ChatGPT: I read the network traffic, not the answers – what changes
Summarize this article with AI
A client call that forces you beyond GEO beliefs
A client calls me on a Wednesday afternoon. He sells garden equipment online. 800 product sheets. Solid organic traffic, 4,000 sessions per month. But for two months now, his most technical pages — lawnmower comparisons, maintenance guides — no longer appear in ChatGPT responses.
« We’ve redone everything, Stéphane. Long-form content, markup, listicles. Nothing. »
I don’t answer him with a GEO hack. I open the network console in my browser. I run the same query his potential customers would: « which electric lawnmower for sloped terrain. » And I read, not the text displayed by ChatGPT, but the JSON that circulated behind the scenes to build that answer.
What I found there changed how I see optimization for generative AI. No magic. Raw signals.
Découvrez le processus technique que révèle l’analyse du trafic réseau. Chaque étape correspond à un champ ou une action observée dans les appels entre ChatGPT et votre navigateur.
Le pipeline invisible des sources ChatGPT
Comment l’IA sélectionne ses citations via Bing et les champs réseau
Read the network traffic, not the visible response
Suganthan Mohanadasan’s study (Snippet Digital, Search Engine Journal) aligns with what I see with my clients. He captured 1,240 source records over a few days in exchanges between the ChatGPT server and his browser. Rather than guessing sources from the final response, he uncovered internal instructions.
I did the same for a garden client. Over a week, I collected network calls for 47 queries across his catégories. I extracted the result_source fields, turn_use_case, calls to Bing APIs, the search_query sent by the AI, and the actual model used (gpt-4o or a derivative).
What you see is raw and often opposite to intuition. The AI doesn’t necessarily cite the best page. It cites the one fetched by a process whose rules appear in these exchanges.
Example: 42% of sources came from a web search query launched directly, mostly Bing. The rest mixed the model’s internal knowledge, structured data, and pages already validated in other sessions. Général popularity isn’t the criterion. It’s the pipeline visible in the JSON that decides what gets found and filtered.
What the machine admits to itself
The most telling field is called result_source. It gives the administrative category of each text snippet used. I logged these values: organic (standard web result), news, knowledge_graph, citation, and internal_store (for product sheets from partner databases).
What jumped out at me: no source labeled organic was a non-crawlable page. Not one. The AI pulls directly. If your page hasn’t been indexed by Bing and returned as a relevant result at the time of the query, it doesn’t exist in this pipeline.
Second discovery: the turn_use_case field classifies the nature of the user’s question. « Comparison, » « procedure, » « definition, » « purchase. » Important. Depending on this tag, the system activates or deactivates web search, queries or doesn’t query the knowledge graph. Content optimized for a « list » will have little chance if the AI tagged the query as « procedure » and triggered a call to a single authoritative source.
Third signal: vendor. This field names the source the AI is about to cite. And this name isn’t extracted from the title tag. It’s linked to the presence of an entity card, a Wikidata page, a mention repeated in validated corpora. That means without a third-party validation layer, your brand can be invisible even in first position on Bing.
These fields are not hypotheses. They’re in the JSON. Verifiable. Reproducible.
Three signals to cement before touching content
From this technical reading, I draw three commandments for an e-commerce site.
1. Strict crawlability for Bing. ChatGPT doesn’t go through Google for its sources. It’s Bing it queries. So robots.txt, meta robots, response time, URL structure — everything that slows Bing indexation becomes a problem. I saw a 2,400-word guide at a client’s: it timed out because of a poorly configured cookie consent script. Result: not indexed by Bing, not cited.
2. Verifiable facts and external validation. The AI loves claims you can verify outside the site. A product page with a certification date, a standards number, a figure cited by a trusted organization — these elements are noted as facts, and they rise in ranking. I call it the « useful truth rate. » No need for thousands of backlinks. Just mentions in databases, professional association sites, high-authority thematic articles.
3. Triggering by the query. The turn_use_case decides. Your content must match the intent the AI detects. For an e-commerce merchant, product sheets alone aren’t enough. You need content answering comparisons, « what criteria for, » buying guides — formats the AI knows how to exploit.
When I audit a page, I ask three questions: Is it crawled by Bing in under 800 ms? Does it contain at least one verifiable fact with an external source? And does it answer an intent ChatGPT classifies as comparative informational?
With these three lights on, the probability of being cited goes up directly.
Suivez le parcours mesuré du client jardin : depuis l’appel initial jusqu’à l’obtention de 17 fiches produits citées par ChatGPT, en passant par l’audit technique et la restructuration des contenus.
6 mois pour passer de zéro à 17 pages citées
Les étapes clés de la transformation d’un site e-commerce jardin
800 products, zero citations… until 17 pages in 6 months
Back to my garden client’s call. I applied the three-signal grid. 62% of his product sheets weren’t properly crawled by Bing in under a second. Rendering issue due to asynchronous loading of descriptions. Pages appeared empty.
Second problem: very few verifiable facts. The sheets said « Robust, » « Powerful, » but no visible CE certification mention, no spare part number, no wattage power value sourced. Words, not facts.
We reworked 47 strategic pages. Added Product structured data with identifier (MPN, GTIN), integrated « Verifiable Tech Sheet » paragraphs with a link to the manufacturer’s site. In parallel, secured a mention in an online garden magazine (not a link, just a brand citation in a comparison).
Third lever: created a comparative guide « Electric Lawnmower for Steep Slopes » structured in questions/answers, FAQ markup. This format matches the turn_use_case « procedure/comparison. »
Result observed after 6 months: 17 pages cited by ChatGPT from the 47 reworked. Zero citations before. Organic traffic up 23%, but especially a direct incoming channel from the AI interface, identified in GA4 via the chat.openai.com referrer.
The client didn’t spend a word more. He just provided the signals the machine was looking for.
The myth of the generic listicle
Since GEO optimization became a topic, I hear the same refrain: « Write listicles, comment on Reddit. » Suganthan’s network traffic analysis and mine tell a different story. Pure listicles, without facts, without internal data, without external validation, are rarely labeledcitation or knowledge_graph. They fall into the organic bucket with a low score. Reddit can appear, but mainly when the query is classified as « opinion » or « personal expérience. » It’s not an e-commerce brand’s strategy to bet on « Reddit sentiment » to sell a hedge trimmer.
What works is a content architecture that gives the AI blocks of factual text, well-isolated, with cross-references. A lawnmower comparison that cites internal product sheets with measurable attributes. A category page that points to « how to choose » guides. A mesh that creates a thematic cluster, detected as a coherent entity by the AI’s parser.
Again, the DOSE framework I teach via BMO Academy works with this logic: Deployment by intent, Structure in clusters, Expansion through external validation. It’s not a volume contest. It’s a signal mechanic.
What this changes for your next update
You have an online catalog. Hundreds of sheets. Maybe already well-ranked on Google. The question isn’t to rewrite them. It’s to make them readable by the generative AI selection pipeline.
Go into your Bing Search Console and check unindexed URLs. Look at your response times. Then ask yourself this: for each page, what external validation is available? A Wikipedia page listing your brand, a press article, an industrial database? If nothing exists, create that layer. Forget about link building. Tell me about fact validation: what the machine stores.
I’m not selling you the method. I’m showing you the pages.
And what if the next source cited by ChatGPT for your product was from a competitor who understood the game is played in the JSON?
An audit to show you what the AI’s JSON sees on your pages
In 55 minutes, I run your pages through the AI’s network traffic lens, like for my garden client. Three signals verified, one concrete roadmap. Reserved for e-commerce merchants ready to weigh in on AI citations.
Book a strategic call — 45 minFrequently Asked Questions
Do I need to optimize specifically for Bing to appear in ChatGPT citations?
Yes. I observe that over 70% of ChatGPT queries go through Bing. The finding is clear: clean indexation on that engine, server response under 800 ms, crawlable structure. It holds up.
How do I know if my pages are crawlable by ChatGPT’s AI?
Use Bing Webmaster Tools’ URL inspection tool. Also check the HTTP code, load time, and rendering without heavy JS. A page that times out or displays empty content after JavaScript execution won’t be seen by ChatGPT.
Do structured data help with being cited?
They help indirectly. I observe that the JSON doesn’t have a dedicated field for schema. But well-filled structured data (MPN, GTIN, factual description) helps Bing extract and validate facts. Result: better chances of being seen as a reliable source.
What types of content does ChatGPT cite most often?
Based on internal labels, comparative formats, factual technical guides, and pages with verifiable claims on authority sites are cited more than generic listicles or purely commercial content. The <code>turn_use_case</code> « procedure » or « comparison » maps to these formats.
How do I get external citations useful for the AI without classic link-building?
Target open databases (Wikidata, professional directories), trade press, mentions in consumer association comparisons. A citation of your brand with a verifiable fact is enough: ChatGPT picks you up, even without a link.
