How to optimize for ChatGPT: 20 ranking factors from 129,000 domains.

How ChatGPT actually picks sources

ChatGPT's search architecture is a multi-stage pipeline: a tiny classification model, a specialized “Thinky” query-generation model, Bing-powered web retrieval, semantic chunk scoring, and final synthesis by the frontier model. The system typically cites 3–5 deeply-read sources per response, with Wikipedia dominating at 7.8% of all citations.

SearchGPT is now used in 46% of ChatGPT interactions. Usage for general information searches tripled from 4.1% to 12.5% in just six months (Feb–Aug 2025). The platform processes 2.5 billion daily prompts with 500M+ weekly active users.

For source selection, ChatGPT uses sequential search queries (not parallel like Perplexity), reviewing multiple sites before aggregating. Six selection criteria emerge from testing: precise keyword matching, search intent recognition (appending terms like “tutorial”), aggressive recency filtering, credibility (E-E-A-T), trustworthiness (author bios, methodology), and variety of perspectives.

The top 20 ranking factors

SE Ranking's analysis of 129,000 domains used SHAP (SHapley Additive exPlanations) values to rank features most predictive of ChatGPT citation. The ranked top of the list:

Brand mentions on Reddit and Quora yielded 4× higher citation likelihood. Articles over 2,900 words averaged 5.1 citations vs 3.2 for under 800 words — longer is better, but only with structure (next chapter).

“Flashy "AI hacks" like LLMs.txt barely have any impact. What drives ChatGPT citations are the fundamentals. — Yulia Deda, SE Ranking”

Structure beats every shortcut

The SE Ranking study quantified the structure effects:

• Content sections of 120–180 words between headings receive 70% more citations than sections under 50 words
• Pages with expert quotes average 4.1 citations vs 2.4 without
• Pages with 19+ statistics average 5.4 citations vs 2.8 with fewer
• Broad topic-descriptive URLs get 2× more citations than keyword-optimized URLs
• Companies with review scores below 70% are significantly less likely to be referred

Perhaps most interestingly: .gov and .edu domains averaged only ~3.2 citations — content quality matters more than TLD. The authority bias is weaker than the SEO industry has assumed.

What does NOT work

The most consequential negative findings, all from controlled testing:

• LLMs.txt has negligible impact. Removing it actually improved the SE Ranking model's predictive accuracy. Multiple independent experiments (Dejan AI, others) confirm no measurable lift.
• FAQ schema markup shows negative correlation. Pages with FAQ schema averaged 3.6 citations vs 4.2 without. ChatGPT does not appear to access schema during grounding.
• Keyword-optimized URLs underperform. Topic-descriptive URLs get 2× more citations. ChatGPT cares about clarity, not exact match.

This is the part most GEO “experts” get wrong. The industry has spent eighteen months selling LLMs.txt and FAQ schema as silver bullets. The data says they aren't bullets at all.

The four crawlers (and how to control them)

OpenAI operates three distinct crawlers, each controllable independently via robots.txt:

• GPTBot — for training. Grew 305% in crawling volume from May 2024 to May 2025, becoming the dominant AI crawler at 30% share.
• OAI-SearchBot — for real-time SearchGPT.
• ChatGPT-User — for user-initiated browsing within ChatGPT sessions.

You can allow some and block others. Most brands should allow all three — blocking GPTBot specifically removes you from training data, which has long-tail visibility implications most teams underestimate.

Tracking ChatGPT referral traffic in GA4: filter for the chatgpt.com referrer. Note that free-tier users often don't send referrer data, creating attribution gaps. Plan for that gap rather than around it.

The execution priority order

Based on SHAP weights, the highest-leverage actions for an existing brand:

1. Audit your domain trust score. If you are below 80, fix the foundational citation graph before anything else. Get into trusted lists, directories, and editorial coverage.
2. Get FCP under 0.5 seconds. The drop-off above that threshold is steep and asymmetric — ChatGPT punishes slow pages much harder than Google does.
3. Earn Reddit and Quora presence. 4× citation likelihood is among the highest single-factor effects in the data.
4. Restructure your top 20 commercial pages. Add statistics, comparison tables, expert quotes, and 120–180 word passage blocks between H2s.
5. Submit a structured Organization entity to Wikidata and apply schema.org markup with consistent identifiers across owned properties.

Every statistic in this piece follows Visigent's research standards: named sources, dated figures, corrections on request to help@visigent.com.