trip.com

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.

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.

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.

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.

Time to first byte is how long your server thinks before it starts answering: the gap between an agent asking for the page and the first byte of HTML arriving. Yours crossed our threshold (under 200ms scores full marks, over 2 seconds fails).

Agents fetch with tight timeouts and many pages in parallel. A server that takes seconds to start answering gets cut off mid-queue, and a page that times out is a page that was never read. Speed here is not polish; it decides whether the content gets fetched at all.

1. Cache the HTML itself. Marketing and docs pages rarely change per visitor; serve them from a CDN or static cache so no application code runs per request.
2. If you are on serverless hosting, cold starts are a common cause: keep the function warm or pre-render the page.
3. Move slow work (database queries, third-party API calls) out of the first response; render the shell fast and fill personalised details after.

• Run: curl -o /dev/null -s -w "%{time_starttransfer}\n" https://yoursite.com and watch the number drop.
• Hit Re-scan above: "Time to first byte" improves a bucket at a time (under 500ms passes).

• Don’t answer fast with an empty shell that loads content by JavaScript. That trades this warning for the much worse "content requires JavaScript" failure, since most agents never run your scripts.
• Don’t special-case known crawler user agents onto a faster path. Serving different performance to different audiences is the pattern crawlers penalise when they detect it.