The Noise & Silence audit · $497 one-time

Every PageSpeed Insights warning, put on trial.

PageSpeed Insights is a screening instrument — built to over-flag, never to decide. Your audit is the diagnosis that follows: every flag it raises leaves with a verdict — Real, Wrong, Trivial or Locked — with the falsifying evidence quoted. Then come the findings that were never in the report at all — the Silent Bottleneck in your field data, the checks PSI never runs — and every surviving fix ships as code with a predicted gain and the re-runs to hold us to it.

Read the framework

Secure Stripe checkout · just your email · your report or your money back

One URL · mobile + desktop · 2 verification re-runs included · delivered in less than 24 hours

Skeptical? Run this page through Google's own test

A verdict, not a score

PSI raised 21 flags:
4 Real · 7 Wrong · 8 Trivial · 2 Locked

4 Real
7 Wrong
8 Trivial
2

▲ only these four deserve engineering time

Noise Rate 81% — and the number-one field bottleneck (TTFB, p75 1.6 s) appeared in none of them.

The verdict line from the framework's worked example: 21 tickets in — four fixes worth shipping, one discovered bottleneck, seventeen documented permissions to do nothing.

In plain words: Google's free speed test lists many warnings. We test each one, say which are real, and give your developer the exact changes to make. The report arrives within one day.

01 — The problem, honestly sized

Speed is a revenue problem. Just not the one being marketed.

≈0.65% per 100 ms — the honest slope

We audited twenty years of speed-to-revenue studies, graded by research design. The defensible causal anchor near a 3-second baseline is ≈0.65% conversion per 100 ms — roughly a tenth of the marketed 3–8%. Still real money on real traffic; it just doesn't need inflating. The evidence audit, open data →

Google ranks by experience

Core Web Vitals (LCP, CLS, INP) are ranking signals measured on real users. And a green lab score is not a pass: 43% of pages scoring 90+ still failed a Core Web Vital in the field. When lab predicts field →

Your PSI report is part noise

PageSpeed Insights is deliberately over-sensitive — a false alarm costs your team's time, not Google's. Acting on every flag burns sprints on warnings that move nothing. The expensive question isn't “what's slow?” — it's “which warnings are real?”

An audit that is honest about the ROI is also honest about the findings. Yours are sized with the defensible number — as ranges, arithmetic shown — never a folklore multiplier.

02 — The method

The verdict table is the product

Most tools — and most audits — treat every Lighthouse flag as true. We hold each one to three tests — truth, materiality, jurisdiction — and it leaves with a verdict and its evidence:

Four specimen PageSpeed Insights flags, each with its evidence and a verdict (Real, Wrong, Trivial or Locked)
PSI flagEvidenceVerdict
Reduce render-blocking resources — “save 2,330 ms”The chain is real — but the promised saving would imply a ~450 ms LCP for real users. Impact recomputed honestly against field data.Real
Image elements do not have explicit width and heightLab CLS 0.122 — field CLS for real users: 0.00. The shift exists only in the lab.Wrong
Reduce unused JavaScript — 909 KiBGoogle's own metricSavings: zero on every metric it models. The bytes are reCAPTCHA.Trivial
Serve static assets with an efficient cache policyTrue and material — and owned by the platform, not you. Effort redirected to the controllable equivalent.Locked

Rows from the worked example in our framework — an illustrative composite of recorded adjudications. Your report carries the full table, with verbatim evidence per row.

The Noise Rate

The share of Google's warnings that should receive no engineering time — stated up front, every dismissal carrying its proof. Knowing which warnings to ignore routinely saves more engineering hours than any single fix. That is the paid feature.

The Silent Bottleneck

The constraint your field data proves and no flag names. The number-one problem produced zero flags. It is found by decomposing the field metric, not by reading the report harder.

Read the full method — the Noise & Silence framework →

03 — Beyond the report

Half of what you get was never in PageSpeed Insights at all

Verdicts clean up Google's list. The rest of the audit is measured, not re-read — probes and cross-checks whose findings no export of PageSpeed Insights can contain, because PSI never took the measurement:

This audit measures
Free PSI stops here PSI lab run CrUX field data

+ seven surfaces PSI never reads:

Your CDN’s configuration Platform & app census Response headers Your font files Your DNS & TLS setup Your JavaScript’s build era The page Lighthouse was served
Findings in the audit report, the measurement each comes from, and whether PageSpeed Insights contains it
Finding in your reportWhere it comes fromIn PageSpeed Insights?
Why your lab score and your real-user numbers disagree — and which one to believeField p75 decomposed against the lab runShown side by side, never explained
The Silent Bottleneck — your number-one field constraintDecomposing the field metric PSI only displaysZero flags — the classic case is real-user TTFB
The causal chain behind your LCP, culprit named — fixing one link of an unbroken chain gains nothingCross-resource dependency reconstructionIsolated point symptoms only
The CDN settings that should change — compression, HTTP/3, caching — with a one-command apply packA live probe of your edge configurationNever read
The specific plugin, app or theme costing you — including dead scripts left by uninstalled appsA census of your platform's actual assetsThird-party bytes are summed, never attributed
Back/forward-cache blockers, font fallback overrides, preload correctness, 103 Early HintsChecks PSI doesn't runAbsent from the report
The 2015-era JavaScript your bundler still ships — polyfill bytes measured as actually delivered, with the modern-target config fixA modern-browser fetch with source-map attributionA fixed-signature insight exists — this adjudicates it, then goes deeper
The connection setup underneath your TTFB — nameserver latency, CNAME chain, TLS 1.3, session resumption, 0-RTTLive probes of your DNS and your handshakeNever measured
Whether Google's tool was even shown the page your users get — the audit's own input, integrity-checkedA user-agent flip judged against the page's load-to-load churn — the same gate as our free cloaking testAssumed, never verified
Which fixes matter for your audience — round-trip savings vs byte savings, weighted by real devices and network latencyThe CrUX device mix and connection RTT profile re-weight the planForm factors shown, never used to prioritize
The interaction-risk model when your field INP is poorStatic main-thread analysisThe lab cannot measure INP

Assembling this yourself is real work: a senior engineer's hours across PSI, the CrUX API, WebPageTest and your CDN dashboard — or a monitoring subscription, a tracking snippet on your pages, and weeks of traffic. Your report hands it to you complete — falsification work on the record — in less than 24 hours.

04 — The deliverable

What lands in your inbox

  • Every flag adjudicated — the full verdict table: Google's claim → Real / Wrong / Trivial / Locked → the evidence, with a documented falsification attempt and a confidence grade behind every finding.
  • Up to five deep findings that exist in no free report — causal chains with the culprit named, the field constraint decomposed, the checks PSI never runs — quantified as ranges with the arithmetic visible.
  • Fixes in your platform's vocabulary — the theme.liquid snippet on Shopify, the cart-fragments rule on WooCommerce, the next.config.js change on Next.js, the exact plugin setting on WordPress — plus the corrected <head> block and the image re-encode table, verbatim. Deep commerce modules for Shopify, WooCommerce, BigCommerce, Magento / Adobe Commerce, PrestaShop and Salesforce Commerce Cloud — each fix aimed at the lever you control on that platform, never at settings the platform owns.
  • An edge apply pack when you're on Cloudflare, Akamai, Fastly, Bunny, Azure Front Door, AWS or Google Cloud — only the settings that should change, as one command, exact dashboard paths, or Terraform.
  • Executive summary + JSON twin — one plain-language page for leadership, and every finding machine-readable via the REST API.
  • 2 verification re-runs within 30 days — the accountability step, not a usage quota.

Already fast? Then that's your report: short, evidenced, and honest that there is little to do. Our quality gate fails any report you could have gotten 80% of free at pagespeed.web.dev — we never pad findings to justify the fee.

Anatomy of one finding

The evidence rules every finding must pass:

  1. Evidence — the verbatim artifact: the actual tag, the actual header, the actual byte count. No artifact, no verdict.
  2. Falsification — the documented attempt to prove the finding wrong before you ever see it.
  3. Prediction — the expected movement as a range, arithmetic visible, confidence graded.
  4. Verification — proven in a controlled lab A/B before you deploy; confirmed in real-user data on its own 28-day clock after.
under-delivered
0% 100% of promise 200%

Each dot is one measured fix: what it delivered, as a share of what was promised. All 8 landed at or above 100% — the calibration study →

05 — The process

Predict. Prove. Confirm.

  1. Buy the audit

    Secure Stripe checkout, just your email — no account setup before payment. You get your report, or you get your money back.

  2. Enter your URL — report in less than 24 hours

    Every check runs, every flag is adjudicated, nothing is skipped — the report ships when the method completes, always inside 24 hours. Delivered to your dashboard and inbox.

  3. Hold the report to its predictions

    Every surviving fix ships with a predicted range. Ship the fixes, then use your 2 included re-runs (within 30 days) to prove the movement — and watch your real-user data confirm it on its own 28-day clock.

More about the methodology →

06 — The research program

We publish the evidence — even when it argues against our own pitch

The cost of a millisecond

The speed-to-revenue literature since 2006, graded by design. The famous 3–8% headlines run ~10× above the causal anchor. Full study register, CC BY.

How many runs prove a change?

Median-of-5 can't prove a 5-point score change — that takes ~16 runs per condition. The statistics, and a free calculator that does them for you.

The cost of a kilobyte

210 controlled Lighthouse runs: 100 KB costs ~0.5 s of mobile LCP whether it's JavaScript, CSS or image — and placement matters more than which bytes. Raw data included.

PageSpeedAudit is built by the team behind the Noise & Silence framework and this research program — preregistered designs, published null results, replication kits, open data. All studies →

07 — Questions

Frequently asked

Is this just a Lighthouse export?

An export treats every flag as true. This audit puts every flag on trial — and in the framework's worked example, 17 of 21 didn't survive, each dismissal quoting Google's own numbers back: a flag whose own metricSavings reads zero, a lab-only CLS of 0.122 against a field CLS of 0.00. And half of the report never came from Lighthouse at all — the field decomposition, the CDN probe, the platform census. The full worked example is public — judge it against any export.

I'm technical — can't I assemble all of this for free?

Much of it, yes, with the right tools and hours: the adjudication and the field-vs-lab work are derivable by a senior engineer from PSI, the CrUX API and WebPageTest. Some of it no free report will hand you at any skill level: your CDN's live settings graded against your plan, dead scripts from uninstalled apps, font fallback overrides computed from your actual font files, back/forward-cache blockers PSI doesn't test — and even proving a fix's impact before shipping is a paid tier at WebPageTest. The audit hands you all of it in less than 24 hours, with the falsification attempts documented. If your time is free and your skills are senior, you're genuinely not the customer — our free tools and research are for you.

What does a $3,000–$8,000 agency audit add?

Mostly calendar time and a workshop call. The evidence standard here is the one a senior consultant holds — verbatim artifacts, documented falsification attempts, graded confidence — executed exhaustively by an engineered pipeline instead of billable hours, which is why it lands in under 24 hours instead of weeks. What you give up is a human presenting the findings; what you gain is every single flag adjudicated, which almost no human engagement does exhaustively. The full comparison — including when not to buy this →

Which page should I audit?

The page that makes you money: your top landing page, your product page, or your checkout entry. One purchase covers one URL — most customers start with their highest-traffic commercial page.

What exactly are the “2 re-runs”?

The accountability step. Every fix in your report carries a predicted range — after your team ships, you re-run the same URL to hold the report to its own predictions, with fresh data. 2 re-runs included within 30 days of your first run; failed runs never count.

What if my page is already fast?

Then the report says so — short, with the evidence, the two or three genuine items, and what to watch. We never pad. Our quality gate fails any report whose value you could have gotten mostly for free at pagespeed.web.dev, so “everything is fine” is a result we're allowed to deliver — and proud to.

Can I integrate audits into my own product or agency workflow?

Yes — everything is available via REST API with machine-readable JSON reports, and agencies can white-label the deliverable. See the API documentation.

Do you offer refunds?

You get your report, or you get your money back — a failed run never consumes your purchase. Refunds for delivered reports are handled case by case; just contact us.

Free 5-day email course

Read your PageSpeed report like an auditor

The method behind the audit, taught — one short email a day, five days, every number from our published worked example:

  1. Day 1The 81% problem — a real report, 21 flags, and why only four deserved effort.
  2. Day 2The four verdicts — Real, Wrong, Trivial, Locked — and the test each flag must pass.
  3. Day 3Falsifying a flag — three places the disproof hides, starting with the tool's own numbers.
  4. Day 4Hearing the silence — the five-minute field-vs-lab drill that outranks every flag.
  5. Day 5Predict, prove, confirm — the sample-size math and the 28-day rule that make a fix provable.

The course lives in email only — it isn't published anywhere on the site.

Double opt-in: we email a confirmation link first. Five emails, then it stops — unsubscribe any time.

Walk-the-talk guarantee

This site practices what the report preaches.

Zero JavaScript on this page, one 22 KB font, compressed everything. Run it through PageSpeed Insights and see for yourself.

Test this page on Google