How Google ranks: NavBoost, information gain and the honeymoon

What is NavBoost and how does click data affect ranking?

NavBoost is one of Google’s most important ranking systems, and for years its existence was only inferred until antitrust testimony and leaked documentation confirmed it by name. In plain terms, NavBoost re-ranks results using how real people have clicked on them, holding a long memory of that behaviour, on the order of thirteen months. It is the mechanism that lets the algorithm learn from the crowd: if searchers consistently skip the page sitting at position one and click the one at position four, NavBoost notices, and over time the popular page rises. Rankings are not set once and frozen; they are continuously adjusted by the aggregate verdict of the people who saw them.

The signal NavBoost cares about is not raw clicks but clicks relative to expectation, an idea often written as COEC, clicks over expected clicks. Every position on a results page has an expected click-through rate, higher at the top, lower as you descend. A page earns a boost when it beats the expected rate for the position it occupies, because outperforming your slot is evidence that the result is more compelling than its rank assumed. This reframes a metric most people treat as vanity: your title and meta description are not decoration, they are the pitch that decides whether you beat or miss the expected CTR for your position, and a result written like an ad by someone who understands the searcher’s intent will out-click a flatly descriptive one every time.

Two related behaviours feed the same system and are worth naming because they are within your control. Dwell time is how long a searcher stays on your page before returning to the results; a quick bounce straight back, called pogo-sticking, tells Google the click did not satisfy the query, while a long visit signals the opposite. The practical lesson is that NavBoost rewards pages built for humans, not for crawlers: a compelling title earns the click, and a genuinely useful page that answers the query fast keeps the visitor from bouncing. You cannot fake these signals at scale, because they are produced by real people, which is exactly why Google trusts them. They sit alongside the off-page authority signals as the parts of ranking that have to be earned rather than placed.

What is information-gain delta?

Information-gain delta is the measure of how much new information your page adds beyond what already ranks for the same query, and it is increasingly decisive. Google has described the underlying idea in a patent on the contextual estimation of link information gain, which models how much a document contributes relative to the ones a user has already seen. The practical consequence is blunt: if the top ten results all make the same five points and your page makes the same five points at greater length, your information-gain delta is roughly zero, and the algorithm has no reason to rank a longer copy of what already exists.

This is the antidote to the volume trap, the belief that a longer article automatically outranks a shorter one. Length is not gain. A 3,000-word page that restates the consensus adds nothing a 1,200-word page that contributes one original data point, one first-hand result or one genuinely missing explanation does not already beat. The honest test before publishing is a single question: what does this page know that the pages already ranking do not? If the answer cannot be stated in one sentence, the page has no delta and will struggle no matter how polished it is, and the fix is research or real experience, not more words.

The same delta governs whether AI answer engines cite you, which is why this concept recurs across the cluster. A model assembling an answer has no reason to quote a source that only echoes what it has already gathered from other pages; it quotes the one that contributes the specific number, the named method or the contrarian, well-argued take. This is the deep link between this algorithm concept and the internal-linking-and-information-gain guide: information gain decides whether a page deserves to rank or be cited, and internal linking decides whether that deserving page is found and understood as part of an authoritative cluster. Get either wrong and the other is wasted.

What are the honeymoon effect and the sandbox?

The honeymoon effect and the sandbox are two sides of how Google treats new content, and confusing them leads to bad decisions. The honeymoon effect is a deliberate test: when you publish a new page, Google often gives it a temporary visibility boost, placing it higher than its established authority would justify, to see how real searchers respond. It is a calibration, not a reward. If the page earns clicks and holds attention during the honeymoon, it tends to settle into a strong position; if searchers ignore or bounce off it, it drops back once the test ends. This is why the launch state of a page matters so much: it should go live in its strongest form, with the best title, the clearest content and the cleanest experience, because the honeymoon is the audition and a weak debut wastes it.

The sandbox is the opposing force, applied at the level of a whole young domain rather than a single page. A brand-new site with no track record is held in a probationary state where even good signals are discounted, because Google has no history to confirm the site is legitimate rather than spam built to rank fast. The sandbox is not a penalty; it is suspended judgement, and it lifts as the domain accumulates a steady, natural pattern of content, links and engagement over time. The mistake that deepens it is trying to force the issue with a sudden burst of pages or links, which reads as exactly the manipulation the sandbox exists to catch, the same unnatural-velocity problem behind the drip-feed launch discipline.

There is a third behaviour that unsettles people who do not know to expect it: ranking fluctuation that has no obvious cause. Google runs continuous experiments, sometimes modelled as contextual bandits, where it deliberately tries results in different positions to gather data on how they perform. A page that bounces between position five and position eight for a week is often not being punished; it is being tested. The disciplined response is to wait and watch dwell and engagement rather than panic and start changing the page, because reacting to an experiment as if it were a signal is how people break pages that were doing fine. Patience during calibration is itself a skill.

Why does a page stall at position 3 (local maximum)?

A page that climbs to position three and then sits there for months, refusing to break into the top two no matter how many small tweaks you make, is almost always stuck at a local maximum. The idea comes from optimisation: hill climbing finds the top of the hill it is on, but the hill it is on may not be the highest hill. Your page has been optimised as far as its current form allows, and the small adjustments, swapping a few words, adding a paragraph, tweaking a meta tag, only move it around the peak it has already reached. They cannot carry it to a higher peak, because reaching that peak requires a different hill, not a better position on the same one.

The reason a stall is so easy to misread is that small tweaks usually do produce small gains, so the instinct is to keep tweaking. But a months-long plateau at position three is a signal that the page has exhausted what incremental change can do, and continuing to fiddle is effort spent moving sideways. The pages ranking above you are not winning on a slightly better meta description; they are usually winning on something structural, a fundamentally more complete answer, a stronger information-gain delta, a better match for the searcher’s actual intent, or far greater authority pointing at them. Diagnosing which of those it is, rather than assuming it is a wording problem, is the whole job.

The fix for a true local maximum is structural, not cosmetic. Instead of editing the existing page again, you reconsider it: does it actually answer the query better than the pages above it, or does it need a different angle, more depth, original data, a format the top results lack? Sometimes the right move is a substantial rewrite; sometimes it is recognising that the query is better served by a different page in your cluster and adjusting your internal linking accordingly. The decision to stop tweaking and rebuild is uncomfortable because it admits the current page has hit its ceiling, but it is the only thing that moves a page off a plateau. Knowing the difference between a page that needs a tweak and a page that needs a new hill is what separates wasted effort from real progress.

Which ranking signals can you actually influence?

The most useful way to end is to separate the signals you control from the ones you only earn, because chasing the wrong ones is how time gets wasted. You directly control content quality and relevance, the largest lever by far, which means genuinely answering the query better than the alternatives and carrying a real information-gain delta. You control the click pitch, your title and meta description, which decide whether you beat the expected CTR that NavBoost watches. You control the on-page experience, the speed, stability and clarity that keep a visitor from pogo-sticking back to the results. And you control the technical health and internal linking that decide whether the page can be crawled, indexed and understood as part of an authoritative cluster at all.

What you do not control directly is the rest: NavBoost’s click verdict is earned from real searchers, not set by you; off-page authority is conferred by other sites; and the honeymoon, sandbox and experiments are Google’s to run. The honest implication is liberating rather than limiting. You cannot game the algorithm, but you do not need to, because every signal you can influence points the same way: make a genuinely better page for a real person, pitch it honestly so it earns the click, keep it fast and stable so the visitor stays, and wire it into a cluster so it is found. Do those, and the earned signals follow, because they are downstream of exactly the things you control.

This is also where honest automation helps, which is the forgehouse angle. An agent cannot manipulate NavBoost or shortcut the sandbox, and any tool that claims to is selling the manipulation Google built these systems to catch. What an agent can do is the monitoring that tells you which lever to pull: track where each page sits and whether it is climbing, stalling at a local maximum or fluctuating in an experiment; compare your click-through against the expected rate for your position to flag weak titles; and surface the pages whose information-gain delta is thin. It turns the invisible behaviour of the algorithm into a readable dashboard of where to spend the next hour. This whole discipline of reading the signals and acting on the right one sits inside the full AI SEO automation workflow.