How to Rank on ChatGPT.
The 10 Factors That Actually Move Citations.

Yes. GPTBot is OpenAI's training crawler and OAI-SearchBot is its real-time retrieval crawler. Check your server logs for those user agents — both should be visible if your robots.txt is permissive. The default Next.js robots.ts in this stack already allows them. The fastest sanity check is grep "GPTBot\|OAI-SearchBot" on your access logs over a 30-day window. If neither shows up, your robots.txt is blocking, your CDN is blocking by user-agent rule (Cloudflare's bot fight mode does this), or your site is too new and small to be in OpenAI's crawl rotation. The first two are fixable in minutes; the third resolves once authority signals build.

It is the single biggest lever, but not the whole story. ChatGPT Search retrieves from Bing's index at query time, so being indexed in Bing and ranking in the top 10 for your seed queries is non-negotiable. But citations are not just a function of Bing rank — OpenAI's reranker biases toward sources that already exist in the training data (Wikipedia, Reddit, large publishers) and toward content with clear question-answer structure. Brands that rank top-3 in Bing but have no Wikipedia presence and no Reddit mentions still under-cite. The hierarchy: Bing index + ranking → corroborating training-data mentions → on-page schema and structure.

More than almost any other single signal. Wikipedia and Wikidata are disproportionately weighted in OpenAI's training data — the model treats them as ground truth. Brands with a Wikipedia article are cited at materially higher rates than brands without one for the same category queries. The threshold for a stand-alone Wikipedia article (notability, independent secondary coverage) is a real bar, but a Wikidata entity with a clean sameAs graph is achievable for almost any operating brand and lifts citation rates noticeably on its own. We treat Wikidata entry as a 90-day deliverable on every GEO retainer.

llms.txt is a plain-text manifest at /llms.txt that tells LLMs which URLs on your site are canonical, what each one is about, and which sections to prioritise. Anthropic and several others have signalled they will read it; OpenAI has not committed but historically follows community standards once they reach critical mass. The cost to implement is hours, the upside is non-zero — it costs nothing to be early. Both /llms.txt and /llms-full.txt are already shipping on this site (auto-generated from blog content). If you are running a content site of any size, ship it.

Recency matters for ChatGPT Search and Browse — the live retrieval modes — and it matters less for default ChatGPT, which leans on training data. Pages with explicit datePublished and dateModified schema, recent content updates, and freshness signals (new publication dates, current-year references in body copy) are preferred by the retrieval layer. For brands competing in fast-moving categories (AI, fintech, ecommerce trends), refreshing pillar content quarterly and updating the dateModified field measurably lifts ChatGPT Search citation rates.

Indirectly, yes. The structural shifts that win citations — short, declarative summary paragraphs at the top of each section, clear H2 questions followed by 40-to-80 word answers, definition blocks before context, lists and tables for comparative data — are the same shifts that improve human readability. Writing for the model and writing for the reader converge on the same prose discipline. What we explicitly do not recommend: the kind of awkward keyword-stuffed prose that early SEO tools used to recommend for AI Overviews. ChatGPT's reranker penalises synthetic, low-information prose just like Google's quality systems do.

Yes — through controlled prompt audits. Pick a fixed set of 50 to 200 category-defining prompts, run them monthly across ChatGPT (free + Plus + Search), record whether the brand was cited, the position of the citation, and the surrounding context. Aggregate the data into a share-of-voice score versus competitors. Paid tools (Profound, Otterly, Goodie, Daydream) automate this; the manual version is feasible at smaller scale. Track the same prompt set across ChatGPT, Gemini, Claude and Perplexity so you can see which engine is moving and which is lagging. We bundle this measurement into every GEO retainer.