The Insights page at authgnosis.com/insights/ is live. It searches 222 records across the whole site: the 28 published articles, the 112 question and answer pairs that sit inside them, the five main pages, the 48 entries in the AI project log, and 29 individual sections of those pages. That last group is the change worth explaining. A search for capital funding readiness now returns that specific service block and opens on it, rather than returning the home page and leaving the reader to scroll for the part they asked about.

• A page is not one destination, it is several. The home page holds six service descriptions, a case study, and a set of frequently asked questions. Treating all of that as a single searchable record means the best a result can do is name the page, which puts the work of finding the actual answer back on the reader. Each addressable region is now its own record with its own heading and its own text, so it competes on what it says rather than on what the page around it says. The addresses were already there and already server rendered; what was missing was anything that indexed them.
• The rule for what counts as a destination is written down and generated, not curated. A page carries far more element identifiers than it has places worth linking to, because the page builder stamps one on every row, column and grid. A hand kept list of the useful ones goes stale the first time someone adds a section and forgets, which is exactly the failure it exists to catch. A destination is now defined by test: it is not framework plumbing, it does not name a structural role, and it has a heading inside it. Running that test against the live pages found a case study added days earlier that the hand kept list had never recorded.
• Adding the new records could not be allowed to move the existing results, so it was made structurally impossible rather than merely checked. The second group of records is scored in its own separate index and rendered in its own group beneath the main results, so it never competes with an article or an answer for position. The claim was then tested rather than asserted: the same thirty queries were run against the previous index and the new one, and the top twenty results were identical in both their order and their scores for every query.
• The harder half was making the links actually land. A link to a region of a page looks like it should just work, and it did not. These pages load their images lazily, so the blocks above the target grow after the browser has already made its one jump, leaving the reader somewhere near the section rather than on it. Six of the destinations are collapsed panels that open on click, so a result promising a sentence of description arrived on a closed heading. Both are now handled on arrival: the scroll re-asserts itself until the position stops moving, and a targeted panel opens itself. An index that is correct and a link that does not land is still a broken feature.