A category manager or digital merchant sits down to review a thousand products. The objective is formulaic: what can we do about all these products to make them sell better on the site? There isn’t time to do more than that. More reviews are coming. There is a production objective to hit.
With the advent of AI, that has changed. What previously might take a merchant days now takes a few hours, and the review is more comprehensive.
This is an approach I have used to enable merchants to make more effective recommendations and updates to product data. And more importantly, it incorporates the way that customers themselves are looking.
Section OneWhy automate merchandising with AI?
This process, and others like it, is not aimed at replacing people. It is aimed at making them more efficient and surfacing contextual issues that might not get touched during a category review.
- The human in the loop might not have familiarity with a group of products at a depth that can match the training of AI. Even if that person has a high degree of experience, AI allows them to reach every corner of product information faster.
- If your business has a high degree of scale, you might be adding or updating more than 1,000 products at a time. This process gives you the confidence that nothing is going to be missed. A problem with 1 out of those 1,000 products will be factored in.
- Merchants and Category Managers do research to understand the importance of particular aspects of product information in a given node, but often they don’t look at the other side of an important coin. That is, “What are people searching for when looking for these products?” This process incorporates search history so that important factor isn’t missed.
Section TwoWhat the review actually costs
Here is what a category review looks like in practice. A merchant pulls the top brands, checks attribute completeness across the node, forms a view of the category as a whole, and makes recommendations at that level. Nobody is going product by product writing recommendations for a thousand SKUs.
That isn’t laziness. It’s arithmetic. Give a merchant three minutes per product — long enough to read the title, scan the description, check what attributes are missing, and glance at how it has performed. A thousand products at that rate is fifty hours, or roughly eight working days for one category, and that is being generous about how much of a day anyone spends in focused review. Nobody has eight days per category. More reviews are coming.
So the review that happens is the one that fits the time available, and everything outside the sample gets the benefit of the doubt. What changes with AI isn’t that the same review gets faster. It’s that the thorough version — always correct in theory, never affordable in practice — becomes the one you can actually run.
Fig. 1 — Cumulative categories of 1,000 products completed across a 250-day working year. Modelled, not measured: three minutes per product, six productive hours per day, and one to two days per category once the process is running — the band covers that range. Analysis only; acting on the recommendations is work that exists either way. Change the per-product figure and the gap changes with it. The shape does not.
Section ThreeInputs for the process
Product & performance data
The approach I take is to give AI quantitative data so that it can weight the importance of the product based on historical performance, and qualitative data about the product’s content including its attributes.
- Product titles & descriptions
- Historic revenue and orders (for a given time period)
- Organic clicks and impressions
- Per-product search queries
- Historical search performance
How you get all of these depends on your setup. Clicks, impressions and queries can all be found in Google Search Console. Historical search performance would come from another SEO platform like Semrush or BrightEdge. Looking at how products have ranked in the past on one of these platforms provides another performance angle for AI that compensates for Search Console.
If your company doesn’t have the historical data regarding search queries — you haven’t ranked anywhere, so you don’t know where you’re showing up — an extra step of keyword research is necessary using an SEO tool or Google Ads Keyword Planner.
Product attribute data
You have collected the relevant data about the products. Now you’ll need to provide AI with data about the category itself. Your approach might vary, but generally speaking this is what I collect:
- Attributes currently assigned to that leaf taxonomy node, along with their level of importance if that is defined.
- Per-SKU breakdown of attribute completeness.
This can be supplemented either with an AI opinion or some quick research about what attributes are most present in search results, indicating their importance.
Section FourPrompt considerations
I’m not giving you my prompt. But directly copying my prompt wouldn’t get you far anyways, because really the prompt should be based on your priorities, your systems and your objectives. So here are a few considerations when developing your prompt.
What do you care about?
Priorities are going to vary broadly depending on the type of company and website in question. The abilities of AI can tempt you to try and “boil the ocean,” but that can often lead to wasted time and a sense of loss of direction. Define and prioritize what you’ll put through this process.
What is your objective?
This is where team structure comes in. You might be the person executing on this, or you might simply be handing off a report for another team to pursue. Do they need all the details, or do they need an understanding of where the problems are?
Specify the output you want based on what you’re trying to accomplish. ChatGPT and Claude are both very capable of formatting information usefully.
What is your next step?
The next step can make a difference in what you ask AI to produce. Perhaps this is a standalone report, and the next step is human-in-the-loop work to further refine that information or do research. On the other hand, this might be just one more step in a fully automated process, and it should be formatted accordingly.
Natural next steps include:
- Product attribute gathering and improvement
- Title and description optimization
- Design/specification updates to product pages
Section FiveFree advice
Keep your eyes open and your brain on.
The world of product data and ecommerce search isn’t cut and dry, especially if your company sells several different types of products. Different product and customer types provide nuance in how people search and browse for products. For some products, a particular dimension might be of key importance. For other products, it might be the series or product-line name. AI is pretty good at picking up on these things, but it may miss some. Moreover, there are always exceptions to rules, and you need to be awake to spot them if you want to maximize the usefulness of this.
Define your outcome.
AI is an enabler of maximalism, and sometimes that’s not a great thing. In early iterations of this process, Claude was producing 10 to 15-page reports per leaf node with stunning charts and scorecards in branded themes. That’s great, but if the person receiving this report is going to be overwhelmed into inaction, then you haven’t succeeded. Know precisely what you want AI to provide you with and be sure to tell AI exactly how you want it.
Create repeatability.
Everyone says, “Wow, this is great,” the first time they see it. Make sure that it’s something you can repeat. Skill files are great for this. Whatever AI you’re using can typically provide you with exact instructions if, after you’ve gotten the result you want, you ask it to provide you with steps and a prompt to do it again. Don’t skip that step.
I’m Not Giving You This Prompt walks through workflows I actually run — the inputs, the judgment calls, the parts that went wrong — and stops short of the prompt itself. Not to be cagey. A prompt is welded to one taxonomy, one data stack and one set of priorities, which makes it the least portable thing in the process and the least useful thing to hand over. Everything around it travels fine.