⭐ Get the Agent OS + join 3,400+ founders inside the AI Profit Boardroom → Join AIPB

GoldieBench Blog · 6 min read

AEO For Ecommerce + Which AI Models To Use For Product Pages, Answer Briefs And Markup, With Real GoldieBench Scores

Product pages, briefs and markup now start as an AI draft. Here is what live GoldieBench scores say about which models to use, and where people still matter.

Illustrated hero: gold leaderboard bars beside a product page wireframe and a shopping bag on a dark navy background

AEO for ecommerce now runs on AI models for much of the drafting: product page copy, buyer-question briefs and Product markup all start as a model's first attempt.

Book a free SEO strategy session →

Which model you hand that work to decides how much fixing follows, so this page uses only scores published on live GoldieBench pages, each quoted with its source.

For the parts no model can do, such as feeds, crawler rules and reviews, book a free SEO strategy session with my team.

How GoldieBench scores a model

The GoldieBench models page says every model gets the same fixed prompts, one shot each, scored 0–10 by Julian.

The Landing task page spells out the rules for that task: a single HTML file, no iteration and no examples in the system prompt.

That makes the scores a fair read of first drafts, which is exactly what a store owner gets when a model writes a product page.

What the bench does not measure is SEO accuracy, so treat these numbers as page-building skill, not proof a page will be cited.

The task closest to a product page: Landing

Landing asks for a modern marketing landing page in one shot.

The task page lists 26 models tested, 21 of 26 scored, and an average of 8.11/10.

Fugu Ultra, Fusion, GLM-5.2 and Grok all scored 9.0/10 on it, according to the same page.

For a store, a product page is a landing page with stricter facts, so this is the nearest signal on the bench.

Scores for the models stores reach for most

ModelLandingAverage (all tasks)Where I read it
Grok9.08.09Landing task and models page
Fusion9.08.59Landing task and models page
GPT-5.6 Sol8.78.16goldiebench.com/models/gpt56
Gemini 3.6 Flash8.67.08Landing task and models page
Claude Opus 58.48.27goldiebench.com/models/opus5

Two readings matter here.

Landing shows how polished a single page comes out, while the overall average shows how steady a model is across very different jobs.

Claude Opus 5 has the second-highest average on the models page at 8.27, even though several models beat it on Landing.

What these scores mean for product page drafts

A high Landing score suggests the model structures a page well on the first try: clear headline, sections, calls to action.

Book a free SEO strategy session →

That helps with layout and readability, but it doesn't check whether your stock, sizes or return terms are right.

Google's generative AI search guide rewards first-hand, non-commodity detail, and no model has that detail until you give it.

So feed the model your own product notes and customer questions, then judge the draft on accuracy before style.

Markup is a separate job, and the bench doesn't score it

GoldieBench has no structured data task, so none of its numbers tell you which model writes Product JSON-LD best.

Use any strong model to draft the markup, then verify it yourself against Google's merchant listing documentation: name, image and an Offer with price and currency are required.

Paste each draft into the Rich Results Test, and reject any rating the model added that isn't visible on the page.

A model pairing for a small store, step by step

Take a pet supplies shop with 80 products.

  1. Export the catalogue and the most common customer questions per product.
  2. Give one high-Landing model the job of drafting answer-first product page layouts.
  3. Give a second model with a high overall average the job of checking each draft against the catalogue file.
  4. Draft markup from the catalogue and check it in the Rich Results Test.
  5. Publish five pages a week, and watch Search Console's generative AI performance reports.

Splitting draft and check across two models catches more errors than asking one model to mark its own work.

Keep a person as the final step, reading each page as a buyer would and fixing anything that sounds like a supplier brochure.

A pet supplies page that says which breeds a harness fits, measured by the owner, beats a polished paragraph that says nothing new.

How to brief a model so the score carries over

A benchmark score is earned on a fixed prompt, so a vague brief from you will drag any model below its number.

Book a free SEO strategy session →

  1. Paste the real product facts first: dimensions, materials, compatibility and what is in the box.
  2. Paste the five questions buyers ask most, copied from support emails.
  3. Tell the model the page must answer each question in two or three sentences near the top.
  4. Ban invented claims outright, and ask it to write [CHECK] wherever it is unsure.
  5. Ask for one suggested photo or short clip that would show the product in use.

Google's guide says images and video can bring more chances to appear beyond plain links, so the shot list is worth the extra line.

This brief turns a model that scores well on Landing into one that writes useful product pages rather than pretty generic ones.

What no model can do for your store

  • Allow OAI-SearchBot and PerplexityBot through robots.txt and your CDN, which OpenAI's crawler page and Perplexity's crawler docs recommend.
  • Keep a Merchant Center feed in sync with your live prices and stock.
  • Collect real reviews from real buyers.
  • Claim your store in ChatGPT Business Hub, which OpenAI's help page says needs an authorised owner and verification.

Those jobs decide whether engines can read and trust your store, and they sit outside any benchmark.

A note on the promo task for store ads

If you also want animated product promos, the Aipbpromo task page lists an average of 7.48/10 and names Qwen 3.8 as the winner.

That lower average is a reminder that motion work is harder for models than static pages.

For product ads, budget extra review time for motion drafts and keep the claims in them identical to your product pages.

Where people still beat models

Nobody can promise a product appears in an AI answer, and Google's hiring guidance says no one can guarantee a #1 ranking.

What a team adds is judgement: which products to fix first, which questions buyers really ask and how to earn links to your best guides.

Goldie Agency covers SEO, GEO, AEO and link building, with manual outreach, no PBNs, and Julian's own published numbers list 240+ clients; the free Goldie Agency strategy session is where every engagement starts.

Book a free SEO strategy session to see which of your products to start with.

Book a free SEO strategy session →

Also On Our Network

Pick a high-Landing model to draft, a steady model to check, and book a free SEO strategy session when you want people on the parts of AEO for ecommerce no model covers.

FAQ

Which AI model writes the best product pages?

On the GoldieBench Landing task, Grok, Fusion, Fugu Ultra and GLM-5.2 each scored 9.0/10. Landing measures one-shot page building, not SEO accuracy.

Does GoldieBench test schema or structured data?

No. There is no structured data task on the bench, so verify any model's Product markup in Google's Rich Results Test.

Why use two models instead of one?

One drafts and the other checks the draft against your catalogue. Splitting the jobs catches more errors than self-review.

Can a model get my store into ChatGPT answers?

No model controls that. Crawler access, feeds, real reviews and a Business Hub claim are owner jobs, and OpenAI says placement isn't guaranteed.

The same stack Julian uses

Run this stack yourself.

Every demo on this bench was built inside the Agent Operating System — one prompt, one shot, single HTML file out. The Agent OS, the prompts, the templates, the weekly walkthroughs and 3,400+ founders shipping with it every day all live inside the AI Profit Boardroom.

3,400+founders
258documented wins
38countries
4coaching calls a week