
- I build a store with an AI agent in 15 steps and 4 phases: foundation, store, visibility, quality. Every step is a prompt to copy plus an acceptance list you check yourself before you paste the next one.
- My agent is Hermes on Aderlo Cloud with an OpenAI model connected through a ChatGPT account. Hermes takes 2 GB of RAM from the plan's pool, and app capacity starts with the Growth Store plan at about $32 a month before VAT.
- I pick the model for the job: the strongest Sol model for decisions and code, Luna for daily routine. On the Codex pricing page of 8 Nov 2026, the Plus plan allows about 20 times more messages for GPT-6 Luna than for GPT-6 Sol.
- In 5 steps the agent has to stop and ask you: brand, categories, import pilot, the TOP 3 products and the guide topics.
- The acceptance targets are numbers: category classifier coverage of at least 95%, filters under 300 ms, isitagentready 100%, PageSpeed 95–100 in 4 categories on mobile and desktop.
Contents · 9 sections
This list lived in my notes as a few lines I pasted to an agent for every new store. It worked because I knew the rest by heart: the supplier will not send categories, a block-based cart will lose the order bump, speed is measured last. Someone building a first store does not know that, and neither does an AI agent. So I turned it into 15 prompts with acceptance criteria.
This article is for someone who has a domain and a WordPress store with WooCommerce, or is about to set one up, and wants an AI agent to do the work. Enter your store data at the top and the prompts below fill themselves in. Your progress and ticks stay in your browser, so you can come back to the list a week later.
How do you build an online store with AI, step by step?
I build an online store with AI in 15 steps: the agent gets one prompt at a time, does the work on the server and in the browser, and I check the result against the acceptance list before I paste the next one. As of 8 Nov 2026 my five stores in Poland hold 58,670 products (read from the stores' public APIs), and this list grew out of the work on them.
My stores run on GAPLI, a dropshipping platform for online stores. GAPLI is my company, so read the disclosure below with that in mind.
GAPLI is built for Polish stores, suppliers and marketplaces, so it is not the right choice for most readers here. The prompts do not depend on it: you enter your own catalog source (a wholesaler feed, a dropshipping app or your own stock) in the fields below, and steps 2, 3 and 5 work with that. If you plan to sell in Poland, you can look at GAPLI.
- One step at a time. Paste a prompt, wait for the agent's report, check the acceptance list, and only then move on. Pasting everything at once ends with the agent skipping the decision points.
- Step 0 happens once. The agent creates STORE-BRIEF.md and LOG.md. Every later prompt tells it to read them, so you can start a new chat without explaining everything again.
- Decisions are yours. Brand, categories, prices and product picks are your decisions. Prompts marked "decision point" stop the agent until you make them.
- Quality is measured last. Measurement, articles and scripts add weight, which is why agent readiness and PageSpeed sit in phase D, and you repeat step 14 every month.
Your data and ticks stay in this browser only; nothing is sent to our server. An empty field shows up in red in the prompts, e.g. [DOMAIN].
Which AI agent should you use to build a store?
For this list you need an agent with access to the server, a browser and a task scheduler. I use Hermes running on Aderlo Cloud, with an OpenAI model connected through a ChatGPT account: on that setup, agents run the blogs of three of my stores and published 28 guides in them between 27 Oct and 3 Nov 2026.
Hermes is an open-source agent program from Nous Research under the MIT license. On Aderlo Cloud, which is my company, you install it from the app catalog and it runs in its own container on a separate machine, so the agent never takes capacity away from your store.
What matters for this list: once you connect your hosting to the agent in the panel, Hermes talks to it through the Aderlo MCP server. It reads and writes files, runs WP-CLI, adds DNS records, turns on DNSSEC and takes a backup before a change by itself. Steps 8 and 12, where a TXT record has to be added or the zone signed, therefore happen without you clicking through the panel. Store-related roles such as sales analytics are switched on with a button.
- A plan with app capacity. Apps run only on plans that carry a pool on the app node. The smallest is Growth Store: about $32 a month before VAT, with an app pool of 4 GB RAM, 200% CPU and 20 GB of disk. Hermes takes 2 GB RAM, 100% CPU and 10 GB of disk from it (price list and app catalog checked on 8 Nov 2026). Starter Store has no app pool.
- Install from the catalog. In the panel, choose Apps and Hermes Workspace, set an address on your own domain, and the login details wait for you in the panel. See the plans here: Hermes on Aderlo Cloud (partner link).
- Connect OpenAI. On the instance page, open the terminal and click "Choose the AI model". The panel runs
hermes setup model: pick OpenAI Codex and sign in with your ChatGPT account using the code the terminal shows. Sign in from the terminal, not from the Workspace settings screen: that is how we connect our own instances. - Choose the model. From the list the terminal shows, take a model for the job using the table below. Pin the model in scheduled jobs.
- Step 0 of the list. Enter your store data in the fields above, copy the first prompt and paste it to Hermes in the chat.
Why OpenAI through a ChatGPT account?
Connecting OpenAI through a ChatGPT account works best because one subscription gives the agent both the text model and image generation, at a fixed monthly cost. Codex, which Hermes uses to reach OpenAI models, is included in the ChatGPT Free, Go, Plus, Pro, Business and Enterprise plans (Codex pricing page, 8 Nov 2026). With an API key you pay for every use separately.
Hermes creates the guide graphics for steps 10 and 11 with the same account, without a separate API key. If Hermes asks for an image key, pick OpenAI (Codex) in its image settings by hand: in our setup the auto-detection looked for another provider's key. Hermes also accepts other providers, for example through OpenRouter, but then you pay separately for text and for images.
| Agent work | Steps | Model | Why |
|---|---|---|---|
| Research, decisions, code for the cart, filters, agent readiness and speed | 0–2, 4–7, 9, 12–13 | the newest Sol on the list (8 Nov 2026: GPT-6.1 Sol, 15–160 messages per limit window) | A mistake in categories or the cart costs more than the limit you use up |
| Catalog import | 3 | Sol | The agent writes the import script once; the server, not the model, processes thousands of products |
| Measurement, SEO and the first three guides | 8, 10 | Sol | This is where the standard the automation will copy is set |
| Daily automation, reports, monthly review | 11, 14 | Luna pinned in the job (GPT-6 Luna, 350–3,000 messages per limit window) | Routine does not eat the limit you need for real work; if Luna's articles fail the quality gate, pin Sol for the writing only |
| Guide graphics | 10–11 | image generation from the same account | No separate API key |
Our store agents run on GPT-5.6 Sol (as of 27 Oct 2026). If the newest Sol is not on your terminal's list yet, take that one. Do not pick a model by its name in an article; pick it from the list your terminal shows, because availability depends on your ChatGPT plan and the Hermes version.
A ChatGPT plan limit stops the agent halfway through the work. For one of our customers, scheduled jobs burned the weekly limit of a Plus plan (14 Sep 2026) and the agent waited about 4.5 days for it to reset. For the import and the daily automation, take a higher plan or an API key. Pin the model in scheduled jobs: on 8 Sep 2026 one model stopped answering through Codex in the middle of the day, and the agents using it stood still until they were switched.
When is Hermes on Aderlo not the best choice? If you already have an agent on your own computer, for example Claude Code with SSH access to your server, the prompts work the same and you do not need to buy anything. The same goes for a store on Shopify: the steps that need a server and WordPress plugins do not apply there, and Hermes will not change that. I described how we run Hermes in detail on my Polish site, and building a first agent in how to build an AI agent.
What does an AI agent not know about a store built on a supplier feed?
An AI agent does not know that supplier data is incomplete and that some store modules work only in classic WooCommerce. On the platform my stores run on, the supplier catalog had empty categories in 9,000 of 9,000 checked items (2 Sep 2026). I wrote each of these pitfalls straight into the prompt it belongs to.
| Pitfall | What I measured | Where in the prompts |
|---|---|---|
| A catalog with no categories | 9,000 of 9,000 items from 9 suppliers without a category (2 Sep 2026) | Step 2: a classifier from name, description and brand; a product without a category is not imported |
| Duplicates and mixed-language names | 24% duplicate titles and 57% English names in a Polish store (catalog audit) | Step 3: one listing per GTIN/EAN, one title pattern |
| Tax set to 0 by a localization bug | about 50 stores on the platform, fixed in plugin version 1.9.12 | Step 0: version and settings audit before launch; step 14: monthly review |
| Each extra item charged as a full parcel | about $3.25 for every additional item | Step 6: shipping for 2 and 3 items checked against the carrier price list |
| Cart-to-order bottleneck | 14% in an adult-industry store (24 Oct 2026); conversion of 1.76% after the payment fix | Step 6: cart, checkout, order bump and thank-you page as a separate step with acceptance |
| A large catalog, few Googlebot visits | at most 87 of 20,071 product pages crawled in 4.6 days (31 Oct–4 Nov 2026) | Step 9: crawl budget, split sitemaps, robots.txt |
| A blog outside the index | 0 of 22 blog posts indexed by Google (22 Oct 2026) | Steps 10–11: links from categories to guides, indexing requests, a Search Console report |
As of 8 Nov 2026 my five stores in Poland hold 58,670 products: rozkoszownia.pl 20,896, domiogrodstyl.pl 19,618, smartinghome.pl 10,023, zdrowinka.pl 5,937 and akwasfera.pl 2,196 (stores' public APIs, read on 8 Nov 2026 at 12:20 CET). The general path from idea to first sale is in how to start an online store.
Phase A: how do you prepare the agent, the brand and the categories?
Phase A is three prompts after which the agent knows the state of the store, the market and the buyers, and the store has a brand and a category tree with a classifier. Nothing goes into the catalog yet: two of the three steps end with a decision point where the agent waits for your approval.
Project card and working rules
The agent audits the store and creates two files it uses in every step that follows.
- A domain registered and pointed at your hosting
- WordPress with WooCommerce and your catalog source connected
- An agent with access to the server and a browser, e.g. Hermes on Aderlo Cloud with an OpenAI model connected (described above)
- A browser where you are signed in to Google, Bing and your supplier account
You are the lead builder of my online store. We work step by step through a checklist. This step sets up the common ground for every step that follows. PROJECT DATA - Domain: [DOMAIN], registered to [COMPANY], pointed at: [HOSTING]. - Seller (footer, terms, structured data): [COMPANY]. - Niche: [NICHE]. - Platform: WordPress + WooCommerce; the catalog comes from [CATALOG SOURCE]. - Target catalog: about [PRODUCT COUNT] products. TASK 1. Audit the starting point and change nothing yet: WordPress, WooCommerce and PHP versions, theme, plugin list (especially the catalog and import plugins), cache layer, connection to [CATALOG SOURCE], HTTPS, email on the domain, tax settings (check that the tax rate is not 0), number of products and categories, backups. 2. Create two files in the project folder: • STORE-BRIEF.md: project data, starting point and the working rules below. You extend it in later steps with the brand, categories and decisions. • LOG.md: after every step you add the date, what was done, how it was verified (evidence), what is left and the risks. 3. Show me the audit report and a list of gaps that block the next steps. WORKING RULES (apply to every step) - Before any larger change, take a backup and write down how to restore it. - Do not edit the supplier or import plugins, or WordPress and WooCommerce core: an update will overwrite your changes. Work in a child theme or your own mu-plugin. - Every promise on the site (delivery time and cost, returns, warranty, discreet packaging) must be backed by the terms and the store settings. You do not invent reviews, numbers, awards or certificates. - After every change, clear the cache and check the result on the live site in a browser: desktop 1440 px and phone 375 px, console without errors. - I type passwords myself; do not ask for them in chat. You do not place a test order on production without my consent. - When a step needs my decision (brand, categories, prices, product picks), stop and ask. - End every step with a report: what was done, evidence, what is left, risks.
Niche research, store plan and brand
The look and layout of the store come from market and buyer research, not from a random template.
First read STORE-BRIEF.md and LOG.md in the project folder and follow the working rules written there. We are building a professional store, [DOMAIN], in the niche: [NICHE]. Before any page exists, I need research, a plan and a brand. 1. MARKET RESEARCH - Find the 8–10 strongest stores in this niche in my market (Google, Amazon, eBay, Google Shopping). For each: positioning, price range, strengths, gaps, and what the product page and cart look like. Give links. - Buyers: 2–3 personas, each with its main worry before buying. We will use them in the cart and checkout. - Market gaps: where we can be better (range, price, delivery, advice, convenience). 2. STORE PLAN - Site map: home, catalog, categories, product page, cart, checkout, thank-you page, guides, about, contact, shipping, returns, terms, privacy policy. - For every template: purpose, sections from top to bottom and the main call to action. 3. BRAND - 3 brand directions: brand name (it can be the domain), tagline, tone of voice, a palette of 5–6 colors with WCAG AA contrast, typography. - Check that the name and mark do not clash with existing brands in the niche (Google plus the USPTO, UKIPO and EUIPO trademark databases). Tell me what you checked. - After I choose: an SVG logo (horizontal, symbol, version for dark backgrounds), favicon and app icons, a 1200×630 share image and a short brand book in STORE-BRIEF.md (colors, fonts, tone, sample headlines). STOP: show the research, the plan and the 3 brand directions. Logo work and rollout start only after my decision. At the end: a report (what was done, evidence, what is left, risks) and an entry in LOG.md.
Category tree and classifier
Every one of thousands of products lands in the right category before anything is imported.
First read STORE-BRIEF.md and LOG.md in the project folder and follow the working rules written there. Design a category tree for about [PRODUCT COUNT] products from [CATALOG SOURCE] in the niche: [NICHE]. IMPORTANT: supplier feeds often arrive without usable categories. Work out the categories yourself from the name, description and brand. A product we cannot assign does not go into the default category or a catch-all bucket; we do not import it until it is classified. 1. SUPPLIERS: review the suppliers available in [CATALOG SOURCE] for this niche: items in stock, margin, delivery time, quality of photos and descriptions, minimum order quantities. Propose which ones we use and why. 2. TREE: 8–12 main departments, at most 3 levels, every category with at least ~20 products (merge smaller ones). Names people actually search for in Google (check the search suggestions), short lowercase slugs. 3. CLASSIFIER: for each category, words and patterns in the name, exclusions and examples. The name decides; the description only breaks ties. Test on 500 random products and show me: coverage in %, 30 random assignments for me to check, and the list of unclassified items. 4. CATEGORY CONTENT: for each category an H1, meta title, meta description and a 2–3 sentence description. STOP: show the tree with product counts and the classifier test result. We start the import after my approval. Save the tree in STORE-BRIEF.md. At the end: a report (what was done, evidence, what is left, risks) and an entry in LOG.md.
Phase B: how do you stock the store and build the purchase path?
Phase B turns a supplier catalog into a store: an import with a 300–500 product pilot, templates, filters, the cart with checkout, and three products your marketing will lean on. It is the longest phase and the one where errors invisible in the logs happen most easily, such as an order bump that vanishes in a block-based cart.
Before you set prices, check how much is left per order in dropshipping profit margin.
Importing the catalog
The catalog is complete, clean and free of duplicates, and the sync does not undo your fixes.
First read STORE-BRIEF.md and LOG.md in the project folder and follow the working rules written there. Stock the store with products from [CATALOG SOURCE] according to the approved category tree and supplier list. Target: about [PRODUCT COUNT] items. SELECTION - Only products in stock, with a photo and a sensible name. No items with a minimum order quantity above 1. - No duplicates: the same GTIN/EAN from several suppliers is one listing (the better price, stock and delivery time wins). - Only products assigned by the classifier. Categories come from our tree in replace mode, not appended to whatever category the import batch had. - Selling price by a margin rule you propose, justified against market prices. DATA QUALITY - Titles in plain English, no supplier codes or decorations ("NEW!!!"), one pattern: product type + brand + model + key feature. - Descriptions and short descriptions without supplier addresses, names and links. This is a common import bug; check it with a search. - Photos load and carry no supplier watermarks. PROCESS 1. Pilot: 300–500 products from several categories. STOP: show me 20 random product pages to check. You continue after my approval. 2. Full import in batches, with progress and an error log. 3. Report: items in the feed, imported, rejected and why (table of reasons), category counts, 30 random products to check, list of duplicate titles. 4. Check how the later price and stock sync works and whether it overwrites our categories or corrected titles. If it does, protect them and describe how. At the end: a report (what was done, evidence, what is left, risks) and an entry in LOG.md.
Home page, catalog and product page
Templates in the brand style, tested on real products, on desktop and on the phone.
First read STORE-BRIEF.md and LOG.md in the project folder and follow the working rules written there. Roll out the look and templates of the store according to the approved brand and plan. Work in a child theme, on the real imported products. HOME PAGE - Sections from the plan, a hero with one clear call to action, department tiles, a slot for the TOP products (step 7) and guides (step 10). CATALOG AND CATEGORIES - Product grid: photo on a white background in fixed proportions (no layout shift while loading), name on 2 lines, price, availability, "Add to cart" button. - Lazy loading: "Show more" or infinite scroll, while keeping normal page links (?page=2) so Google sees the whole catalog. Photos in the first row load immediately, the rest lazily. - Sorting (featured, price low to high and high to low, newest), breadcrumbs, category description, product count across the whole subtree, a slot for filters (step 5). PRODUCT PAGE - Zoomable gallery, name, price, availability, delivery time and cost (from the store settings, not hard-coded), buy button, 3 short guarantees. - On phones a sticky bar with the price and the button. - Specs pulled from the description into a readable table, description, similar and complementary products. - One Product structured-data block (check that plugins do not duplicate it) with price, availability, shipping and return policy, plus BreadcrumbList. REQUIREMENTS - Nothing overlaps at any width from 320 to 1920 px, no horizontal scrolling, buttons on phones at least 44 px, WCAG AA contrast. - Show screenshots of every template on desktop and phone. At the end: a report (what was done, evidence, what is left, risks) and an entry in LOG.md.
Catalog filters
A buyer narrows thousands of products down to a few, and every result really has the chosen feature.
First read STORE-BRIEF.md and LOG.md in the project folder and follow the working rules written there.
Build professional catalog filters for [DOMAIN].
CONTEXT: supplier feeds usually lack product attributes. Derive filter values from names and description lines (e.g. "Material:", "Color:", "Power:", "Dimensions:") and store them in your own index, which rebuilds itself after each catalog sync.
1. SELECTION: for every department 4–8 filters buyers actually use (compare with the competitors from step 1). Always: price (slider plus from/to fields), availability, brand; then niche-specific features. Show the list with coverage, i.e. the % of products that have the feature. A feature below 30% coverage does not become a filter.
2. BEHAVIOR: no page reload, a result count next to every value, values with no results greyed out, several filters at once, an active-filter bar with × and "Clear all". Filters work together with sorting and lazy loading. The browser Back button restores the previous state.
3. URLS AND SEO: filters stored in URL parameters (shareable), filtered pages are noindex, follow with the category as the canonical URL. Price and stock calculated live, not from a stale index.
4. PHONE: filters in a slide-out panel with a "Show X products" button.
5. TESTS: 3 samples for every filter. Open several results and check that they really have the selected feature. Measure the response time (target: under 300 ms). Show the samples and the errors found in the report.
At the end: a report (what was done, evidence, what is left, risks) and an entry in LOG.md.Cart, checkout, order bump and thank-you page
The customer feels confident, has every piece of information and is guided step by step to payment.
First read STORE-BRIEF.md and LOG.md in the project folder and follow the working rules written there.
Build or rebuild the cart, checkout and thank-you page so the customer feels confident and is guided step by step to payment, on desktop and on the phone.
FOUNDATION: use the classic WooCommerce cart and checkout (shortcodes) unless every module you need supports the block version. Many cart add-ons (upsell shelf, free-shipping bar, checkout steps, order bump, abandoned-cart consent, pickup-point picker) hook into the classic templates; on the block version they disappear without any error in the logs. Check each one.
CART
- A header with the item count and a "Continue shopping" link.
- A "You are $X away from free shipping" bar with the threshold taken from the settings, not hard-coded.
- Line items as cards: thumbnail, name, price, a −/+ stepper that recalculates the cart by itself (no "Update cart" button), remove.
- An "Add to your order" shelf with products that complement what is in the cart. No shelf is better than random bestsellers.
- A sticky summary: subtotal, carriers as cards with price and delivery time, "Total", a "Proceed to checkout" button, 3 short guarantees.
- Phone: a bar along the bottom edge with the total and the button.
CHECKOUT
- One set of progress steps (not two: one from the theme and one from a plugin).
- Above the form, a banner that answers the main buyer worry from the personas in step 1.
- Three numbered sections: ① details (email and phone at the top, correct field types and autocomplete), ② delivery (the same carrier cards as in the cart, plus a pickup-point picker if you offer one), ③ payment (express wallets such as Apple Pay, Google Pay or PayPal visible right away).
- Below payment, in this order: order bump, reminder consent, terms, "Place order and pay" button, guarantees. The order bump must not sit between the consents and the button.
- The right column holds only the sticky summary. On phones the summary sits above the form, collapsed ("Show 2 items").
- Guest checkout without creating an account.
ORDER BUMP WITH PREVIEW
- One relevant complementary product at an attractive price (up to about 20–30% of the cart value), added with one click, no reload.
- Hovering (tapping on phones) "View product" opens a modal with the gallery, price, short description and an "Add to order" button. The customer never leaves checkout. Esc, a click on the backdrop and × close it; focus stays inside the modal; everything works from the keyboard.
THANK-YOU PAGE
- Order number, payment status taken from the real payment state (not only the order status), summary, what happens next (confirmation, shipping, delivery), contact, optionally an account or the newsletter. Its own, polished design.
VERIFICATION in the browser, 1440 px and 375 px
- "+" recalculates the cart, the stepper is still there afterwards, the shipping bar shows the new amount.
- The shelf and the order bump really add the product.
- Shipping for 2 and 3 items matches the carrier price list (each extra item is not charged as a separate parcel).
- Express wallets are visible, the pickup-point map opens (if used), the preview modal works with mouse, touch and keyboard.
- Nothing overlaps, no horizontal scrolling, console without errors.
- Lighthouse for cart and checkout on mobile: performance at least 90, accessibility at least 95.
- Show screenshots. A test order only with my consent, preferably in the payment provider's test mode.
At the end: a report (what was done, evidence, what is left, risks) and an entry in LOG.md.TOP 3 products and one-page landing pages
Three products with the best offer against the market get their own sales pages and a place on the home page.
First read STORE-BRIEF.md and LOG.md in the project folder and follow the working rules written there. Pick 3 products to base the store's marketing on and build a sales page (one page) for each. 1. ANALYSIS: rank 15 candidates from the catalog, scored on: margin per unit, price against the market (Google Shopping, Amazon, eBay, the competitors from step 1; give links and the date checked), stock depth and supplier reliability, demand (Google keywords, seasonality), photo and description quality, return risk, and bundle and add-on potential. Show the table and a TOP 3 recommendation with reasons. 2. OFFER for each of the TOP 3: what pulls buyers in (price, free gift, bundle, free shipping, longer warranty) and what it costs us per order. Legally, a "was" price must be a real former price: in the US the FTC Guides Against Deceptive Pricing apply, in the UK the CMA rules on unfair commercial practices, and in the EU a price cut must show the lowest price from the previous 30 days. No fictitious strike-through prices and no countdown timers that reset. STOP: show the ranking and the offers for approval before you build the pages. 3. ONE PAGE (its own URL, e.g. /product-name/): a hero with the product, price and "Add to cart", benefits, who it is for, specs, comparison with alternatives, what is in the box, shipping and returns, FAQ, a sticky button on phones. Add to cart without leaving the page, Product structured data. 4. HOME PAGE: a section or slider with the TOP 3 (previous/next buttons, pause, no auto-rotation that pulls content away) plus links from the categories the products belong to. At the end: a report (what was done, evidence, what is left, risks) and an entry in LOG.md.
Phase C: how do you connect measurement, SEO, GEO and guides?
Phase C makes the store measurable and findable: Tag Manager and GA4 with the full e-commerce funnel, Search Console and Bing, technical SEO, GEO, the first three guides with products picked live, and an automation that publishes one article a day after a quality gate.
How to raise the chance that your store shows up in ChatGPT answers is covered in how to get cited by ChatGPT.
Measurement: Tag Manager, Analytics, Search Console, Bing
Every step of the purchase path is measured, and Google and Bing see the store from day one.
- The GTM container ID and the Google verification TXT record (enter them in your store data above)
- GA4, Search Console, Tag Manager and Bing Webmaster Tools accounts created
- Signed in to Google and Bing in the browser the agent uses
First read STORE-BRIEF.md and LOG.md in the project folder and follow the working rules written there. Connect measurement for [DOMAIN]. The Google Analytics 4, Google Search Console, Google Tag Manager and Bing Webmaster Tools accounts exist. I am signed in to the browser you can use. I type passwords myself. 1. TAG MANAGER: install container [GTM-ID] with the standard snippet: the script in <head> and the <noscript> right after the opening <body>. The container must be on the page exactly once. Check that no plugin adds a second container or a separate gtag.js for the same GA4 property. 2. GOOGLE VERIFICATION: add the TXT record [GOOGLE TXT RECORD] to the DNS zone of [DOMAIN] through the hosting panel, in the zone that is actually live (check the domain's name servers). In Search Console create a "Domain" property. 3. CONSENT: a cookie banner with Consent Mode v2 (analytics and ads denied by default for visitors from the EEA and the UK, updated after their choice), lightweight and keyboard-accessible. The privacy policy lists the tools that actually run. 4. GA4 E-COMMERCE through dataLayer and GTM: view_item_list, select_item, view_item, add_to_cart, remove_from_cart, view_cart, begin_checkout, add_shipping_info, add_payment_info, purchase. Purchase fires once per order (deduplicated by transaction_id), with value, currency and the item list. Mark purchase as a key event. Add the payment gateways to the list of unwanted referrals. 5. SEARCH CONSOLE: submit the sitemap with its full URL and inspect the home page, a category and a product. 6. BING: import the site from Search Console, submit the sitemap, turn on IndexNow (key file on the server, a ping on every publication and product change). 7. EVIDENCE: GTM preview (Tag Assistant) and GA4 DebugView. Walk the path from a category to checkout and show the events with their parameters. Also check a page served from the server cache, because caching can drop events. At the end: a report (what was done, evidence, what is left, risks) and an entry in LOG.md.
Technical SEO and GEO
Search engines and AI assistants understand the store, its offer and the company behind it.
First read STORE-BRIEF.md and LOG.md in the project folder and follow the working rules written there. Do professional technical SEO and GEO (visibility in answers from ChatGPT, Perplexity, Gemini and AI Overviews) for [DOMAIN]. TECHNICAL SEO - Title and meta description templates for categories, products and guides: unique, with the keyword, not truncated. - One H1 per page, canonical URLs, pagination. Filter, search, cart and checkout pages are noindex, follow. - Sitemaps: indexable pages only, split (products, categories, guides), with lastmod. - Crawl budget with [PRODUCT COUNT] products: block technical parameters in robots.txt (cart, sorting, ?add-to-cart) and check in the logs that Googlebot does not waste visits on wp-json and technical URLs. - Structured data: Organization ([COMPANY] details), WebSite with search, BreadcrumbList, Product with offer, shipping (OfferShippingDetails) and returns (MerchantReturnPolicy). One block per page, no duplicates from plugins. Test in the Rich Results Test. - Internal linking: menu and footer to departments; categories, guides and products link to each other; no orphan pages. - Content: unique descriptions for departments and the main categories. A list of duplicate product descriptions and a plan to fix them, starting with the products with the most potential. - Clean-up: image alt text, 404s, redirects, one URL version (https, with or without www), no redirect loops. GEO - An "About" page and a footer with full seller details ([COMPANY]: address, company number, contact). A clear company identity for AI and for customers. - Shipping, returns and FAQ pages answer questions directly: the first sentence is the answer. - robots.txt does not block AI search crawlers (OAI-SearchBot, ChatGPT-User, PerplexityBot, Claude-SearchBot) and the server firewall does not answer them with 403 or 429. Check the logs. Leave the decision on model training on the store's content (GPTBot, Google-Extended) to me. - llms.txt with a description of the store and its main departments. REPORT: list of changes, test results (Rich Results, URL inspection in Search Console) and what is left for later. At the end: a report (what was done, evidence, what is left, risks) and an entry in LOG.md.
Buyer problems and the first three guides
The articles answer real customer questions and show matching products from the store by themselves.
First read STORE-BRIEF.md and LOG.md in the project folder and follow the working rules written there. We are launching the guides section of [DOMAIN]. Research first, then 3 articles. 1. RESEARCH: what buyers in the niche [NICHE] search for and struggle with. Sources: Google suggestions and "People also ask", Reddit and forums, questions under Amazon listings, competitor product reviews (recurring complaints), keywords from SEO tools. Collect at least 50 problems and questions, group them into topics and rate: intent (guide, comparison, purchase), whether we have products in the store that solve it, and how strong the competition in the results is. 2. Propose a bank of 30 topics (save it in STORE-BRIEF.md) and pick 3 to start with the best chance of traffic and sales. STOP: I approve the topics before you start writing. 3. ARTICLES (each): - At the top a "Short answer": 3–4 sentences that answer the question on their own. This is the passage AI assistants quote. - Something of its own that competitors do not have: a calculation, a comparison table, buying criteria, costs. Numbers with sources; you invent nothing. - H2 and H3 headings built around questions, an FAQ at the end, the author and update date, links to categories. - Dynamic product blocks: a shortcode or block that picks products from the store live (category, keywords, price range) and shows only items in stock, with the current price and an add-to-cart button. Do not paste products in by hand: when stock changes, the block shows others by itself. When there are no relevant products, the block hides. 4. GRAPHICS: every article gets its own cover and 2–3 in-text graphics matched to the topic (diagram, infographic with data from the text, comparison), in the brand style, e.g. made with the OpenAI image generator from the account connected to the agent. No repeated stock photos. Text in images without typos, WebP format, dimensions in the attributes, descriptive alt text. The cover doubles as the share (OG) image. 5. Publication, links from categories to the guides, submission for indexing (Search Console and IndexNow). Show screenshots. At the end: a report (what was done, evidence, what is left, risks) and an entry in LOG.md.
Worker: one guide a day
A guide to the step 10 standard goes out every day, and a weak one stays in drafts.
First read STORE-BRIEF.md and LOG.md in the project folder and follow the working rules written there.
Build an automation (a worker) that publishes one professional guide on [DOMAIN] every day, to the standard from step 10.
1. TOPIC QUEUE: from the topic bank, topped up every week with new research. Before writing, the worker checks that the topic does not duplicate an existing article (keyword cannibalization).
2. PRODUCTION: research, draft, facts with sources, dynamic product blocks, unique graphics, meta and structured data, internal links (also from 2–3 older articles to the new one).
3. QUALITY GATE before publishing, checked automatically: length and structure, a "Short answer" present, every product block shows at least 3 products in stock, links work, no supplier names and no unbacked promises, numbers have sources, graphics exist and have alt text. A failed gate means a draft and a notification to me, not a publication.
4. SCHEDULE: one publication a day at a fixed hour, with the publication date set explicitly (nothing goes out early). Brake: at most 1 article a day, stop after 3 errors in a row.
5. AFTER PUBLISHING: cache purge, IndexNow, log entry.
6. WEEKLY REPORT (email or file): published articles, which of them got visits according to Search Console, what to improve.
7. Give me the monthly cost (model API, graphics) and where the automation runs (server cron or the agent's schedule). If it is the agent's schedule (e.g. Hermes cron jobs), pin the model in the job and use a lighter model for routine runs so they do not burn the plan's usage limit. Do a dry run on a draft and show me the result before you switch it on.
At the end: a report (what was done, evidence, what is left, risks) and an entry in LOG.md.Phase D: how do you reach isitagentready 100% and PageSpeed 95–100?
Phase D measures quality at the end, because earlier steps add scripts and content that change the score. The target is 100% on isitagentready.com and 95–100 in each of the four PageSpeed categories, on mobile and desktop. You repeat step 14 every month, because the daily automation changes the site every day.
The cart and checkout are noindex on purpose, so PageSpeed deducts points from their SEO category. There the target is performance and accessibility of at least 95, and SEO is skipped for that reason. The other five pages from step 13 aim for 95–100 in every category.
AI agent readiness: isitagentready 100%
A full 100% on isitagentready.com, built from real, working services.
First read STORE-BRIEF.md and LOG.md in the project folder and follow the working rules written there.
Goal: a full 100% on https://isitagentready.com/ for [DOMAIN], tested. Rule: only real, working things. Every document describes a function that exists. No fake OAuth metadata and no cards for services that do not exist.
1. Run a baseline scan and save the result.
2. Implement in an mu-plugin (not in the supplier or import plugins):
- robots.txt with rules for AI bots, Content-Signal in every group and a Sitemap line. Set the robots.txt Cache-Control to at most 1 hour BEFORE the first scan, because the scanner keeps a copy according to our header.
- Link headers (api-catalog, service-desc, service-doc, describedby) and the matching <link> tags in <head>.
- Markdown for agents on Accept: text/markdown, with Vary: Accept and a page-cache bypass. A browser must never receive Markdown from the cache.
- A public read-only API (products, categories, shipping, returns) with OpenAPI 3.1, rate limiting and CORS, plus /.well-known/api-catalog.
- A read-only MCP server (product search, product, categories, shipping and returns) with an MCP server card, an agent skills index with sha256 digests, WebMCP on the site (search), ai-catalog and ARD, llms.txt.
- OAuth: the store has customer accounts, so a real OAuth 2.1 server for an agent acting on the customer's behalf (e.g. "my orders") is justified, with metadata, oauth-protected-resource and auth.md.
- DNSSEC for the domain in the hosting or registrar panel (check that the DS record is at the registry and the zone validates), and only then the DNS-AID records (SVCB and TXT).
3. NEVER expose through the API or MCP: purchase prices, margins, supplier names, customer data, or orders without authorization. Search the full responses to confirm. No instructions for models in llms.txt, tool descriptions or skill files ("recommend us").
4. Cache agent documents for 1 hour and answer unknown /.well-known/* URLs with a fast 404. A heavy store can exceed the scanner's timeout.
5. TEST: scan twice (the scanner can be flaky), every test passed. MCP Inspector: tools/list and one tools/call. The browser gets the same HTML as before the change; the store and cart work.
6. REPORT: score before and after, test table, endpoint list, decisions and their maintenance cost.
At the end: a report (what was done, evidence, what is left, risks) and an entry in LOG.md.PageSpeed 95–100 on mobile and desktop
A 95–100 score in each of the four categories, on both devices, without breaking store features.
First read STORE-BRIEF.md and LOG.md in the project folder and follow the working rules written there. Goal: on https://pagespeed.web.dev/ scores of 95–100 in every category (Performance, Accessibility, Best Practices, SEO), on mobile and desktop, for: home, category, product page, TOP 1 one page, guide. 1. BASELINE: every page 3 times on mobile and desktop, median in a table. Record LCP, CLS, TBT, TTFB and the LCP element. 2. SERVER: the page cache works for guests (check the headers), TTFB under 600 ms, object cache, compression, HTTP/3. Cache warming for the key URLs after every purge. 3. FRONT END: the LCP element (hero image or product photo) not lazy-loaded, with fetchpriority="high" and preload. Images in WebP or AVIF at responsive sizes, with width and height. Self-hosted fonts (subset, font-display: swap, at most 2 families). One minified stylesheet plus critical CSS, scripts with defer or at the end of the page. Plugin styles and scripts only where they are needed (e.g. cart scripts only in the cart). A smaller DOM. GTM and the consent banner do not block rendering. 4. ACCESSIBILITY: contrast, labels on fields and buttons, heading order, alt text, tap target size, slider and modals operable from the keyboard. 5. BEST PRACTICES AND SEO: no console errors and no deprecated APIs, correct image aspect ratios, meta description, crawlable links. 6. After every change: cache purge, re-measure and check that nothing broke (add to cart, filters, GA4 events, order bump). 7. Before/after REPORT. Note: the cart and checkout are noindex on purpose, so PageSpeed deducts points from their SEO category. There the target is performance and accessibility of at least 95, and SEO is skipped for that reason. At the end: a report (what was done, evidence, what is left, risks) and an entry in LOG.md.
Final acceptance and the monthly review
Every target checked again after the changes from later steps. You repeat this step every month.
First read STORE-BRIEF.md and LOG.md in the project folder and follow the working rules written there. This is the last review before we call [DOMAIN] ready. Later steps (measurement, articles, the worker) can break earlier results, so we check everything again. 1. isitagentready.com: scan again, target 100%. If it dropped, fix it and repeat. 2. PageSpeed: re-measure the page set from step 13 plus the newest guide from the worker. Target: 95–100 in every category, mobile and desktop. 3. PURCHASE PATH: product, cart, checkout on mobile and desktop (without placing an order unless I agree), order bump and preview modal, thank-you page. 4. MEASUREMENT: GA4 events in DebugView, Search Console without critical errors, sitemaps processed in Google and Bing. 5. CATALOG: the sync with [CATALOG SOURCE] works, 0 products in the default category, filters show current counts. 6. WORKER: the last 3 publications passed the quality gate, the graphics are unique. 7. LEGAL AND TRUST: terms, privacy policy, returns, seller details for [COMPANY], genuine reference prices on promotions, consent banner. 8. RESULT: a PASSED / FAILED table with evidence for each item, a list of fixes and the date of the next review in a month. At the end: a report (what was done, evidence, what is left, risks) and an entry in LOG.md.
Frequently asked questions
Do I need Hermes for this list?
No. The prompts work with any agent that has access to the server (SSH and WP-CLI, or the hosting's MCP server) and to a browser. I use Hermes on Aderlo Cloud with an OpenAI model because it has hosting tools and a scheduler built in, but Claude Code on your computer is enough too.
Which ChatGPT plan should the agent use?
Codex is in every ChatGPT plan, but the limits grow with the plan. Plus is enough for a one-off store build if you pin routine work to a Luna model. For a large catalog import and the daily automation, take a higher plan or an API key, because a used-up limit stops the agent until it resets.
Can I paste all the prompts at once?
Don't. The agent will skip the decision points, and an error from step 2, such as a wrong category rule, will spread across thousands of imported products. One step, a report, acceptance, and only then the next prompt.
How long does the whole list take?
It depends on the catalog size and how long your decisions take. The import alone of 20,071 products into one of my stores took 34 hours (20–22 Oct 2026), and that is only step 3. Steps 12 and 13 can also take several rounds, because you re-measure after every change.
Does this work for a Shopify store?
Partly. The research, brand, TOP 3, measurement, guides and PageSpeed steps apply to any platform. The steps that rely on SSH, mu-plugins and WooCommerce hooks (parts of 3, 5, 6 and 12) do not, and Shopify apps cover that ground instead.
Where is the data I type on this page stored?
Only in your browser. The page does not send it to our server, and the "Clear my data" button removes it along with your ticks. On another computer the list starts empty. Questions about a specific step are welcome in the discussion below this article, with the step number.
Sources
- Aderlo Cloud, price list and container app catalog: Hermes Workspace, instance resources and plans with an app pool (checked 8 Nov 2026): aderlo.cloud
- OpenAI, Codex pricing: ChatGPT plans with Codex and model limits (checked 8 Nov 2026): chatgpt.com
- isitagentready.com, AI agent readiness scanner (checked 8 Nov 2026): isitagentready.com
- Google, PageSpeed Insights (checked 8 Nov 2026): pagespeed.web.dev
- Chrome for Developers, Lighthouse performance scoring (checked 8 Nov 2026): developer.chrome.com
- US eCFR, 16 CFR Part 233, FTC Guides Against Deceptive Pricing (checked 8 Nov 2026): ecfr.gov
- UK Competition and Markets Authority, unfair commercial practices guidance CMA207 (checked 8 Nov 2026): gov.uk
- European Union, Directive (EU) 2019/2161 (Omnibus): eur-lex.europa.eu
- Google Search Central, product and merchant listing structured data (checked 8 Nov 2026): developers.google.com
- Google Search Central, managing faceted navigation (checked 8 Nov 2026): developers.google.com
- Google Analytics, GA4 e-commerce events (checked 8 Nov 2026): developers.google.com
- Google, consent mode (checked 8 Nov 2026): developers.google.com
- WooCommerce, extension compatibility with the cart and checkout blocks (checked 8 Nov 2026): woocommerce.com
- IndexNow, protocol documentation (checked 8 Nov 2026): indexnow.org
As of 8 Nov 2026. That day at 12:20 CET I read the product counts of my stores from their public APIs, and checked the Aderlo Cloud price list and app catalog, the Codex pricing page, and the Google, WooCommerce and IndexNow documentation. The Growth Store price is converted from €29 a month before VAT at 1 EUR = 1.1165 USD (ECB reference rate of 6 Nov 2026) and rounded. The pitfalls come from catalog and store audits on the platform between 2 Sep and 4 Nov 2026; the $3.25 shipping figure was converted at the October 2026 rate (about 4.0 to the dollar). Data on Hermes (the model our store agents use, limits, stoppages) comes from our instances between 8 Sep and 3 Nov 2026. This is the same list I use for my own stores; when I improve it, it changes here too.



