Ecommerce Store Audit Checklist: Run It With ChatGPT or Claude

Run the audit in five passes: homepage and navigation, a product page, cart and checkout, policies and trust pages, then search basics. For each pass, give ChatGPT or Claude a screenshot or pasted text plus a short scoring prompt, and ask for findings ranked by buyer impact and effort. It can judge copy and structure; it cannot judge speed or trust for you.

Most store audits fail for one of two reasons: the owner asks "what is wrong with my store?" with nothing attached, or collects forty findings and fixes none. The method below avoids both by feeding the assistant real evidence one page at a time and ending with a ranked list of no more than ten items, each with an owner and a date.

Follow me on Instagram@sagnikteaches

Set up the evidence folder before you open the chat

An assistant is only as good as what you hand it. Spend twenty minutes gathering evidence into one folder, and name files so you can refer to them in prompts. For a typical store you need the following.

Connect on LinkedInSagnik Bhattacharya

• Screenshots, desktop and phone: homepage, one collection page, your best-selling product page, your worst-converting product page, the cart, and each step of checkout up to the payment screen. Take phone screenshots on an actual phone, not by shrinking a browser window.
• Page text: copy the visible text of each page into a plain document. Screenshots are good for layout; pasted text is better for wording, because the assistant reads it exactly.
• Policy text: shipping, returns, privacy and terms pages, pasted in full.
• Exports: a product list as CSV (title, description, price, image alt text, page title, meta description), plus your last 90 days of orders by product if your platform allows it.
• Numbers: sessions, add-to-cart rate, checkout starts and completed orders for the same 90 days, from your store's own analytics.

If you are unsure how well the assistant reads uploads, this guide to what ChatGPT reads from PDFs, spreadsheets and photos lists where it stumbles. The short version: keep each upload to one page or one table, and say what the file is.

Subscribe on YouTube@sagnikteaches

Create a Project (ChatGPT Projects and Claude Projects both work on free plans, with file limits that differ) and paste a standing instruction so every chat starts from the same rules:

You are auditing a small online store. I will give you one page or one export at a time.
For each, return: (1) up to six findings, each tagged CONVERSION, SEARCH, POLICY or TRUST;
(2) the evidence for each finding (quote my text or describe what is visible in the screenshot);
(3) impact on buyers (high, medium, low) and effort for me (minutes, hours, days);
(4) a one-line fix. If you cannot tell something from what I gave you, say "cannot judge from this" instead of guessing.
Never invent traffic figures or benchmarks.

That last instruction matters. Without it, assistants cheerfully quote "industry average conversion rates" that they cannot source. You want findings tied to your evidence, not to folklore.

Pass one: homepage and navigation

A new visitor decides in a few seconds whether this store sells what they want. Upload the desktop and phone screenshots of the homepage and ask the assistant to answer, as a first-time visitor, three questions: what does this store sell, who is it for, and what would I click first. If its answers differ from what you intended, you have found your first problem.

Then paste your menu labels and ask it to flag any that a stranger would not understand. A homeware store with a menu reading "Collections, Essentials, Edit, Journal, About" gives a stranger little to go on. "Candles, Diffusers, Gift sets, Sale, About" does.

Checks for this pass:

1. The headline says what you sell in plain words, not a slogan.
2. There is one obvious primary button above the fold on a phone.
3. Delivery cost or the free-delivery threshold is visible without opening a page.
4. Search is reachable in one tap, and a deliberately misspelled query still returns something.
5. The menu has no more than seven top-level items.

The assistant handles items 1, 2 and 5 well from screenshots. It cannot test item 4, so type a misspelling into your own search box and screenshot the result before asking about it.

Pass two: one product page, then the weakest one

Choose your best-selling product page and your weakest page (traffic but few orders). Auditing both shows you which differences matter. Upload the phone screenshot, the pasted text and the images' alt text from your export. Ask for a table comparing the two pages on: first-screen clarity, price and delivery visibility, quality and number of images, description specifics, size or option guidance, reviews, and the add-to-cart button on a phone.

If you want a deeper, product-by-product method, the product page audit tutorial goes through it. Here you only need enough to rank problems across the whole store.

Four findings from a candle shop's weakest page

Here is an illustrative response for a small candle shop's weakest page, the sort of finding a good prompt produces:

Notice what to correct in this kind of output. The assistant cannot know the real burn time, so "add burn time" is an instruction to you, not a fact it can supply. Never let it fill in the number. Also check each "evidence" cell against the page; assistants occasionally describe a feature that is on a sibling page.

Pass three: cart and checkout, the part AI sees least

Checkout is where most stores lose the largest share of motivated buyers, and it is where the assistant's view is narrowest, because it only sees screenshots. Place a real test order with a cheap product or a 100 percent discount code, screenshot every step, and count the taps from "Add to cart" to "Pay". Then ask the assistant to list every point where a buyer is asked for something unexpected: an account, a phone number, an address lookup, a surprise fee, a shipping choice with no delivery date.

Check these yourself because the assistant cannot:

• Guest checkout works without creating an account.
• Every payment method you list actually completes on a phone.
• Discount code boxes do not push the pay button below the fold.
• Error messages say what is wrong ("Phone number is missing a digit", not "Invalid input").
• The confirmation email arrives within two minutes and names the item.

When you share checkout screenshots, crop out anything personal: your address, partial card numbers or customer names. Use a test order for this reason. Real customer data does not belong in a chat.

Pass four: policies and trust pages

This is where an assistant is most useful and most dangerous. Paste your shipping, returns and privacy pages and ask it to find contradictions and gaps, not to confirm they are legal. Three prompts do the work:

Read my returns policy and my shipping policy. List every place they contradict each other or the product pages (for example, different return windows).
Then list questions a buyer would ask that neither page answers: who pays for return postage, how long refunds take, what counts as faulty.
Do not tell me whether the policy is legally compliant.

A typical catch: the returns page says 30 days while the product page footnote says 14. Another: the shipping page promises 2 to 4 working days, but the checkout shows 5 to 7. These contradictions cost sales and invite disputes. For wording ideas, the returns and refunds policy tutorial has sample phrases you can adapt.

Legal compliance is for a qualified adviser. Consumer-rights rules differ between markets and change, and an assistant will sometimes state them confidently and wrongly. Use it to find inconsistencies and to draft plain-language wording, then have the final text checked if you sell where rules are strict. If you have customers in the EU, mention that to your adviser.

Trust signals worth checking by eye

Ask the assistant to list the trust signals present on the pages you gave it (reviews, contact details, delivery promises, secure payment badges, an about page with a real face or address of business). It will do this well. What it cannot do is tell you whether a stranger would believe them. Show the homepage to two people outside your business for ten seconds each and ask what they remember. That test finds more than any prompt.

Pass five: search basics from your product export

Upload the product CSV (or the first 50 rows) and ask for a search review of titles, meta descriptions and image alt text. Ask for four things: duplicate or near-duplicate titles, titles long enough to be cut off in results, missing or duplicated meta descriptions, and images with no alt text or alt text that says "image1". Shopify's alt text field, for example, accepts up to 512 characters, with about 125 recommended, so a 40-character descriptive line is well within bounds. On titles, Google's Search Central documentation says there is no limit on how long a title element can be, but the title link in results is truncated to fit the device width, so about 60 characters is a sensible working cap rather than a rule.

The assistant can find these patterns across hundreds of rows in a minute, a job that takes an hour by eye. It cannot see rankings, so do not ask whether a title "will rank". For ranking evidence use your store's search report or Google Search Console, and see the one-afternoon SEO audit for a wider pass.

An example: fixing a duplicated-title problem

Say a small plant shop has eleven products titled "Trailing Plant - Pot Included". The assistant flags the duplicate pattern and proposes eleven distinct titles such as "Trailing Pothos in 12 cm Pot" and "Trailing String of Pearls in 10 cm Pot". You check each against the real product, because the assistant will otherwise name a plant you do not stock. Then you upload the corrected CSV in a test import of three rows first, confirm the titles changed and the URLs did not, and only then import the rest.

If you had used AI to write all your descriptions in bulk, also read which AI store pages risk search penalties; thin, near-identical pages are what that policy targets.

What AI cannot judge, so you test by hand

Four things sit outside any chat window, however good the screenshots.

1. Speed. A screenshot looks identical whether the page took one second or eight. Test with your platform's speed report or Google's PageSpeed Insights on your top five pages, on mobile settings, and note the Largest Contentful Paint and Cumulative Layout Shift figures, which PageSpeed Insights reports for both mobile and desktop. You can paste those results into the assistant to interpret them, which it does well.
2. Trust. Whether the store feels safe depends on feel, photography and social proof. Ask real people.
3. Behaviour. Where visitors hesitate needs session recordings or heatmaps, which your analytics tool may offer. Without them, you are guessing at causes.
4. Breakage. Forms, coupons, payment buttons and email sends need real tests. A screenshot of a button proves the button exists, not that it works.

This is the honest limit of the method. The assistant widens what you notice; it does not replace testing.

Turn forty findings into a ranked list of ten

After the five passes you will have a pile of tagged findings. Paste them all into one chat and ask the assistant to merge duplicates and score each one. Use a simple rule so the result is reproducible:

Work down A, then B, then C, and put D in a list with a review date. Cap the first sprint at ten items and give each one an owner and a date, even if the owner is you. A ranked list of ten that gets done beats a spreadsheet of sixty that sits.

Putting a dollar figure on each fix

Say a small online shop selling home fragrance gets 6,000 sessions a month, with 3 percent adding to cart (180), and 1.1 percent of sessions ordering (66 orders) at an average order of $38. That is about $2,508 a month. The audit finds delivery cost appearing only at checkout, with an estimated 25 of the 114 abandoned carts attributed to surprise cost (that attribution is a judgement from exit surveys or customer emails, not a measured fact). If showing the cost earlier recovered even a quarter of those, say six orders, that is roughly $228 a month for a twenty-minute edit. The same shop's reshoot of all images might cost $600 and a weekend for an uncertain lift. The ranking is obvious once the sums are on the page.

Treat all of those figures as illustrative. The point is to put a rough dollar value next to each fix so that effort and payoff sit side by side.

A filled-in checklist for a first sprint

Here is how a ranked sprint might look after an audit of a small tea store, illustrative again:

1. (A) Returns page says 30 days, product pages say 14. Pick one, edit both. Owner: shop manager. 30 minutes.
2. (A) Phone checkout pushes "Pay now" below the keyboard when an address is typed. Test on two phones, raise with the theme developer or app vendor. 2 hours.
3. (B) Add delivery cost and dispatch time under every price. 20 minutes.
4. (B) Move star ratings next to the product title. 30 minutes.
5. (C) Rewrite the nine titles that start with "Tea -". 45 minutes.
6. (C) Add alt text to 60 images in a CSV import, checking three first. 1 hour.
7. (C) Add burn-time-style specifics (here, brew temperature and steep time) to the twelve best-selling descriptions. 2 hours.

Total effort under ten hours, spread over two weeks, with every item traceable to evidence. Re-run the same five passes four weeks later on the same pages and compare the numbers from your store analytics: add-to-cart rate, checkout completion and orders per 100 sessions.

When the assistant gets it wrong

Three failure patterns turn up repeatedly in audits like this, so check for them before acting on anything.

• Confident but unsourced benchmarks. "Your add-to-cart rate is below average" with no source. Delete any sentence like this; the standing instruction above prevents most of them.
• Findings about pages it never saw. If a finding mentions a feature you did not upload, ask "which file shows this?" It will often retract.
• Generic advice that fits any store. "Add social proof" with no detail. Ask for the exact element, its position and what it should say, using your own products. If it cannot, drop the finding.

One more habit: after the assistant proposes new wording, paste it back and ask it to list every factual claim it contains. Then check each claim against your product. Material, size, origin, "handmade" and "organic" all carry obligations. Rewriting to sound better must never change what you promise, and an assistant polishing copy has no way of knowing which of those words you can stand behind.

Which assistant, and how much to upload

ChatGPT and Claude both read screenshots, pasted text and spreadsheets, and either is fine for this job. If you use Shopify, Shopify's own assistant, Sidekick, sits inside the admin and can see store data directly; the comparison in which assistant suits which Shopify job explains when each fits. For this audit, an outside chat has an advantage: it only sees what you choose to give it, which keeps customer data out.

On file limits, Claude's help centre allows up to 20 files per chat at up to 500MB each, but files added to a Claude Project are capped at 30MB each, and free accounts can create up to five Projects. ChatGPT Projects limit the number of files by plan: five on Free, 25 on Go and Plus, and 40 on Pro and the business plans. A store with a large catalogue should upload a sample of 50 to 100 products chosen from best sellers, worst converters and a random handful, rather than every row. Run the pattern checks on the sample, fix what you find, then re-export and run again.

A repeatable monthly version

Once the first audit is done, shrink it to a thirty-minute monthly check. Each month, take screenshots of the same five pages, export the same numbers, and ask the assistant to compare this month's evidence with last month's in the same chat or Project. Ask specifically what changed, what got worse and whether any previously fixed item has regressed. Regression is common: a new app installs a banner, a theme update shifts a button, a price change leaves an old figure in a description.

Keep a one-page log with the date, the findings fixed and the numbers beside them. After three months the log shows which fixes mattered and which did nothing, which is more useful than any benchmark. If the log shows you spending hours on checks that never change results, shorten the checklist. If you would rather map the monthly routine with someone and decide which parts to automate, that is what my AI implementation consultation covers.

Store audit questions owners ask before they start

Can ChatGPT or Claude audit my store without access to it?

Yes, if you hand it the evidence. Paste page text, upload screenshots and attach exports such as a product CSV or a search report. It cannot browse your admin, load your pages on a slow phone or see what visitors do, so everything it says rests on what you supply.

How often should I audit an online store?

A full pass twice a year suits most small stores, plus a quick check of the top five pages after any theme change, large app install or price change. Audit sooner if conversion falls for three weeks in a row with steady traffic.

Should I let the assistant change the store directly?

Not during an audit. Use it to find and rank problems, then make changes yourself or approve them one at a time. Auditing and editing in one step makes it hard to tell which change helped.

Is a free AI plan enough for a store audit?

For one store with a few dozen products, usually yes, because the limiting factor is how much you can upload per chat. Larger catalogues work better in a Project or by auditing a sample of products rather than every one.

What should I do first if the audit finds fifty problems?

Rank them by how many buyers each affects and how long the fix takes. Fix anything that blocks checkout or breaks a policy first, then the top five product pages, then search fixes. Park the rest in a list.

Further reads

• How Can AI Assist in Online Sales? Product Pages, Reviews, Follow-Ups — Use AI on product pages and reviews after the audit.
• AI for Ecommerce Analytics: Shopify Reports, GA4 or ChatGPT First? — Decide which reports to feed the assistant.
• How to Write Shopify Product Descriptions With AI That Convert — Rewrite weak descriptions the audit flags.
• AI Tools for Ecommerce Photography: Backgrounds, Shadows, Consistent Sets — Fix weak product images the audit flags.
• How to Set Up AI Abandoned Cart Emails for a Small Online Store — Recover some of the sales you find leaking.
• AI for Ecommerce Inventory Management: Reorder Points and Stockouts — Fix stockouts that the audit shows hurting sales.
• AI Listing Mistakes That Can Mislead Buyers, and How to Check — Catch claims in rewritten copy that you cannot back up.
• Edit a Shopify Theme With AI: Safe Steps Using a Duplicate Theme — Duplicate the theme, let AI draft the change, test on a phone, publish with a rollback ready, and know the point where a developer takes over.
• Shopify Pre-Order App: Which to Pick, Then Set Ship Dates With AI — Pick a pre-order app by order volume, then use AI to turn supplier lead times into ship-date wording and a delay email.
• Shopify Profit Report With Claude: Add Ad Spend, COGS and Returns — Join four exports in Claude to see profit per order after ads, costs, shipping and returns, with a full worked month.
• Shopify RFM Analysis: Segment Customers With ChatGPT From an Export — Score customers on recency, frequency and spend from a Shopify export, then write one email per segment, keeping personal data minimal.
• Shopify SEO With AI: Fix Titles and Meta Descriptions in Bulk — A CSV round trip for Shopify SEO text: export, rewrite with AI in batches, check length and duplicates, import, and judge it in Search Console.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Shopify Help Center (image alt text), Google Search Central (title links), PageSpeed Insights documentation, Claude and ChatGPT help centres (file uploads and Projects).