Every full Stackra scan gets reviewed by three AI experts. A CMO looks at conversion and trust. An SEO lead looks at how findable the site is. A CTO looks at speed, security and accessibility. Each one sets 30% of its pillar score. The other 70% is measured: crawled pages, Lighthouse runs, link checks, forms, schema and security headers.

Thirty percent is real weight, so the experts don't get to freelance. Before one of their opinions reaches your score, it has to get past three things:

• •A to-do list it didn't write
• •More than 200 written guardrails
• •A round of checks in code after it answers

The example: a plumber's homepage

The easiest way to show how that works is to follow one scan all the way through. Picture a scan of a local plumbing company (an illustration, not a real customer). Here's what the crawl and the checks find on the homepage:

• •A "Book a repair" button up top, with a short quote form under it
• •Five customer quotes with names, in a reviews section
• •A 4.8-star Google badge in the header
• •Two broken links in the footer
• •A slow load on phones: the main image takes over four seconds to appear

Hand a site like this to an AI with no rules and you'll get the usual list. Add testimonials, strengthen your call to action, speed it up. Two of those three are wrong here. Here's what happens to each one.

Stop 1: the facts show up before the expert does

The experts never browse the site. They get a data pack: everything the crawl, Lighthouse and our checks measured, written out field by field. For the plumber it says the main call to action is "Book a repair" and a quote form is present. It also says two links are broken and the page loads slowly on phones. Some fields are written so they can't be misread:

• •Testimonials come through as a level, not a raw count. The plumber's reads "5+ found on the homepage."
• •A heading that says "Reviews" with nothing under it doesn't count as testimonials. Neither does a star badge
• •"None detected" means absent, and present means present. The expert doesn't get a second opinion on either

Stop 2: the to-do list is written before the AI shows up

This is the part people tend not to expect. The expert doesn't decide what's wrong with your site. Seven detectors do that first, in plain code:

• •Broken links
• •Calls to action
• •Security
• •Performance
• •Accessibility
• •Forms
• •Search basics

Each pillar gets a ranked shortlist of up to eight confirmed issues, with the wording already written. Then the rules for the expert are blunt:

• •Recommend fixes from the shortlist only, and copy each one word for word
• •Nothing off the list goes in as a recommendation
• •An empty shortlist means no recommendations and no weaknesses, not "go find something"

For the plumber, the shortlists hold the two broken links and the slow phone load. The booking button passed the call-to-action check, so "strengthen your call to action" isn't there. "Add testimonials" isn't there either, because nothing we measured says this site is short on proof. Two of the three generic suggestions are dead before the expert writes a word.

Stop 3: more than 200 guardrails

Then the expert reads its instructions, and this is where a lot of our time has gone. Between the rules every expert shares and each one's own brief, there are more than 200 written guardrails. Each one exists because a model left alone will pad, hedge, contradict itself or recommend things you already have. Some of the ones doing the heavy lifting:

• •Data first: every claim has to trace back to a field in the data pack
• •Already there? Then don't recommend adding it. Improve it or move it, at most
• •Enough is enough: with three or more testimonials, client logos or calls to action, "add more" is off the table
• •Stay in your lane: the CMO doesn't talk page speed, the SEO lead doesn't talk conversion rates, the CTO doesn't talk brand messaging
• •No contradictions: a strength can't also show up as a weakness
• •Plain words in titles. Technical terms only go in the how-to steps
• •One or two high-priority items per expert, max. If everything's urgent, nothing is
• •Less than 70% sure what kind of business it is? Treat the type as a hunch, not a fact
• •"No action needed" is an allowed answer. A strong site doesn't get filler

For the plumber, the CMO's review calls out the booking button, the quote form and the reviews as strengths. The slow phone load goes to the CTO, whose job it is. A separate model reads a screenshot of the homepage for layout and design, and it gets the same rules. To it, the 4.8-star badge is a badge, not a wall of reviews.

Each expert also gets a scoring brief

The guardrails cover what an expert can say. A separate brief covers how it scores. The CMO's, for example, spells out what to weigh:

The SEO lead and the CTO each get their own version. All three end on the same instruction: score from the measured data, not from a general feel for the site.

Stop 4: checks after the answer

Written rules get followed most of the time. Most of the time isn't good enough for something that moves a score, so every answer gets checked in code before it reaches your report:

• •Anything that doesn't match a shortlist item is dropped
• •What survives gets its title, priority and pain point reset to the detector's wording
• •At most five recommendations per expert make the plan, ten across all three
• •No more than four items marked high in the whole plan, with measured findings first in line
• •When an expert and a detector say the same thing, the detector's version wins
• •Near-duplicates across the three experts get merged
• •An "add more proof" item is dropped when the site's proof is already strong
• •Em dashes get stripped, because models write them even when the prompt says not to

So if "add customer testimonials" had somehow slipped past the instructions, it would die here twice. It isn't on any shortlist, and the proof check can already see five customer quotes on the page.

Stop 5: how far the expert can move the score

After all that, the expert's score gets 30% of the pillar and the measured score keeps 70%. An expert can nudge a pillar. It can't take it over:

Say the plumber's measured Conversion & Trust score is 72. The CMO sees the booking button, the form and the reviews working together, and gives it 82. The pillar lands at 75. A real nudge, earned by things that are actually on the page.

The one lever that only pulls down

There's one more way an expert affects the grade, and it only works in one direction. If the CMO or the SEO lead scores its area under 70, the overall score can't reach an A. Under 50, it can't reach a B. An expert who sees a real problem can keep a site from a grade it hasn't earned. It can't hand one out.

When an expert doesn't come back

Sometimes a review times out or fails. When that happens, the expert doesn't get a vote. The pillar runs on measured data and the report marks its confidence lower. A missing CMO or SEO review also keeps the score out of A territory. We'd rather under-grade on partial data than over-grade on a guess.

The one place they get to think

Everything above is about keeping the experts honest. There's one spot where we want their judgment on purpose: the headline recommendation. The CMO and the SEO lead each get to name one strategic move that isn't tied to the shortlist. It's what an experienced marketer or search lead would say after looking at the whole picture. It still follows every data rule and has to be different from the fixes already in the plan. It leads the Action Plan as that expert's take. For the plumber, the CMO's might be to put the booking button and the customer quotes in the same view on phones. The proof sits right next to the ask.

Why we bother

A score is only useful if it means the same thing every time and holds up when you go look at your own site. The 70% measured core gives it stability. The guardrails make sure the other 30% is judgment about your site, not about websites in general. When a recommendation can't be traced back to something we measured, we treat that as a bug, not a matter of taste.