Expertise: what the agent knows

Search demand & intent#

• Intent — checks whether a page answers what the searcher typed, in the shape that already ranks for it: a comparison, a how-to, a quick fact → the page is written for the query's dominant meaning, and a quick-fact question gets its answer in the first sentence or two.
• Titles & snippets — checks for duplicate or keyword-stuffed titles, shared meta descriptions, and the four defects that make Google rewrite a title: half-empty, stale, inaccurate, boilerplate → every page gets its own accurate title and description.
• Depth — checks whether a subject reads as covered or only mentioned, and whether a page adds anything beyond what is already published → thin pages gain the missing facets, a catch-all page splits into focused ones, and the links between them say what they point to.

Crawl, index & speed#

• Crawl & index — checks that each page can be fetched, rendered and kept in the index: truthful sitemap dates, the 50,000-URL sitemap limit, single-hop redirects, real "not found" responses → when the agent moves a page, the old address redirects to the new one in a single hop, recorded in the same commit.
• Structured data — checks markup against the rich results Google still supports and against what is visible on the page → markup describes only what a visitor can see, and no fix is sold on the promise that markup alone moves rankings.
• Speed — checks pages against Google's published Core Web Vitals thresholds: LCP within 2.5 s, INP within 200 ms and CLS at 0.1 or less, at the 75th percentile → the heaviest pages are trimmed to their own content, and speed is treated as a tiebreaker between equally relevant pages.

Authority & risk#

• Links — checks that every page is reachable through at least one link, and that link text names its destination → orphan pages get linked from the navigation or related pages, and "click here" becomes words that say where the link goes.
• Spam policy — checks for the violations Google names outright: templated pages that differ only by a name, near-duplicate variants built to rank, hidden text, integration lists copied from partners → each templated page gets at least one fact unique to its subject, and duplicate variants are merged or made distinct.
• Core updates — checks the whole site against what already ranks, not only the pages that dropped → after a ranking drop, pages that only restate other sources get original substance, and nobody promises a recovery date.

What answer engines require#

• Eligibility — checks for what actually keeps a page out of AI answers: nosnippet, a tight max-snippet, or Bing's NOCACHE → any of them on a page you want quoted is flagged, and no separate "AI version" of a page is written, because Google and Bing say none is needed.
• AI crawlers — checks which bots fetch your docs and what each one's token controls: model training, a search index, or a fetch a user asked for → blocking a bot is weighed by what that bot actually does, and a crawler is trusted by its published IP ranges, not by the name it sends.
• llms.txt — checks that llms.txt links to markdown copies of your pages and that robots.txt does not block it → the file stays a short index that points at clean markdown, and it is never sold as a way into Google's AI Overviews, which Google says need no such file.

Being quoted by models#

• Passages — checks that each section answers one question on its own: its subject named, its own figures, a stable heading anchor, no table split across headings → sections open by naming their subject and carry their numbers, because an answer engine lifts a passage, not a page.
• Content moves — checks the edits research has actually measured: topic, price, freshness and list position decide whether a page is cited at all, and keyword stuffing scores below doing nothing → pages state their topic, price and update date plainly, and rewrites that research found useless are not made.
• Measurement — checks how AI visibility is read: Search Console's AI report counts impressions only, and bot visits mean different things by purpose → reports keep "an AI bot crawled us" apart from "an answer cited us", and no citation rate is invented.

Page shape & structure#

• Page types — checks that each page does one job: tutorial, how-to, reference or explanation → a getting-started page stops being tutorial and how-to at once, and a how-to past about ten steps splits into sub-tasks. Most rules here are hypotheses, so they arrive as advice, not tasks.
• First screen — checks what a reader gets before deciding to stay: the answer in the first paragraph, the key facts within two screens, titles and headings that lead with the keyword → pages open with the answer, and filler openers like "A guide to" go.
• Navigation — checks that a page can be found once it exists: labels in the words readers use, shortcuts through deep trees, breadcrumbs, search suggestions that return results → vague or invented section names become familiar ones, and deep sections get links that skip straight down.

Readability for humans#

• Plain language — checks for sentences a reader gets through on the first pass, using readability scores as a warning sign, never as a target → tangled sentences are rewritten, without cutting the "because" a reader needed.
• Scanning — checks for more than two or three paragraphs with nothing to break them, sections without subheadings, headings that open with filler, and flat lists over seven items → text gets headings, lists and bold lead-ins at those intervals.
• Accessibility — checks the WCAG criteria a writer decides: alt text, heading order, link text, colour never the only signal, the page's language → images get alt text, headings stop skipping levels, and every link says where it goes.

Trust & evidence#

• E-E-A-T — checks who stands behind a page, and whether it is accurate and open about that → pages that touch money or personal data link to a findable About or contact page, and a self-interested comparison says so.
• Freshness — checks whether a page that was true still is, and whether its date tells the truth → pages about changed features are updated or removed, and a date changes only when the content does.
• Original data — checks that factual and performance claims name their source, and whether you publish data nobody else can → claims get their citation, and your own measurements become pages of their own.

Language & markets#

• Translation — checks that a translated page is translated in full, lives at its own address, links to its other language versions both ways, and is never forced on a visitor by location → no page ships a translated menu around an untranslated body.
• How developers read — checks against research on how developers use docs: examples that match real tasks, the answer findable mid-task, key facts not hidden on concept pages → every API operation gets a real usage example, and task pages carry both a copy-paste example and a link to the full reference.

The agent does not post on your behalf. On these axes it checks your pages and tells you what to do elsewhere.

Social distribution#

• What travels — checks the measured properties of content that gets shared, and keeps shares apart from links → a launch plans its follow-up posts, because a post gets most of its reach in its first two hours, and a share count is never read as a sign of backlinks.
• Per platform — checks each platform's own published rules for how a link is shown and ranked → your pages already ship the four required Open Graph tags with a 1200×630 preview, and posts follow each platform's rules, such as keeping a page's real title on Hacker News.

Communities & Q&A#

• Communities — checks where questions about your product are already asked, and whether search can see that place → a login-only Discourse forum is flagged, because nothing in it gets indexed, and tutorials stay in your docs, since Stack Overflow closes tutorial requests.
• The rules there — checks the published limits on promoting your own product in someone else's community → any post that mentions your product states your affiliation, and paid links carry nofollow or sponsored.