The Click Memory — how a click becomes the memory that ranks the next person

Google's ranking does not just read your page. It remembers what real users did the last time it showed them pages like yours. A US antitrust trial, a granted-patent family dating to 2006, and a leaked internal ranking API now converge on the same mechanism: a behavioural memory that turns clicks into rank. This is the Click Memory — and AI answers inherit it.

For most of the last decade, "do clicks affect rankings?" was the question Google answered most evasively. Representatives downplayed it; the SEO industry argued about it; nobody outside Mountain View could see the machinery. Then the machinery was entered into evidence. This piece does something the click debate never could before 2023: it builds the model from primary sources you can open yourself — court exhibits, sworn testimony, a federal opinion, granted patents, peer-reviewed click models, and the platform documentation — and it closes by drawing the honest line none of that evidence crosses.

L0Examination

Rank and snippet decide whether a result is even seen — attention falls top-to-bottom.

Cascade model · presentation-bias patent

Over-perform your position

L1The click

The click records perceived relevance — attractiveness, not yet satisfaction.

DBN click model · 'Life of a Click' exhibit

Win the click honestly

L2The long click

Dwell separates a good click that stays from a bad click that pogo-sticks back.

goodClicks / badClicks / lastLongestClicks

Satisfy on arrival

L3Re-rank memory

Clicks aggregate into a memorized table, keyed by query × document, sliced by locale and device.

Navboost (sworn testimony)

Be consistent across the slice

L4Decay

The table is a rolling ~13-month window — recent clicks count, aged ones fade.

13-month window · temporal-feedback patent

Refresh the deposit

L5Brand-demand rollup

Navigational, brand-seeking queries roll click behaviour up into a sitewide quality modifier.

Panda / site-quality patents

Manufacture brand demand

L6AI-answer inheritance

AI Overviews retrieve via core ranking, so the memory governs citation — but the answer layer stops feeding it.

'Rooted in core ranking' docs · click erosion

Be the source worth quoting

None of these layers is a keyword. The engine is not scoring your page in isolation — it is replaying what real users did the last time it showed them a set of pages like yours, then promoting the one they were satisfied with. Optimise for the click that gets kept, not the impression that gets ignored.

Framework: Martech LLC · synthesis of US v. Google testimony + exhibits, granted Google patents, the 2024 ranking-API documentation, and the click-model literature

Did the click debate actually end?

It did, and the place it ended was a courtroom. In United States v. Google, the government entered an internal Google training deck titled "Life of a Click." That exhibit lays out what Google internally calls the three pillars of ranking — body (what the document says about itself), anchors (what the web says about it), and user-interactions (what users say about it) — and notes that user-interactions include "clicks, attention on a result, swipes on carousels and entering a new query."[1] One of the three pillars is, in Google's own words, behaviour.

Then it was confirmed under oath. Google's VP of Search testified that the Navboost system memorizes past clicks for past queries, is trained on user data, and retains that click information for every query received in the prior 13 months.[2] And the court itself made it a finding of fact. Judge Amit Mehta's liability opinion concluded that ranking relies heavily on user click-and-query data and that "learning from this user feedback is perhaps the central way that web ranking has improved for 15 years."[3]

Learning from this user feedback is perhaps the central way that web ranking has improved for 15 years.

A 17-year Google search-quality engineer put the years of public hedging more bluntly in his own testimony, as reported by Search Engine Land: the debate was never whether clicks were used, but why anyone pretended otherwise.

Pretty much everyone knows we're using clicks in rankings.

This matters because everything downstream — the patents, the leak, the AI-answer question — is no longer a theory in search of confirmation. The confirmation came first, from the witness stand. The rest of this piece is the mechanism behind it.

What is the Click Memory?

The figure above is the whole argument in one diagram, so read it as a pipeline rather than a list. A search enters at the top as an examination — whether a result is even seen — and exits at the bottom as an entry in a memory that will shape the next person's results. Each layer is independently sourced, and each exposes exactly one lever an operator can actually pull. The crucial reframing: the engine is not scoring your page in isolation. It is replaying aggregate behaviour and promoting whatever satisfied people last time.

The reason this is worth a framework rather than a tip is that the layers interact. Winning a click you can't satisfy (L1 without L2) actively teaches the memory to prefer a rival. Earning satisfied clicks on desktop while neglecting mobile (L2 without L3 consistency) banks signal in a partition that may not be the one a given searcher draws from. The layers are a system, and the system has a direction of flow.

How does the system tell a good click from a bad one?

This is the layer everything else rests on, because not all clicks count the same. The distinction is between a click that is kept and a click that is regretted. A granted Google patent describes tracking click data and transforming it into a "click fraction" used to re-rank future results, explicitly separating long clicks — longer views that imply satisfaction — from short ones.[6] The 2024 leak gave that distinction its production vocabulary. The leaked Content Warehouse API documentation lists explicit click fields — "bad clicks, good clicks, last longest clicks, unsquashed clicks, and unsquashed last longest clicks."[4]

bad clicks, good clicks, last longest clicks, unsquashed clicks, and unsquashed last longest clicks are all considered

Good clickthe memory keeps it

1. 1 Query → your result chosen
2. 2 User stays · reads · scrolls
3. 3 No return to the results page

The session ends here. To the system this page answered the query — the click that lasted longest is the one it remembers.

Bad clickdiscounted, even reversed

1. 1 Query → your result chosen
2. 2 User bounces in seconds
3. 3 Pogo-sticks back · picks a rival

The return-to-results is the tell. A high click-through with a fast bounce is not a win — the next result down the page gets the credit instead.

This is why ranking for a term you don’t satisfy backfires: every regretted click teaches the table to prefer the page that did satisfy it. The behavioural layer rewards the answer, not the headline.

Mechanism: the long-dwell vs return-to-SERP distinction — leaked goodClicks/badClicks fields, the click-fraction patent, and Bing's own engagement documentation

Notice the phrase last longest click. The signal isn't merely that you were clicked; it's that yours was the click the user stayed on longest before the session ended — the result that resolved the query. This is not unique to Google. Microsoft openly documents that Bing uses engagement signals, asking directly whether "users spend time on these search results they clicked through or quickly return to Bing."[10]

The operator consequence is uncomfortable for anyone optimising for click-through rate alone: ranking for a query you don't satisfy is worse than not ranking for it. Every pogo-stick back to the results page is a vote for the next result down. The behavioural layer rewards the answer, not the headline.

Why does your snippet have to beat your rank?

Before any of that, a click has to be possible, and attention is not distributed evenly down the page. The foundational study here established that a "cascade" model — where users read results top to bottom and stop as soon as they find a worthwhile one — best explains why click probability falls with rank.[8] Lower positions are examined less, so they are clicked less, regardless of how good they are.

Rank 1

almost everyone looks

Rank 2

Rank 3

Rank 4

Rank 5

the scan is thinning

Because attention is biased by position, raw click-through can’t be trusted at face value — and Google’s own patents describe correcting for it. The operator consequence is blunt: you are debiased against at every rank below the top, so your title and snippet have to over-perform your position to earn the examination in the first place.

Mechanism: Craswell et al. cascade model (WSDM 2008); Google models and corrects this presentation bias before clicks are trusted (US8,938,463 B1)

Raw click-through, then, is contaminated by position. A serious ranking system has to correct for it before trusting clicks, and the literature is explicit about how. The Dynamic Bayesian Network click model separates a result's attractiveness — which drives the click — from its actual relevance, which governs whether the user stays satisfied afterward.[9] Google's own patents describe modelling presentation bias to recover an unbiased relevance estimate from clicks, and its researchers have published production-scale methods for estimating that bias from ordinary click logs without degrading the user experience.

For the operator, debiasing cuts both ways. You are corrected against at every rank below the top — the system knows you got fewer clicks partly because you sat lower, and it won't fully credit you for that. So your title and snippet have to over-perform your position: earn more examination and more satisfied clicks than your rank would predict, and you give the debiasing math a reason to lift you.

The memory has a window — and that is your opening

A memory that never forgot would ossify; the incumbent that won in 2015 would win forever. It doesn't, because the table is time-bounded. The same testimony that confirmed Navboost confirmed its horizon: clicks are retained on a rolling window of about 13 months. And a granted Google patent describes adjusting relevance by a "temporal element" of user feedback, down-weighting aged selections so that recent behaviour counts for more.[7]

A click is a deposit, and deposits decay. The page that stops earning satisfied clicks doesn't hold its rank — it slides out of the favourable memory while a fresher answer slides in.

This is the most actionable property in the whole model. Decay means the leaderboard is always partly open. A challenger who starts earning the long click now is writing into a window that will be the active one in a few months; an incumbent coasting on a page that no longer satisfies is watching its deposit age out. Refreshing a page is not housekeeping — it is re-depositing into the memory that ranks.

Does brand demand really move sitewide rank?

The layers so far operate per page. The last behavioural layer operates per site, and it runs on a kind of click most teams ignore: the navigational query, where someone searches for you by name. A granted Google patent computes a site quality score from user-demand signals — a ratio built from the count of reference, brand-seeking queries against the count of queries the site is merely associated with through result selection.[14] Co-invented by the engineer the "Panda" quality system is named after, it encodes a simple intuition: a site people seek out by name is, on the evidence of their behaviour, a site that satisfies.

The lever is the hardest and the most durable: manufacture genuine brand demand. Every person who searches your name instead of your category is casting the highest-value behavioural vote there is — and it rolls up into a sitewide modifier that lifts every page you publish, not just the one they were looking for.

What does this mean for AI answers?

Here the evidence gets more interesting and more honest at once. The behavioural memory was built for the ten blue links. Does it reach the AI answer? Google's documentation says the plumbing is shared. Google states that its generative AI features are "rooted in" its core Search ranking and quality systems, and that grounding retrieves pages through those same core systems.[11] If core ranking decides what gets retrieved, and behaviour shapes core ranking, then the Click Memory governs whether you are even eligible to be cited.

But the AI answer also starves the memory it inherits, and the data on that is stark. A Pew Research Center clickstream panel found that when an AI summary appeared, the rate of clicking a traditional result fell from 15% to 8%, and just 1% of users clicked a link inside the summary itself.[12] The longer trend is the same direction. On a tens-of-millions desktop panel, fewer than half of US Google searches in March 2025 ended in an organic click — 40.3% organic click against 27.2% ending with no click at all.[13]

Pew · click on a traditional result

No AI summary shown15%

AI summary shown8%

Link inside the AI summary1%

When an AI summary appears, the rate of clicking a real result falls from 15% to 8% — and barely 1% click the citations inside the summary.

Datos · how a US desktop search ends

Organic click40.3%

No clicks27.2%

Searched again17.1%

Stayed on Google14.3%

Paid click1.1%

Fewer than half of searches now end in an organic click — the raw material the click memory is built from is in slow, secular decline.

Here is the paradox the operator has to hold: AI answers inherit the click memory at retrieval, yet they stop feeding it — fewer outbound clicks means thinner, staler behavioural signal over time. The window to be remembered as the satisfying answer is closing, not widening.

Data: Pew Research Center clickstream panel (Jul 2025, 68,879 searches / 900 US adults); Datos × SparkToro State of Search Q1 2025 (US Google desktop, Mar 2025)

That is the paradox an operator has to hold: AI answers read from the click memory at retrieval but contribute almost nothing back to it. Academic work has started naming the disconnect — generative AI search collects only coarse feedback on the final answer, not the fine-grained per-document clicks that taught web ranking for two decades. The behavioural signal that historically self-corrected ranking is thinning exactly where the answers are moving. The window to be remembered as the satisfying source is closing, not widening.

What the evidence does NOT prove

A model this clean invites overreach, so here is the discipline. None of these sources publishes the weights. The first-hand analysis of the leak was emphatic that the documentation "doesn't show things like the weight of particular elements in the search ranking algorithm, nor does it prove which elements are used in the ranking systems."[5] The court redacted the formulas. The patents are methods, not proof of the live spec. And Google's public posture still omits the mechanism entirely: its official roster of named ranking systems lists BERT, RankBrain, and PageRank — and names no click or engagement system at all.[15]

Confirmed

Sworn / court / official

• Google's deck: clicks are 1 of 3 ranking pillars
• VP of Search, under oath: Navboost memorizes 13 months of clicks
• Federal opinion: user feedback is 'the central way' ranking improved for 15 years

Suggestive

Leak / patent — method, not live spec

• Leaked fields: goodClicks · badClicks · lastLongestClicks
• Patents: click-fraction, dwell-duration, temporal decay
• Shows the fields exist — not their weight

Mechanism

Academic — how, not whose engine

• Cascade & DBN: attractiveness vs satisfaction
• Unbiased LTR: debiasing clicks at scale
• Describes the math, not any one production system

The honest line every tier stops at: none of this publishes the weights.The court redacted the formulas, the leak shows field names but, in the discloser’s own words, “doesn’t show the weight of particular elements,” and the patents are methods, not the live spec. This is a defensible model of what is measured — not a claim about how much each signal is worth.

Ledger: Martech LLC · US v. Google (exhibits, testimony, liability opinion); granted Google patents; the 2024 ranking-API documentation; peer-reviewed click-model literature

Two honest caveats sit on top. First, the trial adjudicated core web ranking; the claim that the AI synthesis layer re-ranks on clicks is inferred from Google's "rooted in core ranking" documentation, not stated in the opinion. Second, the witness who confirmed Navboost also cautioned it is "a factor … by no means the only factor," and that many documents have no clicks at all. The Click Memory is a defensible model of what is measured. It is not a claim about how much each signal is worth — and any vendor who tells you they know the weights is selling you the part nobody outside Google can see.

Run one query through the Click Memory

Models are easiest to trust when you watch them run on something concrete. Take a single commercial query, a single page sitting at rank four, and walk it through all the layers — what each one tests, and the one lever that wins it.

The query

A buyer searches "best b2b attribution software." Your comparison page is one of ten candidates, sitting at rank 4.

Run the Click Memory

1. L0 · examination· you cover

   Your comparison page sits at rank 4 of the visible results.

   tests · are you even examined at this rank?

   What wins this layer

   A specific, benefit-loaded title + snippet that earns the look despite the position.

2. L1 · the click· you cover

   Against nine other titles, the searcher chooses yours.

   tests · does your promise beat the alternatives?

   What wins this layer

   Promise exactly what the page delivers — over-promising poisons the next layer.

3. L2 · the long click· you cover

   They land, find the answer above the fold, and don't bounce back.

   tests · is this the last, longest click of the session?

   What wins this layer

   Answer the intent on arrival so the visit ends here, not back on the results page.

4. L3 · re-rank memory· you absent

   Across thousands of similar queries, satisfied clicks accrue on your URL — per locale, per device.

   tests · is the pattern consistent enough to memorize?

   What wins this layer

   Ship the same satisfying result on mobile and in every market — the table is sliced.

5. L4 · decay· you absent

   Six months on, the query mix shifts and fresher pages start earning the clicks.

   tests · are you still the satisfying answer today?

   What wins this layer

   Refresh the page so it keeps earning the long click — the memory is a rolling window.

6. L5 · brand-demand rollup· you absent

   Buyers who already trust you start searching your name, not the category.

   tests · do navigational queries vouch for the whole site?

   What wins this layer

   Build real brand demand — branded searches roll up into a sitewide quality lift.

7. L6 · AI inheritance· you absent

   An AI Overview now answers the query and cites two sources.

   tests · did core ranking retrieve you as one of them?

   What wins this layer

   Win L0–L5 so core ranking retrieves you — then be the passage worth quoting verbatim.

3/7steps covered

The first three layers are a single session you can win this week — a sharper snippet and a page that actually answers. The last three are the moat: a pattern of satisfied clicks the system memorizes, keeps only while you keep earning it, and then hands to the AI answer at retrieval. That is why behavioural rank compounds for incumbents and decays the moment you coast.

Illustrative walkthrough of the documented mechanism · the citation of record for each layer stays inline in the prose above

The pattern the walkthrough exposes is the whole strategy in miniature. The first three layers are a single session you can win this week: a sharper snippet, an honest promise, an answer above the fold. The last layers are the moat — a pattern of satisfied clicks the system memorizes, keeps only while you keep earning it, and increasingly hands to the AI answer at retrieval. Behavioural rank compounds for whoever satisfies people consistently, and decays for whoever coasts. That is the entire game, and it has been the game, on the court's own finding, for fifteen years.