Most site search returns the document that contains an answer and leaves the reader to find it. The new Insights page at authgnosis.com/insights/ indexes 140 documents: the 28 published articles, and separately the 112 question and answer pairs that sit inside them. Because each pair is its own record rather than a field on its parent, a forty word answer can outrank a three thousand word article when it is the better response, and the result links directly into that answer on the article that carries it. The page also browses and filters without any of that: the article list is static HTML, and the search layer is added on top rather than replacing it.

• Indexing each question and answer pair separately is what makes a short answer competitive. A conventional index treats the pairs as fields on the article, which keeps every result article shaped: a query such as how do you charge returns the page, and the reader still has to locate the paragraph. Holding each pair as its own record changes the ranking arithmetic, because the scoring divides by document length and a short record is no longer diluted by the article around it. It also makes the deep link possible, since each pair carries its own anchor and the linked answer opens on arrival.
• Nothing is downloaded for search until the search box is touched. The index is roughly a quarter of a megabyte, and the large majority of visitors to a library page never search it. Loading it on arrival would charge every one of them for a feature few use, on a page that sits in the main navigation. Measured on the live page, opening it issues no request for the index and focusing the search box issues exactly one. A query typed while that download is still running is held and answered when it arrives.
• The browse layer is server rendered, and no search or filter state is written to the URL. Every article link is present in the page source before any script runs, and the eight topic filters hide entries with CSS rather than removing them, so what a crawler sees does not depend on which filters a reader selected. Keeping the query out of the URL is the same consideration in reverse: parameterized search and filter URLs generate large numbers of near duplicate crawlable pages, which dilutes indexing on a small site. The cost is that an individual search cannot be linked to, which is an acceptable trade at this size.
• Ranking was specified as ten checkable expectations before the page was built, and all ten hold on the live page. Each names a query and the result that should lead, including several that pass only if the word stemming and punctuation handling are right: a search for charges has to match charge, and a question about working in the U.S. has to survive the periods inside the acronym. Writing them in advance is what turns ranking from a matter of opinion into something that can be re-run after any change to the content or the scoring weights.