Which evidence wins
Google’s documentation is reliable for product facts: which features exist, what the policies say, where the thresholds sit.
It is not the last word on what ranks.
Google long played down clicks.
Then a vice-president described NavBoost under oath in 2023, and the 2024 leak listed signals Google had denied.
So for ranking claims we weigh, in order: our own controlled test, court testimony, the leak, Google’s public word, outside studies, then opinion.
A before-and-after with nothing to compare against is not a controlled test.
A leaked field shows Google records something, not how much it counts.
How we decide, and how we stop changing our minds
A page gets changed, someone asks “are you sure?”, and it gets changed back.
These rules stop that loop.
Before changing a page, write down what should move, the number it starts from, what we compare it with, and what would count as failure.
Without a starting number, any result looks like a win.
We checked our own record: 68 SEO changes in seven weeks, and only 3 had a clean starting number.
Evidence: Our own data
A changed page stays untouched for 28 days. Day 14 is a safety check. Day 28 is the verdict.
Search data is noisy week to week, and two changes on one page make both unreadable.
Broken pages and wrong facts are fixed at once.
Evidence: Our judgement
If the result is unclear, extend once by 28 days, then close it with a written “keep” or “undo”. Time passing is not failure.
An inconclusive test is not proof that nothing happened.
Flipping back on a hunch is how the back-and-forth starts.
Evidence: Our judgement
“Are you sure?” means checking it a second way. If both ways agree, the answer is yes. A rule changes only on new evidence or a demonstrated error.
A second, independent calculation catches real mistakes.
It can still share a wrong assumption, so we say how each was done.
Evidence: Our judgement
Compare equal date windows, year on year and against the period before.
Unequal windows once gave us a wrong figure.
Year on year alone cannot isolate one edit either, because the whole market moves.
Evidence: Our own data
Before blaming a change for a drop, check that the change came before the drop.
We have seen page edits blamed for a slide that started months earlier.
Evidence: Our own data
What we judge a page by
Enquiries first, clicks second, judged by what the page is for.
A service page earns its keep in enquiries.
A guide earns it in clicks and usefulness, and never becomes a sales page.
Clicks alone misled us once: one page quadrupled during a haze spell.
Evidence: Our own dataOur judgement
Read Search Console by page and query together for any claim about a page.
Grouped by site, Search Console counts your best result once per search.
Grouped by page, it counts each page.
Mixing the two gives precise-looking numbers that mean nothing.
Source: Google: how Search Console counts (opens in a new tab)
Evidence: Google’s documentation
Before changing a page, look at who now ranks above it and what they offer.
That shows the gap.
It does not prove why you dropped, so it informs the change rather than justifying it.
Evidence: Our own data
Titles and page content
Write titles that are accurate, descriptive and concise, and that describe the same thing as the main heading.
Google may cut a long title to fit the screen.
It may also rewrite a title it finds unclear or mismatched.
Only the second replaces your words.
Source: Google: title links (opens in a new tab)
Evidence: Google’s documentation
Keep the main search’s words in the title, in any order. Treat about 60 characters as a guide, never as a reason to cut a search you already rank for.
Large studies find most top-ten titles contain the search.
On our pages, having the words helped and word order did not.
The leak lists a title-match score.
Still correlation, so we test it.
Source: Backlinko: 11.8 million results (opens in a new tab), iPullRank: leak analysis (opens in a new tab)
Evidence: Outside studyOur own data2024 leak
Do not chase exact word order, keywords in the web address, or a keyword in the first sentence.
We found no source for any of them, and no gain on our own pages.
Never change an established web address just to add a keyword.
Evidence: Our own data
No word-count targets. Answer the question first, put the depth below it.
Google lists writing to a word count “because you’ve heard or read that Google has a preferred word count” as a sign of content made for search engines.
Source: Google: helpful content (opens in a new tab)
Evidence: Google’s documentation
Earn the click, then end the search.
In 2023 a Google vice-president called NavBoost, built on click data, “one of the important signals”.
The leak lists good and bad clicks too.
Source: Trial transcript, 18 Oct 2023, pp. 6405 and 6472 (opens in a new tab), iPullRank: leak analysis (opens in a new tab)
Evidence: Court testimony2024 leak
One clear date per page, changed only when the content really changes.
The leak shows several page dates stored and compared.
Google also asks whether you change dates to look fresh when the content has not changed.
Source: Google: helpful content (opens in a new tab)
Evidence: 2024 leakGoogle’s documentation
When two of your pages show for one search
Two of your pages showing for one search is not a problem on its own.
Google’s John Mueller said in 2025 that several pages in one result is not a problem just because there is more than one.
Source: Search Engine Journal on Mueller, Sep 2025 (opens in a new tab)
Evidence: Google’s documentation
Act only with harm (lost clicks or enquiries) or a clear mismatch, such as a location page winning a search about children’s lessons.
Swapping places alone is not enough.
“Cannibalisation” tools flag dozens of these.
In our case many were sitelinks under our own top result.
Evidence: Our own dataOur judgement
A canonical tag says two pages are duplicates. A permanent redirect says the old page is gone. Neither is for two pages with different purposes.
Google describes both as strong signals, and says it may still choose its own canonical.
Source: Google: duplicate URLs (opens in a new tab)
Evidence: Google’s documentation
Links
To move a page from page two to page one, link to it from pages people actually visit, with varied, natural wording.
A study of 23 million internal links found traffic rose with link count and varied wording.
It is correlation, so we test a few pages against matched controls first.
Source: Zyppy: 23 million internal links (opens in a new tab)
Evidence: Outside study
Earn links from real organisations. Any paid or sponsored link is marked as such.
Google’s spam policies cover link schemes and require paid links to carry rel=”sponsored” or nofollow.
Source: Google: spam policies (opens in a new tab)
Evidence: Google’s documentation
Disavowing links is only for links that have caused, or are likely to cause, a manual action.
That is Google’s own condition for using the tool.
Routine disavowing is not.
Source: Google: disavow links (opens in a new tab)
Evidence: Google’s documentation
Arranged or fake brand mentions do not help. Real coverage does.
Google’s guidance on AI search says inauthentic mentions are not as helpful as they may seem.
Source: Google: AI optimisation guide (opens in a new tab)
Evidence: Google’s documentation
Local guide sites and location pages
Every local site or location page must be a real guide to that place, with facts the others do not have. Never a funnel to one sales page.
Google’s spam policy names “multiple domain names targeted at specific regions or cities that funnel users to one page” as doorway abuse.
Source: Google: spam policies (opens in a new tab)
Evidence: Google’s documentation
Place facts come from official sources. First-hand, dated observations are welcome alongside them.
Original information is what Google’s helpful-content guidance asks for, and it is what separates a guide from a copy.
Source: Google: helpful content (opens in a new tab)
Evidence: Google’s documentationOur judgement
Judge new sites over months, not weeks.
The leak describes a “hostAge” field used to hold back fresh spam.
Whether it slows honest new sites is not shown, so this is our caution, not a proven rule.
Source: Leaked field: hostAge (opens in a new tab)
Evidence: 2024 leakOur judgement
The whole site’s quality counts, so weak pages are improved or removed rather than left to drag on the rest.
The leak includes a site-wide authority score.
That shows site-level quality is recorded.
That pruning weak pages lifts the rest is our judgement.
Source: Leaked field: siteAuthority (opens in a new tab)
Evidence: 2024 leakOur judgement
Google Business Profiles and the map
The map results run on Google Business Profiles, so local effort goes into real profiles: categories, photos, hours, reviews.
An annual survey of local search specialists ranks the profile as the main factor.
Expert opinion, not an experiment.
Evidence: Outside study
Check eligibility before investing in a profile. Classes held at a venue you do not own or have authority to represent do not qualify.
Google rules out “an ongoing service, class, or meeting at a location that you don’t own or have the authority to represent”.
Breaches can suspend profiles.
Source: Google: Business Profile eligibility (opens in a new tab)
Evidence: Google’s documentation
Ask every customer for a review, neutrally. Never ask for five stars, never offer anything for one, never ask only the happy ones.
Google’s review policy forbids incentives and selective asking.
Source: Google: review policy (opens in a new tab)
Evidence: Google’s documentation
No review stars in your own business markup.
Google says pages using LocalBusiness or Organization markup about themselves are not eligible for review stars.
Source: Google: review snippets (opens in a new tab)
Evidence: Google’s documentation
Technical basics
Hygiene checks.
They rarely make a site climb, but failing them can hold it back.
Page speed is good at LCP 2.5 seconds or less, INP 200 ms or less and CLS 0.1 or less, for 75% of real visits.
Those are Google’s Core Web Vitals lines, measured separately for phone and desktop.
Lab scores find causes.
They do not give the verdict.
Source: Google: Core Web Vitals (opens in a new tab)
Evidence: Google’s documentation
Every page meant to rank loads, can be indexed, points its canonical at itself, sits in the sitemap and works on a phone.
Duplicates, redirects and utility pages each get their own treatment, outside the sitemap.
We check this list first whenever rankings drop.
Source: Google: Search Essentials (opens in a new tab)
Evidence: Google’s documentation
Keep the main text in the HTML, not only drawn by JavaScript.
Google renders JavaScript.
Most AI crawlers appear not to, so a script-built page can rank in Google and still look empty to them.
Source: Our GEO guide, JavaScript section
Evidence: Google’s documentationOutside study
When a page moves, redirect it to its real replacement. Where nothing replaces it, let it return 404 or 410.
Redirecting everything to the homepage throws away what the old page ranked for.
Google recommends keeping redirects for at least a year.
Source: Google: moving a site (opens in a new tab)
Evidence: Google’s documentation
Structured data and AI search
Structured data must match what is visible on the page. Expect little ranking lift from it.
Google’s structured data policies require it to describe the visible content.
Source: Google: structured data policies (opens in a new tab)
Evidence: Google’s documentation
Do not add FAQ or HowTo markup to win rich results.
Google stopped showing FAQ rich results on 7 May 2026 and removed the documentation on 15 June.
HowTo results went in 2023.
A visible FAQ is still fine if it helps readers.
Source: Google: Search changelog (opens in a new tab)
Evidence: Google’s documentation
Google’s AI Overviews and AI Mode need no special tricks. Ordinary good SEO is what they draw on.
Google’s own guidance says so.
We do no new llms.txt work.
Source: Google: AI features (opens in a new tab)
Evidence: Google’s documentationOutside study
Allow the search crawlers. Whether to block AI training crawlers is a business choice, not an SEO one.
Google-Extended controls Gemini training and grounding.
Google states it “does not impact a site’s inclusion in Google Search nor is it used as a ranking signal”.
Source: Google: Google-Extended (opens in a new tab)
Evidence: Google’s documentation
What we stopped doing
Merging our notes into one rulebook surfaced advice we had been carrying that was wrong or out of date.
We dropped it:
- Adding FAQ markup for rich results (Google removed them in 2026).
- Review stars in our own business markup (not eligible).
- An llms.txt file as a required build item.
- Disavowing links as routine housekeeping.
- Quotas such as “add two or three unique paragraphs”.
- Presenting a 2.0-second speed target as Google’s line. Google’s is 2.5.
- Treating every overlap between our own pages as a problem to fix.
- Changing web addresses to insert keywords.
Sources and method
We merged six internal SEO documents into this rulebook in October 2026.
They contradicted each other in 12 places.
Every rule labelled “Google’s documentation” was checked against Google’s current pages that month.
Two independent reviews then read the whole rulebook cold, and we fixed what they found.
The leak is the Google Content Warehouse API documentation published in May 2024.
Google confirmed it was genuine, while warning it was “out-of-context, outdated, or incomplete”.
The testimony is from United States v. Google, 18 October 2023.
“Our own data” means Search Console figures for the sites we run.
Found an error?
Send us the source and we will review the rule.