Why the ChatGPT Your Customers Use Cites Differently Than Your Test

GPT-5.6, released in July 2026, does not serve the same AI to everyone who opens ChatGPT. It ships as three distinct models: Luna for free-tier users, Terra for ChatGPT Plus at default effort, and Sol for Plus users who select higher effort, plus all Pro, Business, and Enterprise accounts.

This is not a setting. It is not a toggle. It is a permanent architectural decision that OpenAI has built into the product.

Our July 16, 2026 methodology investigation (Scout session 79, `research-vault/methodology-recs/2026-07-16-gpt56-multi-tier-audit-scope.md`) confirmed the tier structure from multiple independent sources and documented what it means for how ChatGPT cites businesses. The finding is straightforward. Its implications for how you measure AI visibility are not.

What the Three Models Are and Who Gets Them

The access mapping confirmed in our July 16 investigation:

| User type | Model tier | |---|---| | Free / Go plan | Luna | | ChatGPT Plus (default effort) | Terra | | ChatGPT Plus (medium/high effort) | Sol | | Pro / Business / Enterprise | Sol at medium/high; Sol Pro available | | API (high-volume) | Luna |

Luna is designed for speed and efficiency. It begins summarizing instead of citing at context depth -- when a query reaches the point where surfacing specific sources would require additional retrieval work, Luna doesn't do that work. It returns an answer from what it already has.

Sol produces longer, more comprehensive sub-query fan-out. More sub-queries mean a larger citation pool and a higher probability of retrieving a specific business, even one that doesn't dominate signals for the primary query.

Terra sits between them. Our investigation assessed Terra's citation characteristics as likely resembling the GPT-5.5 Thinking tier behavior.

The Citation Gap the Tier Structure Creates

GPT-5.6 empirical citation data by tier is not available yet -- comparison studies are expected in the July 16-28 window. Our July 16 investigation used the GPT-5.5 analog as the nearest empirical reference.

Under GPT-5.5, the Thinking tier (Sol equivalent) produced 47% brand site citation rates. The Instant tier (Luna equivalent) produced approximately 6%.

An 8x difference in citation probability for the same business, from the same query, depending on which model tier was generating the response.

The mechanism is the retrieval architecture. Sol-class models decompose a user prompt into multiple related sub-queries and pull from a larger candidate pool. Luna processes efficiently and returns results with less retrieval depth. For a business with strong citation signals, the gap may be small -- both models retrieve it. For a business with moderate signals, appearing in one of Sol's many retrieval angles but not in Luna's narrower pass is exactly what the 47%-versus-6% data describes.

Our July 16 investigation also documented a 16% silent rerouting rate from GPT-5.5: 16% of free-tier prompts were silently moved to the Thinking tier without the user's knowledge. With three model tiers in GPT-5.6, the rerouting logic across Luna, Terra, and Sol creates additional unpredictability. A free-tier user may occasionally receive a Sol-class response without seeing any indication of it.

How This Stacks with Reasoning Mode

Our July 10, 2026 investigation (Scout session 73, `research-vault/methodology-recs/2026-07-10-chatgpt-reasoning-mode-audit-scope.md`) documented a separate, earlier-confirmed source of variation: reasoning mode within a single tier. A Semrush study with researcher Kevin Indig, run across 100 prompts in 20 buyer journeys and published June 30, 2026, found that standard and extended thinking modes produce only a 25.6% citation overlap for the same queries. Meaning 74.4% of the domains cited in extended thinking are different from what standard mode returns.

GPT-5.6 now layers model tier and reasoning mode on top of each other.

A Sol-tier user with extended thinking enabled is generating 24+ sub-queries from a model with Sol's deeper retrieval architecture. A Luna-tier user on standard reasoning generates 5-6 sub-queries from a model that summarizes at context depth. These are not variations on the same ChatGPT. They produce different citations for the same query from a largely non-overlapping source pool.

When a business owner tests their own ChatGPT presence, they're running one specific combination of tier and reasoning mode. That combination may or may not reflect what their target customers are seeing.

What This Means for How You Measure AI Visibility

If you test your ChatGPT presence as a free-tier user, you're measuring Luna behavior. If your potential customers are primarily professionals or business decision-makers on paid ChatGPT plans, they're running Terra or Sol. The 47%-versus-6% gap from our GPT-5.5 analog suggests you may be systematically underestimating what those paying subscribers see -- or missing entirely that Sol is finding your business when Luna does not.

The reverse also applies. Testing on a Plus account with Sol selected may overstate your visibility to the free-tier majority of users.

Our July 16 investigation identifies the correct audit default: run on a Plus account at Sol, medium effort. This represents the highest-engagement user segment -- the people most likely to act on AI recommendations for professional and commercial decisions. But the audit should disclose which tier was used, because free-tier users (Terra and Luna) may see different citation patterns for the same query.

When a prospect says they find themselves in ChatGPT but an audit doesn't show a citation, our July 16 investigation now documents three explanations to check, in order: query formulation variation (the audit uses structured fan-out; self-tests use a single phrasing), reasoning mode difference (extended thinking searches a 74%-different source pool), and model tier difference (the prospect may be on a paid plan triggering Sol while the comparison was run on a lower tier).

What Doesn't Change: The Underlying Signals

Understanding the tier gap changes how you interpret measurements. It does not change what builds citation probability across all tiers.

The signals that get a business into Sol's wider retrieval net are the same ones that get it into Luna's narrower one. Directory presence across the sources all ChatGPT tiers retrieve from -- Yelp, BBB, industry-specific directories -- consistent entity signals, and third-party mentions that entered the training data pool. Businesses with strong citation infrastructure become effectively tier-independent. Businesses with moderate signals are the ones whose citation presence varies most between free and paid users.

Our platform divergence research (`research-vault/knowledge/platform-divergence-data.md`, updated June 2026) puts a related data point in context: Perplexity's brand citation rate is 13.05% vs ChatGPT's 0.59%, a 46x difference confirmed across three independent studies of 680 million citations. That cross-platform gap dwarfs any within-ChatGPT tier gap -- which is why a business focusing only on ChatGPT optimization is missing the platform where citation rates are most accessible.

The tier structure in GPT-5.6 adds one more reason that a single self-test of ChatGPT doesn't capture your AI citation picture. A Signal Check at sourcepull.ca specifies the platform and runs queries across ChatGPT, Perplexity, Gemini, and Claude simultaneously. If your self-test result and your audit result diverge, the tier difference is the third explanation to check -- and the per-platform breakdown shows which underlying signals are missing, because citation probability at any tier is built on the same infrastructure.