You cannot see your own product page anymore. You have looked at it four hundred times. You know what the product is, who it is for, what the material feels like and why the price is what it is, so your eye slides over every place the page never actually says any of that.
A stranger arrives knowing none of it, and leaves without telling you why.
This is the kit I run on my own pages every month. Six prompts that put AI in that stranger's position, then two checklists for the part that goes wrong when you rebuild a site, then how to decide which findings are worth your afternoon.
Nothing here needs a special tool or a subscription. If you have already built the brand folder from the setup playbook, these prompts get sharper, because the audit runs against your real customer and your real voice rules instead of a generic idea of a shopper. If you have not, they still work today.
What do you need before you start?
One page, and the honesty to run it on a page that matters rather than one you already know is bad.
- The live page. The URL is enough if your AI can read the web. If it cannot, paste the page text, and say so, because a page pasted as text loses the layout and the first prompt depends partly on order.
- The product photos, if you are pasting. Not for aesthetics. For what the images are being asked to explain that the copy never does.
- Your reviews for that product. Optional for the first prompt, required for the second. This is where the real objections live, in your customers' words rather than your guesses.
- Your bestseller, not your problem child. Audit the page that already converts. That is where a fixable point of friction is worth the most money, and it is the page you are least able to see.
What you do not need: heatmaps, session recordings, a CRO consultant, or traffic. This finds the problems that are visible in the page itself, which is most of them, and it finds them before you spend anything.
What is the skeptical-visitor prompt?
One instruction, three questions. This is the whole audit in its shortest form, and if you only ever run one prompt from this page, run this one.
The role matters more than the wording. "Review my product page" gets you a polite list of best practices that would apply to any page on the internet. "You are a first-time visitor with no context and low patience" gets you a person who does not care about your brand yet, which is the only opinion that counts.
The last sentence is the one that makes it usable. Without it you get three observations you cannot act on. With it you get three locations. A finding you can point at is a task. A finding you cannot is an opinion.
What a good run looks like: it quotes real lines from your page, it names a question your page genuinely does not answer, and at least one of the three findings is slightly annoying to read. That is the one worth having.
What a bad run looks like: generic advice with no quotes. Add social proof. Improve your headline. Use urgency. That means it never really read the page, it recited what it knows about product pages in general. Rerun it, and this time paste the page text in directly.
What are the five follow-up prompts?
The first prompt tells you where the page fails. These five tell you why, one failure mode at a time. Run them as separate prompts, not as one long instruction, because a single prompt asking for five different analyses returns five shallow ones.
Objection mapping
The gap between what customers worry about and what the page addresses. This is the highest-yield prompt on the page, because your reviews already contain the answer and almost nobody reads them this way.
The last sentence is not optional. Without it you get plausible objections rather than real ones, and plausible objections are how a page ends up answering questions nobody was asking.
The first-screen test
Everything above the fold has to do four jobs before anyone decides to scroll. Most pages do two of them and assume the rest.
"Implied" is the finding. Stated and absent are both easy to act on. Implied is where you are relying on the visitor to do work she has no reason to do.
Price justification
Not whether your price is right. Whether the page earns it. Those are different problems and only one of them is a copy problem.
The usual answer to that last question is a specification, a process detail, or a review that says the thing the brand keeps claiming about itself. Evidence you already own and never put on the page.
The mobile read
Open your own analytics and check what share of this page's traffic arrives on a phone. Whatever that number is, a page that reads well on a laptop can be an entirely different page on a small screen. Length behaves differently. Order behaves differently.
CTA clarity
The last prompt, and the one that most often finds something in a single line.
What should you check before you move a site?
This is the part of a rebuild that nobody warns you about, and it is the one that costs real money.
Every page you have published has earned its position over time. When a site moves, the old page addresses stop existing. If nothing tells search engines where those pages went, the history does not transfer. It is not a slow decline. The pages are simply not there anymore. I have watched a founder move her store onto an AI builder and lose her search traffic for exactly that reason.
The work is boring and it is entirely preventable. Do it before you launch, not after.
Before you touch anything:
- Export every URL on the current site. All of them, not just the ones you remember. Products, collections, articles, policy pages, old campaign landing pages.
- Mark which ones earn traffic. Your analytics and your search console both know this. These are the ones that must not break.
- Mark which ones have links pointing at them from other sites. A page someone else linked to is worth protecting even if its own traffic is small, because that link is what its ranking is built on.
- Write the map. One row per old address, one new address beside it. Every single row gets a destination. A row you cannot map is a decision you have not made yet, not a row you get to skip.
Then, in the build:
- Point old addresses at the closest equivalent page, not at the homepage. Sending everything to the homepage is the fastest way to throw away exactly what you were trying to keep.
- Use permanent redirects, not temporary ones. A temporary redirect tells search engines to keep the old address. That is the opposite of what you want.
- Keep one hop. An old address should reach its destination in a single step, not through two or three other addresses first.
- Keep your address conventions consistent. Trailing slashes, capitalization, www or no www. Pick one, apply it everywhere, and redirect the other version.
- Publish a new sitemap and keep the old addresses reachable through their redirects rather than removing them.
After launch, on the first day:
- Walk the export. Take your list of old addresses and check that each one lands on the right new page. This is tedious and it is the whole point of having made the list.
- Watch for missing pages in your search console coverage report for the first few weeks. A cluster of them is a mapping row you missed.
- Check the pages you marked as traffic-earning first, before anything else, and check them by hand.
Can AI assistants read your product page?
A cheap, fast, beautiful site is worth very little if an AI assistant cannot work out what you sell. Some of your customers now ask an assistant what to buy before they ever reach a search results page, and an assistant can only recommend what it can read.
Cheap is the easy part. Findable is the work.
On the page:
- Real text, not words baked into an image. This is the single most common failure. Your specifications, care instructions and sizing sit in a beautifully designed graphic, and to a machine that graphic is blank.
- One clear heading structure. One main heading, then sections beneath it in order. Headings are how a machine works out what the page is about and which part answers which question.
- The page answers the question in its own title. If the title promises what something is made of, the page should say what it is made of, in plain text, near the top.
- The specifics as text. Material, dimensions, weight, origin, care, shipping timeline, returns window. These are the exact things people ask assistants, and the exact things brands leave in a PDF or an image.
- A short question-and-answer block using the real questions your customers ask, in their words. Prompt 02 gives you the list.
- Reviews as text on the page, not only inside a widget that loads separately.
Then test what the assistants actually say:
- Ask several of them what your brand sells, in a fresh session with no history.
- Ask what a specific named product is made of, then check the answer against your own page.
- Ask about your returns policy and shipping times. These get invented more than anything else, and an invented returns policy is a customer service problem waiting to happen.
- Ask which brands they would recommend for the thing you sell, and see whether you appear at all.
- Check that your site is not blocking the crawlers you want reading you. Your robots file is a one-line answer to a question most brands have never asked.
Which findings do you fix first?
An audit that produces twenty findings and no order of operations is a list that never gets worked through. Here is how I sort them.
Fix first, in this order:
- A question the page never answers. A missing answer costs you sales from people who were ready. Nothing else on the list is worth more.
- Anything wrong on the first screen. It is seen by everyone who arrives, including the people who leave before scrolling, and they are the ones you are losing.
- The most-repeated objection with no answer on the page. Prompt 02 ranked these for you. Take the top one and write the answer into the page rather than into a support macro.
- Any specification that only exists inside an image. Cheap to fix, and it fixes both the human problem and the machine one.
Ignore, or at least deprioritize:
- Tone preferences. "The copy could be warmer" is not a finding, it is a taste. Yours is the taste that matters.
- Suggestions to add urgency or scarcity you cannot back. If the stock is not low, do not say it is. This is also where a compliance problem starts.
- Any statistic it offers without a source. Percentages about what shoppers do arrive fluently and confidently and should be treated as decoration.
- Anything it cannot point to on the page. This is the master rule and it settles most arguments. If it cannot quote the line it is reacting to, it is guessing.
How often. Monthly on your bestsellers, and once on every page after a rebuild, migration or theme change. Monthly is frequent enough to catch drift and rare enough that you actually work the list. Weekly produces a backlog nobody clears, which quietly teaches everyone that the list does not matter.
What comes next?
These prompts work on any page today. They work considerably better once the AI is running on your actual brand, because then it audits against your real customer, your real voice rules and the things your brand has decided it never says, instead of against a general idea of a shopper.
That setup is one folder, four files and about ninety minutes, and it is written out completely in The Claude Setup Playbook for E-Commerce Marketing. Do that first if you have not, then come back and run this kit against it.
I teach it in public, one piece at a time, and it is free.
How do you audit an e-commerce product page with AI?
You give it the page and you make it a stranger. The instruction that does the work is a role and a constraint: a first-time visitor, no context, low patience. Then you ask three questions. Where do you hesitate, what question never gets answered, and where do you get bored. Everything else in an audit is a follow-up to one of those three. You are not asking it to like your page. You are asking it to fail to buy from it and tell you where.
What should you ask AI to check on a product page?
Six things, in this order. The skeptical walk-through first, then the five follow-ups: which objections the page answers and which it leaves open, what the first screen actually communicates before anyone scrolls, what on the page justifies the price against a cheaper alternative, how the page reads on a phone, and how many competing actions sit between the visitor and the one thing you want them to do. Each one produces a different kind of finding, which is why they are separate prompts rather than one long one.
Why do you lose search traffic when you rebuild your store?
Because the old page addresses stop existing and nothing tells search engines where they went. Every product page and article you have published has earned its position over time. If the new site uses different addresses and the old ones are not pointed at the new ones, that history does not transfer. It is not a slow decline, it is the pages simply no longer being there. I have watched a founder move her store onto an AI builder and lose her search traffic for exactly this reason, and it is the single most preventable thing on this page.
How do you check whether AI assistants can read your product page?
Two checks, one on the page and one on the assistants. On the page: confirm your specifications, materials, sizing, shipping and returns exist as real text rather than as words baked into an image, and that the headings are a clean structure a machine can follow. On the assistants: ask them directly what your brand sells, what a named product is made of, and what your returns policy is. If the answer is vague, wrong, or invented, they could not read it. That gap is the work.
How often should you audit your product pages?
Monthly on your bestsellers, and once on every page after any rebuild, migration or theme change. Monthly is often enough to catch what has drifted and rare enough that you actually act on the findings. Auditing weekly produces a list nobody works through, which is worse than not auditing at all, because it teaches the team that the list does not matter.
