Companies spend a surprising share of their search budget on things no machine reads, and almost nothing on the things every machine reads first. This is a short account of the second category.
What gets read first
The response, before JavaScript runs. Crawlers and AI retrieval systems differ in how much client-side rendering they will wait for, and the ones that wait least are often the ones summarising you. Content that only exists after hydration is content you are gambling with. If a claim matters, it should be in the HTML that comes back from the server.
Headings, in order. Not for keyword reasons. Heading structure is how an
extractor decides which passage answers which question. A page with one h1 and
nine divs styled to look like headings is a page with one addressable idea.
The first sentence under each heading. Passage-level retrieval means the unit of citation is often a paragraph, not a page. A paragraph that begins with context and reaches its point in sentence four gets cited less than one that leads with the point. This is an editing habit more than a technical fix.
Explicit statements of fact. “PageRadar is an SEO and AI Search consultancy based in Lisbon” is extractable. “We’re a different kind of partner” is not extractable by anything, including humans.
Structured data, when it agrees with the page. Schema.org markup is useful because it removes ambiguity: what is a price, what is a person, what is a date. It stops being useful the moment it describes something the page does not say. Marked up claims that contradict visible content are worse than no markup.
What gets read next
- Dates. Published and modified, in a machine-readable format. Freshness is a ranking and citation input, and an undated page is treated as old.
- Authorship. A named author who exists elsewhere on the site, and ideally off it, is stronger than a company byline.
- Internal links. They tell a crawler which pages you consider important and how topics relate. A page nothing links to is a page you have told the system to ignore.
- External corroboration. What other sites say about you carries more weight than what you say about yourself, and it is the part most companies never work on.
What matters less than teams expect
Word count. Length correlates with depth, and depth is what is being measured. Adding words to a thin page produces a longer thin page.
Keyword density. It has not been a meaningful signal for a long time. Writing naturally about a subject produces the vocabulary the systems expect.
Meta descriptions as a ranking factor. They influence whether a result is clicked, which is worth real effort. They do not influence whether it ranks.
The test worth running
Take your most important page. Strip it to the text a machine would extract: headings, paragraphs, structured data. Read what is left.
If a stranger could not tell from that text what you sell, who it is for, and why it is worth more than the alternative, then no amount of technical work will fix what happens when a model is asked about you. The problem is upstream of the markup.