30-second rundown
Key learnings:
- A single default-model swap cut unique domains cited per ChatGPT response by about 20% overnight. The winners were the domains the model already trusted. Treat citation counts as volatile, not permanent.
- The drop was architectural, not editorial. The free tier’s
response_lengthnow defaults to “short,” which pulls fewer sources, and most users are on the free tier. - You are playing two games. Parametric visibility, what the model knows before searching, is the prerequisite for dynamic visibility, what it retrieves live. If the model does not know you, it never goes looking for you.
- “ChatGPT” is several models with different appetites. The premium engines search far harder and cite brand sites several times more often than the free default, and one fallback model has no web access at all. Segment your monitoring or fly blind.
- Position beats fame. Centrality in the model’s association graph, being connected to the right authoritative names, matters more than raw recall.
This week: Run a parametric audit. Open ChatGPT with web search switched off, and ask it plainly what it knows about your brand, who it thinks your competitors are, and which sources it treats as authoritative in your space. That answer, with no live retrieval propping it up, is the floor everything else is built on, and if you are not in it, you have your priority.
Coming 3 months: Build parametric authority deliberately, before the next training cutoff bakes today’s gaps into the weights. That means earned PR, a defensible Wikipedia and Wikidata presence, and editorial coverage on high-authority sites, the signals that decide whether the model considers you a candidate at all. In parallel, stand up model-segmented tracking so you are measuring 5.3, 5.4, and the rest separately instead of drowning the truth in an average.
On March 4, 2026, OpenAI quietly swapped ChatGPT’s default engine from GPT-4o/5.2 to GPT-5.3 Instant, and the French consultancy RESONEO was standing there with a stopwatch and a net.
Working with citation data from the tracking platform Meteoria, they logged twenty-seven thousand responses across four hundred prompts over fourteen weeks. The number of unique domains cited per response fell roughly 20% the moment the switch flipped.
Fewer websites. Same internet. Different machine.
1. The Collapse Was Not Random. It Was Wired In
Before the switch, an average response cited 19.1 unique domains across 24.1 URLs. After, it cited 15.2 domains across 19.1 URLs, while the URLs-per-domain ratio held dead steady at 1.26 and crawl depth per site never budged. The machine was not reading each source less thoroughly, it was simply knocking on fewer doors, and the trusted domains already in its good graces swallowed the difference.
Treat that 20% as a snapshot rather than a verdict, though, because citation counts have since proven volatile and clawed back some ground, which is its own lesson about building on sand.
RESONEO traced the cause into the plumbing. ChatGPT’s old web tool spoke a cramped pipe-separated format with four commands, and the new one is a structured JSON toolset with roughly a dozen operations, including find, click, screenshot, and product_query.
Buried in there is a response_length dial set to short, medium, or long, and here is the kicker: it governs how many sources get pulled, not how long the answer reads, and the free tier defaults to short. With the overwhelming majority of weekly users on the free plan, the default experience now triggers fewer searches and coughs up fewer citations by design.
2. You Are Playing Two Visibility Games at Once, and Most Brands Know Neither Score
RESONEO’s sharpest contribution is a clean split. Parametric visibility is what the model knows from training alone, before any search tool wakes up: stable for months, reproducible, cheap to query, and fed by PR, Wikipedia, Wikidata, and editorial coverage on authority sites. It is the language-model version of reputation baked into the weights.
Dynamic visibility is what the model pulls from the live web in the moment: volatile, model-specific, geography-dependent, and chained to Google and Bing ranking through technical SEO, structured data, and freshness.
Here is the brutal hinge between them. When the model decides what to go fetch, it aims its search queries at sources it already recognizes from training, so a brand absent from parametric memory never even makes the shortlist.
Being unknown to the model means being invisible before the search starts. Parametric awareness is the entry fee for the dynamic game, and most brands cannot tell you where they stand on either board.
3. “ChatGPT” Is Not One Machine. It Is a Fleet, and They Disagree
The same prompt fed to GPT-5.3 Instant, 5.4 Thinking, 5.4 Pro, and the free fallback i-5-mini produces wildly different citations, and i-5-mini is the punchline, because it carries no web tools at all and answers purely from memory. The heavy models fan a single question into many sub-queries and chain five to more than ten rounds of search, while the lightweight default runs a handful and calls it a day.
Independent studies peg the premium models citing brand-owned sites several times more often than the free default. So the model your customer opens without thinking is often the one working least on your behalf.
A monitoring strategy that averages across all of them tells you nothing true about any of them. You need per-model tracking the way you once needed per-channel reporting, and almost nobody has built it.
One eerie footnote on how RESONEO learned all this: they set a honeypot, a hidden page with a hashed filename, then sent volunteers to coax ChatGPT into walking there itself, and it did, disclosing its own tool schemas and operation lists with unsettling consistency, so long as nobody said the magic words “system prompt.” The model will confess almost everything about itself if you ask without spooking it. OWASP has said the quiet part for a while now: the system prompt was never a secret.
The last word belongs to a graph. DEJAN AI’s Dan Petrovic asked Gemini 3 Flash to name a hundred brands at random two hundred thousand times, wired the answers into a directed association network, and ran Personalized PageRank across 2.9 million nodes to measure what he calls associative embeddedness.
The top-ranked non-seed brand was Maison Margiela, a name the model never volunteered on its own, yet one so densely woven into the high-authority luxury cluster that every path ran through it. That is the new game in a sentence.
You do not have to be famous. You have to be standing next to the famous, in a map the model drew long before it ever opened a browser.
