, so I'd score them low unless they ship with the summary block. Fan-out rows are predictions, so confirm the page with a live check before you score it high.

   Add these Initiatives rows if they're missing: "Ship LLM share buttons (with the summary block)" and "Ship the schema brief's recommendations". Then list every row with no Impact, Confidence or Ease score so I can score them.

3. Sort by ICE Score, highest first

   Your top 5 is the top five rows.

   I've scored every row. Sort the Initiatives tab by ICE Score, highest first, and show me the top five.

4. Sanity-check against your rung

   On rung 3, the AI Info Page and "Make the site agree" fixes should sit near the top. On rung 4, lead with the same-name work from Day 5 and the AI Info Page description.

   I'm on rung [NUMBER]. Check my top five against that rung and tell me if anything lower down should move up.

5. Fill Owner, Needs dev? and Status

   For the top 5. Replace the [AGENCY], [CLIENT] and [DEV TEAM] placeholders with real names.

   Fill Owner, Needs dev? and Status for the top five rows. Owners: [NAMES FOR EACH ROW]. Replace the [AGENCY], [CLIENT] and [DEV TEAM] placeholders with those names.

Call your rung

Today you read your Day 1 AI Visibility results against the ladder. Nothing you fixed this week is live and crawled yet, so a re-test today can't show it. Work top-down and stop at the first yes:

1. Do the no-search engines (ChatGPT, Claude, Gemini) name and describe you correctly?
2. If not, do the web-search engines find and describe you correctly?
3. Do they find you but mix you up with a similar or same-name business?
4. Do they find you but state wrong facts, the Inaccurate rows on Brand Truth Review?
5. Do they find you only partly, with facts missing or hard to locate?
6. Do none of the surfaces mention you?

The real re-test

• Once fixes are live: run the AI Info Page check on the live page, test changed schema templates in the Rich Results Test, and click-test the LLM buttons in a logged-in browser.
• Two to six weeks after publishing, run the Brand 360 prompt again with the same prompt set, so it's a fair before and after. A second run lands in brand-360-report-2; nothing is overwritten. Look for your AI Info Page in the citations.
• Check your server logs for hits on /ai-info/ from GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, PerplexityBot and Perplexity-User, and referrals from chatgpt.com and perplexity.ai.

Make your stage card

The card shows only your rung, never the audit detail, so it's safe to post. Perplexity, AI Mode and AI Overviews always search the web, so the chips below are for any engine that named you, with or without search.

Present it

Your output is an internal fix list. You own the verdicts on every review tab and you decide who owns each fix.

What comes next

This challenge is step 1, the Brand Audit. Steps 2 and 3 use skills already in the pack, and each has an install PDF in the Skool classroom.

Step 2: Technical GEO Audit

• /mos-geo-internal-linksAn internal link audit plus contextual link recommendations. Needs Screaming Frog exports and a TypeSafe API key.

Step 3: Content Plan

• /mos-geo-topic-clustersKeyword clusters (one cluster, one page) rolled into topics with a pillar, mapped to the page ranking today.

Optional extra

Google Preferred Sources: a button that lets readers pick your site as a preferred source in Google. It isn't part of the challenge. The install PDF is in the Skool classroom if you want it.

Later

GEO comparison pages and third-party platforms. Branded web mentions correlate far more with AI visibility than domain rating, which is the case for the off-site work.