- Home
- Skills
- E-commerce
- Shopify CRO Audit: every finding cites a crawled page
Works with the AI Agents you already use
Shopify CRO Audit: every finding cites a crawled page
A deterministic crawler captures a bounded set of public pages (home, products, collections, cart, content, robots, sitemap) into a manifest, and Gate A validates that manifest.
$14
Shopify CRO Audit: every finding cites a crawled page
Example session with this skill installed
Run a CRO audit on https://www.graza.co and give me a report I can send to the client. Only tell me things you can back up with the pages you actually crawled.
- Read your context and instructions
- Compiled the shopify cro audit
Summary
Crawled 14 public pages of graza.co (coverage: complete, 0 blocked), wrote the audit to the JSON contract, passed Gate B with 0 errors and 0 warnings, and rendered the client report below. Every experiment cites a captured page ID.
$ python3 run_audit.py --url https://www.graza.co --store graza --skip-write
== Crawling https://www.graza.co (store='graza', max_pages=15) ==
Captured 14 page(s) -> artifacts/graza/manifest.json
== Gate A: manifest ==
[OK] 0 error(s), 0 warning(s)
$ python3 harness/validate.py --audit sample_output/graza/audit.json --manifest artifacts/graza/manifest.json
[OK] 0 error(s), 0 warning(s)
$ python3 harness/render.py --audit sample_output/graza/audit.json --out sample_output/graza/audit.md
Wrote sample_output/graza/audit.md
graza audit
Executive summary
Graza has the hard part solved: a differentiated product story, a 4.9-star review base in the thousands, and a sitewide 25% promotion running at the time of the crawl. What leaks is the handoff between that story and the purchase. The buy box shows list prices while the discount is only promised in a top bar, the subscription widget shows cadence but never says what you save, and the cart page is an empty shell with a single 'Back to Shop' link. Each of these is a shopper standing at the decision point without the one number that would make them click.
The second leak is weight. The Drizzle product page ships 1.2 MB of HTML, 93 scripts, and 68 images with a third of them lazy-loaded, and the homepage autoplays five videos. None of that is visible in a crawl manifest as an error, but on a phone it is the difference between reading the reviews and bouncing.
The experiments below span all five pillars. The first four touch the buy box, cart, and subscription copy and can ship inside a week on the existing theme. The acquisition and performance work follows, and the competitor table points at the one pattern each rival does better that Graza can adopt without changing its voice.
Proposed experiments
| # | Pillar | Experiment | Evidence | Primary KPI | Lift | Conf. |
|---|---|---|---|---|---|---|
| a1 | Conversion | Surface the 25% sitewide discount in the PDP buy box, not only the top bar | pg_690c4505 /products/drizzle | PDP add-to-cart rate | +6-12% | 72 |
| a2 | Conversion | Replace the empty cart shell with a recovery module | pg_db805270 /cart | Empty-cart exit rate | +3-8% | 64 |
| a3 | AOV | Show a free-shipping progress bar in the cart drawer | pg_db805270 /cart | Average order value | +5-10% | 70 |
| a4 | AOV | Pre-select the Duo on the Drizzle PDP | pg_690c4505 /products/drizzle | Units per order | +8-15% | 68 |
| a5 | Retention | Make the subscription option explain its saving in words | pg_690c4505 /products/drizzle | Subscription take rate on PDP | +10-20% | 66 |
| a6 | Retention | Add a 'first bottle ran out?' reorder email with refill cans | pg_690c4505 (new flow) | 60-day repeat purchase rate | +5-12% | 58 |
| a7 | Acquisition | Add og:title and og:description to every product and collection page | pg_690c4505 /products/drizzle | CTR from social referrers | +3-7% | 75 |
| a8 | Acquisition | Turn the Culture blog into use-case landing pages | pg_c8450408 /blogs/culture | Organic sessions that add to cart | +4-9% | 60 |
| a9 | Performance | Cut product-page HTML weight below 400 KB | pg_690c4505 /products/drizzle | Mobile LCP on PDP | +2-5% | 62 |
| aa | Performance | Stop autoplaying five hero videos on the homepage | pg_71c4011e / | Homepage click-through to PDP | +2-6% | 61 |
exp-a1 — Surface the 25% sitewide discount in the PDP buy box
Hypothesis: Shoppers who reach the buy box see list prices ($21 Drizzle, $37 Duo) while the only mention of the 25% promotion is a top banner saying it is 'Applied in cart'; showing the discounted price next to the list price at the decision point reduces hesitation and cart abandonment.
Primary change: Control: list price in buy box, promo only in announcement bar. Variant: strike-through list price with the post-discount price and an 'Applied automatically' label inside the buy box.
Decision rule: Ship if PDP add-to-cart rate rises and checkout completion does not fall; revert if refund or support contacts about pricing rise.
exp-a3 — Show a free-shipping progress bar in the cart drawer
Evidence note: The crawl found the string 'free shipping!' only inside embedded JSON in the page scripts, never in rendered cart or PDP text, so the threshold is invisible to shoppers at the moment they decide quantity.
Hypothesis: A visible 'you are $X away from free shipping' bar in the cart drawer increases the share of orders that add a second bottle or refill can to clear the threshold.
Primary change: Control: no threshold messaging in the cart. Variant: progress bar with remaining amount and one-click 'add Sizzle' suggestion.
Decision rule: Ship if AOV rises by at least $3 with no drop in conversion rate; the $ threshold is the decision gate, not the lift.
exp-a5 — Make the subscription option explain its saving in words
Evidence note: Subscription markup exposes only cadence labels ('Every 2 Months', 'Every 3 Months'); no saving percentage or 'skip anytime' copy appears in the rendered text.
Hypothesis: Shoppers choose subscribe-and-save when the saving and the cancellation terms are explicit; the current widget shows cadence only, so the choice reads as a commitment without a reward.
Primary change: Control: cadence radio buttons. Variant: 'Subscribe & save X%, skip or cancel anytime' label plus a one-line note on how long a bottle lasts.
Decision rule: Ship if subscription share of PDP orders rises by at least 2 percentage points with no increase in 30-day churn.
exp-a7 — Add og:title and og:description to every product and collection page
Hypothesis: Captured product HTML carries og:image but no og:title or og:description, so shared links on iMessage, Slack, and social fall back to generic titles; complete social previews raise click-through on shared links and creator posts.
Primary change: Control: og:image only. Variant: og:title, og:description, og:url, and twitter:card on all product and collection templates.
Decision rule: Ship if social referral sessions per share rise; this is a one-line template change so keep it if neutral.
exp-a9 — Cut product-page HTML weight below 400 KB
Hypothesis: The captured Drizzle PDP document is 1.2 MB of HTML with 93 script tags and 68 images, of which only 24 are lazy-loaded; a lighter document improves Largest Contentful Paint on mobile, where most olive-oil DTC traffic arrives, and lifts conversion.
Primary change: Control: current template. Variant: inline review feed loaded on scroll, lazy-load all below-fold images, defer non-critical app scripts.
Decision rule: Ship if mobile LCP improves by at least 0.5 s and PDP conversion does not fall.
(The other five experiments carry the same fields in audit.json and audit.md.)
Competitor analysis
| Competitor | Domain | Positioning | What they make easier | Store edge | Pattern to adapt | Source |
|---|---|---|---|---|---|---|
| Brightland | brightland.co | Design-led California olive oil at a premium price | Gift sets and seasonal drops are the default landing experience | Graza is cheaper per bottle and leads with a squeeze-bottle utility story | Gift-set landing pages with clear 'who it is for' framing | manual |
| Kosterina | kosterina.com | Greek EVOO with wellness and polyphenol messaging | Subscription saving and cadence are stated in the PDP buy box | Graza has a much larger review base and a stronger freshness story | Explicit subscribe-and-save percentage next to the price | manual |
| California Olive Ranch | californiaoliveranch.com | Mass-market, grocery-first EVOO brand | Store locator and retail availability are prominent | Graza controls its DTC funnel and refill program | Recipe-to-product landing pages that rank for cooking queries | manual |
Technical checks
| Check | Status | Detail |
|---|---|---|
| SSL Certificate | Pass | All 14 captured pages were served over HTTPS without certificate errors. |
| HTTPS Redirect | Warn | Crawl started at an https URL; the http to https redirect was not exercised by this crawl. |
| Sitemap | Pass | sitemap.xml returns a valid index with product, page, collection, blog, and article sitemaps. |
| Robots.txt | Pass | robots.txt allows product and collection crawling and blocks cart, checkout, and account paths as expected for Shopify. |
| Critical Pages Loading | Pass | Homepage, four product pages, two collections, cart, and content pages all returned HTTP 200. |
| Meta Tags & Social Previews | Warn | Product pages carry a meta description and og:image but no og:title or og:description; shared links fall back to generic titles. |
| Structured Data | Pass | Product JSON-LD with Offer, AggregateRating, and OfferShippingDetails is present on the Drizzle PDP. |
| Favicon | Pass | Icon link tags are present on every captured HTML page. |
| Mobile-Friendly | Pass | Viewport meta tag is present and the comparison table declares horizontal scroll for small screens. |
| Page Speed Mobile | Warn | Drizzle PDP HTML is 1.2 MB with 93 script tags; no lab timing was measured in this crawl. |
| Page Speed Desktop | Warn | Homepage references 26 .mp4 sources across 5 autoplaying video elements; no lab timing was measured. |
| Broken Links | Pass | No 4xx or 5xx responses among the 14 captured pages. |
| Image Optimization | Warn | Product page images are served as jpg and png; no webp or avif variants were observed, and only 24 of 68 images are lazy-loaded. |
| Cookie/Privacy | Pass | Privacy Policy and Terms of Service links are present in the footer and a GDPR consent hook is referenced in page scripts. |
| Checkout Reachable | Warn | The /cart page loads but renders only an empty shell in the captured HTML; checkout itself is disallowed by robots.txt and was not crawled. |
Next steps
- Open sample_output/graza/audit.md (this report) and audit.json (the validated contract) in the package.
- Ship a1–a4 first on the existing theme; they touch the buy box, cart and subscription copy only.
- Re-run next month with the same command and diff the two audit.json files.
Connects securely to your tools. The creator never sees your data.
About this skill
What it does
Give your agent a storefront URL. A deterministic crawler captures a bounded set of public pages (home, products, collections, cart, content, robots, sitemap) into a manifest, and Gate A validates that manifest. Your agent then reasons over those artifacts only, writes the audit as JSON to a strict schema, and Gate B rejects the report if any experiment cites a page that was not captured, if the five pillars are not all covered, if dollar thresholds leak into lift estimates, or if confidence is padded above 90. Only a report that passes both gates is rendered to Markdown.
Why this instead of a prompt
Ask an agent to "audit my store" and you get ten confident paragraphs you cannot check. This harness makes fabrication structurally hard. The writer and the evaluator share one validator, so a claim without evidence does not render. Blocked pages are recorded as blocked, not guessed at. The output is a JSON contract first and Markdown second, which means you can diff audits month over month, feed them into a tracker, or hand them to a client with the evidence IDs attached.
Key features
- Bounded crawl (default 15 pages) with page typing, access status, and a scope summary; a blocked store is a valid record of a barrier, not an invented audit.
- Exactly 10 experiments, two per pillar, each with hypothesis, control vs variant, primary KPI, decision rule, expected lift and confidence.
- 15 canonical technical checks (SSL, sitemap, robots, structured data, social previews, image optimization, checkout reachable and more), each with an evidence page ID.
- Competitor table with source provenance for every row.
- Cross-record validators: evidence resolution, pillar coverage and balance, host mismatch, overconfidence, new-page evidence notes.
- Markdown renderer for a client-ready deliverable; a real sample audit of graza.co is included so you can see the exact output before you run it.
- Optional Anthropic-API writer for fully scripted runs; without a key your agent does the reasoning and writing itself.
Who it is for
Shopify agencies and freelancers who sell CRO audits, in-house ecommerce teams who want a repeatable monthly check, and founders who want a second opinion on their own store without hiring one.
How it is built
Python 3.10+ with one dependency (jsonschema). run_audit.py for the crawl and gates, harness/ for crawl, discovery, validation and rendering, schema/ for the two JSON Schemas and the contract documentation, a router SKILL.md plus reason and write sub-skills, an eval rubric, tests and fixtures.
License
Your purchase includes a license to use the skill commercially for your own store or your clients' stores; reselling or redistributing the skill itself is not permitted.
Known limitations
- Audits publicly reachable pages only. It does not log in, add to cart, submit forms, bypass bot protection or read analytics, and it never claims a recommendation caused a lift.
- Collection grids rendered client-side appear thin in the crawl; the report says so instead of guessing products.
- Page speed checks report document weight, script and image counts; no lab timing is measured.
- Screenshots are optional and best-effort (Playwright); they are never required for a valid audit.
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for E-commerce.
Audits publicly reachable pages only: no login, no add-to-cart, no form submission, no bot-protection bypass, no analytics access. Client-side rendered collection grids appear thin in the crawl and the report says so. Page speed checks report document weight and asset counts, not lab timings. Screenshots are optional and best-effort.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before buying a skill.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 27 days ago
- Software
Python 3.10+jsonschema
- Optional
- Playwright
Anthropic API key
Install jsonschema with pip install -r requirements.txt. Playwright is only for screenshots; the Anthropic key only for the scripted writer.
Permissions
Network access is plain HTTP GET to the public pages of the store the user names, capped at 15 pages by default; nothing is logged into, submitted or bypassed. File write is limited to artifacts/<store>/ and sample_output/<store>/ in the working directory. The optional Anthropic API call only happens if the user sets a key.