GPT-5.6 Sol vs. Terra: What the Citation Numbers Actually Measure

Practitioner content circulating since the GPT-5.6 launch cites specific Sol and Terra citation rate differences with notable confidence. Numbers like "Sol achieves 97% citation accuracy" and references to a large Sol/Terra citation gap have appeared in AEO vendor blogs and SEO practitioner posts, framed as evidence of how much more often Sol cites external websites than Terra does.

We traced where those numbers come from. They are real. They just don't measure what practitioners are claiming they measure.

What the Numbers Actually Come From

In our July 17, 2026 investigation (Scout session 80, `research-vault/edge-cases/sol-terra-in-context-vs-web-citation-conflation-2026-07-17.md`), we documented the source of the Sol/Terra citation performance benchmarks circulating in AEO content.

The benchmarks come from in-context long-document retrieval tests. The scenario: a developer loads one or many long documents into the context window, then asks the model to retrieve specific information from those documents. "Citation accuracy" means correctly attributing retrieved content to the right source document within the supplied prompt.

Sol performs well in this test. Luna summarizes rather than cites at long context depth. Terra sits between them. These are legitimate findings from developer benchmark studies.

What they don't measure: how often ChatGPT includes a specific business's website in its response when a user asks "best plumber in Austin."

The Conflation and Why It Matters

AEO practitioners care about web citation behavior -- whether and how often ChatGPT surfaces a specific website when answering commercial queries. An in-context document retrieval benchmark measures something entirely different: whether the model can correctly track which document a piece of information came from when those documents are supplied inside the prompt.

When a developer benchmark says "Sol achieves 97% routing accuracy in primacy-positioned document retrieval," it is measuring Sol's ability to keep track of the right source document in a long supplied prompt. It says nothing about whether Sol is more likely to surface a business's Yelp profile when a user asks which HVAC contractor to hire.

The conflation matters because it leads to decisions built on a false foundation. If Sol's in-context retrieval performance actually described Sol's web citation behavior, the Sol/Terra web citation gap would be large and empirically confirmed. Since the two measurements are different -- as they are -- the Sol/Terra web citation gap remains a hypothesis, not a confirmed finding.

As of July 17, 2026, no published study has measured Sol vs. Terra web citation behavior for commercial queries. The studies expected to fill that gap -- Writesonic opened their data collection window July 16; SISTRIX opens July 21 -- had not published results at the time we filed this.

What IS Actually Confirmed

The absence of Sol/Terra web citation data doesn't mean we know nothing about how ChatGPT tiers cite differently.

Our June 16, 2026 investigation (Scout session 49, `research-vault/methodology-recs/2026-06-16-gpt55-chatgpt-citation-shift.md`) documented confirmed findings from two independent studies of GPT-5.5 behavior. The Writesonic study ran 50 prompts across 16 categories and analyzed 11,469 web search results and 1,257 citations. The SISTRIX study covered 3.8 million German-language ChatGPT responses.

The GPT-5.5 Thinking tier -- the closest architectural analogue to Sol -- produced 47.2% brand site citation rates. The GPT-5.5 Instant tier -- the closest analogue to Luna -- produced approximately 6%. The mechanism was documented: GPT-5.5 reduced its use of the `site:` operator from 40.5% of web searches to 12.6%, meaning brand-owned pages are no longer actively sought in the retrieval layer. They appear only when organic search naturally surfaces them.

These numbers are confirmed from primary studies with documented methodologies. They describe GPT-5.5 behavior, not GPT-5.6 Sol/Terra behavior directly -- but the architectural analogy is strong and the mechanism is consistent with a meaningful gap between tiers.

Our July 16, 2026 methodology investigation (`research-vault/methodology-recs/2026-07-16-gpt56-multi-tier-audit-scope.md`) formalized the current known/unknown boundary: Sol and Terra produce different citation behaviors based on their retrieval architectures. Sol's deeper sub-query fan-out creates a larger candidate pool. Whether the Sol/Terra gap under GPT-5.6 resembles the ~40 percentage point spread seen between GPT-5.5 Thinking and Instant is an inference, not yet a measurement.

Why Confident Numbers That Haven't Been Measured Spread

The AEO field has a recurring pattern: a measurement from one context gets applied to a different context because the framing sounds similar. "Citation accuracy" sounds like it describes how often a model cites web sources in its responses. In-context document retrieval and web search retrieval both involve a model surfacing sources. The categories feel adjacent.

The same pattern appeared earlier in July 2026 with a vendor article claiming GPT-5.6 prompt caching creates a citation moat for early movers -- a mechanism that doesn't exist in the consumer product. And it appeared before that with fabricated Gemini citation thresholds that circulated as authoritative for months before practitioners checked them against actual audit data.

We track these conflations because the wrong mental model leads to the wrong actions. If Sol's confirmed in-context accuracy is taken as evidence that Sol frequently cites business websites, a business might deprioritize the work that actually drives web citation behavior -- directory presence, third-party review platforms, NAP consistency across listings -- in favor of a mechanism the data doesn't support.

What to Do While Empirical Sol/Terra Data Is Pending

The Sol/Terra web citation gap is likely real and likely significant. The architectural inference is reasonable: Sol generates more sub-queries, reaches a larger candidate pool, and has higher probability of surfacing a business with moderate citation signals. That inference is what we work with until confirmed numbers are available.

The practical implication doesn't change either way. A business with strong directory presence and review volume -- consistent signals across the sources all ChatGPT tiers retrieve from -- becomes effectively tier-independent. It appears whether Sol runs eight sub-queries or Luna runs three. The businesses whose citation presence varies most between tiers are those with moderate or inconsistent signals, where Sol's wider retrieval net makes a material difference.

Our audit methodology as of July 16 runs on a ChatGPT Plus account at Sol tier for exactly this reason: it represents the highest-engagement user segment and the widest citation pool. When the Writesonic or SISTRIX Sol/Terra comparison study publishes confirmed numbers, we'll update the methodology if the gap warrants multi-tier audit runs. Until then, the audit discloses which tier was used -- because that disclosure matters when you're interpreting what the results mean for different segments of your audience.

If you want a baseline that tells you where your business actually stands today across ChatGPT, Perplexity, Gemini, and Google AI Mode, Signal Check at sourcepull.ca runs the assessment in about 90 seconds and specifies which platform tier and query formulation was used. That's the starting point for evaluating what any citation rate claim actually means for your specific business.