The VISIBLE-7 audit: how to check whether search engines and AI assistants can actually read your site - and fix it in an afternoon
For business owners, marketers, and developers. Every check below can be run in minutes with free tools, and most fixes are small. No SEO agency required.
Here is a thing that surprises almost everyone who checks: a website can look perfect in your browser and be empty to the machines that increasingly decide who finds you. Google's crawler, ChatGPT's browsing, Claude's web search, Perplexity - when they visit your site, they don't see what you see. They see the raw response your server sends. And on a large share of modern websites, that raw response is a blank page with a note that says, effectively, "run some code to find out what's here." Some machines run the code. Many don't, or do it late, or give up.
The consequence is quiet and expensive. Your beautiful site ranks below an uglier competitor whose text is actually readable. An AI assistant asked "who builds X in [your city]" recommends someone else - not because they're better, but because their site could be read. Your link, shared in a WhatsApp group, shows a gray box instead of a preview card, and nobody clicks.
I learned this the hands-on way, auditing and fixing several of my own and my clients' sites in a single week. Different platforms, same lesson each time: the browser view proves nothing. So this is the audit I now run on every site - seven areas, each with a test you can do right now, ordered by how much damage failing it causes.
One habit before we start, because it's the master key to the whole audit: view source, not the page. In any browser, right-click and choose "View page source" (not "Inspect"). That raw text is what machines receive. If your actual sentences - your services, your prices, your city - are in there, machines can read you. If you find a near-empty page with a few script tags, you have found the problem.
Area 1 - Rendering: is your content in the raw HTML at all?
The test: view source on your homepage and your most important pages. Search (Ctrl+F) for a full sentence you can see on the page - your headline, a service description. Found in the source: pass. Not found: your site depends on JavaScript rendering, and you are invisible or half-visible to a meaningful share of crawlers and AI systems.
A second opinion in one command, from any terminal: `curl -s https://yourdomain.com | grep "a phrase from your homepage"` Silence means fail.
Why it's first: everything else in this audit is decoration on top of this. Perfect metadata on an empty page is a label on an empty box.
The fix (for your developer): server-side rendering or prerendering - the site generates real HTML at build time or request time instead of assembling it in the visitor's browser. Every modern framework supports this (Next.js, Astro, Nuxt, plain static generation). This is usually a configuration-level change, not a rebuild - and it's the single highest-value fix in this document.
Area 2 - Titles and descriptions: does every page say what it is?
The test: in the source, check `<title>` and `<meta name="description">` on your five most important pages. Each page should have its own - specific, human-readable, mentioning what you actually do. Failing looks like: every page titled "Home", missing descriptions, or a template's default text.
Why it matters: the title is your search result's headline and the strongest single relevance signal you control. The description is your search-result ad copy - Google may rewrite it, but a good one usually survives.
The fix: one title and one description per page, written like you're labeling a shelf for a stranger: "AI applications and automations for businesses - Yellow House Digital" beats "Home". 150-160 characters for descriptions.
Area 3 - Link previews: what happens when someone shares your site?
The test: paste your URL into a WhatsApp message to yourself (don't send - just look at the preview). Then check the formal way: opengraph.xyz or the LinkedIn Post Inspector. You want a real image, your title, your description. A gray box or a random page fragment is a fail.
Why it matters: every share - WhatsApp, LinkedIn, Slack, iMessage - renders your Open Graph tags. A branded preview card can double the clicks on the identical link. This is the most visible machine-readability feature to normal humans, and the cheapest to fix.
The fix: Open Graph and Twitter Card meta tags on every page: `og:title`, `og:description`, `og:type`, and `og:image` pointing to a 1200x630 image served from your own domain (images hosted on external CDNs fail on some platforms' scrapers), with `og:image:width` and `og:image:height` declared - some lazy scrapers skip images whose size they'd have to download to learn. After fixing, force a refresh in LinkedIn's Post Inspector; scrapers cache aggressively.
Area 4 - Crawl permissions: are you accidentally telling machines to leave?
The test: open yourdomain.com/robots.txt. Three things to verify: it exists, it doesn't block everything (`Disallow: /` under `User-agent: *` is the classic staging-config-that-went-live accident), and it declares your sitemap. Then decide your AI policy consciously: many sites now block AI crawlers (GPTBot, ClaudeBot, PerplexityBot, CCBot) by default via their platform or CDN - sometimes without the owner knowing.
Why it matters: robots.txt is the front door sign. And the AI-crawler question is a real business decision: blocked means AI assistants can't read, cite, or recommend you. For a publisher monetizing content, blocking may be rational. For a business that wants to be found and recommended, blocking AI crawlers is turning away the fastest-growing referral channel on the internet.
The fix: a minimal robots.txt - allow all, name the sitemap - and an explicit line-item decision on AI crawlers. If you sell to humans who ask AI for recommendations (increasingly: everyone), allow them.
Area 5 - Sitemap and canonicals: can machines find everything, once?
The test: open yourdomain.com/sitemap.xml. It should exist and list every page that matters - including newer sections (a blog added last year is often missing from a sitemap generated at launch). Then view source on two pages and check each has a `<link rel="canonical">` pointing to itself in exactly one form (with or without trailing slash, www or not - one consistent choice).
Why it matters: the sitemap is your table of contents for crawlers; missing pages get found late or never. Canonicals prevent the same page existing at four URLs and splitting its ranking signals four ways.
The fix: generate the sitemap at build time so it can't go stale, submit it once in Google Search Console (free, ten minutes, and the single best "make Google notice my changes" lever), and emit a canonical on every page.
Area 6 - Structured data: does your site introduce itself formally?
The test: paste your homepage URL into Google's Rich Results Test or validator.schema.org. You're looking for JSON-LD blocks: Organization or Person (who you are), WebSite, and - if you publish - Article on each post with headline, date, and author.
Why it matters: structured data is your site's business card in a format machines parse without guessing. It feeds Google's knowledge panels and rich results, and it's exactly the kind of unambiguous signal AI systems lean on when deciding whether you are a real, specific entity worth citing.
The fix: a small JSON-LD block per page type. For a business site: one Organization/Person block sitewide, Article on each post. Any developer can add this in an hour; the schema.org examples are copy-paste-adaptable.
Area 7 - The AI layer: llms.txt and content built for machines that read
The newest area, and the one where being early still stands out.
The test: does yourdomain.com/llms.txt exist? Beyond that, read your key page the way a machine would: is what you do stated in plain sentences ("I build X for Y") - or implied through vibes, hero images, and three-word fragments? Machines don't infer from vibes.
Why it matters: llms.txt is an emerging convention - a plain markdown file at your root that tells AI systems what the site is, who's behind it, and where the important pages are. Adoption is young and its weight is still being proven, but it costs fifteen minutes, and its audience is the systems increasingly answering your customers' questions. Beyond the file, the durable principle: write so a machine quoting one paragraph of your page would represent you correctly. Clear headings, real sentences, one topic per page, your location and offer stated in text rather than trapped in images.
The fix: an llms.txt with a one-paragraph summary, a linked list of key pages with one-line descriptions, and contact info. Then a pass over your core pages replacing implication with statement.
The scorecard
Run all seven tests; score each pass, partial, or fail:
- Rendering - your sentences appear in view-source.
- Titles and descriptions - unique, specific title and description per page.
- Link previews - WhatsApp and LinkedIn show a branded card.
- Crawl permissions - robots.txt exists, allows, names sitemap, AI policy chosen.
- Sitemap and canonicals - sitemap.xml complete; one canonical URL form.
- Structured data - JSON-LD validates for who-you-are and articles.
- AI layer - llms.txt exists; key pages state, not imply.
Fix order if you failed several: Area 1 first, always - nothing else counts until the content is readable. Then 3 (visible to humans immediately), then 2 and 5 together (they share a deploy), then 4, 6, 7 in one afternoon sweep. Total effort for a typical small business site: one to two developer-days, most of it Area 1. Everything else combined is an afternoon.
One closing reframe. This is not "SEO tricks." Every fix above is the same move: taking what your site already says and making it legible to every kind of reader - human, crawler, or model. The businesses that did this for Google over the past twenty years got found. The ones doing it for AI systems now will be the ones those systems recommend. The window where this is a differentiator instead of table stakes is open today; it will not be open long.