Good Abandonment

What it is

A search engine records what happens after a query. The person clicks a result, rewrites the query, or does neither and leaves. The third outcome is called abandonment, and for a long time it was read as a negative signal: the results had nothing worth clicking, so the search failed.

Li, Huffman and Tokuda pointed out that this reading is sometimes wrong. If the results page itself contains what the person wanted, there is nothing left to click. The person has been served, and the log records a query with no follow-up. The authors defined good abandonment as a query abandoned without a click or a refinement because the results page already met the need.

The paper analysed abandoned queries on PC and mobile search across three locales. Its reported findings are of three kinds. Good abandonment makes up a large share of abandoned queries. It is significantly more common on mobile than on PC. And its main categories vary by locale and device, so the pattern is not one behaviour but a family of them. The authors drew a design conclusion from this: search engines could encourage more good abandonment, especially on mobile, by improving search features and result snippets.

The label does one job. It separates two events that share a signature in the data and have opposite meanings for the person. One is the searcher who got what they wanted and left. The other is the searcher who gave up. A count of queries with no click merges them.

In effect

The practical effect is on how a team reads its numbers. Suppose a product team improves the results page so that a common question is answered at the top. Fewer people click, and more queries end in abandonment. A dashboard that treats clicks as the measure of success records the improvement as a decline. The team that built the better answer is told, by its own metric, that it made the product worse.

That is our reading of the paper's conclusion, and the paper's own wording is narrower: abandonment should not be treated uniformly as a negative signal. The reading follows from it. Any metric that takes the absence of an action as a sign of failure will misjudge a product that works by removing the need for the action.

The case belongs in the account of Engagement Metrics because it shows the failure in its cleanest form. Mounia Lalmas, Heather O'Brien and Elad Yom-Tov describe web analytics as powerful at scale and as stripping away the user's motivation and context. Good abandonment is what that stripping costs. The log holds the behaviour, which is no click. It does not hold the reason, which is that the answer was already on the page, and the reason decides whether the behaviour is a success or a failure.

The sources read for this page report that the pattern is more common on mobile. They do not report why, and this page does not supply a cause.

The finding also runs in a useful direction for the other metrics. Time on page, return visits and click counts each have a good reading and a bad reading, and the data alone does not say which applies. Good abandonment is the documented case where the good reading was the common one.

What it does not say

It does not say that clicks are meaningless. In most searches a click on a result is what a successful search looks like, and the paper's point is that abandonment cannot be scored the same way in every case. It also does not say that all abandonment is good. The category is defined by the results page having met the need, and the opposite case, where the person gave up, remains real.

This page gives no figure for the share of abandoned queries that were good abandonment. Only the abstract-level description of the paper was read for the Why We Project, and it states the finding as a large share, not as a number. A reader who needs the figure should take it from the paper.

The data are from 2009, from three locales, and from one search engine's users. Nothing here shows that the same shares hold now, that the pattern is the same on other engines, or that it carries over to products that are not search. The inference to other metrics is ours and is labelled as such above.

The paper's recommendation, that search engines could encourage more good abandonment, is the authors' design suggestion. It is not a tested intervention in the sources read here.

---