Methodology

How we find where AI can improve your store.

Every recommendation comes from what your store actually shows a visitor or an agent today — never a guess, and never anything behind a login. This page explains what we check and how each opportunity gets its rank.

01 — The premise

The honesty rule

Most audit tools have two answers: pass or fail. That forces a verdict even when nothing was actually verified — so “we didn't check this” quietly becomes “you're missing this.” We use four states instead, and the fourth one is the important one.

Already strong

We found real evidence of this on your page.

Room to grow

We found partial evidence — some of it is there, not all.

Big opportunity

We could test this, and found no evidence of it.

Can't tell from here

This isn't verifiable from the outside. We're not guessing.

“Can't tell from here” beats a confident wrong answer. We never call something a big opportunity just because we couldn't see it.

02 — What we grade

How a scan works

Five steps, about ten seconds.

  1. 01

    We load your page in a real browser.

    Not a plain download — a headless Chromium browser, the same engine as Chrome. That matters: things added by JavaScript after the page loads (chat widgets, search bars, personalization scripts, tracking pixels) are invisible to a simple fetch, but we see them the way a real visitor does.

  2. 02

    We check the page didn't lie to us.

    Some sites serve automated visitors a blocking wall, a decoy, or an empty stub instead of the real page. We detect that and try a second route to reach the genuine page. If we still can't read your site, we tell you — we don't report on a page we never actually saw.

  3. 03

    We run the deterministic checks.

    The majority of what we report is pattern-matching against your actual markup: which tools and vendors your page loads, what structured data it publishes, and specific markup signals. Same page in, same answer out — no opinion involved.

  4. 04

    We work out what kind of site this is.

    An online store and a documentation site need completely different advice, so we decide which you are before recommending anything. (see Stores vs. other websites below)

  5. 05

    We assemble the report.

    Findings are mapped onto opportunity areas, each with a plain-language explanation, a sourced benchmark, and real tools to explore.

03 — By method

What we actually check

Almost everything we report is deterministic — read directly from your page's markup, not inferred. Here's the full picture, by method.

Tools and vendors your page loads

We identify the scripts your site runs and match them against known providers, grouped by what they do:

Site searchAlgolia, Klevu, Constructor, Coveo, Searchspring, Bloomreach
Visual searchSyte, Slyce, ViSenze, Vue.ai
TranslationWeglot, Transifex, and others
RecommendationsDynamic Yield, Nosto, Rebuy, Certona, Emarsys
Email & retargetingKlaviyo, Mailchimp, Omnisend, and the major ad pixels
AR / try-onVntana, Zakeke, 3DLook, Avataar
Upsell & bundlingRebuy, Bold, ReConvert
Chat & supportIntercom, Gorgias, Tidio, Zendesk, Drift, Crisp, and others
ReviewsYotpo, Judge.me, Okendo, Loox, Stamped, Trustpilot
AdvertisingGoogle Tag Manager, Meta, TikTok, Pinterest
A/B testingOptimizely, VWO
Behavior analyticsHotjar, Microsoft Clarity, FullStory, Mouseflow
Fraud preventionSignifyd, Riskified, Forter, Sift

Structured data

We parse the Schema.org ↗ markup your page publishes — the vocabulary search engines and AI agents read to understand what a page is. We specifically look for Product, FAQPage, Organization, and BreadcrumbList. This drives your agent-readiness and AI-answer findings.

Markup signals

Specific, checkable properties of your HTML:

  • Meta descriptionpresent, and a usable length
  • Canonical linkset or not
  • Image alt textshare of images carrying real alt text
  • hreflang tagsmulti-language support
  • Search inputpresent, with genuine autocomplete — an attribute, a combobox role, or a known typeahead widget
  • Platform fingerprintShopify, WooCommerce, BigCommerce, Magento, Wix, Squarespace — left blank when it isn't identifiable

One AI-judged check

Exactly one check needs reading comprehension rather than pattern-matching: whether your product copy is specific and descriptive, or thin and generic. It's judged only from text actually visible on the page, and it's instructed to answer “can't tell” rather than guess when there's nothing to judge. Everything else on this page works with no AI at all.

04 — The limits

What we can't see

A scan looks at your site from the outside, as a visitor. Plenty of valuable things simply aren't visible from there — so we mark them “Can't tell from here” rather than inventing a verdict.

These are the capabilities we deliberately never guess at, because confirming them would require access to your backend, your data, or a real purchase flow:

  • Cart recoverydetecting checkout hesitation is backend behavior. A newsletter popup isn't proof of it.
  • Dynamic pricingpricing logic doesn't announce itself in your markup.
  • Personalized homepages and audience segmentswhat one visitor sees tells us nothing about what another does.
  • Predictive analytics and inventory forecastinginternal systems, invisible externally.
  • Agent copilots for support teamsan internal tool, not a public one.
  • Agentic checkoutwould require completing a real agent-driven purchase.
  • Presence in AI answersdepends on live AI systems we don't control or query.
  • Image enhancementwe can't tell a retouched photo from an original.

If one of these shows “Can't tell from here” on your report, it isn't a criticism. It means we respect the difference between absent and unverifiable.

05 — Two paths

Stores vs. other websites

An online store and a blog need different advice. So the first thing we decide is which one you are.

Before recommending anything, we check whether your site is actually an online store — using real evidence: whether you run a commerce platform, whether you publish product structured data, whether there's a genuine cart, and whether your page reads like a storefront.

If you're a store, you get the full personalized audit: nine opportunity areas checked against your actual page, plus a live before/after rewrite of one of your own products.

If you're not a store, we say so plainly — and show you general, high-impact ways AI can still improve almost any website: getting cited in AI answers, becoming readable by AI agents, content and accessibility, support, and findability. We'd rather tell you the store checks don't apply than run them anyway and report gaps that were never relevant to you.

06 — Strict rules

The before/after rewrite

When we show one of your products rewritten by AI, it plays by strict rules — because an “improvement” that isn't clearly one is worse than nothing.

  • Only your facts. The rewrite can use only information already on your product page. Invented materials, measurements, or numbers are automatically rejected by code, not just discouraged in an instruction.
  • Your voice, not generic marketing. We read how your brand actually writes — your headlines, your about page, your other product copy — and the rewrite has to sound like you.
  • No filler. A list of stock marketing phrases (“elevate”, “game-changer”, “must-have”) is blocked outright.
  • It has to name a real fix. The explanation of what changed must describe a concrete, shopper-facing improvement. Vague claims like “improved the tone” are rejected.
  • A second AI has to approve it. An independent review pass — whose only job is to say no — must agree the rewrite is genuinely better, on-brand, and keeps every real detail. We never let the AI that wrote it grade its own work.
  • If nothing clears the bar, we say so. Strong copy stays as-is and we point you at a real gap instead. We never downgrade good writing just to have something to show.

07 — Sourcing

Where our numbers come from

No stat appears on this site without a source you can open and check yourself.

Every benchmark figure links to its origin, and we've checked the number against the source's own wording rather than a secondhand summary.

We also draw a line most tools blur. Some figures come from independent research — Baymard Institute, Pew Research Center, Salsify, and similar. Others come from a vendor reporting on outcomes tied to its own product. Those are useful as direction, but they aren't neutral measurements, so we label them Vendor-reportedand phrase them as “Vendors report…” rather than stating them as fact.

A vendor-reported number is a signal, not a guarantee for your site. We'd rather show you the difference than hide it.

08 — What we build on

What we build on

Only things we genuinely use. If it's not here, we don't check it.

Schema.org

The structured-data vocabulary we parse from your page — it drives your agent-readiness and AI-answer findings.

WCAG 2.2

The accessibility standard real remediation is measured against. We check one signal from it — whether your images carry real alt text — not a full audit. Genuine accessibility work means real fixes, not an overlay widget.

Independent research

Baymard Institute, Pew Research Center, Salsify, Epsilon, Optimizely — the studies behind our benchmark stats.

Emerging agent standards

ACP and UCP are the open standards forming around AI-agent checkout. We reference them as context for where commerce is heading — we don't test your site against them.

09 — Scope

Where this stops

A ten-second scan is a starting point, not an audit. Here's what it isn't.

  • We read your homepage, not your whole site. For stores, we also read one product page for the rewrite. Findings reflect those pages.
  • It's a snapshot. Your site changes; this reflects the moment we looked.
  • Some sites block automated visitors. If yours does, we'll tell you we couldn't read it rather than report on a page we never saw.
  • Detection proves presence, not quality. Knowing you run a search tool doesn't tell us how well it's configured.
  • This isn't a score. There's no grade out of 100 anywhere on this site, by design. It's a map of where AI could help — not a judgment of your work.

See where your store stands.

A full scan takes about ten seconds and reads only your public pages.

Run my scan →
AI Website Improver. Run a scan