This is part 2 of our GPT-5.5 citation study. Part 1 covered the Thinking tier, what changed when ChatGPT replaced GPT-5.4 with GPT-5.5 as the “Latest” Thinking model.

ChatGPT replaced GPT-5.3 with GPT-5.5 as the default Instant model in early May 2026. The change wasn’t announced as a behavioral shift. It was framed as a model upgrade.

The data tells a different story.

GPT-5.5 Instant cites brand websites half as often as GPT-5.3 Instant did. Reddit went from a peripheral source on the previous default to ChatGPT’s most-cited domain by a 3x margin. And 16% of the prompts we ran on Instant got silently routed to the Thinking tier without telling the user, even with the Auto-switch toggle disabled.

The free-tier ChatGPT experience your customers are using right now is fundamentally different from the one they were using two weeks ago. Most brands haven’t noticed.

How we did this

We ran 50 prompts through GPT-5.5 Instant. Same prompt set as our broader 5.5 vs 5.4 vs 5.3 study, covering SaaS, ecommerce, healthcare, finance, travel, education, home, food, legal, marketing, productivity, fitness, shopping intent, head-to-head comparisons, and trends.

42 of the 50 conversations ran cleanly on GPT-5.5 Instant (model_slug = gpt-5-5). 8 escalated mid-conversation to GPT-5.5 Thinking even though we hadn’t enabled Auto-switch. We re-ran those 8 in fresh sessions with the toggle explicitly disabled. 7 of 8 escalated again. Routing is content-based, not random.

The 42 clean conversations form the basis of this analysis.

For each conversation, we pulled the full payload from ChatGPT’s /backend-api/conversation/<id> endpoint. Every fan-out query, every web result, every cited URL. We classified each cited URL as first-party (the brand the user asked about) or third-party (review sites, blogs, Reddit, retailers, media outlets) using Claude Haiku 4.5.

Now here’s what we found.

Brand citations halved

GPT-5.5 Instant cites brand websites 6% of the time. GPT-5.3 Instant cited them 13.4%.

On 5.5 Instant, ~94 of every 100 cited URLs go to a third-party source. Not a brand site. That’s down from ~87 of every 100 on 5.3 Instant.

Both Instant tiers cite zero pricing pages across all conversations. The Instant search infrastructure simply doesn’t reach into brand pricing structures. If pricing-page traffic from ChatGPT matters to your business, it’s coming from Plus and Pro users on Thinking, not free-tier users on Instant.

The per-category breakdown for 5.5 Instant shows how stark the brand drop is:

For most prompt categories, the model cites zero brand websites. Only Finance and Marketing reach 18%. Compare this to GPT-5.5 Thinking on the same 50 prompts, where the average is 47% first-party and most categories see brand citations regularly.

The brand visibility your customers had on the free tier under GPT-5.3 was already small. On GPT-5.5 Instant, in most categories, it’s gone.

Reddit is now ChatGPT’s most-cited domain on Instant

This is the single most striking pattern in the data. Reddit is GPT-5.5 Instant’s most-cited domain by a wide margin.

Reddit citations on Instant grew 6x from one model version to the next. From a peripheral source on 5.3 (cited 6 times, ranked 9th) to dominant on 5.5 (38 citations, more than 3x the next-most-cited domain).

Forbes, the most-cited domain on 5.3 Instant by a large margin (21 citations), got knocked out of the top three entirely. The shift isn’t subtle.

For brands relying on user-generated content visibility, this is the single most consequential change in the Instant tier. Reddit threads, including older ones, comparison threads, “best X for Y” recommendation threads, are now disproportionately surfaced when free-tier users ask ChatGPT for product or service recommendations.

If your category has active Reddit threads, those threads are now your free-tier visibility surface.

8 of 50 prompts auto-escalate from Instant to Thinking, even with the toggle off

The most operationally important finding in the dataset: ChatGPT’s Instant tier silently routes complex prompts to the Thinking tier, regardless of user settings.

Of our 50 prompts, 9 conversations came back with both gpt-5-5 (Instant) and gpt-5-5-thinking slugs in the conversation payload. The model started the response on Instant and was rerouted to Thinking partway through.

We re-ran those 9 prompts in fresh sessions with Auto-switch to Thinking explicitly disabled. 8 of 9 escalated again. Routing is content-based, not random.

The 8 prompts that re-escalated:

• “What are the biggest trends in ecommerce for 2026?”
• “How is AI changing the recruiting and hiring process?”
• “Best online learning platforms for professional development in 2026”
• “Compare Coursera vs Udemy vs LinkedIn Learning for tech skills”
• “Best coding bootcamps for career changers in 2026”
• “Notion vs Obsidian vs Roam Research for personal knowledge management”
• “What are the top cybersecurity threats businesses should prepare for in 2026?”
• “How is AI changing the legal industry in 2026?”

The pattern is clear. Broad-recommendation prompts, multi-vendor research prompts, and open-ended trend prompts get classified as “too complex” for Instant and silently rerouted. Tighter head-to-head comparisons (e.g. “Compare HubSpot vs Salesforce vs Pipedrive”, which has multiple vendors but a tight question) often stay on Instant.

For brands auditing ChatGPT visibility, this matters in a specific way. Instant-tier traffic on broad recommendation prompts isn’t actually Instant-tier behavior. It’s Thinking-tier behavior delivered to a free user. The response will look like Thinking output (more brand sites, more site: queries, ~85% Google-disconnected) even though the user picked Instant.

If your client’s measured ChatGPT visibility is bouncy on certain prompt categories, this is likely the cause. Check the conversation’s model_slug field. It tells you which model actually answered.

How GPT-5.5 Instant differs from GPT-5.5 Thinking

GPT-5.5 Instant and GPT-5.5 Thinking share the same underlying model generation and the same release timeline. They behave very differently.

The Thinking tier issues 7x more search queries, reaches 6x more pages, cites brand sites 8x more often, and reaches into pricing pages on ~9% of citations. The Instant tier is faster, leaner, third-party-heavy, and doesn’t expose publication-date metadata at all.

Citation domain overlap between 5.5 Instant and 5.5 Thinking is just 8.7%. Even on the same 50 prompts, the two tiers pick almost entirely different sources.

In other words: a user on the free tier and a user on Plus, asking the same question, see categorically different answers backed by categorically different domains.

How Instant compares to Google rankings

We cross-referenced 30 prompts against Google US and Bing US top-10 organic results via SerpAPI:

GPT-5.5 Instant tracks closer to Google’s index than GPT-5.5 Thinking does, by 11 percentage points. Both Instant tiers (5.3 and 5.5) hover around 70-72% absent from Google. Both Thinking tiers hover around 84-87%.

For brands that historically optimized for Google rank, this is the version of ChatGPT where SEO investment most clearly carries over. A meaningful share of cited domains on Instant do appear in Google’s top 10 for the same query. The Thinking tier is much further removed from search-engine consensus.

If you’re already investing in Google SEO, that work is partially paying off on the free-tier ChatGPT experience. It’s not paying off on the Plus tier.

What this means for brands

1. The free-tier ChatGPT experience is now Reddit-heavy. If your category has active Reddit threads, positive, negative, or mixed, they’re being surfaced to free-tier users disproportionately. Audit your Reddit footprint this week. Consider Reddit-native presence if your category lives there. We wrote a full Reddit playbook for AI visibility, and the strategies in there matter even more now.

2. Pricing pages get zero free-tier visibility, regardless of version. Both Instant tiers cite zero pricing pages across all conversations. If pricing-page traffic from ChatGPT matters to you, it’s coming from Plus and Pro users on Thinking only.

3. Brand visibility on Instant has halved between versions. Pages of yours that used to land in 5.3 Instant answers (already rare) are now even rarer. The Instant tier has moved toward third-party content sources at almost every category.

4. SEO investment still helps on Instant, partially. 27% of GPT-5.5 Instant’s cited domains appear in Google’s top 10 for the same query, vs 16% for GPT-5.5 Thinking. Traditional ranking still pulls some weight on the free tier in a way it doesn’t on Plus.

5. Watch for silent escalation. ~16% of broad-recommendation prompts on Instant get rerouted to Thinking. If a client’s measured ChatGPT visibility is bouncy on certain prompt categories, content-based escalation may be the cause. The conversation’s model_slug field tells you which model actually answered.

6. Search runs nearly 100% of the time on Instant. GPT-5.5 Instant ran web search on every single one of its 42 conversations. The only model in the study where the no-search behavior is fully extinct. Assume any prompt in your category will trigger search on the free tier.

We built free-tier vs paid-tier ChatGPT visibility tracking into our AI visibility platform. Track citation share by model tier, monitor the Reddit threads cited for your category, and see which queries on Instant get silently rerouted to Thinking. See it in action →

How to verify which model your conversation actually used

You can audit any single ChatGPT conversation directly from the browser. After sending a prompt:

Step 1: Open the console

Cmd + Option + J on Mac. Ctrl + Shift + J on Windows. Switch to the Console tab.

Step 2: Paste this script

const cid = location.pathname.split("/c/")[1];
const session = await (await fetch("/api/auth/session")).json();
const r = await fetch(`/backend-api/conversation/${cid}`, {
  headers: { Authorization: `Bearer ${session.accessToken}` },
});
const data = await r.json();
const slugs = new Set();
for (const node of Object.values(data.mapping)) {
  const s = node?.message?.metadata?.model_slug;
  if (s) slugs.add(s);
}
console.log("Model slug(s):", [...slugs]);

What you’ll see

If you see only gpt-5-5, you got pure Instant. If you see both gpt-5-5 and gpt-5-5-thinking in the same conversation, your conversation auto-escalated to Thinking partway through. The user-facing UI doesn’t tell you when this happens. The payload does.

Questions we’re still investigating

Run-to-run variability. ChatGPT is non-deterministic. Single-run measurements like this give directional reads, not statistical certainty. The 8 escalated prompts had a dedicated re-run; the 42 Instant conversations did not. We’re planning a multi-run pass on the Instant set.

Reddit thread selection logic. Of the 38 Reddit citations, which threads got picked? Is it engagement-weighted? Recency-weighted? Topic-relevance-weighted? Reading those 38 conversations side-by-side to characterize the selection logic would clarify what brands should optimize for.

The escalation gradient. ~16% of prompts re-escalated cleanly. Is there a fuzzy middle category (prompts that escalate sometimes and not others)? A multi-run pass against the same 50 prompts would tell us.

International tier behavior. This study ran from a US ChatGPT account. Whether GPT-5.5 Instant behaves the same in non-US markets, or whether OpenAI A/B tests different routing logic by region, is open.

Methodology snapshot

50 prompts spanning 16 categories, derived from a representative cross-section of consumer and B2B research queries. Run on May 6, 2026 from a single ChatGPT free-tier account in the United States. 42 conversations ran cleanly on Instant. 8 auto-escalated to Thinking and were re-run in fresh sessions with the Auto-switch toggle disabled. 7 of 8 escalated again.

Conversation payloads pulled directly from ChatGPT’s /backend-api/conversation/<id> endpoint with browser-session authentication. Every payload includes the full message tree, all search_model_queries (fan-outs), all search_result_groups (web results), all content_references (citations), and model_slug per message.

Citation classification: Claude Haiku 4.5 with detailed instructions and 50+ in-prompt examples. Each cited URL classified independently as FIRST, THIRD, or UNCLEAR. Calibration check against our broader 5.5 vs 5.4 vs 5.3 study, where 5.4 Thinking’s first-party rate measured 56.8%, in line with previously observed benchmarks.

Search-engine cross-reference: 30 prompts × Google US and Bing US top-10 organic results via SerpAPI.

Limitations: single user account, single run per prompt (8 escalated prompts had a dedicated re-run), single point in time. Repeat runs would produce slightly different results due to ChatGPT’s non-determinism. The 5.5 Instant sample is 42 conversations rather than 50 because of the auto-escalation pattern.

TLDR

• GPT-5.5 Instant cites brand websites 6% of the time. About half as often as GPT-5.3 Instant did (13%). For most prompt categories, brand citations are 0%.
• Reddit went from peripheral to dominant. Reddit citations grew 6x from 5.3 Instant to 5.5 Instant. It’s now the single most-cited domain by a 3x margin.
• Pricing pages: zero, both Instant versions. The Instant search infrastructure does not reach into brand pricing structures, regardless of model version.
• 8 of 50 broad-recommendation prompts auto-escalate to GPT-5.5 Thinking even with the Auto-switch toggle off. Routing is content-based and persisted on retry.
• GPT-5.5 Instant tracks Google’s index more closely than GPT-5.5 Thinking does. 27% of cited domains in Google top 10 vs 16%. SEO investment still carries some weight on the free tier.
• Citation overlap between 5.5 Instant and 5.5 Thinking is only 8.7%. Same generation, same prompt, almost entirely different cited sources.
• Search runs 100% of the time on 5.5 Instant. The only model in the study where no-search responses are extinct.
• For brands: audit your Reddit footprint, accept zero pricing-page visibility on Instant, segment ChatGPT measurement by tier, and check model_slug to know which model actually answered.

If your AEO measurement isn’t segmented by tier, you’re averaging two fundamentally different surfaces. The free-tier experience your customers see most often is now Reddit-heavy, brand-light, and not necessarily on the model the user thinks they’re using. Start with a free Writesonic account or book a demo to track free-tier and paid-tier ChatGPT visibility separately in your dashboard.