Perci vs. ChatGPT: the copy was never the hard part

Writing the title, bullets, and description is one part of publishing an optimized listing. Here is the rest, for every ASIN:

• Pull the listing's current content and backend data.
• Research competitors and Amazon keywords for that product.
• Find the exact attribute fields Amazon expects for its product type.
• Find accepted, truthful values for those fields.
• Check the copy against Amazon restricted and caution phrases.
• Structure variation families so parents and children stay consistent.
• Publish everything to Amazon and confirm it landed.

None of that is writing. All of it is per-ASIN. A general AI tool helps with roughly one item on that list unless you build and maintain the surrounding data connections yourself.

Most of these are not cases where ChatGPT is simply worse. They are cases where the answer depends on external Amazon data. Without that data connected, ChatGPT can still answer, but it cannot derive the missing numbers from language alone.

It doesn't have keyword data, so it invents keywords

Ask ChatGPT for target keywords without providing Amazon data and it returns a clean, confident, plausible list. It is not derived from search volume, your Search Query Performance report, or what competitors rank for. It is a language model producing words that sound like what Amazon shoppers might type.

Twenty-five of the 36 keywords had zero search volume. On the long-tail list, 15 of 18 were zero, including claw grip wireless gaming mouse, ultra lightweight mouse 80g, 1ms wireless mouse, and hero 25k sensor mouse. They are accurate product descriptions that shoppers were not searching in the measured dataset.

It did get some right. Lightweight gaming mouse at 7,004 monthly searches, ambidextrous gaming mouse at 1,246, rechargeable gaming mouse at 1,004, and rgb gaming mouse at 1,400 were all real. The list was not fiction. It was roughly 31% real.

Nothing in the output identified which 31%. Every term arrived with the same confidence and formatting. About 84% of the usable volume came from wireless gaming mouse, a broad term with 82,261 monthly searches. The full list returned about 97,850 in combined monthly volume; remove that one term and the rest represented roughly 15,600 between them.

What the data found that ChatGPT did not

The same product, researched with Amazon keyword and competitor data, surfaced separate groups of buyers that the generated list missed.

If you would rather not use Perci, use your Search Query Performance report or another keyword tool with real Amazon data. ChatGPT can help analyze an uploaded export. Do not ask a language model to guess the dataset.

It doesn't have your category's schema, so it approximates one

Start with one field.

We asked ChatGPT for the attributes on a Hands Free Lights listing. It returned recommended browse node 49980143031, described as "currently shown publicly." That node does not exist in Amazon UK.

The answer was not hedged. It was a specific eleven-digit number presented alongside otherwise sensible advice. Paste it into a flat file and you would be debugging the rejection rather than questioning the source. That is the shape of the problem: the output looks the same whether the model knows or not.

Then look at the scale of the list.

Amazon's actual schema for this category contains 27 required and 18 recommended attributes: 45 fields. ChatGPT returned around 70. More than two dozen suggestions did not exist in this category's schema, including material, shape, light colour, wattage, battery capacity, charging time, connector type, switch style, beam angle, mounting type, water resistance level, target audience, and occasion type.

A plausible superset is not usable. To act on 70 suggested fields, you must check all 70 against Amazon's real schema. You end up doing the research anyway, with extra rows to disprove.

It also missed fields with consequences: Dangerous Goods Regulations for a lithium-battery product; the package hierarchy fields; and Is Fragile, Power Plug Type, Colour Map, and Product Subcategory. It returned one flat list without distinguishing required fields from recommended ones.

The response contained at least 15 hedges such as "if applicable," "if known," "if field exists," and "depending on template." Those caveats read as thoroughness while handing verification back to the seller.

It gets plenty right. In the same answer, ChatGPT correctly named external product ID, country of origin, number of items, package dimensions and weight, energy efficiency class, control method, voltage, finish type and colour temperature. It is not useless. It is approximate; Amazon's flat files are not.

To be clear about the alternative: you can get the real fields for free. Download Amazon's category template and the exact schema is sitting in the header row. No model required, no guessing. If your catalog is small, that's what you should do.

What neither the template nor ChatGPT does is fill it in. The template tells you there are 27 required fields for this category; it doesn't find the values, doesn't cover the other categories you sell in, and doesn't do any of it five hundred times.

Perci reads the schema from Amazon's API, including the exact fields for that category, the values Amazon accepts, and whether each is required or recommended, then fills them across the catalog. The schema was never the expensive part. The filling is.

It does not maintain Perci's restricted-phrase bank

Perci maintains a curated bank of restricted and caution phrases based on the wording Amazon investigates or suppresses. ChatGPT can follow a policy list you provide, but a general conversation does not include Perci's maintained bank or its update process.

It does not bring the Amazon workflow with it

ChatGPT supports file uploads, data analysis, and connected or custom apps. But without an Amazon-specific integration, you still assemble current listings, competitor research, SQP data, and category templates; then upload flat files, read error reports, and fix rejected rows.

The copy inside the cells gets better. The process surrounding those cells remains yours to build and maintain.

If you have less than 10 listings, use ChatGPT.

Write the copy with AI. Pull keywords from Search Query Performance or a keyword tool rather than asking ChatGPT to invent them. Fill the attributes from Amazon's category template and upload the result manually. It works, it can cost nothing beyond your time, and buying specialized software may be overhead at that size.

The tradeoff is not copy quality. It is time per listing multiplied by listings.

Doing it yourself, per listing:

30 minutes

Assemble current listing, competitor, and keyword inputs

5 minutes

Prompt and edit the listing copy

25 minutes

Find and fill the category's attributes

25 minutes

Build flat files and fix upload errors

Perci: approximately 45 seconds per listing, with up to about 500 listings in a run.

At 10 listings, the difference is an afternoon. At 200, it is the difference between a week of work and a coffee break. At 1,000, doing it by hand is not a plan.

One credit. No card.