What Sol's '97% Citation Accuracy' Actually Measures (And What It Doesn't)

A specific number has been circulating in AEO practitioner content since mid-July 2026: Sol, the top tier of GPT-5.6, "achieves 97% routing accuracy" -- cited as evidence that Sol-tier ChatGPT dramatically outperforms Terra and Luna for web citation of businesses.

The 97% number is real. The application is wrong.

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 traced this figure to its source. It comes from in-context long-document retrieval benchmarks -- a developer tool capability test that measures something entirely different from whether ChatGPT cites your business when a user searches for a plumber.

What "In-Context Document Retrieval" Actually Tests

The Sol/Terra benchmark scenario that produces accuracy numbers like 97% works like this: a developer supplies a long document -- or many documents -- inside the model's context window. The model is then asked to retrieve specific information from those in-context materials. "Citation accuracy" in this benchmark means correctly attributing retrieved content to the right source document within the developer's supplied prompt.

This is a capability test for developers building applications on top of GPT-5.6. It answers: can the model keep track of which document said what when I give it 200 pages of context? That is a real and practically important question for software developers using the OpenAI API to build enterprise tools.

It does not answer: when a user in Austin asks ChatGPT "best HVAC contractor near me," does Sol surface businesses more often than Terra does?

Those are different measurements, conducted in different environments, for different purposes. What gets conflated is that both involve the word "citation."

What We Know About Web Citation Behavior

Web citation behavior -- how often each model tier surfaces external websites when answering user queries -- has not yet been measured empirically for GPT-5.6 as of July 17, 2026. Writesonic opened their comparison study window on July 16; SISTRIX's window opens July 21. Until those results publish, no confirmed Sol vs. Terra web citation rate exists.

What we do have is the GPT-5.5 analog -- the closest empirical reference point, documented in our July 16, 2026 methodology investigation (Scout session 79, `research-vault/methodology-recs/2026-07-16-gpt56-multi-tier-audit-scope.md`).

Under GPT-5.5, the tier structure produced a measurable web citation gap:

- GPT-5.5 Thinking tier (the Sol equivalent): 47.2% brand site citation rate in controlled studies - GPT-5.5 Instant tier (the Luna equivalent): approximately 6% brand site citation rate

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

The mechanism behind that gap is architectural. Sol-class models decompose a user query into multiple related sub-queries and retrieve from a wider candidate pool. More sub-queries mean a larger retrieval sweep. A business with moderate citation signals -- present in some directories, but not dominating Yelp and BBB for its category -- may appear in Sol's wider retrieval pass and not in Luna's narrower one.

The July 16 investigation assessed Terra's behavior as likely resembling GPT-5.5's Thinking tier -- meaning Terra may carry much of the citation probability advantage that practitioners are attributing specifically to Sol. The Sol-over-Terra gap, when actually measured, may prove smaller than the Sol-over-Luna gap. We will not know until the web citation studies land.

Why the Wrong Number Costs You

When practitioners cite in-context document retrieval benchmarks as evidence for web citation advantages, they install a mental model that leads to wrong conclusions.

If Sol's web citation advantage over Terra were actually 97% vs. some lower Terra number, it would mean that paying for ChatGPT Plus is almost a prerequisite for your business to appear in AI search results at all. That would be a product decision that OpenAI would almost certainly have publicized. They haven't -- because the in-context retrieval number describes document tracking capability in developer applications, not consumer-facing search citation rates.

The same pattern showed up in a separate investigation. Our July 11, 2026 edge case (Scout session 74, `research-vault/edge-cases/vendor-prompt-caching-mischaracterization-2026-07-11.md`) documented an AEO vendor claiming GPT-5.6's prompt caching feature created "a moat backed by compute economics" for brands that got into early cached answer patterns. The vendor also cited a "34% citation gap between rivals in a single model transition" -- with no methodology, no sample size, no primary source.

Two different mischaracterizations, both published within days of a major ChatGPT model release, both using specific-sounding numbers to describe a mechanism that doesn't work the way described. The pattern is useful to recognize: model transitions generate vendor content that races ahead of the evidence.

What the Tier Gap Means Right Now

The Sol/Terra web citation gap is real in expectation -- the architectural analogy to GPT-5.5 is strong, and the mechanism (sub-query fan-out depth) is consistent. Our July 16 investigation set the methodology trigger: if the Sol/Terra gap in the Writesonic or SISTRIX data exceeds 20 percentage points, tier-specific audit runs become operationally justified.

Until that data exists, the practical position is clear.

The signals that determine citation probability are tier-agnostic. Directory presence, NAP consistency, review volume, and on-page content feed the candidate pool that each tier retrieves from. Sol retrieves more broadly from that pool; Luna retrieves more narrowly. A business with strong citation infrastructure lands in both. A business with thin presence misses both.

Our July 16 investigation also documented that 16% of free-tier ChatGPT prompts were silently rerouted to the higher-reasoning Thinking tier under GPT-5.5. With three tiers in GPT-5.6, the rerouting logic across Luna, Terra, and Sol adds additional unpredictability. A free-tier user occasionally receives Sol-level responses without any visible indication of it.

This is relevant for how you interpret your own testing. If you test your ChatGPT presence on a free account and find yourself, that does not mean you're appearing for Terra users -- it may mean that particular query was silently routed to Sol. If you test on a Plus account and don't find yourself, that combination of tier and reasoning mode may not reflect what a Plus user with default settings sees. Neither self-test is the measurement.

What Audits Specify

Our current audit default -- documented in the July 16 methodology rec -- is ChatGPT Plus at Sol tier, medium effort. This represents the highest-engagement user segment and the most actionable tier for business-class decisions. Audit reports now include disclosure language noting which tier was queried and that Terra-tier users may see different citation patterns.

When the Writesonic and SISTRIX Sol/Terra web citation comparison publishes, we will update methodology based on the actual measured gap. If it is large enough to justify running audits at both tiers, we will do that. If the Sol/Terra gap turns out smaller than the GPT-5.5 Thinking/Instant gap suggested, a single-tier audit with disclosure remains appropriate.

In the meantime, the number to use for the ChatGPT tier citation gap is 47.2% vs. ~6% -- from GPT-5.5 studies with disclosed methodology. Not 97%, which measures something different in a context that has no bearing on whether your business shows up when someone searches for you.

A Signal Check at sourcepull.ca runs at the Sol tier and shows your citation position across ChatGPT, Perplexity, Gemini, and Google AI Mode. When the per-platform breakdown is visible, the tier question becomes secondary to the infrastructure question: which specific signals are missing, and which platform gaps are largest. That is where the optimization work actually lives.