Inside the website audit

PSI stands for PageSpeed Insights, a free testing service run by Google. It loads the site with Lighthouse, Google’s own website-auditing engine, while simulating a mid-range phone on a mobile connection, and grades it in four categories, each 0–100. Because it’s Google’s own tool, the numbers carry weight with clients: this is how Google itself sees the site.

The mobile friendly line in the breakdown also comes from this test: Lighthouse checks whether the page declares a proper mobile viewport, the signal that it adapts to phone screens instead of showing a shrunken desktop page. It’s pass/fail, and counts for real on its own, because for local businesses most visits come from phones. If the PageSpeed test fails to run, the crawler’s own viewport check fills in.

All four PSI numbers are measured on the mobile version of the test, which is stricter than desktop; the same site typically scores lower on mobile. That’s deliberate: it matches how customers actually reach a local business.

Metadata is the invisible layer of a page that search engines and social apps read: the title in the browser tab, the description under a Google result, the preview card that appears when the site is shared in a text or on social media. The crawler runs a set of checks, and each one the site fails pulls the metadata score down.

Every failed check is listed under Metadata issues on the business page, so the pitch can be specific: “when someone texts your website to a friend, it shows up as a bare link with no picture.”

The crawler collects the links on the homepage that point to other pages of the same site and tests them to see if they still work. A link counts as broken if the page it points to errors out (a 404 “not found”, a server error) or doesn’t respond at all.

The penalty is deliberately harsh: even a small share of broken links drives the score down quickly, and a homepage riddled with them bottoms out. That’s intentional: broken navigation is one of the most visible “nobody is maintaining this” signals a site can send. A homepage with no internal links to check is treated as neutral rather than penalized.

Up to ten of the broken links are saved with their error codes and shown on the business page: concrete, demonstrable defects you can pull up during a pitch.

This is the rule-based half of the design grade (it stands in for the whole design grade when the AI review is unavailable). It never looks at the page visually; instead it inspects the page’s code for verifiable, dateable technology choices. It starts from a clean slate and marks the site down for each red flag it finds, the more serious the flag the bigger the mark:

Everything it finds is listed under Site signals on the business page, including a non-scoring note about what the site was built with (WordPress, Wix, Squarespace…) when detectable, which is useful for anticipating what you’d be replacing.

Because these checks are mechanical, they can’t judge taste: a site can pass every one and still look terrible. That is what the AI review is for.

This is the human-eye half of the design grade (it stands in for the whole grade if screenshots couldn’t be captured). During the audit, a real browser loads the site twice, once at desktop size (1440×900) and once at phone size (390×844), and takes a screenshot of each. Both images are then sent to an AI model that can analyze images, with instructions to act as a senior web designer evaluating a redesign prospect. The desktop screenshot is kept and shown on the business page (click it to zoom), so the evidence behind the design notes is always one glance away.

The model rates the visual design 0–100 against a fixed rubric:

The instructions constrain it to judge only what is visible (layout, typography, color, imagery quality, visual hierarchy, and how the mobile version renders) and to ignore the business itself and the quality of the writing. Along with the score, it returns exactly three short observations, which appear as Design notes on the business page. They’re written to be quotable in a pitch: specific things a designer would point at, not generic praise or complaints.

Why both design scores? They cover each other’s blind spots. The objective checks are perfectly consistent but blind to aesthetics; the AI review sees what a person sees, but it is a judgment call. Weighting the visual review more heavily reflects that “this looks dated” is usually the argument that actually sells a redesign.

The audit’s findings feed two things on the business page, both built for the moment you actually reach out.

Top fixes. The findings are ranked by how much each one dragged the website score down, weighing how heavily the component counts against how far the site fell short on it, and the worst are shown as Top fixes a redesign should address: up to three concrete, prioritized problems. This is your talking-points list, in the order that matters.

The pitch. Each completed audit also gets a short sales pitch, written by an AI model from the full audit picture. It follows a fixed brief: open with the business’s strength (its Google reputation), name the two or three most damaging problems, each grounded in a number the audit actually measured, and close with what a redesign gets them. Plain language throughout (“loads slowly on phones”, never “poor PSI score”), 60 to 110 words, and honest: a site in decent shape gets a lighter-refresh pitch, not manufactured outrage.

When the lead has a primary contact, the pitch greets them by first name. From the business page you can copy the pitch or open it as a pre-filled email draft in your own mail client. If the AI writer is unavailable, a templated pitch built from the top fixes fills in, so there is always something to send.

Audits run in a background queue, one business at a time. While the results page is open it refreshes itself every few seconds, so scores fill in live as each audit completes. But a large search can take a while, and there is no reason to babysit the tab.

Close it. When the last audit of a search run finishes, Sifter sends you a single email: how many audits completed, how many failed (failed audits still count as finished, so the email never stalls waiting for them), and a link straight back to the results page.

It is always exactly one email per search run, never one per business, and a run whose results were all audited previously sends nothing at all: no news, no noise. Kick off a big search over lunch and let the inbox tell you when the numbers are ready.

Not every measurement succeeds on every site, and the scores are designed to degrade honestly rather than guess:

Failed audits can be re-run from the business page with Re-run audit. Scores are a snapshot of the day the audit ran; if a prospect ships a new site, re-run to refresh the numbers.