You will have seen this claim: 91% of pages ranking first pass Google’s speed checks.
It is true. It is also the most misused statistic in this field.
It does not mean speed put those pages first.
Well-run sites tend to be fast and to rank well. Both follow from the same care.
So this article separates two kinds of evidence.
Correlations, which suggest. And controlled tests, which demonstrate.
The controlled tests are strong enough that you do not need the correlations. ⚡
🎁 Compare Hosting and Site Tool Trials →
🧾 Principaux résultats en bref
| Mesure | Chiffre | Evidence type |
|---|---|---|
| Rakuten 24 revenue per visitor | Up 53.37% | Controlled A/B test |
| Rakuten 24 conversion rate | Up 33.13% | Controlled A/B test |
| Vodafone Italy sales | Up 8% | Controlled test |
| RedBus sales after INP work | Up 7% | Controlled test |
| Sites passing all three vitals | 33% | Survey |
| Mobile sites passing all three | 42% | Survey |
| Position-one pages passing | 91% | Correlation only |
| Good LCP threshold, March 2026 | Lowered to 2.0 seconds | Google change |
🔬 The Evidence That Actually Proves It
Three companies ran controlled experiments and published the results.
A controlled test compares two versions with real users, changing one thing.
That design can show cause, which no survey can.
🔬 Measured results from controlled tests
Rakuten 24
Improving largest contentful paint raised revenue per visitor by 53.37%.
Conversion rate rose 33.13% in the same test (Digital Applied, 2026).
This is the strongest published evidence in the field.
Vodafone Italy
A 31% improvement in largest contentful paint produced 8% more sales.
Smaller than Rakuten, and from a very different starting point.
Both directions agree, which is what matters.
RedBus
Work on interaction responsiveness produced a 7% sales increase (MigrateLab, 2026).
That metric measures how quickly a page responds when tapped.
It is the one most sites neglect.
Why these three carry weight
Each compared against a control group at the same moment.
Seasonal effects, marketing campaigns and traffic mix affected both versions equally.
That design removes the usual explanations for a change in sales.
⚠️ The Statistic You Should Stop Repeating
Now the other kind of evidence, and its limits.
91% of pages in position one pass all three vitals.
Only 47% of page-two pages do (Bloggersideas, 2026).
That gap looks decisive. It is not.
| What the gap could mean | Plausible? |
|---|---|
| Speed causes the ranking | Partly, it is a minor factor |
| Well-resourced sites do both well | Very likely |
| Popular pages get cached and optimised | Likely |
| Better sites hire better developers | Very likely |
| High traffic funds performance work | Likely |
Four of five explanations have nothing to do with speed causing rankings.
Google has been explicit that speed is a small ranking factor.
Why the misuse matters
Sites spend heavily chasing a ranking effect that is modest.
Meanwhile the genuine effect, on conversion, gets less attention.
You are optimising for revenue, not for a score.
The correct framing
Speed work pays for itself through conversion, not through rankings.
Any ranking benefit is a bonus rather than the business case.
The controlled tests above justify the spend on their own.
🔎 Test Hosting Free Before Switching →
💰 The Numbers Behind the Headlines
Several widely quoted figures deserve context before you use them.
| Claim | Status |
|---|---|
| 1 second delay costs 7% of conversions | Widely cited rule of thumb |
| 100ms costs about 1% of conversions | Consistent across studies |
| 53% of mobile users leave slow sites | Survey based |
| Amazon loses $4.8bn per 100ms | An extrapolation, not a measurement |
| Retailers lose $2.6bn annually | Modelled estimate |
The Amazon figure appears everywhere and is worth examining.
It applies an old internal finding to a current revenue number.
Nobody at Amazon measured $4.8 billion. It was calculated afterwards.
Why that distinction matters
An extrapolation assumes the original relationship still holds at today’s scale.
That assumption may be reasonable. It is still an assumption.
Use the number as illustration, never as evidence.
The figure that does hold up
Roughly 1% of conversions per 100 milliseconds recurs across independent studies (Colorlib, 2026).
It is modest, consistent and easy to apply to your own numbers.
That makes it far more useful than the dramatic ones.
🧮 Working Out Your Own Number
Generic statistics do not tell you what to spend. This calculation does.
| Étape | Example |
|---|---|
| Annual online revenue | $200,000 |
| Current load time | 4.2 seconds |
| Realistic target | 2.0 seconds |
| Amélioration | 2,200 milliseconds |
| At 1% per 100ms | 22% theoretical uplift |
| Applying a cautious half | About 11% |
| Estimated annual gain | Around $22,000 |
The halving is deliberate, because the relationship is not linear at every point.
Even halved, the figure justifies serious investment.
Why to be conservative
Gains are largest when moving from very slow to moderately fast.
Going from 2.0 seconds to 1.5 seconds returns less than 4.0 to 3.5.
Early improvements are cheap. Later ones are not.
Where to stop
Reaching the good thresholds is worth doing.
Chasing a perfect score rarely pays for the engineering time.
Diminishing returns arrive sooner than most teams expect.
📉 Most Sites Are Failing
The pass rates put the opportunity in perspective.
📉 How many sites pass each check
Only 33% of sites pass all three checks.
The weakest is main content load time, at 58%.
The competitive read
Two thirds of your competitors fail at least one check.
This is an unusually easy area in which to be above average.
Most improvements are configuration rather than rebuilding.
Progress is slow
The pass rate moved from 22% in 2021 to 33% in 2026.
Eleven points in five years is not rapid.
The advantage available here will persist for some time.
📅 What Changed in March 2026
Google tightened one threshold, and it moved the goalposts for many sites.
| Seuil | Avant | From March 2026 |
|---|---|---|
| Good main content load | 2.5 seconds | 2.0 seconds |
Sites sitting between 2.0 and 2.5 seconds moved from passing to failing.
They did not get slower. The standard got stricter (Idea Fueled, 2026).
Que faire face à cela
Re-measure rather than assuming last year’s result still applies.
A site that passed in 2025 may not pass now.
Check the current threshold before celebrating an old score.
The direction of travel
Thresholds have tightened over time and will likely continue to.
Build with some margin rather than sitting just inside the limit.
Aiming comfortably below the threshold protects against the next change.
🛠️ What Actually Makes Sites Faster
Ranked roughly by effect per hour of work.
| Réparer | Effort | Typical effect |
|---|---|---|
| Compress and resize images | Faible | Often the largest single gain |
| Use modern image formats | Faible | Substantial |
| Enable caching | Faible | Substantial |
| Remove unused plugins and scripts | Faible | Moderate to large |
| Use a content delivery network | Modéré | Large for distant visitors |
| Reduce third-party tracking | Modéré | Often underestimated |
| Better hosting | Modéré | Large if currently poor |
| Rebuild the front end | Très haut | Large, rarely necessary |
The first four are configuration changes, not development projects.
Most sites can reach passing scores without rebuilding anything.
Images first, always
Oversized images are the most common cause of slow pages.
A photograph uploaded straight from a phone may be ten times larger than needed.
Fixing that alone often moves a failing site into passing range.
The third-party script problem
Analytics, chat widgets, advertising tags and social embeds all cost time.
Each was added for a reason, and each reason is rarely revisited.
List them, then remove the ones nobody uses.
📱 Mobile Is Where the Gap Sits
Only 42% of mobile sites pass all three checks.
Mobile devices are slower, and connections are less reliable.
| Facteur | Ordinateur de bureau | Mobile |
|---|---|---|
| Processing power | Haut | Beaucoup plus bas |
| Connection stability | Usually good | Variable |
| Screen size | Grand | Small, images still load |
| User patience | Modéré | Inférieur |
53% of mobile users abandon sites that load slowly (Colorlib, 2026).
Test on a real phone, not on a desktop simulation.
The common mistake
Developers test on fast machines with fast connections.
The site feels fine, and the data says otherwise.
Borrow an older phone and try it there.
⏱️ Où va réellement le temps
Before fixing anything, it helps to know what a browser is doing.
⏱️ A typical slow page, broken down
Server response
The time before anything appears at all.
Poor hosting shows up here, and no amount of image work fixes it.
If this number is high, hosting is your problem.
Images
Usually the single largest share of a slow page.
Also the cheapest to fix, which is a rare combination.
Compression and correct sizing often halve total page weight.
Scripts
Analytics, chat tools, advertising and tracking pixels.
Each blocks the page briefly while it loads and runs.
Removing three unused ones frequently beats weeks of other work.
Fonts
Custom fonts delay text appearing, which feels slow even when it is not.
Limiting yourself to two weights makes a visible difference.
System fonts remove the cost entirely, at some design expense.
🔧 A Practical Order of Work
Doing these in sequence avoids wasted effort.
| Order | Action | Why here |
|---|---|---|
| 1 | Measure on a real phone | Establishes the baseline |
| 2 | Compress every image | Largest gain per hour |
| 3 | Turn on caching | Usually a single setting |
| 4 | Audit third-party scripts | Removes hidden weight |
| 5 | Re-measure | Most sites now pass |
| 6 | Consider a delivery network | If visitors are distant |
| 7 | Consider better hosting | If server time is still high |
Steps two to four are configuration, not development.
Most sites finish at step five without spending money.
Why measuring first matters
Without a baseline you cannot tell whether anything worked.
Teams routinely spend weeks and never confirm an improvement.
Record the starting numbers before touching anything.
The hosting decision
Upgrade hosting only after the free fixes are done.
Faster hosting cannot compensate for oversized images.
Notre répertoire des essais lists hosting you can test before committing.
🛒 What Speed Does to Buying Behaviour
The mechanism behind the conversion numbers is worth understanding.
| Moment | What delay does |
|---|---|
| First page view | Visitor leaves before seeing anything |
| Browsing products | Fewer pages viewed per visit |
| Adding to basket | Tap feels broken, gets repeated |
| Commander | Highest-cost abandonment point |
| Payment step | Doubt about whether it worked |
Delay at checkout costs far more than delay on a product page.
The visitor had already decided to buy.
Why responsiveness matters at checkout
A button that does not react immediately feels broken.
People tap again, then worry they have paid twice.
That is the moment RedBus fixed for a 7% sales gain.
The compounding effect
Fewer pages viewed means fewer products seen.
Fewer products seen means fewer added to baskets.
Small delays multiply across every step of the journey.
📊 How to Measure Honestly
Two kinds of measurement exist, and they disagree constantly.
| Type | Qu'est-ce que c'est | Trust for |
|---|---|---|
| Lab testing | Simulated load in a tool | Diagnosing causes |
| Field data | Real visitors on real devices | Judging reality |
Lab tools tell you what is slow and why.
Field data tells you what your actual visitors experienced.
Why they disagree
Lab tests use one simulated device on one connection.
Your visitors use hundreds of devices across varied networks (Huppen, 2026).
Google grades you on field data, not lab scores.
The practical approach
Use field data to decide whether a problem exists.
Use lab tools to find out what is causing it.
Then re-check field data weeks later to confirm the fix reached real users.
The reporting delay
Field data is collected over a rolling period, so improvements appear slowly.
A fix made today may take weeks to show in the numbers.
Do not conclude the work failed after three days.
🏢 What Speed Work Costs
The revenue case is only half the decision. Here is the other half.
| Approche | coût typique | Who it suits |
|---|---|---|
| Compress images yourself | Free, a few hours | Tout le monde |
| Caching plugin or setting | Free to modest | Tout le monde |
| Remove unused scripts | Free, an afternoon | Tout le monde |
| Content delivery network | Modest monthly fee | International visitors |
| Better hosting | Moderate monthly increase | High server response times |
| Developer engagement | Significant | Complex or custom sites |
The first three cost nothing but time and fix most problems.
Exhaust the free options before paying anyone.
When paying is justified
Work through the calculation earlier in this article first.
If the estimated annual gain is $22,000, a $2,000 engagement is obviously worth it.
If your site earns very little, the same spend is not.
The trap to avoid
Agencies sell speed audits as recurring retainers.
Most sites need one round of work, not a monthly subscription.
Notre subscription audit method applies here too.
📈 Setting a Target Worth Hitting
Pick a number before starting, or the work never ends.
| Goal | Reasonable target |
|---|---|
| Main content visible | Under 2.0 seconds |
| Response to a tap | Feels immediate |
| Layout stability | Nothing jumps while loading |
| Mobile performance | Tested on a real phone |
| Passing all three checks | The finishing line |
Passing all three puts you ahead of two thirds of the web (Idea Fueled, 2026).
That is a defensible place to stop.
Build in some margin
Sitting at 1.99 seconds means the next threshold change breaks you.
Aim for 1.5 seconds so a tightening standard does not undo the work.
Margin is cheaper than repeating the project.
Review it occasionally
Sites get slower over time as content and tools accumulate.
Check the numbers twice a year rather than once and never again.
Adding one heavy plugin can undo months of careful work.
The three checks in plain language
The official names are unhelpful, so here is what each one measures.
The first asks how long until the main thing appears on screen.
The second asks whether the page reacts quickly when someone taps it.
The third asks whether content jumps around while loading.
Which one to fix first
Main content load fails most often, at 58% passing.
It is also the one most improved by compressing images.
So the most common failure has the cheapest fix.
Layout jumping is usually caused by images without stated dimensions.
Setting width and height attributes fixes it in minutes.
A note on scores versus experience
A tool score is a summary, not the thing itself.
Two sites with identical scores can feel very different to use.
Load your own site on a phone and watch what happens.
If it feels slow to you, the score is not the authority.
Trust the experience, then use the tools to explain it.
🚫 Ce que ces données ne vous disent pas
Case studies are chosen for success. Failed experiments are rarely published.
Rakuten’s 53% is not typical. It reflects one site at one starting point.
Correlation stats prove nothing. The 91% figure is the clearest example.
Some sources sell speed services. Read their emphasis accordingly.
Speed cannot fix a weak offer. Fast pages selling the wrong thing still fail.
The publication bias problem
Companies publish speed work that succeeded.
Projects with no measurable gain never become case studies.
So the published average overstates the typical result.
The ceiling nobody mentions
If your site already loads in 1.8 seconds, further work returns little.
The large gains in these studies came from genuinely slow starting points.
Measure first, then decide whether there is anything to win.
🏁 La version courte
Ignore the 91% ranking correlation. It does not show cause.
The controlled tests are the real evidence, and they are strong.
Rakuten 24 measured 53.37% more revenue per visitor after speed work.
Vodafone Italy gained 8% more sales, RedBus 7% (MigrateLab, 2026).
Only 33% of sites pass all three checks, so the advantage is available.
Google tightened the main threshold to 2.0 seconds in March 2026.
Start with images and caching, then measure on a real phone. ⚡
❓ Questions fréquemment posées
Does page speed affect Google rankings?
Yes, but only slightly. Google has described it as a minor factor. The 91% correlation is not evidence of cause.
What is the strongest evidence that speed matters?
Controlled A/B tests. Rakuten 24 measured 53.37% higher revenue per visitor after improving load time.
How much does one second cost?
Around 7% of conversions by the common rule of thumb. The more reliable figure is about 1% per 100 milliseconds.
Is the Amazon $4.8 billion figure real?
It is an extrapolation from an old internal finding, not a measurement. Treat it as illustration only.
What changed in March 2026?
Google lowered the good threshold for main content load from 2.5 seconds to 2.0 seconds.
How many sites actually pass?
Only 33% pass all three checks, and 42% of mobile sites do. Most competitors are failing something.
What should I fix first?
Images. Oversized photographs are the most common cause of slow pages and the cheapest thing to fix.
Do I need to rebuild my site?
Usually not. Compression, caching and removing unused scripts move most sites into passing range.
Why does mobile perform worse?
Phones have less processing power and less reliable connections. Test on a real device, not a simulation.
When should I stop optimising?
Once you are comfortably inside the thresholds. Chasing a perfect score rarely repays the engineering time.
Will speed fix poor sales?
No. It removes friction from an offer people already want. It cannot create demand that is not there.
📚 Références
Digital Applied. (2026). Page speed statistics 2026: performance and revenue impactConsulté le 8 août 2026 sur https://www.digitalapplied.com/blog/page-speed-statistics-2026-revenue-impact
MigrateLab. (2026). How page speed affects conversion rates: research, studies and revenue dataConsulté le 8 août 2026 sur https://migratelab.com/resources/page-speed-affects-conversion-rates-research
Bloggersideas. (2026). 80+ page speed and Core Web Vitals statistics 2026Consulté le 8 août 2026 sur https://www.bloggersideas.com/page-speed-core-web-vitals-statistics/
Colorlib. (2026). 50+ site speed statistics: page load, Core Web Vitals and performanceConsulté le 8 août 2026 sur https://colorlib.com/wp/site-speed-statistics/
Idea Fueled. (2026). Core Web Vitals 2026: fix speed or keep losing trafficConsulté le 8 août 2026 sur https://ideafueled.com/blog/core-web-vitals-2026-explained/
Huppen. (2026). How page speed affects e-commerce conversionsConsulté le 8 août 2026 sur https://huppen.com.np/public/index.php/blog/page-speed-ecommerce-conversions
Lectures complémentaires sur ce site
Notre website builder analysis covers what platforms really cost to run. See also our software spending researchet le Répertoire des essais gratuits for testing hosting and site tools free.
À propos de cette analyse
This article separates controlled experiments from observed correlations, and labels each figure accordingly. The widely repeated claim that 91% of top-ranking pages pass Core Web Vitals is correlation and is presented as such rather than as proof. The Amazon revenue figure is identified as an extrapolation rather than a measurement. Figures were checked on August 8, 2026.
-
Page Speed and Revenue: The Evidence - septembre 24, 2026 -
AI Thumbnail Generators: Why They Lose to Your Own Photos - septembre 24, 2026 -
Website Builders in 2026: Market Share and Real Costs - septembre 21, 2026