Can an AI fix Core Web Vitals for you?

Yes, for the theme-level causes. My AI coworker Viktor read both PageSpeed reports, edited the Shopify theme, and took desktop performance from 79 to 92 and mobile LCP from 11.7 seconds to 5.0 seconds in about 46 minutes. It handled duplicate font loading, render-blocking scripts, image preload, lazy loading and image resizing. What it could not do was remove platform-level JavaScript that Shopify injects on every page.

Why can't a Shopify site score 100 on PageSpeed?

Shopify injects its own analytics and checkout JavaScript through content_for_header, which you cannot remove. Core stylesheets usually have to stay render-blocking for the page to render correctly. The mobile test also simulates slow 4G on a low-end Android device, which is punishing for any image-heavy page. On a visual Shopify theme, high 80s to low 90s on desktop and low to mid 60s on mobile is a realistic ceiling without moving to a headless storefront.

What are the fastest Core Web Vitals wins on a Shopify theme?

The five that moved the needle for me: remove duplicate font loading across theme.liquid and header.liquid, delete fonts the CSS overrides, make third-party font scripts non-render-blocking, add preconnect to the CDN plus a preload for the LCP image, and lazy load every gallery image below the fold while serving all images at real display size via CDN width parameters.

Is a headless rebuild worth it for a better PageSpeed score?

In my case, no. Moving to Hydrogen or Next.js was the only path to 90+ on mobile, and that means weeks of work plus ongoing maintenance in exchange for points on a synthetic lab test. Fixing the theme-level issues captured most of the real-visitor benefit in under an hour.

Viktor is an AI coworker that lives in Slack. It can analyze data, write and deploy code, edit websites, run marketing analysis and handle ongoing business tasks with access to your real tools. You can try it with $100 in free credits plus $50 off your first purchase at ref.viktor.com/roger-parent.

Back to blog August 03, 2026

I Told My AI Coworker to Get My PageSpeed Score to 100. It Told Me No.

AIAI CoworkerCore Web VitalsFirst-HandPageSpeedSEOShopifyViktor

Google's PageSpeed report on an agency site I co-own was ugly. Desktop 79. Mobile 60. Mobile LCP at 11.7 seconds, which is the kind of number that makes you close the tab.

I sent my AI coworker the two report links on a Saturday morning and told it to get the scores to 100.

Slack message asking Viktor to read the Core Web Vitals analysis and fix everything possible

Eleven minutes later it had already shipped a first round of theme changes and told me one of them backfired.

Viktor reporting desktop score jumped 79 to 92 but total blocking time spiked

Desktop was up 13 points and mobile LCP had dropped by more than half, but Total Blocking Time went the wrong way. It flagged that itself instead of reporting the good half and stopping. That is the behavior I care about more than the score.

Where the 46 Minutes Went

Viktor's list of eight theme fixes with before and after Lighthouse metrics

Chart showing desktop score 79 to 92, mobile 60 to 64, mobile LCP 11.7s to 5.0s

Nothing on the list is exotic. That is the point. Every item was already sitting in the theme, and none of it needed a developer ticket:

Fonts were loaded twice. Typekit and Google Fonts were both called in theme.liquid and again in header.liquid, so every visitor paid for the same requests twice.
A font was being downloaded that the CSS immediately overrode. Nobody sees it, everybody waits for it.
Typekit was render-blocking. Making it async saved about 1,960ms of blocking on mobile.
The largest homepage image had no preload. Adding one, plus a preconnect to the CDN, is most of the LCP win.
Eight hero gallery images were loading eagerly. Only the first one is visible. The other seven now lazy load.
Images were being served at full size and scaled down in the browser. Adding CDN width params cut roughly 30% off each, and one collage image went from 1.46MB to about 300KB.

Mobile LCP finished at 5.0 seconds on good runs, down from 11.7. Desktop TBT went from 300ms to 70ms. Desktop landed at 92, in the green.

Then It Told Me No

I asked for 100. It came back and said 100 is not happening on this stack, and showed its reasoning.

Viktor explaining why a 100 PageSpeed score is not realistic on a Shopify theme

Shopify injects its own analytics and checkout JavaScript through content_for_header and you cannot strip it. Two stylesheets have to stay render-blocking or the page renders wrong. The mobile test simulates slow 4G on a low-end Android, which punishes any image-heavy page. Getting to 90+ on mobile would mean a headless rebuild in Hydrogen or Next.js.

That answer saved me more money than the speed fixes did. A headless rebuild is weeks of work and a permanent maintenance tax, bought with a lab score that is not the same thing as what real visitors experience.

The Second Question in That Message

I had also asked whether every new blog post was actually making it into the sitemap I submitted to Search Console. Answer: Shopify adds published articles to sitemap_blogs_1.xml automatically, so they were in there, but nothing was telling Google to come look.

So it changed the publishing routine to resubmit the sitemap after every new post. Small thing, and it is the kind of small thing that quietly never gets done.

What This Cost Me

Four Slack messages and 46 minutes of clock time on a Saturday, most of which I spent not at my desk. No agency retainer, no dev sprint, no Lighthouse plugin subscription.

The honest limitation: these are lab scores from a synthetic test. Field data in Search Console moves later and moves less. Desktop 92 and a mobile LCP cut in half is a real improvement for real visitors, but I did not turn the site into a rocket in one morning and I am not going to claim I did.

If You Want to Run This on Your Own Site

Pull your PageSpeed report for desktop and mobile separately. They fail for different reasons and the mobile one is where the money is.
Search your theme for duplicate font loads. Multiple themes and apps stacking over the years makes this common.
Check whether your hero image has a preload and whether everything below the fold is lazy loaded.
Look at the actual byte size of your biggest images. A 1.4MB image on a homepage is not rare.
Decide up front what score is good enough, so nobody talks you into a rebuild over 25 points on a simulated phone.

Viktor is the AI coworker that did this. It lives in my Slack, and it had access to the theme files and the PageSpeed data at the same time, which is why it could read a report and ship the fix without a handoff to anyone. It also rebuilt this website and audited the analytics on that same agency site, where it found most of the traffic was bots.

If you want to point it at your own site, you can get $100 in free credits plus $50 off your first purchase here.

More real examples, screenshots and all: my full Viktor review.

Written by

Roger Parent

Founder of Chief Marketing Dad · Fractional CMO

Roger has spent 15+ years helping ecommerce brands grow, these days as a fractional CMO working with a handful of brands at a time. He digs into the numbers most people skip and cares way more about what actually moves revenue than what looks good in a deck. When he is not buried in an ad account, he is usually with his wife and kids or on a pickleball court somewhere in Connecticut.

More about Roger →