FAT AgentGoogle's rulebook
Doorway pages, scaled content, keyword stuffing, hidden text, sneaky redirects and back button hijacking, the policy Google started enforcing in June 2026.
P1Templated near-duplicate pages (doorway page risk)
Fix · Audit · Test
He's big ondetail.
FAT Agent is the heavyweight website auditor for Claude Code. He crawls every page, checks your site against Google's latest rules, and doesn't clock off until the punch list is clean.
/plugin marketplace add spruikco/fat-agent-skill
Paste it into Claude Code. Two more lines and he's on the case. Full install
241things he checks
Minutesnot a two-week agency audit
$0free, no subscription
Exhibit 01: the stakeout
The trailer
Roll the tape.
Sound on.Every frame drawn in code by Claude.
What he checks
He's been over every inch of it.
Twenty-three modules and one standard: find the problem before Google does. Every finding lands on a punch list, ranked from P0 (the site is broken) to P3 (polish).
Every page, every link
A fast crawl into a local database. Broken internal links, orphan pages, duplicate titles, redirect chains and sitemaps full of dead URLs.
P0Internal links point at broken pages
Query fan-out
AI Mode splits one question into many. He checks whether your site answers the whole cluster, then who actually gets cited for it.
P2Low query fan-out coverage: 'seo agency melbourne'
Real clicks, not guesses
Full Search Console exports with no row caps. Every templated page gets a verdict: keep, improve or prune, backed by its own clicks.
P2Indexable pages with no search impressions
The timeline
Lines your traffic up against every Google core and spam update since March 2024, so "why did we drop?" gets a date and a suspect.
P1Traffic drop across the August 2026 spam update
The paperwork
Titles, descriptions, canonicals, structured data, and markup for rich results Google has retired. FAQ results went dark on 7 May 2026.
P3Markup for retired rich results
The back door
Security headers and CSP quality, cookie flags, secrets left in page source, and an opt-in probe for leaked .env and .git files.
P0Exposed sensitive path: /.env
Everyone gets in
Alt text, form labels, focus states and landmarks, plus layout shift, render-blocking scripts and pages over Googlebot's 2MB limit.
P2Images missing width/height (CLS risk)
Main street
Name, address and phone, Google Business Profile, LocalBusiness schema, product markup and merchant listings for shops.
P2LocalBusiness schema missing NAP details
Jurisdiction
He works every beat.
If it's live on the web, he can audit it. Hosted builders, headless stacks, the lot. Stores switch on the e-commerce checks automatically.
BEAT 01ShopifyStores, products, schema
BEAT 02WooCommerceWordPress stores
BEAT 03WordPressThemes, plugins, Yoast
BEAT 04WixHosted sites and stores
BEAT 05SquarespaceHosted sites and stores
BEAT 06WebflowVisual builds
BEAT 07BigCommerceStores at scale
BEAT 08Next.jsPlus Astro, Nuxt, Remix
BEAT 09NetlifyPlus Vercel, Cloudflare
BEAT 10Anything elseAny live URL
Every beat gets the full audit. Fixes come with step-by-step guides written for the platform you're on, so the advice matches the buttons you can actually press.
Beyond SEO
He checks the locks too.
FAT Agent is a security tester as much as an SEO auditor. He looks at your site the way an attacker would, from the outside, grades every defence on quality rather than presence, and ranks what he finds from P0 to P3.
P0Leaked keysStripe, AWS, GitHub, Slack, Anthropic and OpenAI keys in your page source
P0Exposed files.env, .git, database dumps and 20 more, probed only with your go-ahead
P1Security headersCSP and HSTS graded, not just ticked
P1Email spoofingSPF, DKIM and DMARC, so nobody phishes in your name
P1ConsentTrackers that fire before anyone says yes
P2Cookies and scriptsSecure, HttpOnly, SameSite, and third-party code without integrity checks
New department
He can tell it was generated.
Scores can be green while the screen still reads as generated. FAT Design runs six lenses over what the visitor actually sees: what it does, what it says, how the type is set, the space and colour, how it behaves under your hands, and how it moves. Every weak decision gets a name, a price and a brief.
P0ProductThe wrong control, an unnamed consequence, a state nobody designed
P1CopyButtons that name a mechanic, claims with no evidence, errors with no way out
P2TypeFaux bold, a voice that argues with the words, sizes that nearly match
P2CraftA look never chosen, gaps that nearly match, border soup, six greys
P0InteractionForms that wipe your work, invisible focus, targets you cannot hit
P1MotionA hero blank on first paint, movement with no job, reduced motion ignored
Case file
The vanishing traffic.
A real audit, run on real Search Console data. Client details withheld.
Agent's notes, typed up for the client.
Where did the impressions go?
Case 2026-0917Subject: Australian digital agency · Pages: 1,875 indexable · Symptom: leads stopped
Traffic didn't slide. It fell off a cliff. Weekly Google impressions dropped from about 29,000 to about 1,100 in a single week.
- P1The drop lines up with Google's August 2026 spam update, which began on 18 August. Clicks fell 87% across the update window.
- OKEvery key page was still "Submitted and indexed". Google had the pages. It had simply stopped showing them.
- P1674 pages were near-identical templates: suburb and city pages with the place name swapped. Together they earned 57 clicks in 90 days.
- P2Every suburb page claimed to be run by "local experts" but listed the same interstate phone number.
31Keep
55Improve
462Prune
Verdict: consistent with a doorway and scaled-content demotion. Prune the templates, strengthen the hub pages, and watch for recovery at the next core update.
Triage by page diffing and real Search Console clicks. We also tested a small local AI judge on this site: it made the calls worse, so model judges stay advisory until they prove themselves. Correlation, not a confession: FAT never claims Google penalised you unless Search Console shows a manual action.
Second file on the desk. A store this time.
The empty checkout.
Case 2026-1031Illustrative case · composite figures
Subject: Shopify store, homewares · Pages: 4,812 crawled · Symptom: organic sales down 62%
Sales didn't dip. They vanished. He lined up the usual suspects and questioned every one.
3,400Near-duplicate collection and filter pages, all indexed
1,200Product descriptions copied word for word from the manufacturer
212Products with no price in their structured data
38Internal links to checkout pages that 404
Verdict: the duplicates did it. It's always the duplicates. Consolidate the filter pages, rewrite the copied descriptions, fix the product markup and the dead links. Organic sales were back within two core updates.
This one is a composite, built to show how he works a store. The first file is real.
Method
Fix. Audit. Test.
Three letters, one routine, and a punch list that remembers what it found.
Fix
He doesn't just leave a report on your desk. With your repo open he makes the changes himself: headers, schema, redirects, sitemaps and content briefs.
Audit
He crawls the site, runs every relevant module, pulls your Search Console data and ranks what he finds from P0 to P3.
Test
After you deploy he re-crawls. Fixed items close themselves. Anything that crept back gets reopened, so nothing slips through twice.
In practice he audits first. But nobody wants to hire an AFT Agent.
Informants
Every good detective has a few.
Some calls need judgement, not pattern matching. Is this suburb page real local content, or the template with the name swapped? Does this page actually answer the question? He asks an informant, and every one answers the same way: a calibrated yes or no.
Default
Claude
Nothing to set up. Claude judges a sample of pages as part of the audit, and was the most accurate judge in our own testing.
Free · works everywhere
Hosted
Jev by TypeSafe
A decision model built for exactly this: typed answers with calibrated confidence, in milliseconds, across every page.
Bring a key · cents per site
Local
OpenJev
An open, Jev-compatible model on your own graphics card. Free and private, and advisory until you validate it: in our test a small local model made worse calls than Claude.
Needs an 8GB GPU · free
No API key required. Ever.
FAT HQ
Keep him on retainer.
The plugin is free and always will be. FAT HQ is where he keeps the case files after you close the laptop: every audit on one board, the score over time, your Search Console clicks against Google's updates, re-checks on a schedule and an email when something breaks.
Case filesEvery audit from the plugin or HQ, with what got fixed, what's new and what came back.
Search ConsoleWeekly clicks and impressions with Google's core and spam updates marked, so "why did we drop?" has a date.
StakeoutsWeekly or daily re-checks of the pages you pick. No laptop required.
Tip-offsAn email the moment a P0 appears, a fix comes undone or the score drops five points.
Client reportsA clean report link for each site. On Agency, with your name, logo and colour.
Install
Put him on the case.
Inside Claude Code, run these two lines. Then point him at a site.
/plugin marketplace add spruikco/fat-agent-skill/plugin install fat-agent@fat-agent-marketplaceRun
/fat-audit https://your-site.com, or just say "run FAT agent on my site".
He's not fat.
He's thorough.