grok.com

We gave an AI model the visible text of your page and asked it two questions: what does this page offer, and what is the price or main action? It could not answer from what your page serves.

This is the same task a shopping assistant performs before recommending you. If a model reading your page can’t name your offer and how to act on it, an assistant answering a customer can’t either, and it routes the recommendation somewhere it can.

1. State what you offer in plain text near the top of the page: a sentence a stranger could quote.
2. Put the price (with currency) or one clear primary action ("Start free trial", "Book a demo") in server-rendered text, not only inside images or scripts.
3. If the probe found your offer but no price or action, that is the half to add. The check’s finding above says which half was missing.

• Re-scan: "Primary action reachable" flips when the model can name both the offering and the price or call to action.
• Read your page’s first screen of text aloud. If it doesn’t say what you sell and what to do next, neither does the HTML.

• Don’t hide a text summary for bots that humans never see. The probe reads the same visible text extraction as every other check, and divergence between audiences reads as cloaking.
• Don’t bury the primary action under five equal-weight buttons; a model (and a customer) should be able to tell which action is THE one.

Of everything your server sends for this page, very little is actual readable text. The rest is code, styling, and markup wrapper. Agents fetched a lot of bytes and found few words.

Agents work with retrieval budgets. A page that is 2% text either gets skimmed (and mis-summarised) or skipped. More signal per byte means more of your actual message survives into the agent’s summary.

1. Confirm your real copy is server-rendered. See the JavaScript check; these two usually fail together.
2. Cut boilerplate wrappers: deeply nested divs, inline SVG logos repeated per section, base64 images inlined into HTML.
3. Move large inline styles and scripts into external files.
4. Check how much of the document is inline script: your report now says. If most of it is, that is your framework serializing state for hydration, often a second copy of text already in the HTML. Send only the props a component cannot recompute, and render statically where a page does not need to hydrate.

• Re-scan: "Text-to-markup ratio" improves, and "Page weight" often improves alongside.

• Don’t pad the page with keyword text to inflate the ratio. The ratio is a proxy for substance, and stuffing is the opposite of substance.

Headings (H1, H2, H3…) are the page’s outline. Yours skip levels, repeat H1s, or use styled text instead of real heading tags, so the outline machines extract is scrambled.

Agents use the heading tree to navigate long pages and to decide which section answers a question. A broken tree means quotes get attributed to the wrong section, or sections are missed entirely.

1. Exactly one H1 per page: the page’s actual title.
2. Nest without skipping: H2 for sections, H3 for subsections. Never jump H1 to H4 for visual size.
3. If a heading looks right but is a styled <div>, make it a real <h2> and adjust CSS. Machines only see tags.

• Re-scan: "Heading hierarchy" flips to PASS.

• Don’t pick heading levels for their font size; style with CSS and keep levels semantic.

Your page HAS structured data, but it is malformed: broken JSON syntax, or types and fields that don’t exist in the schema.org vocabulary. Machines can see the block but can’t parse it.

Invalid JSON-LD is worse than it looks. The agent spends its attention budget on the block, fails to parse it, and falls back to guessing from prose anyway. You pay the cost of structured data without getting the benefit.

1. Paste the failing page into validator.schema.org. It names the exact line and field that breaks.
2. The most common breaks: a trailing comma (invalid JSON), a typo’d type name, or a price written as "$79" instead of "79" with a separate priceCurrency.
3. Fix in place, redeploy, revalidate.

• validator.schema.org reports zero errors for the page.
• Re-scan: "Structured data validates" flips to PASS.

• Don’t delete the block to make the error disappear. That trades "invalid" for "absent" and fails the presence check instead; fix it.
• Don’t hand-edit generated JSON-LD in the page. Fix the template or plugin that generates it, or the error returns on the next publish.

HTTP responses can carry a Link header: the same canonical and alternate-language information your HTML declares, but delivered in the response envelope itself. Agents that only send a HEAD request, or that decide what to do while the page is still streaming, read the header without parsing any HTML. Yours is missing (or carries none of the relations agents use: canonical, alternate, describedby).

An agent triaging fifty URLs doesn’t want to download and parse fifty pages to learn which are duplicates of which. The Link header answers at the cheapest possible layer. Sites that provide it get correctly de-duplicated and correctly language-routed even by the most minimal fetchers.

1. Emit a Link header alongside each page, mirroring what your HTML head already declares. Example: Link: <https://yoursite.com/page>; rel="canonical".
2. Most stacks set this in one place: Next.js headers() config, an nginx add_header line, or a CDN response-header rule.
3. If you serve translations, add rel="alternate" entries with hreflang, matching your HTML.

• Run: curl -sI https://yoursite.com | grep -i "^link:" and confirm the relations appear.
• Hit Re-scan above: "Link response headers" flips to PASS.

• The header must AGREE with the HTML. A Link header pointing one place while the in-page canonical points another gives machines two contradictory answers, which is worse than one missing answer.
• Don’t inject the same site-wide canonical on every page via a blanket CDN rule; each page names its own clean URL, exactly as in the HTML tag.