seo-geo
Skill1 starsChanged 48 days agoMITView on GitHub
SEO and AI-search audit for a site or page. Use when asked why a site isn't ranking or cited by ChatGPT, Perplexity or AI Overviews, or to check robots.txt, AI crawlers, sitemaps, schema or llms.txt.
58 other projects use a version of this.
Adopted by
Most recently taken up by
- FlorentCoulonBS/claude-skillstheir own version, first seen Nov 16, 2026
- martinbattaglia-commits/Claude-skillstheir own version, first seen Nov 16, 2026
- yotamwiesel/claude-skillstheir own version, first seen Nov 16, 2026
- joaoguirunas/team-ostheir own version, first seen Nov 15, 2026
- theexperiencecompany/gaiatheir own version, first seen Nov 13, 2026
- YerayRodri/claude-skillstheir own version, first seen Nov 13, 2026
Used with
Projects with it also use seo-agentic, seo-competitor-pages and 10 more.
Other files people use alongside this one.
Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
AgriciDanielSkillseo-agentic
Audit and fix agent readiness: the Lighthouse Agentic Browsing fraction, accessibility tree for agents, robots.txt and Content-Signal for AI agents, WAF treatment of agent traffic, llms.txt, Markdown delivery, ai-catalog.json, /.well-known discovery files, and WebMCP tools. Use when the user asks whether AI agents can use or browse the site, or about llms.txt or WebMCP. Exclude AI citability and brand signals (seo-geo) and commerce protocol depth (seo-ecommerce).
Also in 5 projects that use this
megastepSkillseo-competitor-pages
Generate SEO-optimized competitor comparison and alternatives pages. Covers "X vs Y" layouts, "alternatives to X" pages, feature matrices, schema markup, and conversion optimization. Use when user says "comparison page", "vs page", "alternatives page", "competitor comparison", or "X vs Y".
Also in 22 projects that use this
timruebenachSkillseo-google
Google SEO APIs: Search Console (Search Analytics, URL Inspection, Sitemaps), PageSpeed Insights v5, CrUX field data with 25-week history, Indexing API v3, and GA4 organic traffic. Provides real Google field data for Core Web Vitals, indexation status, search performance, and organic traffic trends. Use when user says "search console", "GSC", "PageSpeed", "CrUX", "field data", "indexing API", "GA4 organic", "URL inspection", "google api setup", "real CWV data", "impressions", "clicks", "CTR", "position data", "LCP", "INP", "CLS", "FCP", "TTFB", or "Lighthouse scores".
Also in 14 projects that use this
megastepSkillseo-sitemap
Analyze existing XML sitemaps or generate new ones with industry templates. Validates format, URLs, and structure. Use when user says "sitemap", "generate sitemap", "sitemap issues", or "XML sitemap".
Also in 22 projects that use this
timruebenachSkillseo-image-gen
AI image generation for SEO assets: OG/social preview images, blog hero images, schema images, product photography, infographics. Powered by Gemini via nanobanana-mcp. Requires banana extension installed. Use when user says \"generate image\", \"OG image\", \"social preview\", \"hero image\", \"blog image\", \"product photo\", \"infographic\", \"seo image\", \"create visual\", \"image-gen\", \"favicon\", \"schema image\", \"pinterest pin\", \"generate visual\", \"banner\", or \"thumbnail\".
Also in 13 projects that use this
- +3 more from megastep
- +4 more from timruebenach
Other versions58 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 58
- AgriciDaniel/claude-seo★ 30ksame name
- ReScienceLab/opc-skills★ 1.9ksame name
- theexperiencecompany/gaia★ 327same name
- eunomia-bpf/eunomia.dev★ 244same name
- seranking/seo-skills★ 201same name
- zhtyyx/codex-seo★ 152same name
- PostPlusAI/postplus-skills★ 27same name
- chuhaiqu/skills★ 22same name
- coopersimson96/ai-content-system★ 17same name
- whyashthakker/agent-skills-marketing★ 15same name
48 more not listed
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository2 other files it gives its agents.
- AGENTS.mdAGENTS.md
- content-reviewSkill
The file
Show its rules and text
We split this skill into the rules it gives an agent
14 rules. 0 are common, 9 are distinctive, and 5 are very specific to this project.
0
Common
Most of the 53 edited versions we read keep these.
9
Distinctive
This author's own choices. Worth a look.
5
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 20
Before you repeat a number from this skill to anyone, open its source.
- Very specific: line 36
Every script takes a saved file as well as a URL where that makes sense, prints a plain report and supports --json.
- Distinctive: line 60
Auditing a whole site?: Read §18a first. It covers what to do before trusting any finding: check it against the live HTML, never recommend …
- Very specific: line 73
Every image has alt, width and height?
- Very specific: line 77
Do the visible date, dateModified and sitemap lastmod agree?
- Very specific: line 83
Search generative AI setting?
- Distinctive: line 161
Only about 11% of domains are cited by both ChatGPT and Perplexity (Profound, the same 100,000 prompts run on each, July 2025 data), and …
- Distinctive: line 201
Keep the answer to each heading visible on load.
- Very specific: line 205
If JavaScript changes the canonical or the robots meta, the raw HTML and the rendered page must agree.
- Distinctive: line 213
Put the answer at the top.
- Distinctive: line 229
Write a 40 to 60 word direct answer near the top that survives being quoted with no context.
- Distinctive: line 231
Prefer "X is a Y that does Z" over a chain of pronouns, and write plainly.
- Distinctive: line 291
Do not split one argument across a heading.
- Distinctive: line 299
Stop spending on these.
“Common” is measured: kept by at least 6 in 10 of the 53 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
SEO and GEO, graded by evidence
Optimisation for classic search (SEO), for generative engines (GEO) and for answer engines (AEO), where every rule carries a tag saying how strong the evidence behind it is. Controlled experiments are kept apart from correlation studies, vendor claims are marked as vendor claims, and there is a list of tactics that were measured and found wanting.
Last verified against primary sources: 29 September 2026. Crawler rules, index composition and vendor guidance all change, and one widely quoted overlap figure moved 38 points in seven months. Before you repeat a number from this skill to anyone, open its source.
Quick start
Plain Node 18 or later, nothing to install. Paths are from the repository root; in an installed
skill the scripts sit in this skill's own scripts/ folder.
node skills/seo-geo/scripts/audit-page.mjs https://example.com/pricing # one page, every mechanical rule below
node skills/seo-geo/scripts/crawl-site.mjs https://example.com # a whole site: broken links, orphans, duplicates, redirect chains
node skills/seo-geo/scripts/check-crawlers.mjs https://example.com --live # which AI crawlers robots.txt and the CDN let in
node skills/seo-geo/scripts/check-sitemap.mjs https://example.com # every sitemap URL: 200, indexable, self-canonical
node skills/seo-geo/scripts/analyze-gsc.mjs ./gsc-export/ # Search Console exports, read offline
node skills/seo-geo/scripts/check-logs.mjs access.log --verify # which crawlers really visited, verified by IPEvery script takes a saved file as well as a URL where that makes sense, prints a plain report and
supports --json. Exit codes: 0 clean, 1 errors found, 2 usage error, 3 could not check.
Where to look
This file holds the rules that apply to every page. The detail is in the reference files, and section numbers carry across, so "§9b" always resolves.
| Read | For |
| references/site-changes.md | §18 auditing a whole site, migrations and redirects, hreflang, faceted navigation, programmatic pages, a first 30 days |
| references/measurement.md | §16 Search Console, Bing's AI Performance report, analytics and logs, and building a prompt panel |
| references/page-fields.md | §8 document structure, §9 head elements, dates, authorship, schema by page type, §13 what structured data is for |
| references/authority-and-links.md | §4 writing something worth citing, E-E-A-T, entities; §6 off-page; §7 freshness; §10 links |
| references/technical-delivery.md | §11 images, §12 Core Web Vitals, §14 crawler control by purpose and CDN policy, §15 accessibility and agents |
| references/commerce.md | §17 AI shopping: feeds, ChatGPT and AI Mode, checkout protocols |
| references/verification.md | copy-paste curl checks, priorities by page type, and every source by evidence grade |
Data files: data/crawlers.json (documented AI crawlers, by purpose, with the vendor's own words on robots.txt), data/myths.json (widely repeated claims that are false, each with its source), data/google-changes.json and data/ai-platform-changes.json (dated changes since February 2026, each confirmed on the platform's own pages).
Auditing a whole site? Read §18a first. It covers what to do before trusting any finding: check it against the live HTML, never recommend what already exists, and never state a number you did not read from data you were given.
60-second pre-ship check
- Is the content in the raw HTML?
curl -s <url> | grep -q "<a distinctive sentence>" - One
<h1>, no skipped heading levels? - The answer in the first 100 or so words, and at the top of each
<h2>section? - Does each
<h2>section name its subject rather than saying "it", and would it make sense quoted alone? - A unique title that says plainly what the page is, echoing the H1? A description written for a person?
- Every image has
alt,widthandheight? The main image above the fold hasfetchpriority="high"and is not lazy-loaded? - Structured data matching the visible copy, with every fact that must be quotable also in the visible text?
- Do the visible date,
dateModifiedand sitemaplastmodagree? (§9b) - Three to five contextual internal links out, with descriptive anchor text?
- Any real statistics, first-hand data, quotes or named sources, or is it prose any model could have written?
- Can the search crawlers get in, in robots.txt and at the CDN: Googlebot, Bingbot, Applebot,
OAI-SearchBot, Claude-SearchBot, PerplexityBot? Is the site set to Include in Search Console's
Search generative AI setting? Is there a stray
noarchivethat takes pages out of Copilot? (§14)
How to read this: evidence grades
Most SEO and GEO advice has never been measured. Three things make that worse:
- Google's public statements have been contradicted by its own internal documents.
- Most GEO research is published by vendors selling GEO tools.
- "Cited pages have X" is not "adding X gets you cited". Almost every correlation in this field is confounded by one variable: being a large, well-run site. Big sites have schema, fast pages, author bios and llms.txt files, and they also get cited. The only way through is intervention: someone adds X to real pages and measures the change against controls. When teams do that, they keep finding nulls where correlation studies found lifts.
So every rule is tagged:
| Tag | Meaning | Trust |
| [E] | Controlled experiment or split test: X was added, the change measured against controls | Highest |
| [P] | Peer-reviewed or academic research | High, if the setting matches reality |
| [M] | Directly measured behaviour: crawler logs, API instrumentation, court records | High |
| [S] | Large-sample correlation study, sample size given | Medium: direction, not dose |
| [G] | Official platform documentation (Google, Microsoft, OpenAI, Anthropic, Apple) | Medium, see below |
| [V] | Vendor study: directionally useful, conflicted | Low |
| [X] | Tested and found not to work, or failed replication | Act on this |
Where evidence conflicts, prefer [E] and [M]. And keep one base rate in mind: across all the SEO split tests SearchPilot ran in 2022 and 2023, about 75% were inconclusive, about 15% won and 7 to 8% actively hurt [E]. SearchPilot later re-scored a sample with a more sensitive model and counted nearer 29% as wins, which still leaves most changes doing nothing measurable. Price any tactic sold without a test against that.
Official guidance is the starting point, not the last word
The May 2024 Google Content Warehouse leak (2,596 modules, 14,014 attributes) contradicted years of public statements. Google confirmed the documents were real while disputing how they were being read [M]:
| Google said publicly | The documents showed |
| There is no "domain authority" metric | A siteAuthority feature |
| Clicks are not used for ranking | NavBoost, click-driven re-ranking, also described in sworn testimony at the US v. Google trial |
| There is no sandbox for new sites | A hostAge attribute used "to sandbox fresh spam" |
| Chrome data is not used in Search | A chromeInTotal site-level signal |
The leak proves these attributes exist in a storage schema, not their weights or current use. Public guidance is written to discourage manipulation, so it tends to understate anything that can be gamed.
Sometimes Google is right and the industry is wrong. Google says structured data is not needed for AI features [G], and the evidence agrees (§13). Its AI optimization guide (updated 10 July 2026) says there is "no requirement to break your content into tiny pieces" [G], which matches what has been measured for Google. For other engines, measured retrieval behaviour (§3) shows passage structure does matter. The two fit together: one clear subject per section is not tiny pieces.
Google also warns about tools that sell this kind of advice. Its third-party SEO guidance (June 2026) names "AEO" and "GEO" tools and says third-party tools "don't have access to our internal ranking data" and "can't guarantee performance" [G]. Any score a tool computes, ours included, is that tool's own measurement.
0. The retrieval map: "AI search" is not one channel
Each engine retrieves differently, which decides how much classic SEO carries over. As of September 2026 [M] [G]:
| Engine | Retrieves from | What carries over, and the controls |
| Google AI Overviews and AI Mode | Google's index, through "query fan-out" (§3). Grounding uses a faster, cut-down semantic ranking described in the court's findings in US v. Google (September 2025), citing Google's own testimony. Since Google I/O (May 2026) an AI Overview can lead straight into an AI Mode conversation | Google SEO carries over most. The page must be indexed and snippet-eligible, and the site included in Search Console's Search generative AI setting (defaults to Include). nosnippet and max-snippet limit what AI answers may use |
| Microsoft Copilot | Bing's index, which since June 2026 also grounds third-party agents through Microsoft's Web IQ APIs | Bing SEO carries over directly, and Bing says grounding weighs freshness and accuracy more than ranking did. Bing uses noarchive to keep a page out of Copilot answers entirely, and nocache to limit it to title, URL and snippet, while Google ignores both |
| ChatGPT search | OpenAI's crawler (OAI-SearchBot) plus licensed and third-party data. In Profound's July 2025 data only about 8% of ChatGPT's citations matched Bing's top results and about a third matched Google's | Partly. OpenAI says sites that block OAI-SearchBot "will not be shown in ChatGPT search answers, though can still appear as navigational links"; only noindex removes the link as well, and the crawler must be allowed to see it |
| Perplexity | Its own index (over 200 billion URLs, as Perplexity described it in September 2025), with its own rerankers | Partly. Google rank correlates but is not the mechanism |
| Apple (Siri and Search answers) | Applebot. Siri AI, released September 2026, answers from web information | Apple says nosnippet keeps a page out of its AI-generated answers, and isAccessibleForFree set to false keeps it in search but out of AI context |
| Claude | Anthropic's Claude-SearchBot index plus a third-party search provider that Anthropic does not name (Brave appears on its subprocessor list). Since September 2026 the new Claude experience searches whenever Claude judges it useful, with no toggle | Least documented |
What follows from this:
- Only about 11% of domains are cited by both ChatGPT and Perplexity (Profound, the same 100,000 prompts run on each, July 2025 data), and across other pairs of engines Profound found between about 6% and 16%. Academic audits in 2026 find overlap as low or lower. Optimising for "AI" as one thing is a category error [S].
- Google's AI retrieval is semantic first: relevance to the sub-query beats site authority there [M].
- Where organic rank still gates citation is contested and moving (§6 has the numbers). The stable version: being retrievable somewhere each engine looks is the floor.
1. Rendering: the rule that outranks every other
Every word that must rank or be cited belongs in the server-rendered HTML.
- The crawlers that feed ChatGPT, Claude and Perplexity executed no JavaScript when Vercel and MERJ measured them across Vercel's network (published December 2024, about 1.3 billion AI crawler fetches in a month): GPTBot and ClaudeBot downloaded JavaScript files in 11.5% and 23.8% of requests and ran none of them, and CCBot does not render either [M]. That is still the best public measurement, and it is 21 months old. OAI-SearchBot now sends a full Chrome user agent, which does not prove it renders, and one October 2025 test saw a JavaScript-only price turn up in Perplexity's results. Re-test with a sentence that only JavaScript inserts, and check your logs, before relying on either answer.
- Three major crawlers do render: Googlebot, Bingbot and Applebot. Googlebot's rendering also serves Gemini; Bing's guidelines ask sites to "allow Bingbot to crawl and render content efficiently" and warn that "content that cannot be reliably rendered may not be indexed or selected for grounding results", which covers Copilot; and Apple's documentation says "Applebot may render the content of your website within a browser" [M] [G]. So never write that AI crawlers as a group skip JavaScript: name the crawlers, or say "every crawler except Googlebot, Bingbot and Applebot".
- Browser agents are different. ChatGPT's agent, Perplexity's Comet and Claude in Chrome drive real browsers and do render, but those are single-user sessions, not index crawls. Retrieval and citation still run on the raw HTML [M].
- Retrieval pipelines turn HTML into plain text or Markdown before a model sees it. Visible body text
survives that. Text living only in
<script>tags is unreliable: in searchVIU's test no engine extracted a price that existed only in JSON-LD, while in another ChatGPT and Perplexity read an address from a JSON-LD block as plain text (§13) [M]. - Google may not render pages that return a status other than 200, so an error page that builds its content in JavaScript may never be seen [G].
What to do:
- Server-render or statically generate every piece of meaningful content. Use client-side JavaScript for interaction, never for content that must be found.
- Keep the answer to each heading visible on load. Collapsed text inside the HTML still counts for
indexing, but Google's "read more" deep links, which jump a searcher to the matching section,
favour content that is visible on load (April 2026). Collapse detail, not answers, and do not strip
#fragmentsfrom the URL on load [G]. - If JavaScript changes the canonical or the robots meta, the raw HTML and the rendered page must
agree. Never put
noindexin the raw HTML and remove it with JavaScript [G].
audit-page.mjs reports how much visible text is in the raw HTML. If that is close to zero, fix it
before anything else.
2. Answer first
Put the answer at the top. Several kinds of evidence point the same way:
- Lost in the Middle (Liu et al., TACL 2024): language models use information at the start or end of their context much better than in the middle. Newer long-context models narrow the gap but, in 2026 re-tests, have not closed it [P].
- About 44% of ChatGPT citations come from the first 30% of a page (Kevin Indig, 44.2% of 18,012 verified citations). The same work found cited passages are about twice as likely to use definitional phrasing ("X is", "X refers to") and read at a simpler grade level [S].
- Question-format
<h2>headings gave +12% organic sessions in a SearchPilot split test (the headings also gained the product name). Brand selling-point modules near the top of pages gave +18% traffic from AI assistants, with the effect on Google inconclusive (SearchPilot with Omio) [E]. - Google's grounding is extractive: it lifts near-verbatim sentences that stand on their own [M].
So:
- Open the page, and each
<h2>section, with the answer. Do not build up to it. - Write a 40 to 60 word direct answer near the top that survives being quoted with no context. Make it a required field in your CMS, not a habit.
- Prefer "X is a Y that does Z" over a chain of pronouns, and write plainly.
The one controlled GEO test pair we know of is a warning [E]. The same SearchPilot programme that found +18% from selling-point modules also tested a "key takeaways" bullet block. It raised AI referral traffic and cost an estimated 6.5% of Google organic sessions, so it was not rolled out. GEO and SEO effects can pull in opposite directions on the same change, so measure both before changing templates. And do not split content into micro-pages: Google says it understands multi-topic pages, and fragmenting has no evidence behind it.
3. Query fan-out and passage-level retrieval
This is how AI search retrieves, and it moves the unit of optimisation from the page to the passage.
Fan-out. "Query fan-out" is Google's own term, introduced with AI Mode in March 2025 and now in its AI documentation. The engine splits a question into subtopics and runs many searches at once; Google says its Deep Search mode runs "hundreds of searches". It has never published a count for an ordinary query, so treat any specific number you read as made up [G].
What the evidence says fan-out rewards:
- In a 1.4-million-prompt ChatGPT study (Ahrefs, published April 2026), similarity between the page title and the sub-queries was among the strongest citation correlates, stronger than similarity to the original prompt. Descriptive, natural-language URL slugs also correlated with citation [S].
- 88% of Google AI Mode citations are not in the organic top 10 for the visible query (Moz, about 40,000 queries), because they rank for the sub-queries instead [S].
- Sub-queries are generated and vary between runs, padded with words nobody typed ("best", "reviews", the current year). You cannot target them exactly. You can cover them. Bing's AI Performance report shows the actual "grounding queries" that retrieved your pages (§16).
What to do:
- Give each sub-question its own
<h2>, phrased the way a person would ask it. Question-format headings have causal support: +12% in a split test [E]. - Answer the obvious neighbouring questions on the page rather than sending the engine elsewhere.
- State comparisons, specifications and criteria explicitly. Fan-out reliably generates commercial and comparative sub-queries.
- Cover sub-questions on the page; do not spin out a page per fan-out variant. Google's AI guide says creating "separate content for every possible variation", fan-out queries included, to manipulate rankings or AI responses "violates Google's scaled content abuse spam policy" [G].
Passage mechanics, measured. Dan Petrovic instrumented Google's Gemini grounding API (7,060 queries, 883,000 snippets) [M]:
- Google allows roughly 2,000 words of grounding per query, split across the top sources by relevance. The top source contributes around 530 words, the fifth around 266.
- The median page contributes about 377 words, and extraction levels off around 540 words per page. Pages under 1,000 words get about 61% of their content used; pages over 3,000 words, about 13%. Density beats length.
- Extraction is near-verbatim, chosen for sentences that stand alone.
Mike King's chunking demonstrations point the same way. Splitting a two-topic paragraph improved one half's similarity to its query by about 19%, and in a separate example putting the heading in front of a passage improved it by about 17% [M]. They are different measurements, so do not add them. Chunk sizes inside ChatGPT, Perplexity and Claude are not public.
At passage level:
- Repeat the subject's name instead of "it" at the start of each section. "Acme Analytics tracks..." survives extraction; "It tracks..." does not.
- One subject per passage: each
<h2>section should answer its own heading completely. - Do not split one argument across a heading.
- Front-load each section; the start and end of a passage carry the most weight.
- Past about 500 dense words on one subtopic, most of it is never extracted.
(§4 onwards lives in the reference files: see the table under "Where to look".)
5. What has been measured NOT to work
Stop spending on these.
| Tactic | Finding | Evidence |
| llms.txt | 97% of the files received zero requests in server logs from 137,000 domains (Ahrefs, May 2026 data). Of the bots that did ask, SEO audit tools were the largest group and AI retrieval bots about 1%. SE Ranking found no citation difference across 300,000 domains. Google says Search does not use it and that it "will neither harm nor help". No engine documents reading it | [M] [G] [X] |
| JSON-LD as an AI-citation lever | 1,885 already-cited pages that added schema, compared 30 days before and after against about 4,000 that did not: no upward movement. ChatGPT +2.2% and AI Mode +2.4% are within noise; AI Overviews −4.6% was statistically significant but small. A matched natural experiment, not a controlled test. §13 covers what schema is still for | [S] [M] [X] |
| FAQ and HowTo rich results | HowTo rich results ended in 2023. FAQ rich results stopped showing on 7 May 2026. Visible question-and-answer copy still gets extracted; the markup wrapper does nothing in Google | [G] [X] |
| Keyword density and chasing content scores | Keyword stuffing lost ground on the GEO paper's main visibility measure (about −8%), and a 2026 survey of the replications grades it null or negative. Surfer's own study: content score to rank correlation of 0.28. No controlled test of score-driven edits exists. Use coverage tools as checklists | [P] [X] [V] |
| Generic "GEO rewrites" at scale | Significant in only 3 of 54 method and domain combinations in C-SEO Bench; another 2026 benchmark found rewritten pages lost about 9% at retrieval and 16% at reranking; gains shrink as everyone adopts them | [X] |
| A page per fan-out query | Google names it scaled content abuse (§3) | [G] [X] |
| Author markup as a ranking tactic | Google's Danny Sullivan: bylines "aren't something you do for Google". No positive test in years of it being standard advice. Write author bios for readers (§4) | [G] [X] |
| Tracking AI rank with single prompts | SparkToro and Gumshoe (600 volunteers, 2,961 runs): less than a 1 in 100 chance two runs of the same prompt return the same list of brands. Aggregate rates across many prompts are the defensible measure (§16) | [S] [X] |
| Buying Reddit or forum mentions | Planting text works in a lab: a Cornell team got a 13-word planted passage cited in 38% to 51% of deep-research agent reports in a sandbox. In the wild, a thread's search rank rather than its karma predicts citation, Reddit says it catches about 25,000 spammy posts and comments a day, and bought placements disappear with the accounts | [S] [M] |
| Chasing domain authority | For AI Overview visibility, the number of backlinks (0.218) correlated far more weakly than plain brand mentions (0.664) across 75,000 brands (§6) | [S] |
| Length for its own sake | Extraction levels off around 540 words per page; pages over 3,000 words get about 13% used. Covering sub-questions (§3) is what "comprehensive" should mean | [M] |
| Speakable schema for AI search | Google still documents it, as a beta, for one narrow use: Google Assistant reading news aloud on Google Home devices set to US English. No AI answer engine, AI Overviews and AI Mode included, documents reading it | [G] |
| Serving different content to AI bots ("AI shadow sites", "edge AI delivery") | Cloaking. Since 15 May 2026 Google has stated that its spam policies also apply to generative AI responses in Search | [G] do not |
On llms.txt: ship one only if you want a Markdown map of your docs for developer tools. Do not ship
it expecting ranking or citation gains, and treat any audit that leads with llms.txt as a warning
sign; an "AI visibility audit" built on llms.txt plus schema plus a proprietary score is a familiar
sales pattern. Serving Markdown versions of pages is a separate question: some developer tools do
ask for them (Cloudflare reports Claude Code and OpenCode send Accept: text/markdown, and server
logs show Cursor and Copilot doing the same), while ChatGPT, Claude.ai, Gemini and Perplexity do
not. If you negotiate, do it at the edge: Cloudflare "does not consider vary values in caching
decisions" by default, so origin-side negotiation behind a cached HTML route can serve Markdown to
browsers. Cloudflare's own Markdown for Agents converts at the edge and sets Vary: Accept itself,
but see §14 for the content signal it adds [G].
For reviewing a draft before it ships (sources, freshness, machine-sounding prose), use the content-review skill.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

