On 3 November 2026, Practical Ecommerce published a piece by Ani Nandi, founder of EcomChief and NIR Systems, on using AI to find site errors before an ad campaign goes live. The method is plain. Pull page content with a tool like Firecrawl or Tavily, or take screenshots, hand them to ChatGPT or Claude with a specific prompt, and have it flag broken links, conflicting offers, misleading availability, missing product information and mobile layout problems. Google Lighthouse and Screaming Frog cover speed and crawl errors, and Screaming Frog crawls up to 500 URLs free. The findings are grouped by customer impact, then a person tests them before each campaign.

The need is real, and it is older than AI

The article’s starting point is one I recognise from twenty years of running campaigns for clients. A merchant with many products cannot check every detail by hand. The example given is a clothing store advertising 30 garments while some of them are sold out. As the piece puts it, “An ecommerce business can launch an ad campaign without realizing there’s a glitch on the site.”

So what changed? The checking got cheap. A pass over every landing page used to cost a person a day, so it got done once, or never. Now it costs a prompt and a crawl.

What didn’t change matters just as much. Someone still has to fix the error. Someone still has to decide which errors matter and whether the offer behind the page is right. The article is honest about this and keeps a human test before every campaign. That is the right shape.

Why checking is the right first job for AI

Most of the AI conversation in marketing is about writing. Product descriptions, ad copy, blog posts. I run content at volume myself, so I am not against it. But writing needs judgement at every step. What to say, who it is for, what position it carries. Give that to a machine with no frame and you get generic output.

Checking needs no judgement at all. Does the price on the page match the price in the ad? One answer. Does the image load? One answer. Is the size chart there? One answer. Thousands of questions, each with one answer, is exactly the work a machine should do and a person should not.

The cost side settles it. A mediocre product description loses you a little. A page with a dead image or a product that cannot be bought loses the click you paid for and the shopper who clicked. The error is charged to your ad budget, over and over, until someone notices.

My rule has been the same for a while. Let the machine do the counting, the checking and the drafting. Keep the judgement. Applied to a catalogue, the order is checking first, drafting second.

What it looks like in a real shop

Take a store with two thousand products, ads running on Google and Meta, and a marketing team of two. Nobody on that team opens two thousand pages before a launch. Nobody should.

The setup is a crawl of every page the ads land on, fed to the AI with a specific brief. The price shown matches the feed. The main image loads. Stock status matches what the ad promises. The variant list is complete. A mobile screenshot shows the menu and the add-to-cart button. The output comes back as a list, grouped by what costs the most. Sold out but advertised as available goes to the top.

Then a person verifies the list. Trust nothing unverified cuts both ways. The AI’s list will carry false alarms, so a human checks the flagged pages before anything is fixed. If the crawl flags forty pages out of two thousand, the human’s job is forty pages. That is the whole trade, and it is a good one.

Where the machine stops

The AI does not fix the error. It does not decide whether the offer is right, or whether the product belongs in the campaign at all. Those stay with you.

It also does not run once. Pages drift. A supplier pulls an image, a feed price changes, a variant gets deleted after the ad was written. The check belongs before every campaign, as a step in the checklist, not as a project someone did last quarter.

I work this way on my own builds. Claude does the building, the counting and the checking. I make the decisions. A catalogue is no different.

What to do

  1. Before the next campaign, crawl every page the ads send people to. Screaming Frog is free up to 500 URLs, and the article names Firecrawl and Tavily for pulling content.
  2. Write the brief as a list of exact checks, with the ad and the product feed as the reference. Vague prompts get vague lists.
  3. Ask for the output grouped by customer impact, worst first.
  4. Have a person verify only the flagged pages. Fix them, then launch.
  5. Make it a fixed step in the campaign checklist, every time, before anyone asks the AI to write a single line of copy.

Source: Practical Ecommerce: Find Site Errors with AI before Ads Run