Update at 01/27/2026 by Luiz Gustavo
You probably have a small website and you're wondering: how the heck do I get ChatGPT to recommend my business?
I'll be honest with you, it's not super easy, but it's definitely possible. While traditional SEO focuses on Google rankings, generative engine optimization (GEO) is about getting cited by AI search tools like ChatGPT, Perplexity, and Claude. ChatGPT does prioritize high-authority websites in most cases, but there are outliers. And today, we're exploring exactly those small sites that cracked the code for SEO for ChatGPT.
I want to start with the fun part, the actual sites. These sites are getting recommended because they're doing specific things right. Let me break down each one so you can see exactly what's working.
Although the site doesn't have the most appealing visual design, it achieved the highest score by combining three rare elements:
<p> tags correctly.png)
Second-best overall score (81.45/100) driven by exceptional vocabulary density:
Competitive edge: Most technical vocabulary of any site analyzed. Despite moderate length (2,779 words), achieves maximum density with specialized affiliate marketing terminology.
Third-place score (75.84/100) dominated by industry jargon mastery:
Competitive edge: Industry-specific jargon signals deep expertise. Uses YouTube-specific terminology (CPM, RPM, CTR, etc.) extensively, proving insider knowledge.
Fourth-place score (71.44/100) with efficiency and local expertise:
Competitive edge: Short but dense. Packs Miami real estate expertise into 1,175 words without dilution, maintaining high vocabulary diversity (TTR 0.423).
.png)
Fifth-place score (71.31/100) with solid research foundation:
Competitive edge: Balanced profile across all dimensions. No single exceptional metric, but consistent strength in research (8/10), expertise (17/20), and vocabulary density (22.53%).
Sixth-place score (68.66/100) with perfect nuance:
Competitive edge: Exceptional analytical balance. Perfect nuance score (5/5) shows sophisticated treatment of Instagram algorithm complexities, presenting multiple perspectives.
Seventh-place score (59.83/100) with maximum vocabulary efficiency:
Competitive edge: Ultimate efficiency. Achieves TTR 0.537 (highest in top 10) in only 663 words, proving that density matters more than length for LLM citations.
Eighth-place score (59.03/100) with balanced fundamentals:
Competitive edge: Reliable fundamentals. No exceptional metrics, but maintains good baseline across originality, expertise, and vocabulary density.
.png)
Ninth-place score (57.8/100) with jargon expertise despite no proprietary data:
Competitive edge: Proves that original research isn't mandatory. Compensates for lack of proprietary data (0/10) with exceptional technical jargon (109 acronyms) and expertise (16/20).
Tenth-place score (56.3/100) with extreme vocabulary efficiency:
Competitive edge: Maximum compression. Achieves extraordinary TTR 0.6243 in only 527 words, proof that ultra-concise, vocabulary-dense content can compete with longer articles for LLM citations.
Before diving into the numbers, here's how we've made this research (for the full methodology, check the end of this article):
Understanding ChatGPT SEO optimization requires a different approach than traditional search engine optimization. We analyzed 60 low-authority websites (all with Domain Rating < 15) across 74 different metrics to decode what makes content citable by AI search engines.
The setup: For each of 30 different questions (like "What supplements are popular in 2025?" or "Which affiliate niches are most profitable?"), we found:
We then analyzed these 60 sites across 74 metrics including originality, vocabulary density, HTML structure, schema markup, answer completeness, recency, and traditional EEAT signals.
So, now that our research is done, what are the main patterns that we found out? Let's break it down.
This is the big one. When we scored sites for originality, the top performers averaged 77.1% (61.7 out of 80 points) and this single factor carried more weight than anything else. But what does "originality" actually mean to an LLM?
Case study: robertyoung.consulting (#1 ranked)
78/80 Points (97.5%)
Why This Site Won
robertyoung.consulting achieved the highest originality score in our entire study by combining three rare elements:
The Contrast
Compare this to the average random site: ~35/80 points (43.8%). The difference? Random sites curate existing information. This site created new knowledge. That's what originality means to an LLM.
We broke it down into four concrete criteria:
1. Unique Angle (out of 20 points)
2. Proprietary Data (out of 10 points)
3. Nuanced Analysis (out of 5 points)
4. Demonstrated Expertise (out of 20 points)
Real Example: What Originality Looks Like
Let's compare two approaches to "What supplements are popular in 2025?"
Generic approach (low originality):
Original approach (high originality):
The difference? The second one couldn't be written by someone who spent an afternoon Googling. It required genuine expertise, original research, and a unique perspective.
The Takeaway
Originality isn't about being contrarian or clickbaity. It's about:
Optimizing for search engines is not enough. Start having something new to say, backed by real data, explained with actual expertise. That's what will help you to get cited in the age of AI search.
This was our 2nd biggest finding. 83.3% of cited sites were published in 2025 (the current year), compared to only 23.3% of random sites (answering the same questions). That's a +257% differencethe single biggest gap we found in the entire study.
Publication year distribution: Cited vs Random sites
⚠️ Zero citations
Even 1-year-old content struggles
+257% Difference
Current year content (2025) has a massive citation advantage. Even content from 2024 received zero citations in our study, making freshness the single biggest factor for AI search visibility.
But here's the kicker: even content from 2024 (just one year old) barely gets cited. We found zero cited sites from 2024, while random sites had 8. ChatGPT wants content from this year, not just "recent" content.
The takeaway: If your article says "In 2023..." or "Last year...", you're already losing. Update that date, refresh your examples, and make it scream "2025" or "2026."
This makes sense when you think about it: when ChatGPT uses its AI search tool, it's looking for fresh information that's not already in its training data, making recency a critical factor in any generative engine optimization strategy.
Sites that got cited used 15.76% more diverse vocabulary (measured by Type-Token Ratio). They also used:
Vocabulary density comparison across all metrics
Key Insight
Cited sites consistently show 15-25% higher vocabulary density across all metrics. They use technical terminology naturally, signaling genuine expertise rather than surface-level knowledge.
And here's the plot twist: cited sites were actually shorter on average (1,493 words vs 1,960 words for random sites). They just packed more punch into fewer words.
Cited sites are shorter but score higher
The Takeaway
Cited sites are 23.8% shorter but score 57.8% higher. They pack more value into fewer words, proving that density and expertise matter more than hitting arbitrary word counts. Quality over quantity, every time.
Think about it, when you read something written by a true expert, they use precise terminology naturally.
The takeaway: Don't dumb down your writing. Use technical terms when appropriate. Show your expertise through language sophistication, not word count.
Random sites actually had more HTML structure, more headings, more paragraphs, more everything. But cited sites had better structure. The key metric? P/Div ratio.
Cited sites scored 0.8 on P/Div ratio vs 0.2 for random sites. That's a +300% difference the second-biggest gap in our entire study.
Cited sites use semantic HTML that's easy to extract
Semantic HTML with proper <p> tags
Generic <div> soup with no semantic meaning
What is P/Div Ratio?
P/Div ratio measures semantic HTML quality by dividing the number of <p> tags by <div> tags.
Why it matters: LLMs can extract content more easily when it's properly tagged. Use <p> for paragraphs, <article> for main content, and save <div> for layout.
Yeah, we said it. Schema markup showed only a +0.8% difference between cited and random sites. Statistically irrelevant.
Only 33.3% of cited sites even had Article schema, yet they still got cited. This doesn't mean schema is bad, it's just not a deciding factor for LLM citations.
However, 83% of the cited websites use JSON-LD, so I'll keep adding it here just in case.
The takeaway: If you have to choose between writing great content and perfecting your schema markup, write great content. Every time.
After analyzing 60 sites and 74 different metrics, clear patterns emerged. Here's what the top-performing cited sites have in common.
100% of our top 10 sites scored perfect 10/10 on answer completeness. They didn't bury the lede or make you scroll forever. They answered the question in the first few paragraphs, then backed it up with depth.
This isn't about being superficial, it's about being structured. Give the direct answer, then provide the context, data, and nuance that proves you know what you're talking about.
The top sites used industry jargon naturally. One site had 109 acronyms in a single article about drones. Another had 29.33% long technical words in an affiliate marketing piece.
They didn't explain every term like they were writing for beginners. They assumed their audience had basic knowledge and went straight to the advanced stuff. That confidence signals expertise.
Cited sites averaged 20.8 lists per article (vs 18.6 for random sites). They broke information into scannable bullets, numbered steps, and organized hierarchies.
Why? Because it makes information actionable. "Top 10 AI Tools" naturally becomes an ordered list. "Benefits of X" becomes bullet points. LLMs can extract this structured information cleanly.
Remember: cited sites were 23.8% shorter than random sites (1,493 vs 1,960 words), yet they had:
The random sites stuffed in extra words to hit some arbitrary length target. The cited sites said exactly what needed to be said, then stopped.
Here's a mind-bender: the shortest site in our top 10 was 527 words long. The longest was 5,685 words. Both got cited.
What they shared? Vocabulary efficiency. The 527-word article had a Type-Token Ratio of 0.6243, the highest in our entire dataset. Every word counted. No fluff, no repetition, just dense, expert-level information.
9 out of 10 top sites were published in 2025. The one exception was robertyoung.consulting (our #1 ranked site), which had no publication date listed, though given its high performance and the nature of its content (discussing supplements "popular in 2025"), it's likely recent even if we couldn't verify the exact publication date.
Just as interesting as what works is what doesn't work:
All those traditional EEAT "trust signals" that work for Google seem not to be that relevant for AI search citations, at least for these small websites with low domain authority.
Based on our composite scoring (which predicted citations with decent accuracy), here's the rough formula:
What actually matters for getting cited by ChatGPT
Plus a multiplier: Is it from 2025/2026? If no, your odds drop by ~75%.
If you want ChatGPT to cite your content, your SEO for ChatGPT strategy should focus on:
<p> tags, <article>, lists)In summary, write like an expert, publish fresh content, be technically precise and make it extractable. This generative engine optimization approach differs from traditional SEO. It's about content quality that AI search engines can recognize and extract, not about gaming algorithms with backlinks and keywords.
Focus on three core factors: publish fresh content (2025/2026), demonstrate unique expertise with original insights or data, and make it easily extractable with semantic HTML. Our research found that 83.3% of cited sites were from the current year, and they averaged 77.1% on originality scores (unique angles + proprietary data). Write like an expert explaining something to a peer, not like you're optimizing for a search algorithm.
Yes, domain authority still matters for LLM citations. According to the Savannabay 60-keyword study, which analyzed 2,410 citations across Google Search, Google AI Overview, ChatGPT, and Perplexity, the average domain authority requirement is almost identical among all platforms (only a 5.1-point difference). For competitive short-tail queries, LLMs still favor high-authority domains (DR 75–80+), much like traditional SEO. But for mid- and long-tail queries, authority requirements drop sharply (12–30%), allowing smaller or niche sites to be cited when their content is specific, recent, and directly answers the question.
However, ChatGPT shows the highest variability of all platforms, with a very large standard deviation (±30–36 DR points). This means it can cite both extremely low-authority sites (DR 0–1) and very high-authority sites (DR 90+) for the same type of query.
That's where small websites can thrive. If you write original content demonstrating expertise, with a fresh date, a 20% vocabulary density, you are in the right path.
No. Schema markup showed only a +0.8% difference between cited and random sites, statistically irrelevant. Only 33.3% of cited sites had Article schema, yet they still got cited.
That said, 83% of cited sites use JSON-LD markup (various types, not just Article schema), suggesting it's become standard practice. Since it's technically easy to implement and doesn't hurt, include it. Just don't prioritize it over writing great content.
Priority order: Original content > Fresh dates > Technical vocabulary > Clean HTML > Schema markup
There's no magic number. The shortest cited site in our top 10 was 527 words. The longest was 5,685 words. Both got cited.
What they shared was vocabulary efficiency: every word counted. The 527-word article had the highest Type-Token Ratio (0.6243) in our entire dataset, meaning it used incredibly diverse vocabulary without repetition or fluff.
Cited sites averaged 1,493 words vs 1,960 for random sites, 23.8% shorter, yet they packed more technical depth, richer vocabulary, and higher information density. Quality over quantity, every time.
Current year only. 83.3% of cited sites were published in 2025 vs 23.3% of random sites, a +257% difference, the biggest gap in our entire study.
Even content from 2024 (just one year old) barely got cited: zero cited sites vs 8 random sites. ChatGPT's web search specifically looks for fresh information not in its training data.
Action items:
Short-term wins (do this today):
<article>, proper <p> tags for content paragraphs)Medium-term improvements (do this this week):
Long-term competitive advantage (do this this month):
The highest-impact action? Publish something original with data no one else has. Sites with proprietary data scored 8-10/10 on that dimension and massively outperformed generic content.
Yes, but focus on semantic quality over quantity. Cited sites scored 0.8 on P/Div ratio vs 0.2 for random sites (+300% difference).
Cited sites also used 20.8 lists per article vs 18.6 for random sites. Lists make information actionable and extractable, perfect for LLM parsing.
Based on our composite scoring model that predicted citations with decent accuracy:
Plus a critical multiplier: Is it from 2025/2026? If not, your odds drop ~75%.
This isn't a perfect formula, but sites scoring 70+ using these weights had significantly higher citation rates than sites scoring below 50.
Research basis: Analysis of 60 low-authority websites (DR < 15) across 74 metrics, comparing 30 sites cited by ChatGPT vs 30 non-cited sites answering identical questions. Dataset: oficial.csv, 2026-01-22.
We analyzed 60 low-authority websites (Domain Rating < 15) using ChatGPT's web search feature: 30 sites ChatGPT cited and 30 control sites from organic search answering identical queries. All sites verified with DR 0-14 (median: 3.5), word count 527-5,685, publication years 2019-2025, collected January 15-18, 2026.
Originality (Manual): Unique angle (0-20), proprietary data (0-10), nuanced analysis (0-5), demonstrated expertise (0-20). Inter-rater agreement: 87%.
Vocabulary Density (Python/NLTK): Type-Token Ratio (cited: 0.4619 vs random: 0.399), long words % (17.87% vs 14.94%), acronyms (28.7 vs 23.03), content word ratio.
HTML Structure (PowerShell): P/Div ratio (0.8 vs 0.2), semantic tags, heading hierarchy, list usage.
Schema Markup (Python): Article schema completeness (5.23/9 vs 4.93/9, not significant).
Recency: Publication dates extracted from schema/meta tags. 2025: 83.3% cited vs 23.3% random (χ² = 24.8, p < 0.001).
Answer Completeness (Manual 0-10): Direct answer (0-4), completeness (0-3), specificity (0-3).
EEAT Signals: Author bios, credentials, external links (cited: 1.3, random: 1.9). Random sites had MORE EEAT signals but got cited LESS.
Weighted model: 40% Originality (largest differences) + 25% Expertise (top sites averaged 17/20) + 20% Density (15-25% differences) + 10% Structure (+300% difference) + 5% Completeness (less discriminating). Top 10 cited: 56.3-89.32 points (mean: 69.1) vs random: 31.2-58.4 (mean: 43.8). Predictive accuracy: ~78%.
T-tests for normally distributed data, Mann-Whitney U for non-parametric, Chi-square for categorical. Significance threshold: α = 0.05. Largest effects: Recency +257% (p < 0.001), P/Div Ratio +300% (p < 0.01), Long Words +19.6% (p < 0.05), TTR +15.76% (p < 0.05).
Sample size: n=30 per group (power ~0.65), confidence intervals ±10-15%. Selection bias: English queries only, informational focus, B2C niches. Temporal validity: January 2026 data, ChatGPT-specific. Causation: Observational study identifies correlations, not causal relationships.
While our sample size limits statistical power, the observed patterns are strong enough (particularly for recency and P/Div ratio) to provide actionable insights. The +257% recency difference and +300% P/Div difference show effect sizes large enough to be practically significant even with n=30.
Luiz Gustavo is full-stack developer in Savannabay and Gobrunch, Computer Science student
Richard Lowenthal is founder of Savannabay, co-founder of GoBrunch and Live University, AI Search & GEO enthusiast
Savannabay. 2026 All rights reserved.
MENU