NavBoost remembers pages, Glue remembers the whole SERP
NavBoost ties each query to the documents clicked for it. Glue applies the same memory to every block of the page: images, videos, People Also Ask, local pack, AI Overview.
NavBoost
Query + documentFor each query–document pair, NavBoost keeps the history of its clicks, broken down by country or region and by device. It brings this accumulated experience into the ranking of web links.
total_long_clickslong clicks: the page answeredtotal_short_clicksshort clicks: a quick return to the SERPclicks_offclicks on another result of the page
Glue
The whole SERPGoogle itself calls Glue “just another name for Navboost that includes all of the other features on the page”. Each result carries 34 families of signals, from AI Overview interactions to image shares and app installs.
block_level_goodnesssatisfaction measured for the whole SERP block that holds the resulttotal_searches satisfied by a SERP blockquery_ event_ with_ good_ interaction_ on_ feature total_searches satisfied by a web linkquery_ event_ with_ good_ interaction_ on_ web
Public references: NavBoost described by Eric Lehman; Glue in the ruling of 2 September 2025; Glue’s role in presenting results.
Six families of signals describe a search
Clicks are one family among six. Each answers a concrete question, for a single result and for the query as a whole.
What did we choose?
Raw and weighted clicks, short, medium and long ones, moves to the next page.
clicks_onclicks on the resulttotal_long_clickslong clicks: the page answeredtotal_short_clicksshort clicks: a quick return to the SERPtotal_next_page_clicksmoves to the next page: the first one was not enough
What could we see?
Views, full views, time on screen and its share: visibility is measured in seconds and in pixels.
total_full_viewsimpressions seen in fullview_time_ontime spent on screenpct_view_time_onshare of the page’s viewing time spent on this resultheightheight of the block, in pixels
How did we explore?
Hovers, scrolls, hesitations and the “More results” button trace the path through the page.
total_hovershovers over the resultlog_time to the first scrolltime_ to_ first_ scroll total_back-and-forth scrolling: the searcher hesitatesback_ and_ forth_ scroll more_clicks on “More results”result_ button_ clicks
Did it help?
Good and bad clicks, satisfaction and dissatisfaction, for a result, its block or the query.
good_clicksgood clicks (the leak’s goodClicks)bad_clicksclicks judged unsatisfyinggoodnessshare of interactions judged goodbadnessmeasured dissatisfaction
What happened next?
The clicks and satisfaction that follow place each gesture in a longer search.
downstream_clicks on results afterwardsresult_ clicks downstream_best satisfaction reached afterwardsmax_ goodness downstream_num_eventsinteractions that followmax_depthdepth of the journey that follows
How did the question evolve?
Words added or removed, a new query on the same topic, a switch of tab, a query Google rewrote.
followed_a query retyped after an abandonment: the page did not answerby_ manual_ refinement followed_a reformulation on the same topicby_ semantic_ manual_ refinement toolbelt_mode_signalsswitches to tabs: Images, Videos, Newsfrom_searches that come from a query Google rewroterewritten_ query_ events
A link or a block: its exposure, the gestures on it and what follows.
The whole search: cumulative interactions, abandonments, the lowest position that still satisfies.
A direct answer can satisfy as soon as it is read; a complex search can take several visits. The unit Google measures is the whole experience, not the click.
A click is read against the chance to click
A link at the top gets immediate exposure; another, three screens down, depends on scrolling. Glue measures how long each result stays on screen, in seconds and in pixels, and relates its clicks to that exposure.
Two results, two levels of exposure
Fictional exampleAt the top of the page
Visible every time the page is shown.
Further down the page
Visible after scrolling, in some searches.
Among its on-screen views, B gets 12% clicks, compared with 6% for A. Visibility changes how these counts are interpreted.
Rate relative to appearances on the page
Rate relative to on-screen views
Position
For web links, satisfaction is corrected for position: a result placed low keeps the credit for what it delivers. Glue also records how deep satisfaction and attention go.
position_satisfaction corrected for positionunbiased_ query_ goodness lowest_position of the lowest result that satisfiesresult_ with_ good_ interactions_ position pct_view_time_belowshare of time spent further down
Presentation
Each block is described by what it offers: images, videos, height, click and swipe targets. Time on screen is divided by height, so that a two-line link and a carousel can be compared.
view_time_on_per_pixeltime on screen relative to heightnum_imagesimages in the blocknum_click_targetsclick targetsnum_swipe_targetsswipe targets
Ads
The ads above a result and the height they take are stored with its satisfaction: a result pushed below four ads is judged at its fair place. A separate set measures organic satisfaction during ads experiments.
num_ads_above_topads at the top of the pagevcap_height taken on screen, in pixelsrendered_ height_ px_ sum ads_the same measures during ads testsexp_ ad_ aware_ signals
A first organic position under four ads and a first position visible at once do not offer the same chance of a click. Glue records the difference and judges each result at its fair place.
Satisfaction is read over the whole journey
Time spent, a return to the SERP and a reformulation make sense together. Glue gives each everyday scenario its own counters.
An immediate answer
Opening hours or a match score satisfy on Google’s page. Glue counts these searches apart, so that an abandonment without a click is not read as a failure.
abandoned_abandonments because the SERP answeredfor_ webanswers followed_a reformulation after a SERP answer: it was not enoughby_ manual_ refinement_ for_ webanswers
A request becomes more specific
After the main steps, the searcher looks for a detail. When that refinement leads to a good result, the page before it gets a share of the credit.
downstream_satisfaction that follows, weighted by rankmrr_ goodness followed_a reformulation on the same topicby_ semantic_ manual_ refinement
A search that needs another result
The first result answered part of the question. The click that ends the search, the last and longest one, tells which page did the job.
clicks_offclicks on another result of the pagelast_longest_clicksthe last and longest click: the one that ended the search
goodness · badnessSatisfaction and dissatisfaction are measured separately, for the result, for its block and for the query, with variants restricted to the searches where the result shows.
block_level_goodnesssatisfaction of the whole blockgoodness_satisfaction, only over searches where this result showsfor_ query_ when_ result_ triggered
p_good · next_gen_p_goodp_good is the probability of a good click: a satisfied click, as opposed to a click followed by a return to the SERP. Two generations of these measures sit side by side on every result and every query: Google runs the new model next to the old one.
next_gen_goodnessthe new generation of satisfactionnum_searches whose satisfaction exceeds 0.3query_ events_ with_ goodness_ gt_ p3_ int
The leak’s goodClicks are one branch among 34. The click model the 2024 leak made famous, CrapsSignals, sits on each result next to satisfaction scores, journeys, ads and AI counters. Satisfaction also travels: backsmeared_goodness completes a result’s satisfaction with data borrowed from neighbouring searches, a technique Google calls smearing.
Every block of the SERP has its own counters
Glue does not stop at blue links. People Also Ask, images, apps, local actions, tabs and even missing query words have dedicated signals.
People Also Ask
Interactions inside the block and below it are counted apart: a PAA box that holds attention changes how the links under it are read.
explicit_interactions inside the blockinteractions_ in_ block explicit_interactions on it or below itinteractions_ on_ or_ below
Images
For an image, Glue counts the click to the page that hosts it, the good ones among them, shares and saves. In the image grid, clicks are corrected for position.
good_cthp_clicksgood clicks through to the host pagesavesaves of the imageposition_clicks corrected for the thumbnail’s placeunbiased_ query_ clicks
Apps
For an app result, Glue follows what happens after the click: installs, quick uninstalls, revenue.
acquisitioninstallsquick_uninstalluninstalls shortly afterrevenue_micro_usdrevenue, in millionths of a dollar
Local
Directions, calls, opening hours and popular times count as interactions with a local result, with or without a visit to the site.
num_redirect_to_maplinks to the map (directions)num_click_to_callcall buttonsnum_opening_hoursopening hours
Tabs
Switching to Images, Videos or News is a signal in itself: a query where people switch to Images calls for images on the main page.
image_mode_switch_rateshare of searches that switch to the Images tabtotal_satisfaction reached by switching to the tabmode_ switch_ goodness_ sum
Missing words
When a clicked result lacks one of the query’s words, Glue keeps the word with its impressions and clicks: the word was incidental to the need.
missing_unigramsquery words missing from the resultmissing_for each missing word, impressions and clicksunigram_ interactions_ map
Clicks count twice. They decide whether a block shows for a query, then which results rise within it. For [oak coffee table], the clicks of earlier searchers decide whether an image block appears, then which images lead it.
Each result is judged against the CTR its presentation predicts
Observed clicks are set against expected ones. A model predicts the click, the abandonment and the mere view of each block from its presentation; another predicts from the query alone whether the search will end without a click.
Observe
Clicks, time on screen, scrolling and the rest of the journey.
Contextualise
Query, result, block, country, device and period.
Aggregate
Statistics per result and per query.
Adjust
Normalise, correct for position, apply time decay.
Decide
Ranking, blocks shown and their order.
What was observed
Raw, weighted, normalised and position-corrected counts give several readings of the same events. 49 signals also keep a version without time decay.
weighted_click_countclicks, weightednormalized_satisfaction for the query, normalisedgoodness_ for_ query output_the option that keeps a version without decayundecayed_ version
What was expected
The gap between real and predicted clicks tells whether a result does better or worse than its format promises. A link without a thumbnail next to an image carousel is judged against the CTR its presentation led to expect.
clickpredicted probability of a clickabandonmentpredicted probability of abandonmentclicklesspredicted share of zero-click searchesctrpredicted CTR of the query
Google also holds blocks back on purpose. Next to the page it served, it keeps the block that should have shown and did not: the difference in clicks on the rest of the page measures the block’s own effect.
base_featurecounterfactual_featuretest_codes
A total also says how many people stand behind it. Glue statistics carry a count of distinct voter tokens, kept as a 256-bit bitmap. The 2024 leak describes it as a lower bound on the number of distinct users: Glue knows whether a thousand clicks come from a thousand people or from a handful.
distinct_voter_tokensdistinct contributors behind a statisticvoter_token_bitmapthe 256-bit set of those tokens
unsquashedClicks, in the 2024 leak, are the same clicks, before squashing. The leak says so: NavBoost keeps two copies of its click counters, one squashed, one unsquashed. Squashing flattens large values so that a spike of clicks does not dominate the other signals; the unsquashed copy keeps the real magnitude. These raw clicks, read in 2024 as reliable clicks resistant to manipulation, are the same measure on another scale. Glue does the same with time: 49 signals exist with and without decay.
Thirteen months of memory, and click spikes caught within minutes
A search about breaking news moves faster than a recipe. NavBoost keeps the long history; Instant Glue follows the last twenty-four hours.
NavBoost’s history
The window of interactions kept for each query–document pair.
Ruling of 5 August 2024 · Pandu Nayak interviewInstant Glue’s window
Glue’s signals from the last twenty-four hours, aggregated per locale.
Google document, 2021Instant Glue’s latency
The delay before a fresh interaction reaches the statistics.
Same documentInstant Glue ties recent statistics to the query, the documents and each locale, with a timestamp and a flag for rare queries. It spots, per locale, when a click spike starts and the share of recent clicks. On the day of the Nice attack, Instant Glue removed the image block from [nice pictures]: news had become the main intent.
fresh_start of the recent click spikeclicks_ peak_ start fresh_clicks_ratioshare of recent clicksnormalized_the query is rarequery_ is_ rare live_impressions of live sports resultsresult_ sports_ impressions
The date shown in the snippet matters too. For each broad intent, satisfaction is summed by the age of the displayed date: under a day, three days, a month, three months, a year. For each query, Google knows whether searchers are satisfied with yesterday’s pages or with year-old ones.
result_satisfaction on results dated less than a day oldgoodness_ sum_ byline_ age_ lt_ 1d result_… less than a month oldgoodness_ sum_ byline_ age_ lt_ 30d result_… less than a year oldgoodness_ sum_ byline_ age_ lt_ 365d result_… on results that show a dategoodness_ sum_ has_ byline_ age
Query volumes are also counted over 30, 90 and 180 days, and 49 signals keep a version without time decay next to the decayed one: past experience and recent behaviour each keep their weight.
AI Overview counters decide what Google sends first
magi is Google’s name for the AI in its results pages, the one behind AI Overviews and AI Mode. Glue counts its answers like any other block, for the query and for each result.
magi in Glue
From the answer showing to the next question
Expanded, collapsed, suppressed, declined, flagged as sensitive, and the reasons it did not show.
num_searches with the AI Overview expandedqueries_ with_ expanded_ magi num_searches where the AI declines to answerqueries_ with_ punt_ magi non_reasons it does not show, countedtriggering_ reason_ count
At the top or further down, its share of the time, and the scrolls back up to it.
num_AI Overview at the top of the pagequeries_ with_ magi_ at_ top percent_time_on_magishare of time spent on the AI answernum_scrolls back up to reread itviewport_ regressions_ on_ rendered_ magi_ block
“Generate”, “Show more”, and the opening of the sources that back the answer.
generate_button_clicksAI Overview on demandshow_clicks on “Show more”more_ button_ clicks corroboration_openings of the supporting sourcesexpansion_ clicks
Reformulations are counted with and without an AI Overview, and Search Labs testers apart from other searchers.
num_reformulations after an AI Overviewqueries_ with_ magi_ followed_ by_ add_ or_ delete_ manual_ refinement lab_the counters of Search Labs testersmagi_ query_ level_ signals
A built-in test: reformulating with and without AI
If searchers reformulate as often after an AI Overview as without one, the answer did not answer. The gap measures what the AI saves. The counters for searchers outside Search Labs also travel in Instant Glue: AI Overview behaviour is followed within minutes.
The AI decides what loads first. To save time, Google can send the first result before the rest of the page: First Result Out. It gives this up for about a hundred reasons, eleven of which name the AI. When many searchers expand the AI Overview on a query, or when the query asks to plan a trip, a weekend or meals, the AI comes before the first link.
INELIGIBLE_MAGI_ EXPANDED_ RATIO_ ABOVE_ THRESHOLD INELIGIBLE_MAGI_ TRIP_ PLANNING INELIGIBLE_MAGI_ WEEKEND_ PLANNING INELIGIBLE_MAGI_ MEAL_ IDEAS INELIGIBLE_AIO_ GLUE_ HIGH_ TRAD_ IMP INELIGIBLE_LATE_ BAILOUT_ MAGI_ ENABLED
The trial points the same way: Tangram, the engine that lays out the SERP, uses Glue to decide whether to show an AI Overview. Glue data meant for LLMs also travels with the query and with each result: click signals feed the models that write the answers. For publishers, visibility includes the experience around the AI answer: how it is expanded, which sources are opened, what people search next.
LlmGlueSignalData
Public reference: Plaintiffs’ remedies proposed findings of fact, §86.
Clicks vote for the meaning of a query
The same words can cover several needs. Clicks, tab switches and satisfaction are counted per interpretation and per broad intent.
An ambiguous query
“jaguar”
If the top pages and their clicks are about the car, the car interpretation comes first and brings its SERP blocks with it. This is where brands named like a common word win or lose their query.
Intent Brain goes one step further. For each query, language and country, it estimates about twenty broad intents, each measured by clicks: satisfaction on a result of the page, or after switching to the matching tab. [late-night pharmacy] calls for a map, [mullet haircut] for photos, [car registration fee] for a government site.
LOCALIMAGESHOPPINGSHORT_VIDEOUGC_DISCUSSIONFRESHINFORMATION_SOURCE_ GOVERNMENT INFORMATION_SOURCE_ MERCHANT
A query too rare for statistics of its own borrows the intents of its neighbours.
total_searches borrowed from similar queriesqueries_ from_ smearing intent_the intents those searches bringsmearing_ signal_ map
Clicks reach well beyond the ten blue links. They propose intents of their own, supply the alternative names of entities and brands, weigh on the terms that define a page, feed autocomplete suggestions and filter the Shopping catalogue.
KE_NAVBOOSTFRESH_NAVBOOSTSRC_NAVBOOSTENTITY_NAVBOOSTSUBTYPE_NAVBOOST_D2Q
How much a click score is trusted
In KScorer, which chooses between interpretations, clicks estimate the score quality raters would give (user_interaction_score), weighted by this confidence. On a rare query, clicks weigh little; on a frequent one, fully. Four of the six features that decide whether a card such as a knowledge panel shows come from these click scores; a fifth is the average score raters give knowledge panels.
What NavBoost and Glue reward
The page that ends the search weighs more than the page that draws a click. Three consequences for SEO.
Hunt down disappointment
Glue reads a short click, a quick return to the SERP, as a disappointment, and credits the last and longest click with the search. In your analytics, find the landing pages from Google that bring the most unqualified visits: sessions too short to even read the page. Judge each page against its own type, since some intents are resolved in ten seconds and others take a minute: the signal is the gap from the page’s average session length, not an absolute threshold. Then fix the welcome so that the visitor finds at once what the result promised.
Example. A buying guide whose sessions last 1 min 30 s on average, but where a third of Google visits end within ten seconds: the price and the comparison table sit three screens below the introduction.
Show a real date when freshness matters
Satisfaction is summed by the age of the date shown in the snippet. On a query that wants recent results, a page without a visible date drops out of these counters.
Example. A price comparison or a news explainer shows its update date, and the update is real.
Answer the next question
A refinement that leads to a good result gives part of the credit back to the page before it. Details, variants and links to the next step keep the journey going.
Example. After the general tutorial, the page offers the version for electric bikes.
A click is never counted alone
The context of a gesture
Time on screen, pixels, ads above and the block’s format: each click is read against the chance it had.
The whole journey
Abandonment, reformulation, the last and longest click, the satisfaction that follows: the unit is the search.
Every block, AI included
PAA, images, apps, tabs and AI Overviews have their counters, and those counters decide what Google sends first.
NavBoost and Glue turn billions of searches into a collective memory: what earlier searchers saw, skipped and kept shapes the page the next one gets.