GPT-5.6 Luna Scores 41.3% on Long-Context Recall — Half of Sol's 91.5%

GPT-5.6 Luna — now the default model for ChatGPT Free and Go tiers since August 6–13 — scores 41.3% on MRCR, a long-context recall benchmark that measures how reliably a model retrieves and cites facts spread across multiple documents. Sol scores 91.5% on the same benchmark. Terra scores 89.6%. The gap between the model serving the majority of ChatGPT users and the next tier is not marginal — Luna's long-context recall is less than half of Sol's. For GEO strategy, this is the most consequential model change since GPT-5.3 Instant.

What the Benchmark Data Shows

MRCR stands for Multi-Record Citation Recall. It measures a model's ability to retrieve and accurately attribute information spread across multiple source documents — the exact capability required to produce a multi-citation answer to a complex query. BenchLM.ai published Luna's profile on August 17, 2026, using benchmark data from the GPT-5.6 release (July 9, 2026) and Vellum.ai's tier comparison (August 3, 2026).

The GPT-5.6 tier breakdown on MRCR:

• Luna (Free/Go default): 41.3% — the model reaching the largest ChatGPT user population
• Sol (Plus/Team): 91.5%— 2.2x Luna's recall score
• Terra (Pro): 89.6% — near-parity with Sol on this benchmark

Luna's score is consistent with the citation concentration pattern first observed during the GPT-5.3 → GPT-5.3 Instant transition in May 2026, when smaller, faster model variants began favouring fewer, higher-authority citations over diverse source sets. Luna confirms the pattern is structural, not a one-off.

Sources: BenchLM.ai Luna profile (August 17, 2026); Vellum.ai GPT-5.6 tier comparison (August 3, 2026); OpenAI GPT-5.6 release notes (July 9, 2026).

Why It Matters for GEO Strategy

1. Luna is the model most of your prospects' customers are using.Free and Go tier users represent the majority of ChatGPT's user base. When they ask complex questions, Luna handles the answer — and at 41.3% MRCR, Luna systematically misses lower-prominence sources that Sol would include. Being the second or third citation on a topic is no longer a reliable fallback position for the dominant ChatGPT population.

2. Below the top-1 source on a topic cluster, citation probability drops sharply.Luna's recall cliff means it tends to anchor on the most authoritative source per sub-topic and drops the rest. If your brand is not the dominant source for a specific claim — not just "relevant," but definitively the top source — Luna is unlikely to cite you on multi-document queries. Sol would. Luna won't.

3. Topical authority is now a survival requirement, not a differentiator.GEO strategies built on being "one of several cited sources" are exposed. The diversified citation model (appear in many answers with many co-citations) worked when Sol-class models handled the user base. Luna's 41.3% recall collapses that model. You need to be the primary source, not a secondary one.

4. ChatGPT Plus/Team users still get Sol — the split audience problem.If your target audience includes business buyers on paid ChatGPT tiers, they're getting Sol's 91.5% recall. That's a different citation landscape. GEO strategy now has to account for audience tier: Luna users vs Sol users will see materially different citation sets for the same query.

5. Separate data confirms ChatGPT's index is neutral to site size.BotRank.ai research (August 14, 2026) found no meaningful difference in how ChatGPT's in-house search index serves licensed vs. unlicensed sites. Retrievability — whether ChatGPT can crawl and index the page — matters more than publisher deals or domain authority inherited from Google. This means Luna's citation concentration is a recall capability constraint, not an access constraint. If Luna can find your page, it can cite you. But at 41.3% MRCR, it has to want to cite you — and it does that only when you're the clear top source.

What to Do

1. Audit your citation posture per topic cluster

For each core topic your brand targets, ask: Are we the definitively dominant source, or are we one of several sources? The Luna threshold is binary. Being "one of several" means Luna drops you. Being the dominant source means Luna keeps you.

• Use Profound, AirOps, or BotRank to pull your current citation share by topic cluster in ChatGPT responses.
• For each topic cluster where you appear as a secondary citation: either consolidate resources to become the primary source, or deprioritize that cluster entirely.
• For topics where you're already the top citation: protect them. Luna's favouring of the top source means topical dominance compounds. Your lead widens, not narrows.

2. Consolidate thin content into authoritative cluster pages

Luna's lower recall means fragmented content across multiple pages loses its advantage. Sol could retrieve fact A from page 1 and fact B from page 3 and synthesize. Luna more likely picks the single page that best covers the topic and cites only that.

• Identify topic clusters where you have 3+ thin posts. These are candidates for consolidation into a single authoritative page.
• Target semantic completeness of 8.5/10 or higher on the consolidated page. At 41.3% MRCR, Luna is selecting on page-level quality, not portfolio breadth.
• Redirect the thin pages after consolidation. You want crawl budget and link equity pointing at one strong page, not split across several weak ones.

3. Ensure crawlability for OAI-SearchBot

Since ChatGPT's citation index is neutral to publisher deals and domain authority (BotRank.ai, August 14), retrievability is the entry gate. Luna won't cite what it can't find.

• Check robots.txt. Confirm OAI-SearchBot is not blocked. The default ChatGPT crawl agent is OAI-SearchBot. Any disallow rule for this agent removes you from Luna's index entirely.
• Add llms.txt. This file signals AI readiness and helps ChatGPT's crawler prioritise your content. Place it at yourdomain.com/llms.txt with a structured list of your key pages.
• Optimise page load speed. BotRank.ai identified page load speed as a retrievability signal. Pages that load slowly get lower crawl priority. LCP below 2.5 seconds is table stakes.

4. Watch the August 20–22 data window

Luna became the Free/Go default between August 6–13. The first Profound, AirOps, and BotRank citation redistribution data covering the post-Luna period is expected August 20–22. That data will confirm whether citation concentration has shifted — and by how much. If Luna citation redistribution shows more than 10% drift from the pre-Luna baseline, recalibrate your brand authority and AI citability investment immediately.

• Set a calendar reminder for August 22. Pull Profound or BotRank data that day.
• Compare your citation share on complex multi-document queries before and after August 6.
• If you see a drop of more than 10% in secondary citation appearances, the consolidation strategy above becomes urgent, not optional.

Chad's Take

The uncomfortable truth is that most GEO strategies are built for a Sol-class world that no longer exists for the majority of ChatGPT users. Sol's 91.5% recall supports the diversified citation model: produce good content on multiple angles of a topic, appear across many answers, accumulate visibility. That strategy is sound when the model actually retrieves from multiple sources. At 41.3% MRCR, Luna doesn't.

What Luna's recall cliff actually rewards is focus. One killer page per topic cluster, optimized to 8.5/10+ semantic completeness, fully crawlable, with entity disambiguation schema in place. Brands executing that model are not just better positioned — they're the only ones Luna consistently finds space for. The content sprawl strategy that worked for Google rankings and for Sol is a liability at Luna scale.

The citation data lands August 20–22. Until then, treat the 41.3% number as your planning assumption and start the consolidation audit.

Sources

• BenchLM.ai — GPT-5.6 Luna model profile with MRCR benchmark scores, August 17, 2026. benchlm.ai
• Vellum.ai — GPT-5.6 Sol, Terra, Luna tier comparison with long-context recall data, August 3, 2026. vellum.ai
• OpenAI — GPT-5.6 release notes and benchmark data, July 9, 2026. openai.com
• BotRank.ai — ChatGPT search index retrievability analysis: licensed vs. unlicensed sites, August 14, 2026. botrank.ai