The DTC Revenue Leak ReportFive leaks, priced.

The finding

Every revenue leak deletes the evidence of itself. The shoppers it costs you never reach the dashboard where you would count them, so the worse a leak gets, the healthier some of your numbers look.

That is not a figure of speech. It is a description of how the plumbing works, and it is why the five leaks below stay open for years on stores run by people who check their analytics every morning.

Three of them have been separately measured by three organisations that do not talk to each other. Nobody has ever added the columns together. We did:

• 76% of Shopify origins on mobile pass all three Core Web Vitals.
• 38% of ecommerce sites have a mobile product page rated “decent” or better.
• 37% of inner pages on mobile carry JSON-LD — the format that decides whether a machine can read your catalogue.

Multiply them and roughly one store in nine is clean on all three. Even if you assume those three virtues travel together — and they mostly do — no more than 37% of stores can be clean on all three, because that is the strictest single rate. So between 63% and 89% of storefronts fail at least one of three checks that are already public. That is before anyone looks at your JavaScript errors or walks your checkout.

What a Revenue Leak Score measures

One number, so the five leaks can be compared and argued about. It is a diagnostic composite, not a conversion-rate prediction, and the difference matters — a store can score badly and still sell well, which usually means it is being carried by demand rather than by its site.

The grade is the useful part, because a grade is a status ladder and you will locate yourself on it before you finish reading the row.

Leak 1 — Speed. Everyone measures it against the wrong number.

Class 1 Deleted The most-written-about leak in ecommerce, and the industry has agreed on a threshold that is roughly one full second past where the money is.

Google’s bar for a “good” Largest Contentful Paint is 2.5 seconds at the 75th percentile. Pass it and every tool you own turns green. Portent measured the other thing — what shoppers actually did — across 5.6 million sessions on 20 sites in a 30-day window. On the B2C ecommerce sites, the highest conversion rates sat between 1 and 2 seconds: an average 3.05% at a one-second load, falling to 0.67% by four seconds. Their summary line is the one to keep: a site loading in one second converts at 2.5× the rate of a site loading in five.

Read those two facts side by side. Google’s pass mark and the commercial cliff are not in the same place. A store sitting at 2.4 seconds passes Core Web Vitals and is on the wrong side of the curve. That is not a failure of Google’s threshold, which was set for a different purpose. It is a failure of using it as a business target.

What this is built on

No crawl, no survey, no first-party benchmark. This edition is a synthesis and a method. Every population figure comes from a named third party, listed at the end with the date it was read, and every dollar figure is arithmetic on five declared inputs you should replace with your own.

What we did not measure. We have not scanned a population of stores and cannot tell you what share of them fail any individual check. There is no Mtrix figure anywhere in this report. Our own platform data cannot support a market benchmark — it is dominated by a handful of storefronts belonging to a single organisation, which makes it useful for building a product and useless for describing an industry, and we would rather say so than publish it.

The samples do not match each other, and it matters. The Core Web Vitals figure covers 567,932 Shopify mobile origins. The product-page figure covers 155+ manually benchmarked storefronts, most of them large and well resourced — which means 38% is a generous read for the median DTC store, not a harsh one. The JSON-LD figure covers inner pages across the whole web rather than ecommerce specifically. Stacking them produces an order of magnitude, not a measurement, and we say so beside the chart rather than in a footnote.

What we can defend. The five leak definitions, the sub-score formulas, the coefficient bands and the arithmetic. Those are ours, they are published in full here and in the methodology at the end, and anyone can run them against their own store and get the same answer we would.

Full method, exclusions and every source at the end.

Leak 2 — Breakage. The only leak with no dashboard at all.

Class 2 Misfiled Every other failure in this report has a tool that watches it. This one has a gap in the middle of the room.

Here is what physically breaks. A JavaScript error throws in the browser, on one device, on one template, on one variant. The page still returns HTTP 200. The server logs nothing, because nothing reached the server. Your uptime monitor pings the URL and gets a page. Lighthouse loads the page in a lab, on a desktop-class CPU, without your shopper’s ad blocker, extensions, locale or the specific product that triggers it — and passes. The shopper taps Add to cart, nothing happens, they tap again, and they leave.

Nothing in a standard stack is watching the button. The uptime monitor watches the door. The lab audit watches a rehearsal. The server log watches requests that arrive. A broken add-to-cart generates none of those signals, and the one system that would notice — your conversion rate — reports the outcome without the cause.

Leak 3 — The product page you are fairly sure is fine

Class 2 Misfiled The leak that presents as a traffic-quality problem, gets blamed on the ad account, and is fixed nowhere.

Baymard has scored this more carefully than anyone: 30,000+ usability scores across twelve product page topics, on 155+ manually benchmarked storefronts. Only 48% of desktop and 38% of mobile ecommerce sites reach “decent” or better on product page UX. Read the inverse, because it is the useful direction: 62% of mobile product pages are mediocre or worse, on a benchmark whose sample is weighted towards large, well-funded brands. If the biggest stores in the world are at 38%, the median DTC store is not above it.

But the aggregate score is not the finding. The finding is which things are missing, and what they map onto.

Line up Baymard’s product-page gaps against Baymard’s own abandonment reasons and they interlock:

Two-thirds of stores withhold, until the cart, the single number that causes four in ten abandonments. That is the whole chapter. The shopper is not surprised by shipping costs at checkout because shipping costs are surprising. They are surprised because the product page had the answer and did not give it.

Look at the ten stated reasons carts get abandoned and five of them are questions the product page could have answered before a cart existed: total cost, delivery speed, trust, returns, and visibility of the full price. The cart is where the abandonment is recorded. It is not usually where it is caused, and optimising the cart to fix a product-page problem is the most common wasted quarter in this category.

Leak 4 — The shopper who never arrives

Class 1 Deleted The only leak in this report we refuse to price, and the only one whose cost is going up every month.

What physically happens: someone asks an assistant for a recommendation in your category. The assistant reaches for stores it can crawl, parse and quote. If your catalogue is not machine-readable at the moment it looks, you are not in the answer. There is no impression, no click, no bounce and no row. It is the purest class-1 failure in the report — the visit does not fail, it does not occur.

For a long time this did not matter much, because the traffic was small and it converted badly. Both of those facts have flipped.

Adobe, working from more than a trillion visits to US retail sites, has the crossover on record. In July 2024, AI-referred traffic converted at roughly half the rate of everything else. By February 2025 the gap had closed to 9% worse. Over the 2025 holiday season, AI referrals converted 31% more than other traffic sources, with revenue per visit up 254% year to date and conversion 54% higher than non-AI on Thanksgiving. Traffic to retail sites from generative AI tools grew 693% year over year across November and December 2025.

Read that as one sentence: the fastest-growing traffic source in retail overtook the rest of your traffic on conversion rate, in about eighteen months. It is not a channel to evaluate next year. It is a channel that now converts better than the traffic you are currently paying for.

And almost nobody is set up to receive it. Across roughly 12.9 million sites in the HTTP Archive’s July 2025 crawl, llms.txt is present on 2.13% of desktop sites — the file whose entire job is to tell an assistant how your site is laid out. Structured data has reached 50% of home pages, with JSON-LD at 43%; on inner pages, the ones that carry your actual products, JSON-LD drops to 37% on mobile. And only 2% of crawls find structured data injected by JavaScript — which is the good news, because markup that only exists after your JavaScript runs is markup a non-rendering crawler never sees.

Leak 5 — Checkout. The most expensive metre in the store.

Class 3 Offstage Every shopper here has already chosen to buy. That is what makes the losses cost more than anywhere else on the site, and what makes them so hard to watch.

The headline number is old and everyone knows it: 70.22% of carts are abandoned, an average across 50 separate studies. Familiarity has drained it of meaning, so invert it. Seven in ten people who told you they wanted the thing did not get it. Not seven in ten visitors — seven in ten people who put it in a cart.

The reasons are where the work is, and Baymard publishes them:

Five leaks, one store

Everything above, on one hypothetical store. These five inputs are declared assumptions, not measurements. Substitute your own and every number below moves with them.

Its scan comes back like this. Mobile LCP 3.2 seconds — 0.7 over. One high-severity JavaScript error on the purchase path. Product-page purchase UX 4 out of 10. No llms.txt, no Product JSON-LD in the initial HTML. Checkout has no express-pay option and asks for an account.

Score your own store

Five inputs. Nothing here needs a developer and nothing needs a login.

1 · Speed (0–100). Run your best-selling product page through any mobile performance audit and take the performance score. Then take your mobile LCP in seconds and hold it against 2.0, not 2.5.

2 · Errors (0–100). Open your store on a real phone with your ad blocker on. Add the second variant of three products to the cart. Start at 100 and subtract 12 for every one that fails on the purchase path, 4 for anything visibly wrong elsewhere, 1 for cosmetic breakage, 3 for every image or request that fails to load. Floor at 5.

3 · Product page (0–100). Ten checks, ten points each: total delivered cost visible in the buy box · return policy linked in the main content · price per unit where quantities vary · buttons not a dropdown for variants · an in-scale or human-model image · review stars and count near the title · add-to-cart visible without scrolling on mobile · a sticky add-to-cart on scroll · an express payment option · a concrete delivery estimate.

4 · Machine readability (0–100). 25 points for llms.txt returning a real file · 25 for Product and Offer JSON-LD in the initial HTML response · 30 for product copy that answers a question rather than describing a mood · 20 for your brand being named when you ask an assistant for a recommendation in your category.

5 · Checkout (0–100). Start at 100. Subtract 25 if guest checkout is not the default path · 25 for no express payment option · 10 for every required field before payment beyond six · 15 if the total delivered cost only appears at the final step.

Your grade. Mean the five sub-scores. A ≥ 90 · B ≥ 75 · C ≥ 60 · D ≥ 45 · F < 45. Then do the money version: apply the bands from the table above to your own monthly revenue, take the midpoint of the total, and map it — 0% at risk = 95, 25% or more = 30, linear between. The two numbers will not match. The gap between them is the most useful thing on this page, because the sub-score mean tells you how tidy your store is and the money version tells you how much that tidiness is worth.

Fix this → Monday

Ordered by dollars recovered per hour of work, not by size of the leak. The first three cost nothing but attention.

1. Monday, 20 minutes — the two settings

   Make guest checkout the default path. Turn on an express payment option. Between them they address 18% and 9% of your stated abandonment reasons, they are configuration rather than code, and they are the cheapest money in this report.

2. Monday, 11 minutes — the six checks

   The list above: cellular load, second variant, buy-box total, llms.txt, a cold checkout, then your analytics. Write down what you find before you fix anything, because the record is what makes the second scan mean something.

3. Tuesday, the buy box

   Put an estimated delivered total inside it on your three best sellers, and link the return policy from the main page content. One template change against the largest stated reason your carts are abandoned.

4. Wednesday, the errors

   Whatever the second-variant test surfaced. This is the only leak in the report with a binary outcome — no test, no significance, no wait. Ship the fix and move on.

5. Thursday, the markup

   Product and Offer JSON-LD in the initial HTML response, then an llms.txt. Half a day, and it puts you in the 2% of the web that is legible to the fastest-growing traffic source in retail.

6. Friday, the speed number

   Reset your internal target from 2.5 seconds to 2.0 and find out what that costs. Expect the estimate to be uncomfortable; the whole point of the threshold gap is that the last half-second is the expensive one.

Where Mtrix fits — and where it doesn’t

Most of you should close this and go and run the six checks. Mtrix does not shorten that list by a single item.

This is for the smaller group who reached the worked example and asked a different question. Not which sub-score is worst, but why did it take a report to tell me any of them, when I already pay for four tools that watch this store?

What we do about exactly that. The signals in this report sit on one session rather than in five products: the performance your shoppers actually lived, the JavaScript errors that fired inside the session, the session itself, and the order it did or didn’t produce, in Purchases. The dip and its cause are one click apart, not one integration apart. What you find, you can test on the store you already have.

The concessions, which matter more than the pitch:

• Mtrix does not compute this report’s score inside the product. The Revenue Leak Score is produced by our scanner from outside your store. Nothing in the platform grades you A–F.
• We do not do the machine-readability leak. No llms.txt generation, no AI-visibility monitoring, no assistant-ranking product. Leak 4 is real, we scored it, and we don’t sell the fix.
• On a store that stays on Shopify, Mtrix measures and tests — it does not rebuild. The builder edits pages Mtrix generates and hosts. It does not edit your Shopify theme in place, vary your prices, or run inside Shopify’s hosted checkout. Leak 5’s diagnosis is ours; some of its fixes are not.
• Session replay masks credentials and payment fields by type and name before they leave the browser — not everything. Email, name, phone and address are captured in clear unless you exclude them. Know that before you turn it on.
• Web only. No native iOS or Android SDK. Replay and errors kept one month; events and purchases three years.

If your problem is that you don’t know which leak you have, a platform is not the answer — a scan is, and ours is free. Mtrix is for the store where you already know, and where the reason nothing gets fixed is that the number and the session behind it live in different products.

Methodology (in full)

What was measured: nothing, by us. No crawl, no survey, no panel, no sample. Edition one of this report is a synthesis of four named public datasets plus a published scoring method. We do not know what share of DTC stores fail any individual check in it and we do not claim to. Edition two ships against an original crawl and will say so on the cover.

No Mtrix data appears anywhere in this report. Our platform holds tens of thousands of ecommerce sessions, and almost all of them belong to a small number of storefronts operated by a single organisation. The honest methodology sentence for a benchmark drawn from that would be “across three storefronts operated by one brand,” and a sentence like that kills a report on the page it appears. Breadth is what makes Baymard’s and HTTP Archive’s numbers worth citing, and breadth is exactly what we lack. So we cite them and price them instead.

The sources, and what each covers.

Core Web Vitals and structured-data figures come from the HTTP Archive’s Web Almanac 2025, drawn from its July 2025 crawl of roughly 12.9 million sites; the ecommerce CWV table covers 567,932 Shopify mobile origins and 995,782 WooCommerce ones. Product-page figures come from Baymard Institute’s product page UX benchmark — 30,000+ manually reviewed usability scores across 155+ benchmarked storefronts. Cart-abandonment figures come from Baymard’s aggregation of 50 separate studies, last updated 22 September 2025. Speed-to-revenue coefficients come from three independent studies: Deloitte and 55 for Google (37 brand sites, 30M+ sessions, 30 days, 2019 data), Akamai (~10 billion visits to top online retailers, 2017), and Portent (5.6 million sessions across 20 sites, 30 days, 2022). AI-referral figures come from Adobe Analytics, based on more than a trillion visits to US retail sites.

Sample mismatch, stated plainly. Those populations are not the same population. Baymard’s benchmark is weighted towards large, well-resourced retailers, so its 38% mobile figure is a generous read for the median DTC store. The JSON-LD figure covers inner pages across the whole web rather than ecommerce specifically. The CWV figure is platform-scoped to Shopify. The compound finding in The finding multiplies three rates from three of those samples, which is an order of magnitude and not a measurement — it is labelled as such on the chart, in the caption, and here.

The coefficient bands. Speed is priced at 7–15% of monthly revenue at risk per second of mobile LCP over the 2.5-second threshold, capped at 25% (low) and 35% (high) for a single site. Purchase-path errors at 2–5%, applied only when a high-severity error is observed on the path to purchase. Product-page purchase UX at 2–6%, applied only below 5 out of 10, and labelled directional. Checkout at 1–3% for a missing express-pay option and 1–2% for forced account creation, combined and capped at 4%. Totals are capped at 30% (low) and 40% (high) of monthly revenue. Machine readability carries no dollar band by design.

What we deliberately did not use. The bounce-probability ladder — “as load times go from 1s to 3s the probability of bounce increases by 32%” and its five siblings — appears nowhere in this report. It is a small set of borrowed, thinly sourced numbers recycled by many vendors, and quoting it marks a document as a follower. We also did not price Leak 4, and we did not use any vendor’s “X% of ecommerce sites have unresolved JavaScript errors” figure, because the ones in circulation are marketing claims from companies selling error monitoring, with no published method.

The arithmetic. Every dollar figure derives from one illustrative store whose five inputs are declared in Five leaks, one store: $100,000 monthly revenue, 1,250 orders, $80 average order value, 69,400 sessions, 1.8% conversion. Those are hypotheses, not measurements. They exist so the method is inspectable; substitute your own and every conclusion moves with them.

Exclusions. Marketplace sellers, B2B and wholesale storefronts, subscription-only businesses whose conversion happens off the storefront, and native mobile apps. The abandonment figures are US/EU weighted. The AI-referral figures are US retail only.

Use this report

Share it, quote it, put the charts in your deck. Attribute to Mtrix with a link to this page. Want the underlying working — the coefficient bands, the score formulas, the full source list with read dates? Ask us and we’ll send the raw cuts.