Web Design & Development

Sitewide Search: Learning What Visitors Couldn't Find

The challenge

adroll.com had no sitewide search. That left two problems: visitors who knew what they wanted but not where it lived had only the navigation (and on mobile, search is how many people navigate), and we had no way to see what visitors looked for and found, looked for and didn’t find, or found and didn’t click.

Every failed search is a visitor telling you, in their own words, what you haven’t published or can’t be found by. That information was never being collected.

Nobody asked me to fix this. I spotted it while working on other marketing site needs. I was supporting the marketing department’s web needs, and I wove this experimental work in between their requests for web updates.

The approach

Start with what already worked. Our CMS had a native Algolia integration, and it was already performing well on the blog. Expanding it was the lowest-risk way to get sitewide search before committing real resources.

Index what people look for. I audited the site and mapped in blog posts, resource hubs, product pages, news, webinars, case studies, and the glossary. I left out landing pages and SEM-specific tracks, which are built for campaigns and would have crowded results. Inclusion is managed entry by entry in the CMS, so the content and SEO teams can run it without a developer.

Design every state. The search interface has three, and I designed the layout of each.

The AdRoll site header with the search field focused and a panel open below it. A Recent Searches list (what is d2c, retargeting) and a Popular Searches list (dooh, testing, lifetime) sit beside Popular Search Results: six curated links to Pricing, Managed Services, Connected TV Advertising, Account-based Marketing, a Snowflake customer story, and the State of Digital Advertising Report. — Default, before anyone types: recent and popular searches beside six curated links to high-intent pages. It's the most-seen screen in any search interface, so I designed it as a routing tool, and curating it also let us point traffic at a new page or case study.

The search panel after typing what is d2c. A Recent Searches list shows what is d2c beside Results for what is d2c: six blog posts with matches highlighted and a See All Results button below. — Results: searches are saved to Recent Searches, and results are labeled by content type with matches highlighted, plus a link to the full results page.

The search panel after typing developers. It reads No results for developers, lists tips (double-check the spelling, use fewer words, search for a less specific term), and links to popular terms: Retargeting, Programmatic advertising, and Display advertising. — No results: tips and evergreen topics, so a failed search doesn't end in a dead end.

Read the logs. Analytics only work if someone is reading them. I analyzed the Algolia exports myself, and my goal was to own the reporting that would carry the findings to the content, revenue, and product marketing teams, so failed searches could become content briefs. That reporting, including automating it into the marketing team’s Slack channels, hadn’t started when my role ended in June 2026.

Reading the logs also turned up a bug. The search input was appending a trailing space to over 90% of queries, so Algolia treated "remarketing " and “remarketing” as different searches: the clean term returned roughly 25 times as many results as the spaced one, which also broke our synonyms and split the analytics. I wrote it up in a January 2026 report, and with Gemini’s help fixed it by trimming the query before it reaches Algolia.

The result

Thirteen weeks, March 31 to June 28, 2026, from the raw Algolia exports:

First-party search data the company didn’t have before.
46% of searches were empty queries, the default state. Nearly half of all search activity happened on the screen I designed before anyone typed.
95% of the tracked clicks from that state landed in the six curated links.
1.4% of searches returned no results.
Hundreds of distinct queries returned results but got no clicks. That’s the content-gap signal.

A query that’s searched often and never clicked is demand already proven against content that’s missing or can’t be found. Counts per query were modest, so I read them for patterns:

Brand unification: searched often, plenty of results returned, zero clicks, and no dedicated page.
Programmatic spend forecasts: repeated searches across several phrasings, results returned, no clicks. Visitors wanted a forecast we didn’t have.
Developers: “developers” and “tech.adroll.com” returned nothing, and “api” returned only a handful of results. That’s a developer audience with no entry point.
Intent mismatches: “D2C” returned plenty of results and no clicks, because the acronym didn’t match how our content was written. A customer case study got no clicks even when searched by its own name.

It also pointed the other way. Several queries that were searched repeatedly had a 100% click-through rate, including “contextual ad targeting” and “demand side platform”, which marked content strong enough to promote elsewhere.

The next steps

Where I planned to take it:

Start the reporting. Own a regular report for the content, revenue, and product marketing teams, and automate it by sending Algolia analytics into the marketing team’s Slack channels with our corporate Gemini pipeline.
Broaden what’s searchable. Index more content per entry so results match what visitors expect. Each page type stores its content differently.
Keep tuning synonyms. Keep matching the words visitors use interchangeably, like ctv, connected tv, and ott, or abm and account-based marketing.
Optimize the rest of Algolia. Check the six curated links against how each page performs, and add rules for high-intent queries that dead-end today, like routing “competitors” to a customer story.