OpenAI has never published a ranking formula for ChatGPT Search. Its help documentation says only that results are ranked on multiple factors meant to surface relevant, reliable information, and that placement is not guaranteed. Every list of official signals you read is therefore inference. The useful question is narrower: what do the largest independent studies show ChatGPT rewarding in practice? This guide separates what OpenAI confirms from what researchers have measured. You will see how a page moves from retrieval to citation, which signals hold up, which popular tactics do not, and what your team should change first.
A page gets cited when it clears three gates. OpenAI must be able to crawl it. It must sit high in the results for the searches ChatGPT runs behind the scenes. Its title and headings must then match those searches closely enough to be picked over the alternatives. In large studies, retrieval position and title relevance are the strongest observed signals, which makes them the closest thing to ChatGPT ranking factors that anyone can currently evidence. All of this is correlation, not a confirmed algorithm.
OpenAI’s public guidance is short, and most of it concerns access. Its crawler documentation confirms four points:
OpenAI also says ChatGPT sometimes works with other search providers and rewrites a prompt into targeted queries.
Think of the process as hiring. Retrieval builds the longlist, your title and snippet act as the CV, and only a few candidates reach the interview. Four stages matter:
That gap is the reason titles and opening lines deserve so much attention. Your page has to look worth opening, or carry its point in the little text the model sees.
Four signals have the strongest support for ChatGPT citations across independent datasets. OpenAI has confirmed none of them.
AirOps and Kevin Indig analyzed 16,851 queries and 353,799 pages. A page in the top retrieval position was cited 58% of the time, against about 14% at position 10. The study names retrieval rank as the single strongest predictor of citation, which is why classic search work still comes first.
Ahrefs studied 1.4 million prompts using semantic similarity, a measure of how close two pieces of text are in meaning. Cited titles scored 0.602 against the original prompt, while uncited titles scored 0.484. The match rose to 0.656 against fan-out queries. Search results with natural language URL slugs were also cited more often, at 89.78% versus 81.11%.
AirOps also compared each query with the best-matching heading on a page. Even among pages ranked near the top, stronger heading matches lifted citation rates from 55.9% to 75.3%. Headings tell the model which section answers which sub-question, so vague labels waste a real opportunity.
Exact rankings do not explain everything. Grow and Convert found that only about 40% of cited sources appeared in the first 10 pages of Google or Bing for the visible fan-out queries. When the test widened to any page on the same domain, Google overlap rose from 27% to about 50%. The authors read this as a sign that ChatGPT may favor domains with broad authority on a subject, which supports building depth across a topic cluster.
Several common recommendations look weaker once you test them against the data.
| Tactic | What the evidence shows | Better move |
|---|---|---|
| Adding JSON-LD for ChatGPT | RESONEO found structured data stripped when pages were converted for ChatGPT’s reading cache. | Keep schema for Google and entity clarity, and put key facts in visible text. |
| Optimizing only for Bing | RESONEO traced paid Thinking mainly to scraped Google results and free modes mainly to OpenAI’s own index. | Check Bing, compete on Google, and stay crawlable by OpenAI. |
| Covering every fan-out query on one page | AirOps rates fan-out coverage as a weak signal. | Answer the main question in depth, then build supporting pages. |
| Treating freshness as the main lever | Ahrefs found a median cited age of about 500 days in search results, with the newest pages often dropped. News is the exception. | Refresh when facts change, and prioritize recency on news-led topics. |
| Posting on Reddit for citations | Ahrefs recorded a 1.93% citation rate on Reddit’s dedicated retrieval channel. | Use Reddit to learn buyer language, then earn citations on your own pages. |
| Polishing meta descriptions for ChatGPT | RESONEO found OpenAI’s index ignores them, though Google-fed pipelines use them about one time in three. | Keep them for Google, then fix your opening lines. |
For more on why markup alone does not win selection, read our breakdown of what actually gets an answer selected.
ChatGPT is not a single search engine. RESONEO’s August replay found free Think drawing 74.7% of its results from OpenAI’s own index, while paid Thinking drew 75.3% from scraped Google results. Two buyers can ask the same question and receive answers built from different sources.
Google rankings help with paid Thinking, yet they do not guarantee entry to the free pool. You need pages that perform in classic search and also read well to OpenAI’s crawlers. The study’s authors warn that these mechanics shift quickly, so build on durable habits rather than any single quirk.
These steps suit a marketing or SEO lead working with one developer. Start with your 10 to 20 most commercial pages.
When buyers shape a shortlist inside an AI answer, a missing brand leaves no trace in your dashboard. You only see the pipeline that never arrived.
The risk of starting is low. Clearer titles, tighter openings, and crawlable HTML also help human readers and classic search. The real limit is certainty. No agency can promise placement, and the same prompt can return different sources on different days. Sound ChatGPT search optimization therefore judges progress by the trend across several months, never by one answer.
If you want outside support, our ChatGPT AI SEO services combine prompt research, answer-ready content, entity and schema work, and share-of-voice reporting. Our answer engine optimization services extend that work to other platforms, so the effort feeds your wider AI search visibility too. TIS is headquartered in Noida and serves clients across India, the US, and Europe.
ChatGPT appears to reward pages that are reachable, rank well in the systems it draws on, and state their subject clearly in the few places it looks first. Those are findings from outside research, not OpenAI policy, so expect details to shift. The priority order should hold: fix access, strengthen rankings, then sharpen titles, headings, and openings.
Run that sequence on your most commercial pages, measure for a quarter, and expand what works. To get a clear starting point, ask TIS for a citation audit covering your priority prompts and competitors.
No. OpenAI states that results are ranked on multiple factors intended to surface relevant, reliable information, and it says placement is not guaranteed. Its public guidance focuses on access, mainly allowing OAI-SearchBot and confirming that your host permits OpenAI’s published IP addresses. Everything else you read about specific signals, including this guide, comes from independent studies. Treat those findings as strong correlations rather than rules, and retest them as the product changes.
OpenAI says it sometimes partners with search providers, but it does not publish how results are blended. Independent research suggests a mixed picture. RESONEO traced most free-tier citations to OpenAI’s own index and most paid Thinking results to scraped Google pages. For practical planning, rank well in Google, check your Bing visibility, and make sure OpenAI’s crawler can reach every page you want cited.
No. OpenAI treats the two crawlers independently. GPTBot relates to training its foundation models, while OAI-SearchBot governs inclusion in search answers. You can disallow the first and allow the second. The bigger risk is a broad bot rule or firewall setting that catches both by accident. Review your robots.txt file, test your CDN rules, and allow about 24 hours for a robots change to register.
There is no published timeline, and ChatGPT’s reading cache adds another variable. RESONEO found that cached copies refresh according to how often users ask about a page, and it reported copies served more than 90 days after fetching. Popular pages can update quickly, while quiet ones may lag. Expect gradual change, then monitor a fixed set of prompts for several months before judging any rewrite.
It is useful, but current evidence does not show it working as a direct ChatGPT lever. RESONEO observed JSON-LD being stripped when pages were converted for the reading cache, so the model may never see it on that path. Schema still supports Google and describes your organization, products, and authors clearly. Keep it accurate, and make sure every important fact also appears in visible text.
Yes, although authority helps. Retrieval position favors sites that already perform well in search, so a smaller domain will struggle on broad category terms. Narrow, specific questions offer a better opening, especially where your team holds real expertise. Build a focused cluster on one subject, use your buyers’ language in titles and headings, and answer each question directly. Depth on a defined topic is usually the more realistic route.
How to Build AI-Ready Content That Gets Cited by ChatGPT and Perplexity
© 2026 TIS. All rights reserved.
