Updated
Google's ranking factors are not a fixed list of 200. Google runs a chain of separate systems, and how much each one counts depends on the search. Sworn testimony from the US antitrust trial, the 2024 leak of Google's internal API documents and Google's own guides now let you name most of them. Sort them by who controls them and three groups fall out: things you set on your own site, things you earn from other people over months, and things only Google decides. Most ranking advice mixes the three up. Below, each one is graded by its evidence and paired with the action that moves it.
Paste one page from your site. I'll check ten things Google can read in its HTML and match each one to the systems on this page.
Leave your name and email and I'll send you a full review of this page and what I would fix first. I'll only use your email to reply about this.
Thanks, it's with me. I'll reply by email.
of click and query data that NavBoost remembers for every query, per Google's Pandu Nayak under oath [11]
top results RankBrain re-reads and may re-score after the first pass [11]
attributes across 2,596 modules in the leaked API docs, by Mike King's count, none with a weight [16]
the year the "200 signals" figure appeared, in a Google blog post by its chief economist [9]
The number comes from a Google blog post in March 2008. Hal Varian, then Google's chief economist, wrote that its information retrieval experts had "added more than 200 additional signals to the algorithms". It has been repeated ever since as if it were a checklist. [9]
Google's own explanation today reads differently. Its How Search Works page says: "The weight applied to each factor varies depending on the nature of your query." Freshness counts for more on a news search than on a dictionary definition. A fixed, ranked list cannot describe that. [2]
Nor is the leak a bigger list. Mike King counted 14,014 attributes in the leaked documents, but an attribute is a stored field, not a factor, and nothing in the documents says how much any of them counts. Notes of a 2025 call between the US Department of Justice and a Google engineer put it plainly: the leaked documents "don't go into specifics of the curves and thresholds". [16][14]
A better model is a chain of systems, each doing one job at one stage, each with its own evidence. Google's March 2024 core update post puts it plainly: "There's no longer one signal or system used to do this." [3]
I pulled the UK Google results for "google ranking factors" through DataForSEO twice on 12 November 2026 (desktop, English), and the results changed between the two checks. Both times the page was mostly list articles, and at least one snippet still said Google uses over 200 factors. In the second check, two of the nine titles still said 2024. Treat it as a snapshot, not a trend. [23]
Google describes three stages: crawling, indexing and serving. Court records split serving into several passes, so I use five gates. A page has to clear each one in order, and each has its own ways to drop you. Pick a gate.
Googlebot downloads the page, Google analyses it and stores it in its index. Google's own guide says "not all pages make it through each stage" and that indexing isn't guaranteed. A page that is not in the index is not in the race. [22]
Google matches the query to pages by words and by meaning. Notes of a DOJ call with a Google engineer name three base signals, Anchors, Body and Clicks, as "the key components of topicality". At the trial, Pandu Nayak agreed that RankEmbed, a deep learning model, adds "a few more documents" to those found by traditional retrieval. [14] [11] [2]
The candidates are scored for this query. The engineer's notes describe three high-level buckets: topicality, NavBoost and quality, with quality "largely related to the site rather than the query". The 2025 ruling describes PageRank as "a single signal relating to distance from a known good source" that feeds the quality score. Most of the leaked quality fields, such as siteAuthority, sit here. [14] [13] [10] [4]
Nayak testified that RankBrain "looks at the top 20 or 30 documents and may adjust their initial score", and that NavBoost "memorizes past clicks" for all queries from the past 13 months. The leak also refers to "twiddlers" that adjust results at serving time. [11] [10]
The final page mixes web results with maps, video and AI Overviews. Google generally shows no more than two listings from one site in its top results. It says AI Overviews "are built to only show information that is backed up by top web results". [1] [7] [8]
How Google ranks a page. Five gates. Each one can drop you.
1. Crawl and index. Google fetches your page and decides whether to store it. Not indexed means nothing else counts.
2. Retrieval. Each search pulls a short list of candidates. Your words must match what people type.
3. Scoring. Each candidate gets a relevance and quality score. Quality is largely about the whole site.
4. Re-ranking. RankBrain re-reads the top 20 or 30 results. NavBoost remembers 13 months of clicks.
5. Results page and AI answers. AI Overviews only show what top web results back up. A page that does not rank is unlikely to be quoted.
Three kinds of signal. You control. You influence. Google's side. Work on the first two.
Each one has the evidence behind it and the action that moves it. Google's internal names are shown where a source gives one. Filter by who controls it, or by the gate where it acts.
You control: settings and content on your own site. You can change it this week.
You influence: earned from other people over months, such as links, clicks, searches for your name and reviews.
Google's side: models and adjustments you cannot change. Knowing them stops you paying for them.
Showing 26 of 26
Nothing matches both filters. Set one of them back to All.
Google only ranks what it can fetch and store. Its own guide says not every page makes it through crawling and indexing.
What you do: Check the page in URL Inspection and clear any noindex, robots.txt or canonical problem.
Body (B)Google calls matching keywords "the most basic signal that information is relevant". The engineer's notes list Body, "terms in the document", as one of three base signals.
What you do: Use your customers' words in the title, the H1 and the first paragraph.
Passage rankingGoogle can judge one section of a page on its own, so a clear heading with a direct answer under it gets its own chance.
What you do: Give each follow-up question a heading and answer it in the first sentence.
titlematchScoreA site-level score described as "how well titles are matching user queries".
What you do: Write titles in the words people search, not clever ones nobody types.
OriginalContentScoreGoogle says it shows original content "ahead of those who merely cite it". The leaked field exists only on pages with little content, so it reads as a thin-page check.
What you do: Publish something only you have. Do not rewrite what already ranks.
contentEffortSince October 2026 Google's helpful content page tells raters to judge Effort, "the extent to which human work went into creating the content or the systems powering it". The leak has an "LLM-based effort estimation for article pages". Google has not said the two are linked.
What you do: Show the work: a test you ran, a calculation, a table or tool you built.
Google asks whether it is "self-evident to your visitors who authored your content" and calls made-up author profiles a form of deception.
What you do: Put a real, named author on advice pages, linked to a page that shows why they know.
SpamBrainPages made at scale to manipulate rankings break the scaled content abuse policy "no matter how it's created". Hosting third-party pages you do not oversee can break the site reputation policy.
What you do: Do not publish pages in bulk to catch keywords, written by AI or by people.
exactMatchDomainDemotionGoogle says it does not "give too much credit" to domains built to match a search. The leak stores a demotion for them.
What you do: Pick a brand domain. A keyword domain will not carry you.
clutterScoreGoogle says usability helps "when all other signals are relatively equal". The leak has a site-level penalty for "distracting/annoying resources". In Cyrus Shepard's 50-site study, a fixed footer ad was the feature most linked to losing traffic.
What you do: Remove pop-ups, sticky ads and anything between the reader and the answer.
lastSignificantUpdateGoogle's freshness systems favour newer content "for queries where it would be expected". The leak records a page's last significant update, not just its date.
What you do: Update fast-moving pages with real changes. A new date on an unchanged page is not an update.
siteFocusScore, siteRadiusTwo leaked fields measure how much a site sticks to one topic and how far each page strays from the rest. In Shepard's 2026 study of 400+ sites, tight topical focus went with traffic gains (correlation 0.250). That is an association, not a test of these fields.
What you do: Keep the site on what you sell. Move or remove pages far from it.
productReviewPUhqPageGoogle's reviews system rewards "insightful analysis and original research" by people who know the topic. The leak stores the chance a page is "a high quality review page".
What you do: If you review products, show you used them: photos, measurements, what went wrong.
PageRankGoogle says PageRank "continues to be part of our core ranking systems". The 2025 ruling quotes it as "distance from a known good source", used as an input to the quality score.
What you do: Earn links from sites in your field: suppliers, trade bodies, local press, partners.
anchorMismatchDemotionThe leak stores a demotion when the words in links pointing at a page do not match the page.
What you do: Let people link in their own words. Never buy keyword-rich links.
NavBoostNayak confirmed NavBoost memorises past clicks for all queries from the past 13 months. Google's How Search Works page says it uses "aggregated and anonymised interaction data". The leak names goodClicks, badClicks and lastLongestClicks.
What you do: Earn the click with an honest, specific title, then answer so the searcher stays.
siteAuthority, Q*The engineer's notes say quality is "largely related to the site rather than the query". The leak has siteAuthority, "applied in Qstar". Google says third-party authority scores "don't correspond to any of Google's own signals".
What you do: Raise the whole site, not one page: cut thin pages, earn mentions, keep the topic tight.
chromeInTotalTwo exhibits in the 2025 ruling suggest Google's popularity signal is based on "Chrome visit data" and "the number of anchors", though the judge said the court "can say no more". Mark Williams-Cook's research reported brand visibility feeding a site quality score.
What you do: Give people reasons to search for you by name: email, events, PR, customers who come back.
Google says local results are "mainly based on relevance, distance and popularity", and that more reviews and positive ratings can help.
What you do: Complete your Business Profile, ask every happy customer for a review and reply to each one.
Google describes it for news: how notable a source is for a topic or place, original reporting and source reputation. It has not said it applies to every page.
What you do: If you publish news, cover your patch consistently and break stories others cite.
FastSearchGoogle says AI Overviews only show information "backed up by top web results". The ruling says Gemini is grounded with FastSearch, which is "based on RankEmbed signals".
What you do: Rank first. Then give each key number its own short sentence that can be lifted whole.
RankBrain re-reads the top 20 or 30 results. DeepRank is BERT applied to ranking. Nayak agreed both are trained in part on click and query data.
What you do: Nothing to tune. Get into the top 30 and write for meaning.
RankEmbedDeep learning models that match queries and pages by meaning. The ruling says RankEmbed is trained on "1/100th of the data" used by earlier models.
What you do: Cover the topic in plain language. You cannot game a meaning model with repeated keywords.
Google "generally won't show more than two web page listings from the same site" in its top results. When pages are very similar, it shows "only the most relevant results to avoid unhelpful duplication".
What you do: Do not build several near-identical pages for one search. They compete with each other.
Google uses "your location, past Search history and Search settings" to shape results. NavBoost data is also split by locale and device.
What you do: Check rankings from where your customers are, on the device they use.
twiddlers, hostAgeThe leak refers to "twiddlers" that act at serving time, after scoring. The hostAge field, the date a site was first seen, is "used in twiddler to sandbox fresh spam in serving time".
What you do: Expect a new domain to be watched for spam. Publish real pages from day one.
Leave your name and email and I'll send you a full review of this page and what I would fix first. I'll only use your email to reply about this.
Thanks, it's with me. I'll reply by email.
Google's guides say what Google wants you to know. The leak shows what existed in an internal API, not what is used or how much. Court records are sworn testimony and a judge's findings. Independent tests show associations, not causes. A signal backed by more than one kind is the safest bet.
A filled dot means a source of that kind names the signal, and I opened it. Hover or focus a dot to see which source.
| Google docs | 2024 leak | Court record | Independent tests | |
|---|---|---|---|---|
| Original content | Yes | Yes | No | No |
| Effort | Yes | Yes | No | No |
| Spam policies | Yes | No | No | No |
| Exact match domains | Yes | Yes | No | No |
| Clutter and usability | Yes | Yes | No | Yes |
| Freshness | Yes | Yes | No | No |
| Topical focus | No | Yes | No | Yes |
| Links | Yes | Yes | Yes | No |
| Clicks | Yes | Yes | Yes | Yes |
| Site quality | Yes | Yes | Yes | Yes |
| Popularity and brand searches | No | Yes | Yes | Yes |
| AI Overview sources | Yes | No | Yes | No |
| RankBrain and DeepRank | Yes | No | Yes | No |
| RankEmbed and neural matching | Yes | No | Yes | No |
The leaked Content Warehouse documents are an API reference. They list fields and one-line descriptions. Google told The Verge the documents were real, and warned against "inaccurate assumptions about Search based on out-of-context, outdated, or incomplete information". [15]
Three limits matter when you act on them.
So use each source for what it is good at. The leak confirms that a system exists. Google's guides tell you what it rewards. Court testimony tells you how it works. When all three agree, as they do on links, clicks and site-level quality, you can act with confidence.
Google says the model behind AI Overviews "is integrated with our core web ranking systems" and that AI Overviews "are built to only show information that is backed up by top web results". [7]
The 2025 remedies ruling adds the plumbing. To ground its Gemini models, Google uses FastSearch, which "is based on RankEmbed signals" and returns a shorter, ranked list of web results. Google does not use FastSearch for its normal results page. [13]
So the systems above are also the entry ticket to AI answers. That can include the extra searches Google runs behind an answer, which helps explain how a page can be cited for a question it does not rank for directly. A page that cannot get retrieved and ranked is unlikely to be quoted. For what happens after that, see why a site is not in AI Overviews and how query fan-out turns one question into several searches.
The same systems weigh differently by search and by page. Pick yours for the four levers I would pull first. This is my reading of what each system rewards, not a list Google publishes.
Why: Google says local results are mainly based on relevance, distance and popularity.
Do this: Fill in every field, pick exact categories and add real photos.
Why: Google says more reviews and positive ratings can help local ranking.
Do this: Ask every happy customer for a review and reply to each one.
Why: The local help page says prominence takes in how many websites link to your business.
Do this: Get listed by your trade body, suppliers and local press.
Why: Body text is one of the three base signals.
Do this: One page per service, in your customers' words, with areas and prices.
Why: Google shows original content ahead of pages that merely copy it.
Do this: Write your own descriptions and use your own photos, not the manufacturer's.
Why: NavBoost remembers which results people chose for 13 months.
Do this: Put price, stock and the key spec in the title and description.
Why: The leak has a site-level penalty for distracting resources.
Do this: Cut pop-ups and sticky banners on product pages.
Why: Scaled content abuse covers pages made to catch searches, however they are made.
Do this: Do not auto-generate thin category or filter pages to target keywords.
Why: The leak measures how far each page strays from the site's topic.
Do this: Keep the blog on the problems your product solves.
Why: In Shepard's 2026 study, proprietary assets went with traffic gains (0.357).
Do this: Publish your own data, calculators, templates and demos.
Why: Court exhibits suggest popularity draws on Chrome visits and links.
Do this: Run the webinars, reports and PR that make people search for you.
Why: PageRank still feeds the quality score.
Do this: Get cited by partners, integrations and trade press.
Why: Google asks whether it is clear who wrote a page and why they know.
Do this: Put a qualified, named author on every advice page.
Why: Google's reliable information systems "help surface more authoritative pages and demote low-quality content".
Do this: Cite the rule, regulator or study behind each claim.
Why: Page quality is "largely related to the site rather than the query".
Do this: Remove thin pages that do not match your expertise.
Why: Rules, rates and thresholds change.
Do this: Review advice pages whenever the rules change, and say when you did.
Why: The leak's ContentAttributions record is "used during ranking to promote the attributed page".
Do this: Break your own stories and data, so others credit you.
Why: Google rewards news sources notable for a topic or place.
Do this: Cover your topic or patch every week, not now and then.
Why: Google's raters now judge how much human work went into a page.
Do this: Test, measure and photograph things yourself.
Why: The leak includes a "small personal site promotion" score.
Do this: Write in the first person about what you did and found.
Ten questions about one page and its site. Answer honestly. Nothing is saved or sent anywhere.
On 12 November 2026 I scanned all 157 pages in this site's sitemap. For each section I counted the pages that link to an outside source, carry a byline and show a visible date. Nobody else has this data, because it is my site.
The pages I write as articles do well. All 40 insight pages cite outside sources, 37 carry my byline and all 40 show a date. The commercial pages do not. Only 4 of 31 service pages cite a source, and none of the 20 location pages has a source, a byline or a visible date.
Off the site it is weaker again. DataForSEO counts 50 sites linking to this one, 13 of them with nofollow links, and it misses links it has not crawled. Search Console listed ten of my pages as crawled but not indexed on 9 November.
Which Google systems score these pages cannot be seen from outside Google. Matching a system to something visible on the site is my inference, not a measurement.
So I have two jobs: add sources and real examples to the service and location pages, and earn links and mentions. If your own check comes out the same way, the self-check above will say so: high on proof, low on reputation.
I built this page against the signals you can control. Here is the checklist, and where each one shows.
There is no fixed number. The figure of 200 comes from a 2008 Google blog post. The 2024 leak lists 14,014 attributes by Mike King's count, but those are stored fields, not weighted factors. Google's own How Search Works page says the weight of each factor varies with the query.
Yes. Google's Pandu Nayak testified in October 2023 that NavBoost memorises past clicks for all queries received in the past 13 months. The court's 2024 findings say the same, and Google's How Search Works page says it uses aggregated and anonymised interaction data to judge relevance.
Not the third-party scores. Google says those "don't correspond to any of Google's own signals". It does use site-wide signals, and the leak includes a field called siteAuthority. Notes of a DOJ call with a Google engineer say page quality is largely related to the site rather than the query.
No. The documents list fields and short descriptions, with no weights or thresholds. Some fields are marked deprecated. Google warned against assumptions based on "out-of-context, outdated, or incomplete information".
Yes. Google says AI Overviews are built to show only information backed up by top web results, and the 2025 remedies ruling says Gemini is grounded with FastSearch, which is based on RankEmbed ranking signals. A page has to be retrieved and ranked before it can be cited.
Independent UK SEO, GEO and AI consultant. In search since 2007. Portfolio includes 3M, BlackRock, Unilever and E.ON. Owner of SEO Moves Ltd since 2014.
More about Jason →I opened every source below on 12 November 2026. The labels and grouping on this page are mine. RESONEO's catalogue of Google's internal signal names prompted it and is worth reading alongside.
A fixed-price SEO audit checks these signals on your own site and hands you a plain English fix list, in order of impact.