Search for Amazon listing optimisation and you will be sold copywriting. Better bullets, punchier titles, keyword-rich descriptions. We have rewritten plenty of Amazon copy over fifteen years and it does move numbers. It is also the second-order fix. In Amazon’s own published guidance, what decides whether a shopper ever sees your listing is the structured attribute data behind it. Amazon states the consequence of leaving those attributes empty in one unambiguous sentence.
Why Amazon listing optimisation became a copywriting service
The conventional advice is not stupid. It exists for three good reasons.
Copy is visible. A seller can look at a rewritten title and see the work. Nobody can look at a populated material_type field and feel the same thing.
Copy is sellable. An agency can quote a fixed price per ASIN for a rewrite. Quoting for “fix your attribute model across 40,000 SKUs” is a harder conversation and a longer engagement.
Copy is where the keywords live, and keyword matching is the part of Amazon search that sellers understand. If a shopper types “18mm birch plywood” and your title says “18mm Birch Plywood Sheet”, the connection is obvious.
So the market optimised for the visible thing. The result is thousands of listings with immaculate bullet points that no shopper reaches, because the shopper never typed anything. They clicked a filter.
What Amazon says about attributes
Amazon publishes listing and content guides per category. The guide for its business, industrial and scientific supplies categories states the position plainly. Product attributes populate the product specifications section. Refinements are populated by attributes. And then this:
If a customer filters by a refinement, your product will not show up unless you provided data for that refinement.
Read that again, because it is the whole argument. It is not a ranking penalty. It is not a soft signal. It is exclusion from the result set.
The left-hand filters on any Amazon category page are refinements. Voltage. Thread size. Material. Capacity. Load rating. Every one of them is a structured attribute, and every one of them is a wall your listing either passes through or does not. A shopper who filters to 24V, IP67 and DIN rail mount has just removed every product that did not populate those three fields. The copy on those excluded listings is irrelevant. Nobody is reading it.
We see this most sharply in industrial and trade categories, where shoppers filter before they search because they already know the specification. It applies in retail too, in any category where a shopper has a constraint: shoe size, tog rating, screen size, capacity.
A worked example makes it concrete. Take a 24V DIN rail power supply sold by a distributor. The listing has a good title, five clean bullets and three images. It carries brand, model and a description. It does not carry output current, IP rating, mounting type or operating temperature range, because none of those exist as attributes in the source PIM.
Every one of those four is a refinement on that Amazon category page. The listing is therefore invisible to a shopper who filters on any of them, which is how most buyers of that product shop. Rewriting the bullets changes nothing, because the bullets are never reached. Populating four fields changes everything, and the fields already exist in the manufacturer’s datasheet.
The three places attributes decide the outcome
The node sets the required list
Amazon’s guidance is to “choose the most specific leaf node possible”. That reads like housekeeping. It is not.
The node determines which attributes Amazon expects, and therefore which refinements your product is eligible to appear in. Choose a shallower node because the required list is shorter, and you have quietly opted out of every filter that lives further down the tree. This is the single most common self-inflicted wound we find on marketplace accounts, and it is invisible in every reporting dashboard.
The schema decides whether the listing exists
Amazon defines the data requirements for each product type as a JSON schema. It publishes them through its Product Type Definitions API. When you submit a listing, Amazon returns ACCEPTED or INVALID. Its own developer guidance tells integrators to validate against the schema before submitting, so that required attributes are present and values are valid.
That is a product data pipeline, described in Amazon’s language. There is no copywriting stage in it. The copy fields are simply more attributes, subject to the same validation as everything else.
Variation families are built from attributes
Parent and child relationships on Amazon are defined by variation attributes. Get them wrong and forty sizes of one product list as forty unrelated items. That splits reviews, sales history and the ranking signal forty ways.
No amount of bullet point rewriting repairs that. It is a data model problem, and it usually traces back to a PIM with no product level above SKU.
Amazon will rewrite your copy. It will not invent your attributes
This is the part that should settle the argument for anyone still weighing a copywriting retainer.
In August 2024 Amazon announced updated bullet point requirements, restricting special characters, emojis and certain phrases such as refund-related guarantees. It also said it would use AI to remove non-compliant content and generate revised bullet points for sellers to review. Amazon is, increasingly, in the business of normalising your copy.
Its title guidance is heading the same way. Amazon’s seller communications now describe a 200 character allowance split between an Item name of 75 characters and Item highlights of 125 characters. The advice on what to put in the Item name is telling. Keep the brand name there, and include the most important variating attributes. Amazon is instructing sellers to build titles out of attribute values.
Meanwhile nothing on Amazon’s roadmap invents a missing thread pitch, a missing IP rating or a missing load capacity. Those come from you, from your suppliers, or from nowhere.
What Amazon’s research suggests
Amazon Science published COSMO. It builds a commonsense knowledge graph over the Amazon store, using large language models on customer interaction data. The graph encodes relationships between products and the human contexts they sit in. Their functions, their audiences, the situations they are used in. Relationship types include used_for_function and used_for_audience.
Amazon reports material gains from it. In relevance ranking with fixed encoders, the model using the COSMO graph achieved a 60% increase in macro F1 over the baseline. With fine-tuned models it held a 28% edge in macro F1 and a 22% edge in micro F1 over the best-performing baseline.
Be careful what you conclude from that. It does not say attributes rank you. It says Amazon’s relevance work is moving toward structured, machine-readable relationships between products and contexts. That direction of travel does not reward adjectives. It rewards products that can be described precisely enough to be placed in a graph.
The Amazon listing optimisation test you can run this afternoon
Do not take our word for it. This takes about an hour and it is the most persuasive thing you can put in front of a board.
Pick one category where you sell on Amazon. Open the category page as a shopper. Write down every refinement in the left-hand navigation.
Now export your own catalogue for that category. For each refinement, count what proportion of your SKUs carry a populated, valid value for the matching attribute.
That percentage is your ceiling. It is the maximum share of filtered traffic you can compete for, before ranking, before pricing, before a single word of copy.
Most catalogues we run this on land somewhere between 30% and 60% on the refinements that matter, and nobody in the business knew the number. The gap is almost always concentrated in a handful of attributes that were never in the original data model. That model was designed for a website with no filters.
Run the same exercise across marketplaces generally and the picture repeats, because the underlying cause is the same one every time.
When the copy really is the problem
Intellectual honesty matters here, and there are cases where the copywriters are right.
If your attributes are complete and your conversion rate is still poor against comparable listings, that is a content problem. Fix the images first, then the A+ content, then the bullets.
In a category with few meaningful refinements, search text does most of the work. Some consumables and gifting lines sit there, and copy carries more weight.
If your brand voice is genuinely differentiating and your competitors read like translated datasheets, that is worth money.
What does not work is buying the copywriting first when the attribute completeness is at 40%. You are paying to polish a listing that is excluded from most of the ways a shopper could reach it. We have seen clients spend a full year on content rewrites and report almost no change. Three months of product content enrichment then produced a step change that nobody outside the data team could see.
Amazon listing optimisation, in the order that works
Fix the node. Confirm every SKU sits on the most specific leaf node available, and record it as a governed attribute rather than a listing-time choice.
Pull the required attribute list per product type from Amazon’s schema. Treat it as a target for the PIM, not a checklist for the listing team.
Map the refinements. Every filter in the category navigation is a revenue-carrying attribute. Measure completeness against them and report it weekly.
Fix the variation families next, because they compound. Forty split children take forty times longer to rank than one properly grouped parent.
Then, and only then, rewrite the copy. Build the title from the attributes Amazon asks you to put there. Write the bullets against the specification rather than around it.
That sequence is unglamorous and it is what moves the numbers. It is also the sequence we use on every marketplace programme we run, whatever the channel. It is also portable. The same attribute set feeds eBay item specifics, the Mirakl marketplaces and your own site search. That is why we treat it as one piece of work rather than an Amazon project.
Key takeaways
- Amazon states that a product will not appear when a customer filters by a refinement unless you provided data for it.
- Category leaf node choice determines the required attribute set, and therefore which filters you are eligible for.
- Amazon validates listings against a published JSON schema per product type and returns INVALID when required attributes are missing.
- Amazon now normalises non-compliant copy automatically, and instructs sellers to build titles from variating attributes.
- Attribute completeness against the category refinements is a ceiling on filtered traffic. Measure it before buying any content work.
- Copy still matters. It matters second, and it matters most once the attributes are complete.
If you want to know where your ceiling sits before commissioning another content project, book a thirty minute call. We will run the refinement exercise on one of your categories and show you the number. Where the gap is large, product content enrichment against the channel schema is the fastest route to closing it. The attribute model underneath is what makes the gain permanent.