AI e-commerce sources: the data SEO teams overlook

Avant de plonger dans le détail, prenez trente secondes pour visualiser la carte complète des sources IA. Trois d’entre elles restent massivement sous-exploitées en e-commerce, et c’est précisément là que se joue votre prochaine croissance.

The blind spots I see on 15 e-commerce sites every week

I look at 15 sites a week. They all share the same problem.

They track Google. Sometimes Bing. They watch their rankings, their clicks, their impressions. Meanwhile, answer engines pull from elsewhere.

An e-commerce client with 1 200 references calls me on a Tuesday morning. He invested 8 000 $ in content production over seven months. Result: 4 000 organic sessions per month. A plateau.

The root cause was the sources.

An analysis by Chris Green published on Search Engine Journal on September 22, 2026 shows that AI Search relies on a source hierarchy far broader than the Google/Bing pair. Many SEO teams face what he calls « source myopia ».

Search is more than « is it in Google » or « Bing »? Where should you ensure you’re getting mentioned?

In other words: showing up in Google or Bing is just the start. What matters is where your brand gets cited, and how it’s structured.

For e-commerce, this changes everything.

Chris Green’s table, filtered through my e-commerce deployments, shows three blind spots you can act on right away.

So we have two types of sources: those that answer in real time, and those that trained the model. The first get fixed in days. The second get prepared over months.

First observation: most sites have a solid Google Merchant Center feed. But the OpenAI feed remains absent in 9 out of 10 cases. That’s the scale I see with my clients.

Second observation: Yelp and Google Business Profile local listings are rarely linked to product pages.

Third observation: historical corpora like Common Crawl or C4 require a different approach. You can leave a clean trace there starting today.

Traditional tools show your Google rankings. Add your citations in AI Overviews. Add your listings picked up by Copilot. Add your products cited by ChatGPT.

That’s the starting point of the audit.

Why is this urgent?

Because AI answers move beyond the results page. They aggregate feeds, listings, databases. When your source is clean, the AI cites it. When your source is messy, the AI cites another source. Your page stays out.

We now fight to be the source. The click becomes secondary.

OpenAI product feeds: the most profitable blind spot

Blind spot.

The first blind spot is the OpenAI product feed. Separate from Google Merchant Center. OpenAI accepts CSV or JSON feeds refreshed as often as every 15 minutes, according to the documentation cited by Search Engine Journal.

What does that mean?

ChatGPT can display your product listings with price, stock, lead times, images. A click on your site becomes optional. Clean data alone does the work.

Most e-commerce merchants set up their Google Merchant Center feed because Google Shopping requires it. Good. But who sets up the same feed for ChatGPT?

Almost nobody.

Order of magnitude: only 1 out of 10 sites I observe among my clients has a dedicated feed for conversational agents.

The mechanism is simple.

• A stable product identifier.
• A title built from the real query, distinct from the internal name.
• A factual description under 160 characters.
• A tax-inclusive price matching the page.
• A clean canonical URL.
• Availability status.
• Estimated delivery time.
• GTIN or EAN when available.
• Brand.
• Main image.

Ten fields, all simple. What pays off is rigor.

15 minutes. That’s the refresh frequency OpenAI accepts. Your feed stays alive. A price change, a stockout, a restock: everything shows up almost in real time.

Move past indexing delays. Here you connect a direct source.

Watch one detail: the feed title is separate from your title tag. It’s a distinct field. You can put the short brand name, the reference, the main feature. It all depends on the target query.

A good feed title = product name + a distinguishing feature + a use case. Example: « Miele Complete C3 Silence EcoLine 890 W canister vacuum » rather than « C3-890-W ».

In practice: when a user asks, « which quiet vacuum for 80 m²? », ChatGPT has a clean listing to cite, instead of a page to figure out.

An e-commerce client with 800 references moved from 0 product mentions in ChatGPT to 47 citations in six weeks. The site pages stayed untouched. Number observed on a recent deployment.

Mechanism and result.

The product feed is the new title tag.

Another point: the OpenAI feed compléments Google Merchant Center. It adds a new channel. The two sources coexist in tier 1 and tier 2 of Chris Green’s table.

Synchronize the data, and price or stock discrepancies vanish. The same information everywhere, all the time.

Google Merchant Center and local data: the tier-1 layer you can win in 48 hours

Chris Green’s table groups sources into four levels. Tier 1 includes confirmed, current sources, used in RAG, grounding, or direct actions.

In tier 1, you find Google Search, Bing Search, Google Merchant Center, Google Maps, Google Business Profile, and Yelp.

For e-commerce, that means three things.

• Your Merchant Center feed must be flawless. More than active: flawless.
• Your Google Business Profile listing must link to your product pages, with recent photos, accurate hours, and consistent service descriptions.
• Your Yelp listing, if you have a store or in-store pickup, must carry the same information as your Google listing. Full alignment.

The common thread: these are sources AI systems consult today, in real time, to build an answer.

The gap is clear. Teams spend hours on page metas. They treat their Yelp listing as secondary. Yet, for e-commerce with pickup, Yelp is a tier-1 source in Chris Green’s analysis.

And this goes beyond the US. The principle holds for any high-authority local directory: Yellow Pages, TripAdvisor, industry directories.

The good news: activating these sources takes 48 hours.

Check Merchant Center feed fields. Fill the missing attributes: availability, delivery time, strikethrough price, GTIN, brand. Harmonize local listings.

Simple example: a bike e-commerce with 3 stores. On its Google listing, it says « sales and repair ». On Yelp, it says « sales only ». On its site, the store page says « on-site workshop ». The AI receives three versions. It waits for consistency.

Consistency decides.

Observed result for a client with 12 stores: 23 local queries moved into AI Overviews visibility within three weeks. The harmonization alone produced the result, using the existing pages and blog content.

Content was fine. The tier-1 source consistency was the lever.

It’s the fastest win for e-commerce.

Why? Because local data already exists. You just make it consistent. AI systems need a clean listing to cite, more than they need to discover your site.

Another point: link your local listings to your product pages with clean URLs. Each listing becomes an entry point to a category, a product, an offer.

A price discrepancy between your listing and your page creates friction. Conversational agents favor consistent data.

Common Crawl and C4: historical data you still influence

Third category: historical corpora. Common Crawl, C4. These drive the pre-training of large models.

Search Engine Journal picked up the analysis: GPT-3 used filtered Common Crawl for about 60% of its sampling mix. LLaMA 1 reports 67%. And C4, a cleaned version of Common Crawl, accounted for 15% of LLaMA’s pre-training.

Three numbers, one trend.

Common Crawl requires a different approach. You influence the quality of what these corpora capture.

How?

• Clean HTML, with only useful tags and controlled duplicate content.
• Valid product structured data on every relevant page.
• Internal linking that connects important pages together.
• Fast, accessible pages, allowing crawlers on the content you want cited.
• Consistent price and availability archives over time.
• An up-to-date sitemap, readable by crawlers.

The trap: many sites block corpus crawlers (GPTBot, CCBot) out of fear. They vanish from training sources. Worse, they leave draft versions indexed.

Blocking GPTBot means choosing to exclude a training source. Make that choice consciously. Block everything, and you give competitors the entire field.

The reverse approach: allow crawlers on clean public pages, block test zones, carts, accounts. Exactly as for Googlebot.

Everything happens here.

An e-commerce with clean HTML and valid product schema earns more reuse in synthetic answers, whether or not a dedicated feed exists. That’s what I observe with my clients.

The mechanism is simple: corpora capture, filter, clean. A clear page passes the filter. A polluted page vanishes.

A 1,500-page site with consolidated semantic architecture saw its indirect mentions in AI answers climb 34% in four months. The product feed remained unchanged. Just structural cleanup. Number observed on a recent deployment.

You earn a slot in Common Crawl. You build what stays there.

And it pays over time. Each clean page added to the corpus increases the probability your brand gets picked up, rephrased, cited.

Product schema is the ID card of your pages for parsers. A valid Product Schema: name, reference, price, availability, image, reviews. Fill every field.

Corpora read your tags. That’s the only signal they use.

Marketplaces and Shopify feeds: the rising tier-4 source

Chris Green’s table places marketplace feeds and Microsoft Merchant Center in tier 4. Strong evidence, with final confirmation pending.

But there’s more. Shopify already integrates with ChatGPT for shopping, according to the analysis. Shopify catalogs serve as a product data source.

So if your e-commerce runs on Shopify, you already have an entry point. The condition: your catalog must be tight.

What does that involve?

• Each product variant with a distinct SKU.
• meta_title and meta_description fields filled on every listing.
• Images with descriptive alt text, distinct from « product-image-01 ».
• Consistent product types, aligned with Shopify catégories.
• Accurate price and availability tags.

A clean Shopify catalog becomes a usable source for conversational agents. Beyond Google, other systems read it.

A Shopify client with 200 poorly named variants renamed them, aligned SKUs, fixed alt attributes. Three weeks later, their products started appearing in ChatGPT shopping answers. The ad budget stayed the same.

Tier 4 is a secondary priority. But it’s a low-effort move if your Shopify foundation is already in place.

Same for marketplaces. If you sell via Amazon, Cdiscount, Fnac, your marketplace listings get pulled by intermediaries. Tier-4 sources can become tier-2 tomorrow.

Prepare for it now.

Microsoft Merchant Center follows Google Merchant Center logic. Chris Green’s analysis places it in tier 4, strong evidence. It’s still open whether it will become a direct source. But preparing your Microsoft feed takes an hour if your Google feed is already clean. It’s a near-zero-cost bet.

Passons à l’exécution. Voici les cinq étapes à dérouler dans l’ordre pour activer vos sources AI Search, du plus rapide au plus structurel.

The 5-step action plan to activate your AI Search sources

Audit synthesis. Five steps, in order, for e-commerce.

Most teams stop at step 0: Google Search Console.

It’s comfortable. It’s only the beginning.

Tier-1 sources take 48 hours. OpenAI feeds take one workday. HTML cleanup takes two weeks. Historical corpora take ongoing work.

The result shows up in citations, indirect clicks, presence in conversational answers. Average position becomes secondary.

One final number: 15 minutes. That’s the refresh frequency OpenAI accepts for product feeds, according to the documentation cited by Search Engine Journal.

The three classic traps:

• Believing SEO stops at Google.
• Confusing traffic with citations.
• Waiting for sources to stabilize.

Each one delays activation. Meanwhile, a competitor structures their feeds. A catalog becomes a clean source.

Here, consistency beats complexity.

A consistent product feed. A complete local listing. Clean HTML. That’s all.

Which of your competitors already plugged it in?

We build systems that run. Reports become secondary.

I share the method. I show you the pages.

And you, which neglected source is your next competitor already leveraging?

Stéphane Jambu

SEO & AI Engineer

I build growth systems / AI / Neuroscience | 650+ clients · 80 LinkedIn testimonials · 30 years of expertise · 15 years of systems running without me.

Follow on LinkedIn