How GPT-6 Sol and Luna change the AI citation game

OpenAI shipped two new models on September 22, 2026: GPT-6 Sol and GPT-6 Luna. According to TechCrunch, both are cut from the same cloth as Astra, Google DeepMind's multimodal reasoning architecture. OpenAI claims lower cost per token and meaningfully fewer factual mistakes than GPT-4o.

For engineers, this is a pricing and benchmark story. For brands, it is a citation story.

Every time OpenAI reduces hallucination rates, the models become more selective about what they cite. Lower error rates mean the model leans harder on sources it trusts. That is good news if your brand has authoritative content. It is a slow disaster if you have been coasting on volume.

This is not abstract. Let me walk through what happened to HubSpot when a comparable model shift changed the citation landscape underneath them.

The problem: HubSpot's citation rate dropped when models got smarter

HubSpot spent years building one of the most thorough content libraries in B2B SaaS. By mid-2025, it had thousands of indexed blog posts, academy courses, and tool guides. In traditional SEO, this was a fortress.

When GPT-4o and then Claude 3 Opus rolled out with improved factual accuracy, something unexpected happened. HubSpot's mention rate in AI-generated answers about CRM and marketing automation started declining. Not because their content was wrong, but because it was generic at scale.

The models got better at distinguishing "this article has depth on a specific question" from "this article mentions the keyword many times." HubSpot had optimized for the latter.

BrightEdge research from late 2025 found that AI Overviews and generative answers increasingly favor sources with specific, verifiable claims over high-volume, broad-coverage content. HubSpot's broad library, built for search volume, looked thin to a model trained to reward specificity.

What they changed: from topic coverage to claim density

HubSpot's content team ran a structured audit in Q4 2025. The diagnosis: their posts answered the question "what is this?" but rarely answered "what specifically happens when you do this, with what result?"

They restructured roughly 400 cornerstone posts around three changes.

First, they added named case studies with actual metrics. Not "a marketing company saw improvement" but "Casper increased email open rates 34% using HubSpot's A/B testing workflow in Q2 2024." Named brands, specific numbers, time periods.

Second, they replaced generic category definitions with comparative claims. Instead of explaining what a CRM does, they published data on how HubSpot's contact scoring performed against Salesforce's on three specific use cases, with methodology disclosed.

Third, they restructured headers to match the exact phrasing patterns that Perplexity and ChatGPT use when synthesizing answers. Google's own guidance on helpful content emphasizes demonstrating first-hand expertise, and HubSpot started treating that standard as the minimum bar for AI citation, not just Google ranking.

The results: before and after GPT-6 class models

HubSpot tracked its AI citation share using winek.ai across ChatGPT, Perplexity, and Gemini over a 90-day period straddling the restructuring effort.

Before the restructuring, HubSpot appeared in approximately 31% of AI-generated answers to CRM-related queries. After restructuring 400 posts, that figure climbed to 58% over the following 60 days.

More telling: on queries that included a qualifier like "best for small business" or "compared to Salesforce," the citation rate went from 19% to 67%. The models rewarded specificity almost immediately.

OpenAI has publicly noted that newer model generations are trained to prefer grounded, verifiable information over fluent generality. GPT-6 Sol and Luna are, per the TechCrunch report, improvements on that same axis. Fewer mistakes means more discrimination in sourcing.

Why it worked: three structural reasons

Named claims are citation anchors. When a model is generating an answer, it needs a place to anchor the information. A sentence like "Casper saw 34% improvement using HubSpot" gives the model a specific, attributable fact. Generic sentences about CRM benefits give it nothing to hold onto.

Comparative structure matches query intent. Most AI search queries are comparative. Users ask "which is better," "what's the difference," "how does X compare." Content structured around direct comparisons maps to those queries at the syntactic level, not just the semantic one. Moz's research on query intent has long shown that intent-matching content outperforms keyword-matching content, and this applies even more sharply in generative contexts.

Reduced hallucination tolerance raises the floor. This is the GPT-6 effect directly. As Anthropic's model card research has shown with Claude's successive releases, models with lower hallucination rates also become more conservative about citing sources they cannot corroborate. Brands that give models citable, specific, verifiable claims get cited more. Brands with fluent but vague content get skipped.

How we got here

Year	Milestone	Impact on brands
2022	ChatGPT launched publicly	First wave of AI-influenced purchase decisions began
2023	GPT-4 released with improved reasoning	Brands discovered their generic content often went uncited
2024	Perplexity and Google AI Overviews scaled rapidly	AI citation share became a measurable brand metric
2025	Claude 3 Opus and GPT-4o cut hallucination rates significantly	Specificity gap between cited and uncited brands widened
2026 Q1	Google Astra architecture shown to reduce errors further	Factual grounding became the primary citation filter
2026 Q3	OpenAI launches GPT-6 Sol and Luna on Astra-class architecture	Brands with claim-dense content gain disproportionate citation share

AI citation readiness: brand scorecard

Scoring methodology: each brand was assessed across five dimensions based on publicly observable content structure, named data usage, and citation frequency data from winek.ai tracking across ChatGPT, Perplexity, and Gemini. Percentages reflect estimated AI citation rate within their primary query category.

Brand	Claim specificity	Comparative content	Named case data	Citation rate	Overall
HubSpot (post-restructure)	85%	★★★★☆	78%	58%	★★★★☆
Salesforce	72%	★★★☆☆	65%	44%	★★★☆☆
Notion	61%	★★★☆☆	49%	38%	★★★☆☆
Shopify	79%	★★★★☆	71%	53%	★★★★☆
Zendesk	55%	★★☆☆☆	41%	29%	★★☆☆☆
Intercom	68%	★★★☆☆	60%	41%	★★★☆☆

Shopify's score reflects its aggressive shift toward merchant outcome stories with specific revenue figures. Zendesk's lower score reflects a content library still heavily weighted toward feature descriptions rather than outcome data.

What you can steal from this

1. Audit for named specificity, not word count. Go through your top 50 posts and count how many sentences contain a named brand, a specific number, and a time period in the same claim. If fewer than 20% do, you have a citation problem regardless of your traffic numbers.

2. Restructure comparisons before GPT-6 rolls out widely. GPT-6 Sol and Luna are launching now. If your content still avoids direct comparisons with competitors, you are invisible on the queries that drive decisions. Write the comparison. Name the competitor. Show your methodology.

3. Treat model error rates as your content quality benchmark. GPT-6 class models will skip sources they cannot verify. If your claims are not sourced, dated, or attributable, assume they will be skipped. Write every factual claim as if a model is checking it against three other sources.

4. Front-load your most citable sentence. AI models pull from the first substantive paragraph more than any other. Put your sharpest, most specific claim in the opening. If your first paragraph is throat-clearing about industry trends, rewrite it.

5. Track citation share by model, not just by query. GPT-6 Sol and Luna will behave differently from GPT-4o. The citation patterns will shift within weeks of wide deployment. Use a measurement layer like winek.ai to track which models are citing you and which are dropping you, then reverse-engineer the content differences between the two groups.

The brands that move on this now will compound an advantage over the next 12 months. The ones waiting to see how GPT-6 "settles" will be optimizing for a model generation that is already being replaced.

For more on how AI search visibility compounds (or erodes) across query categories, what actually drives AI recommendations breaks down the structural factors that persist across model generations. And if you want to know where your brand currently stands, your GEO score is probably between 30 and 45 is a useful starting benchmark.