# Google rankings use clicks, but the leak gives no formula - Episode: P1E01 (Part 1: Modern SEO) - YouTube: https://www.youtube.com/watch?v=lOd05xvP9Vk - Page: https://irongoo.com/videos/google-ranking - Duration: 15:23 - Published: 2026-10-22T09:00:00+05:00 Google uses click data in rankings, but the leak does not reveal a formula you can copy. Learn how to find the stage blocking your page before spending weeks on the wrong SEO fix. ## What this episode covers - Seven stages from crawling and rendering to ranking, re-ranking and presentation. - What the leaked documentation and court record reveal about Navboost, links and site quality. - Why good clicks, bad clicks and satisfied searchers matter more than a raw click count. - How to use Search Console to check crawling, rendered HTML, indexing and canonicals. - What the evidence says about author reputation claims, new sites and human quality raters. - Why leaked fields reveal neither ranking weights nor a fixed checklist, and how to turn claims into testable hypotheses. ## Key moments - [00:00 Intro](https://irongoo.com/videos/google-ranking?t=0) - [00:51 The factory line](https://irongoo.com/videos/google-ranking?t=51) - [03:59 What Google stores](https://irongoo.com/videos/google-ranking?t=239) - [06:31 What was said in court](https://irongoo.com/videos/google-ranking?t=391) - [08:39 Said vs stored](https://irongoo.com/videos/google-ranking?t=519) - [09:43 One page, traced](https://irongoo.com/videos/google-ranking?t=583) - [12:18 What we still don't know](https://irongoo.com/videos/google-ranking?t=738) - [13:12 Three verdicts](https://irongoo.com/videos/google-ranking?t=792) ## Transcript ### Intro For years, people speaking for Google played down the idea that clicks feed rankings. Then a federal judge wrote, from sworn testimony, that one of Google's most powerful ranking systems learns from thirteen months of user clicks. In the same year, a big chunk of Google's internal documentation leaked. Two windows into the machine opened at once, and today we look through both. By the end you'll be able to point at any SEO problem and say which part of the machine it lives in, and you'll know which famous claims just died. We're also putting three claims on trial. Google ranks pages with two hundred fixed factors. The leak handed us the formula. Clicks matter. Only one of them survives, and even that one comes with a catch. ### The factory line Picture a factory line. A page goes in at one end, and every station along the way can knock it off the belt. Google's own documentation describes three stages, crawling, indexing and serving, and admits that not every page makes it through. We split that into seven stations, because each one breaks in its own way and needs its own fix. Station one is the crawl. Googlebot has to find your URL, be allowed in, and get a working response. Station two is rendering. Google says it runs your JavaScript in a recent version of Chrome. If your main content only appears after a script runs, and the script fails, you've built a very fast, very empty page. Station three is the index. Google checks whether your page duplicates another one and picks a single URL to keep as the canonical. Its documentation names the usual suspects for pages that never get in: low quality content, robots rules that block indexing, and site designs that make indexing hard. A page that isn't indexed can't rank for anything at all. Station four is retrieval. For every search, Google first grabs a shortlist of candidates from the index. The court record describes a model built for this, called RankEmbed, which puts queries and documents in the same mathematical space and looks for close neighbours. Miss the shortlist and ranking never even meets you. Station five is ranking. Notes from an interview with a Google ranking engineer, filed as evidence in the antitrust case, say relevance is built mainly from three raw signals, nicknamed A, B and C: anchors, meaning links pointing at the page, body, meaning the words on it, and clicks. On top sits a quality score called Q star, which the notes describe as largely static and tied to the site more than the query. Station six is re-ranking. Google's vice president of search described functions called twiddlers, which reshuffle results that have already been picked, and the leak is full of boosts and demotions like that. Station seven is presentation: snippets, rich results and AI answers. Google says a page has to be indexed and eligible for a snippet before its AI features can use it. There's no side entrance. Why start with stations? Because a lot of what lands on our desk labelled as a ranking problem actually lives at stations one to three. You can't fix an indexing problem with better copy, and you can't link-build your way past a broken render. So the first question is always the same. Which station is it? Here's a real one. A patent and trademark firm, in business since 1999, had its site rebuilt by another team with no redirects at all. Every old URL turned into a 404, and organic traffic slid for a long time. That was an index problem dressed up as a ranking problem. So we went to work at the right station. We rebuilt the site, redirected every old URL to its right new page, and fixed the technical basics. Organic traffic went from about four hundred visitors a month to nearly four thousand in under three months. ### What Google stores Now window one. In May 2024, internal documentation for Google's Content Warehouse API went public: two thousand five hundred and ninety six modules, fourteen thousand and fourteen attributes. Google confirmed the documents were genuine, and warned that they could be out of context, outdated or incomplete. Both halves of that statement are true, which is annoying for anyone who wanted a simple story. Start with clicks. A module tied to Navboost has fields called good clicks, bad clicks and last longest clicks, described as the clicks that were last and longest in related searches. Read those names again. Google does not treat a click as a click. A visit followed by a sprint back to the results is a very different event from the click that ended the search. Next, a sitewide authority score. A field called site authority sits among compressed quality signals, and the documentation says it is applied in Q star, the same quality score from the court notes. So yes, a sitewide quality measure exists. What goes into it, and what it weighs, is not in the documents, so anyone selling you its formula is selling fan fiction. Then the famous sandbox. A field called host age stores the earliest date Google saw any page on a domain, and the documentation says it is used to sandbox fresh spam at serving time. Notice the words fresh spam. This is a spam control, not proof that every new site sits in a penalty box for months. So for a new site, the advice is boring and correct. A new domain with a clean history, real content and a few genuine mentions has nothing to fear from a spam sandbox. A new domain that launches with thousands of thin pages on day one is exactly what that control was built to catch. Then focus. Fields called site focus score and site radius describe how much a site sticks to one topic, and how far each page drifts from the site as a whole. We put them to work in the topical map episode. There's also a field for site level views in Chrome, which lines up with the court notes describing a popularity signal built partly on Chrome data. And one thing that's missing. You may have read that the leak revealed an author reputation score. We looked, and we couldn't find one. There is a flag saying whether a recognised entity is the author of a document, which is a label, not a score. And when bylines were put through a proper controlled test, twice, they made no detectable difference. ### What was said in court Window two is the antitrust case the United States brought against Google. The judge's 2024 opinion, based on testimony, says Navboost pairs searches with documents by memorising user click data. It was trained on eighteen months of data before 2017 and thirteen months since. The court also noted that thirteen months of Google's data equals more than seventeen years of Bing's. Sorry, Bing. Those A B C signals come from interview notes filed in the remedies phase of the case. Anchors are links from one page to another. Body is the words in the document. Clicks are described as how long a user stayed on a page before bouncing back to the results. Together they make up topicality, Google's judgement of how relevant a page is to a query. The same notes say almost every signal, apart from the deep learning ones, is hand-crafted, so engineers can see what broke and fix it. They describe PageRank as distance from a known good source, feeding the quality score. Links didn't stop counting. They became one input among many, and we spend the whole of part two on them. What about quality raters? The court found that the RankEmbed models are trained on search logs plus scores from human raters. Google's rater guidelines say no single rating can directly move a page up or down. Both are true: raters don't rank your page, they teach the models that do. One more line, from notes of a call with Google's vice president of search, also filed with the court. Google uses over one hundred raw signals, and it avoids simply predicting clicks, because clicks are easy to manipulate and a poor proxy. Keep that sentence handy for the next person who offers to sell you clicks. The case also shows what Google treats as precious. In September 2025 the court ordered Google to share certain data with qualified competitors: the data behind a system called Glue, which includes Navboost data, and the data used to train RankEmbed. The models themselves stay private. When a court decides that interaction data is what rivals lack most, you know where the moat is. ### Said vs stored Now put both windows next to Google's public line. Google's representatives had played down clicks and Chrome data as ranking inputs, while the documents store them and the court record describes using them. On raters, no single rating moves a page, yet rater scores train ranking models. On factors, Google says hundreds, and the record says over a hundred raw signals rolled into a few top-level scores. Here's the interesting bit. Most of those public lines were technically true about something narrow. A single rater really can't push your page down. A raw click count really is noisy. The industry heard a narrow statement and stretched it into a broad one, so read every official line for exactly what it says, and not one word more. On our evidence ladder, the leak sits on the top rung and the court record right below it. Google's documentation sits further down: very reliable about what Google wants from you, less reliable as a description of the machine. When the rungs agree, build on it. When they disagree, the higher rung wins. ### One page, traced Enough theory. Let's push one real page down the line. It's a guide to running a trademark search, on the site of that same patent and trademark firm, with the domain blurred. The tool is URL Inspection in Search Console, and each part of its report maps onto a station. Crawl first: crawl allowed, yes, and page fetch, successful. Last crawled on the fourteenth of October 2026, by Googlebot smartphone. Google found the page through the sitemap and through a referring page, the English version of the same guide. Station one is green. Render next. Click test live URL, then view tested page, and you get the rendered HTML and a screenshot, which is the page as Google's renderer sees it. Now search that HTML for the sentence the page exists to deliver. If it isn't there, stop right there, because nothing further down the line can fix it. Then the index. Indexing allowed: yes. The canonical the site declared and the canonical Google picked are the same URL, so Google agrees about which version counts. Status: submitted and indexed. Three stations, three greens, and this is where URL Inspection runs out of road. Stations four to seven can't be inspected directly, but the Performance report shows what came out the other end. For the main Turkish query this page targets, a search for trademark registration lookups, it collected forty eight thousand nine hundred and three impressions between June and mid October 2026. This page is very much in the shortlist. So for this page, the next gains live at stations four to seven: how well it answers, how its title earns the click, and how it shows up in AI answers. That is a far shorter to-do list than redo the SEO. Diagnosis first, work second. That's the diagnosis map in four lines. Not crawled: look at robots rules, server responses and links. Not rendered: compare the raw and the rendered HTML. Not indexed: look at quality, duplicates and canonicals. Indexed but not winning: now, and only now, you're doing ranking work, and the printable version is free on the episode page. One habit worth stealing. Before you change a word on a page, check that Google sees the page you think it sees. In audits we keep meeting pages where the copy was fine and the real story was a canonical, a robots rule or an empty render. Ten minutes in URL Inspection beats weeks of rewriting the wrong thing. ### What we still don't know Now the honest part. The leak shows what Google stores, not what anything weighs. The engineer's notes filed in court say the documents don't go into the curves and thresholds. And nobody outside Google knows which of those fourteen thousand attributes are live today. So when someone says a leaked field proves you must do something to rank, ask them what that field weighs. They can't know, which makes the claim a hypothesis to test, not a rule. The documents also describe Google as it stood in 2024. Treat them as a strong snapshot, not a live manual. The same caution goes for the court record. It comes from a competition case, not an SEO masterclass, and engineers answered what lawyers asked. That makes it reliable on what it covers and silent on everything else, and we never fill the silence with guesses dressed up as testimony. ### Three verdicts Verdict time. Claim one: Google ranks pages with two hundred fixed factors, and if you tick them all off, you win. The two hundred figure goes back to a 2007 newspaper profile of Google's then head of ranking, and Google repeated it for years. Its documentation now says hundreds of factors. The court record describes over a hundred raw signals rolled into a few top-level scores, tuned by engineers, with a re-ranking layer on top. There is no fixed checklist to tick. Placebo. Claim two: the leak gave us Google's ranking formula, so now we know exactly what to optimise. The leak gave us the names and descriptions of fields. Not one weight, not one scoring function, and a Google engineer's notes filed in court say the documents don't cover the curves and thresholds. A list of ingredients is not a recipe. Ask anyone who has tried to bake from one. Placebo. Claim three: clicks matter for rankings. This one holds. The court record says Navboost memorises thirteen months of click data, and the leak stores good clicks, bad clicks and last longest clicks. But Google also says it avoids simply predicting clicks, because they're easy to manipulate. So clicks matter, just not the way click sellers mean: what counts is a searcher who found what they came for and stopped searching. Works. In practice it splits cleanly. Earn the click with an honest title, then answer so well that nobody needs to go back. Buying clicks means taking on a system with a field literally named bad clicks, run by people who told a court they don't trust clicks for exactly that reason. Here's the episode on one screen. Seven stations, and most problems solved by asking which one. A leak that shows what Google stores, a court record that shows how its engineers describe it, and public statements best read narrowly. Three claims, three verdicts. The which station is it diagnosis map is free as a one page download on the episode page at irongoo.com. Next time, we take the evidence ladder to the most repeated SEO advice and find out how much of it moves anything at all.