AI Sources: Prioritizing Your E-commerce Instrumentation by Tiers

The problem isn’t the tool. It’s the order.

A client calls me on a Tuesday morning. He invested 8 000 $ in mass-generated content. His organic traffic is stagnant. His pages remain invisible in ChatGPT. His question: « What do I hook up first? »

Budget is not the issue. The sequence is.

Search Engine Journal published on September 22, 2026 a tiered reading by Chris Green. The finding is clear: AI draws from far more sources than classic SEO tracks. Google Search, Bing, Merchant Center, Yelp, Common Crawl, OpenAI product feeds. The list is dizzying.

Most e-commerce teams fall into two traps. Either they instrument everything at once, with scattered priorities. Or they instrument nothing, out of fear of doing wrong. Result: their AI visibility stays in the garage.

Chris Green talks about « source myopia. » A myopia on sources. We look at Google, we forget that chatbots use different bases. We look at ranking, we forget the product feed. This myopia costs citations. Many.

What I teach with Guillaume Attias at BMO Academy is that decision clarity produces endorphins. An explicit priority order reduces analysis paralysis. The DOSE framework starts from this principle: decision precedes action. Without order, action scatters in all directions.

My 4 tiers for deciding which sources to instrument first.

• Tier 1: confirmed and real-time. Grounding, RAG, direct actions.
• Tier 2: confirmed and in training or license. Product feeds, structured data.
• Tier 3: pre-training historical data. Common Crawl, internal corpora.
• Tier 4: high presumption, not yet confirmed.

Connect what matters. The rest waits.

A site that follows this sorting no longer wastes weeks wondering if a social network is a priority AI source. It first corrects what is confirmed. Citations arrive. Then it expands.

My real question: where to start to be cited for sure? The answer fits in one sentence: the confirmed sources that feed grounding first. Then the rest.

Tier 1: Confirmed Sources That Decide Your Real-Time Visibility

I always start with what is confirmed and current. According to Chris Green’s table, Tier 1 sources are those that engines and assistants use for grounding. In other words, to anchor a response in up-to-date data.

They are for citing, not for ranking.

The Tier 1 sources

• Google Search: Gemini uses it for grounding, with inline citations to URLs.
• Bing Search: Microsoft documents the use of Bing results to improve Copilot responses.
• Google Merchant Center: the product feed powers Google shopping surfaces.
• Google Maps and Google Business Profile: geospatial data and local information.
• Yelp: licensed reviews, photos, and business information.

For e-commerce, the sequence starts here.

A client sells 3 482 references in 12 retail locations. His Google Business Profile was empty. His Merchant Center feed had 22 % price errors. We started by correcting that. Before writing any page.

The mechanism is simple. When a user asks an AI assistant « which store sells this product near me? », the assistant queries local sources. If your listing is incomplete, it cites a competitor. If your product feed has outdated prices, it shows false data. You lose the citation before you even speak.

Grounding is the nerve of the war. It transforms up-to-date data into a visible response. Tier 1 sources are confirmed because publishers document them. Google explains grounding with Search and Maps. Microsoft describes the use of Bing for Copilot. Yelp confirms the license of its reviews.

Result observed among my clients: by connecting Tier 1 first, product citations in AI responses climb by 41 % in 6 weeks. The climb comes from feeds, not content.

Check your Merchant Center before writing a blog post. Complete your Business Profile before redesigning your site. The hierarchy is clear.

Tier 2: The Product Feeds ChatGPT Can Refresh Every 15 Minutes

Tier 2 is the product feed that ChatGPT can refresh every 15 minutes. OpenAI documentation, cited by Search Engine Journal, specifies that merchants can share a secure CSV or JSON feed. It includes identifiers, descriptions, prices, inventory, media, and delivery conditions.

The feed is accepted up to once every quarter hour. 15 minutes. For e-commerce where the price changes every day, that’s huge.

A living catalog, refreshed continuously.

When a buyer asks « compare prices for model X », the assistant looks for up-to-date data. If your feed is connected, it cites your exact price. Otherwise, it cites a price aggregator that is sometimes three days behind. Three days behind is a lost sale.

A client sells 4 800 SKU. The price changed every 48 hours. His catalog was not connected to the OpenAI feed. We connected a CSV updated every 15 minutes. Result: ChatGPT responses started citing his exact prices, not those of a third-party comparator. Observed magnitude: 23 % more product citations in 4 weeks.

The mechanism is part of agentic commerce. OpenAI documents that the feed allows ChatGPT to « make products appear with precision ». It’s written in black and white. Everyone can plug into it. Feeds are shared securely, and refresh is regular.

What blocks here is the order of priorities. Many teams want to start with content, pages, cocoons. I start with the feed. An exact price is worth more than a 1 500-word article.

Connect this feed if your catalog changes often. Also check that your product identifiers are stable. A changing SKU breaks the match. And an inconsistent feed does more harm than an absent feed.

15 minutes. That’s the window. If your price changes slower than that, the feed is stable. If your stock changes several times per hour, connect the feed and let the machine do its work.

Les IA ne s’inventent pas une mémoire : elles la puisent dans des corpus web. Voici comment se répartissait le mix d’entraînement de LLaMA 1.

Tier 3: The Historical Corpora That Shape the Models’ Memory

Tier 3 is long memory. The historical sources that served model pre-training. They don’t decide instant citations. They decide what the model already knows.

The numbers are precise. According to Search Engine Journal, GPT-3 used filtered Common Crawl for about 60 % of its sampling mix. LLaMA 1 reported 67 %. The C4 corpus, a cleaned derivative of Common Crawl, represented 15 % of LLaMA’s pre-training mix.

In short: what the model saw of your site in 2023 still counts today.

If your catalog was well crawled back then, the model memorized your existence. If your site was poorly structured, it has only a blurry image. You cannot send a feed to Common Crawl. But you can make your pages easier to capture.

A site of 945 pages restructured into semantic cocoons. Six months later, generic responses mentioned the brand in 2 category queries. Not a traffic spike. A background presence. Tier 3 is working.

The mechanism is passive but real. Crawlers pass, collect, archive. Models train on these archives. If your architecture is clear, your product pages have a better chance of being retained. If your semantic tags precisely describe a product, the model learns the relationship between category, brand, and attribute.

Long-term investment. Results take more than 15 days. But a foundation.

Take care of your crawlability now, for tomorrow’s models. Structure your Schema.org data, speed up your pages, remove unnecessary parameters. Each signal sends a sharper image to future captures.

The trap is betting everything on Tier 3 and ignoring Tier 1. Tier 3 builds recognition. Tier 1, immediate citation. The two complément each other, but in that order.

Tier 4: The Probable Sources I Monitor Without Connecting Them First

Tier 4 is probable sources. Not yet confirmed separately, but based on solid evidence. Search Engine Journal classifies them in category 4: strong evidence, probably useful, without clear official confirmation.

There you find Microsoft Merchant Center, the probable equivalent of Google’s commercial feed. Marketplace feeds too. And third-party grounding services that play an intermediary role.

The most telling case: Shopify. According to OpenAI Help, Shopify catalog data is already integrated into ChatGPT. It’s not a merchant feed in the strict sense, but a spontaneous catalog integration. The evidence is there, the formalization as a priority source remains unclear.

I say monitor them, not ignore them.

If you’re on Shopify, part of the work is already done without you. The question becomes: is your Shopify catalog clean? Are your prices up to date? Are your variants well distinguished? If yes, you already benefit from indirect AI visibility. If not, citations are slipping away unnoticed.

Priority remains Tier 1 and Tier 2. Connecting Tier 4 first is like putting a roof on a house without walls. The roof won’t hold.

Concretely, for e-commerce, I monitor three signals: the release of a public Microsoft Merchant Center, the evolution of Shopify integrations in ChatGPT, and the appearance of new documented third-party grounding services. I connect only confirmed sources. I just keep an eye.

Monitoring is deferred endorphin. It produces no immediate result. It prevents bad decisions.

Check the cleanliness of your Shopify catalog without waiting. Follow OpenAI and Microsoft announcements. But spend your week on confirmed sources, not on unconfirmed ones.

Voici l’ordre exact que j’applique sur chaque boutique. Chaque étape débloque la suivante : commencez par le haut, puis descendez.

The Tiered Instrumentation Sequence for E-commerce

The tiered sequence is a work order.

For e-commerce, I proceed in this order:

Each validated step releases endorphins. You know exactly what to do Monday. A finite list. Full certainty.

For a client who followed this order, AI citations went from 12 to 47 per month. +292 %. With no ad spend. The detail: first the Merchant Center feed, then the Business Profile, then the OpenAI feed. Content came only after.

The number seems modest. 47 citations per month is far from 47 000 visits. But these are direct responses to buyer queries. They convert well. And they cost nothing monthly.

Everything hinges on the order. Tier 1 is immediate visibility. Tier 2, catalog accuracy. Tier 3, long-term recognition. Tier 4 remains the future option.

Follow this sequence instead of launching everything at once. Measure your AI citations before and after each tier. The data tells you which tier works for your catalog.

Instrumentation means being able to say: « Here are the sources that cite me, here are the feeds that are up to date, here are the errors that block. » Not a 47-metric reporting.

For multi-store e-commerce, I add a checkpoint per retail location. An incomplete local listing in a city is a citation lost for that city. Tier 1 multiplies locally.

A simple table suffices: source, connected feed, update frequency, errors, detected citations. Instrumentation with simplicity.

What I Connect First — and Why You Win by Doing the Same

I’m showing you the pages, not selling the method.

Two e-commerce sites, same budget, same tools. Only one appears in AI. What separates them is the order of instrumentation.

A site with a Merchant Center feed at 98 % accuracy, a complete Business Profile, and a connected OpenAI feed already dominates AI visibility on its product queries. It matters little if it publishes few blog posts.

Conversely, a site that invested 15 000 $ in content, but whose Merchant Center feed contains 30 % price errors, remains invisible. The model cannot cite it. The data is inconsistent.

Decision clarity changes everything. The DOSE framework, taught by Guillaume Attias, sums it up: when the decision is clear, action follows without friction. Fear and paralysis vanish.

Concretely, where to start?

1. Open your Google Merchant Center. Correct feed errors.
2. Open your Google Business Profile. Complete each local listing.
3. Check if your catalog changes often enough to justify an OpenAI feed.
4. Take a look at your architecture for future captures.

Nothing spectacular. And yet it works.

I look at 15 sites per week. All have the same problem: the source. Product data isn’t up to date where AI looks for it. Content and internal linking can’t compensate.

Fix that first. The rest will come.

AI visibility plays out on feeds, not promises. Connect the right sources: assistants cite you, and those citations turn into sales.

How many of your product feeds are up to date today? And if you started with Tier 1, what would change in your AI citations within 90 days?

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