Clicks, Attention, Satisfaction: How User Behavior Verifies Rankings

Search engines have been candid about the reason. An internal presentation put it plainly: the ability to understand documents directly is minimal, so engines watch how people react to documents and memorize the responses. At web scale, behavior is the cheapest quality signal available.

"Today, our ability to understand documents directly is minimal. So we watch how people react to documents and memorize their responses."

— Eric Lehman, Google — internal presentation, reported in DOJ trial testimony

A vote can be bought and reflects nothing but volume. A verdict is different: it is conditional, continuously re-checked and reversible. A high position with bad behavior underneath is a countdown, not an asset — the same system that promoted a page will demote it when satisfaction is not sustained.

NavBoost is Google's confirmed, click-based re-ranking factor, described under oath by chief of ranking Pandu Nayak. It is a hybrid factor combining clicks, topicality and PageRank, and it aggregates behavior by country and by device. Its scope is candidly huge: it watches what a billion people do after they click.

Glue widens the lens from the ten blue links to the entire results page. It aggregates interaction across every module — knowledge panels, featured snippets, local packs — including hovers and scrolls, and feeds into how the results page itself is laid out. An impression that holds attention is data, even without a click.

• `goodClicks`: clicks that end the search, implying the need was met.
• `badClicks`: clicks that send the user straight back to the results page.
• `lastLongestClicks`: the click that finally satisfied the searcher.
• A 13-month rolling window: engagement is a rented asset, re-earned continuously.
• Country and device slices: behavior is aggregated per slice, not pooled globally.

The distinction matters because the metrics disagree. A click that ends a session is a satisfied need; a fast return is a failed prediction; the longest-held click reveals what the searcher settled for. Re-ranking weighs the pattern, not any single session.

The CAS (Clicks, Attention, and Satisfaction) model exists because modern results pages are non-linear. An attention model predicts the probability of a user examining an item from its rank, its type — Web, News, Weather, Currency, Knowledge Panel — and its geometry: offset, width, height. A click model then assumes a document must be examined and attractive before it is clicked.

The model's quiet achievement is that satisfaction can be predicted without sacrificing click prediction, because utility can accumulate from items the user never opened. Satisfaction may come from snippets, not just from clicked results — which leads directly to good abandonment.

Good abandonment is the confirmed positive case of the no-click search. The need is met on the results page itself — an answer panel, a snippet, a conversion module — and the user leaves satisfied. Mouse movements are a strong signal for identifying these sessions; a currency conversion query may involve almost no movement, yet the user reports full satisfaction.

For publishers, good abandonment reframes zero-click results. Content structured so engines can extract and cite it still earns credit when a module satisfies the searcher. Responsiveness means satisfying all possible needs behind a query, and the results page can be one of the satisfaction surfaces a well-structured document feeds.

Pages that delay the answer lose to pages that deliver it upfront, because responsiveness — satisfying all possible needs behind the query, not merely being topically relevant — is what the machinery rewards. Position tracking alone misleads: a stable rank with deteriorating engagement underneath is decay in progress.

Patience is part of the discipline. Ranking effects lag months, because current positions often reflect engagement from at least six months earlier inside a 13-month window. Bad history is not deleted; it is overwritten by accumulating stronger good signal. The practical loop is unglamorous: publish, observe, improve, repeat.