To get content cited in Google AI Overviews and ChatGPT search, you first need the same things classic search needs: the page must be crawlable by the right bots, indexed, and genuinely useful. Google states that there are no extra technical requirements or special markup for AI Overviews, and OpenAI says any public site can appear in ChatGPT search as long as its search crawler is not blocked. What you can improve beyond that is how easy your page is to quote: direct answers near the top, self-contained sections, specific facts with sources, and headings that match the questions people actually ask.
This guide separates what the platforms officially say from what is reasonable inference, gives you four copy-ready prompts for reworking a page, and ends with a way to measure whether any of it changed your traffic.
What GEO means and how it differs from SEO
GEO (generative engine optimization) is the practice of making content more likely to be used and cited by AI systems that write an answer instead of returning a list of links. The term comes from a 2023 research paper, GEO: Generative Engine Optimization, by Aggarwal and colleagues, which tested how rewriting source pages changed their visibility in AI-generated answers.
The practical difference from SEO is the unit of competition. In classic results, a page competes for a position. In an AI answer, a passage competes to be quoted or summarized, and the answer may draw on several pages at once. Google describes a technique it calls “query fan-out”, in which AI Overviews and AI Mode can run several related searches to gather a wider set of supporting pages. A page that answers one narrow sub-question clearly can therefore be cited for a broader query it would never rank for on its own.
What does not change: AI features on Google are built on Google’s index, and ChatGPT search relies on OpenAI’s own crawler plus other sources. If a page is not crawlable or not indexed, no amount of rewriting helps. GEO is a layer on top of SEO, not a replacement for it.
What Google and OpenAI officially say
Both platforms publish documentation for site owners, and it is shorter and less mysterious than most GEO advice suggests. The table below summarizes the official position on each question, based on Google’s AI features and your website page and OpenAI’s crawler documentation and publishers FAQ.
| Question | Google AI Overviews / AI Mode | ChatGPT search |
|---|---|---|
| Which crawler must be allowed? | Googlebot (normal Search crawling) | OAI-SearchBot |
| Basic eligibility | Page indexed and eligible to show with a snippet | Public site, OAI-SearchBot not blocked |
| Special markup or AI-specific files needed? | No; Google says no special schema.org markup or AI text files are required | No special markup described |
| Training opt-out, separate from search | Google-Extended token (does not affect Search) | Block GPTBot (does not affect search inclusion) |
| Limit what is shown from a page | nosnippet, data-nosnippet, max-snippet, noindex | noindex meta tag, which the crawler must be able to read |
| Where traffic shows up | Search Console Performance report, “Web” search type, combined with other search traffic | Analytics referrals tagged utm_source=chatgpt.com |
| Published ranking factors for AI answers | None beyond normal Search fundamentals | None published |
Two consequences follow. First, a robots.txt rule written to stop AI training can quietly keep a site out of ChatGPT search if it blocks every OpenAI user agent, including OAI-SearchBot. OpenAI treats each bot’s setting independently, so you can block GPTBot and still allow OAI-SearchBot. Our comparison of robots.txt, meta robots and X-Robots-Tag for AI crawlers walks through the exact rules.
Second, files such as llms.txt and AI-specific schema are not requirements on either platform. If you are weighing whether to add one anyway, read what the llms.txt spec requires and what Google says about it before spending time on it.
What the research suggests about extractable content
Since neither platform publishes ranking factors for AI answers, the best evidence available is experimental. The GEO paper built a benchmark of queries and measured how often rewritten source pages were used in generated answers. In that study, adding relevant quotations and adding statistics performed best, citing sources and improving fluency helped moderately, and keyword stuffing did not help. The authors also report that results varied by domain, so a tactic that helps a history page may not help a product page.
Treat this as directional. It is one research setup, run on generative engines as they existed in 2023 and 2024, and production systems change often. The defensible reading is narrow: specific, evidenced passages are easier for an AI system to use than vague ones. That also happens to be what Google’s helpful content guidance asks for, so the overlap is reassuring rather than coincidental.
One warning that matters more than any tactic: “add statistics” means add real statistics you can source. A model asked to make a page more quotable will happily invent numbers. Every prompt below includes an explicit rule against that.
Four prompts for reworking a page
These prompts restructure an existing page without changing its facts. Run them one at a time on a single page, compare each output with your original and keep only the changes that make the page clearer for a human reader. Square brackets mark what you replace.
Prompt 1: A direct-answer opening
The opening is the passage most likely to be lifted into an answer, so it should answer the page’s main question on its own.
Below is the opening of an article titled "[TITLE]". The main question
a reader arrives with is: [QUESTION].
Rewrite the opening as 2-3 sentences that:
- answer the question directly in the first sentence,
- make sense if quoted alone, without the title or the rest of the page,
- define any technical term the first time it appears,
- keep every fact from the original and add none.
If the original opening does not contain enough information to answer
the question, say what is missing instead of filling the gap.
Original opening:
[PASTE]“Make sense if quoted alone” is the core instruction: AI answers often show a passage without its surrounding context. “Add none” and the final fallback prevent the model from inventing an answer when your page does not actually have one, which is useful information in itself.
Prompt 2: Headings that match real questions
Headings act as labels that help both readers and retrieval systems find the right passage. Vague headings such as “Overview” or “Key considerations” tell neither what the section answers.
Here are the H2/H3 headings of my article, each followed by the first
two sentences of its section:
[PASTE]
Here are real queries people use for this topic, from Search Console,
People Also Ask and customer questions:
[PASTE QUERIES]
For each heading:
1. Say which query (if any) the section answers.
2. Propose a heading that describes what the section actually answers,
using the reader's wording where it fits naturally.
3. Flag sections that answer no query on the list.
Do not turn every heading into a question; use a question only where
the section is a direct answer to it.Supplying real queries keeps the headings grounded in demand rather than in what the model imagines people ask. The last instruction avoids a common over-correction where every heading becomes “What is…?” or “How do I…?”, which reads badly and adds nothing. For background on matching wording to intent, see semantic SEO and search intent.
Prompt 3: Self-contained sections
A self-contained section starts with a sentence that answers its heading, then supports it. Sections that rely on “as mentioned above” or “this” referring to a previous paragraph lose meaning when extracted.
Rewrite the section below so that:
- the first sentence answers the heading and can stand alone,
- pronouns and phrases like "this", "it" or "as mentioned" that point
outside the section are replaced with the thing they refer to,
- each paragraph makes one point,
- steps or options that are currently buried in prose become a list,
only where that makes them easier to follow.
Rules:
- Do not add facts, numbers, examples or sources that are not in the
original.
- Where a claim would be stronger with a specific number or a source,
insert [SOURCE NEEDED: what kind] instead of inventing one.
- Keep the author's voice; do not add marketing language.
Heading: [HEADING]
Section:
[PASTE]The [SOURCE NEEDED] placeholder turns the research finding about statistics into an honest workflow: the model shows you where evidence would help, and you supply real data or leave the claim modest. The voice rule matters because a common side effect of “optimize this” prompts is promotional filler.
Prompt 4: An evidence and gap audit
The final prompt checks the whole page against the questions an AI answer would need to cover, which is where query fan-out becomes useful to think about.
Read the full article below. The main query is [QUERY].
1. List 5-8 follow-up questions a person asking [QUERY] is likely to
ask next.
2. For each, quote the sentence in the article that answers it, or
write NOT ANSWERED.
3. List every factual claim that has no source, number or concrete
example behind it.
4. List any statement that may be outdated (versions, prices, product
names, dates) so I can check it against current official pages.
Do not rewrite the article. Report only.
[PASTE ARTICLE]Asking for a quoted sentence rather than a yes or no forces the model to show its evidence, which makes false “yes, it’s covered” answers easy to spot. The output becomes your edit list. If many follow-up questions are unanswered, a broader content gap analysis against competing pages is the next step.
Before and after: one paragraph
The example below is illustrative, written for this guide, on a topic where the facts come from the official sources linked above.
Before: “There are lots of things to think about when it comes to AI search. As we discussed, bots are a big part of it. You’ll want to make sure you’re not accidentally blocking anything important, because that could hurt your visibility a lot.”
After: “To appear in ChatGPT search, a site must allow OpenAI’s OAI-SearchBot crawler in robots.txt. Blocking GPTBot, the crawler OpenAI uses for model training, does not remove a site from ChatGPT search, because OpenAI applies each bot’s rule separately. A blanket rule that blocks all OpenAI user agents, however, blocks both.”
The second version names the crawler, states the condition, explains the distinction a reader is likely to get wrong and needs no surrounding text. Nothing in it is invented, and every claim can be checked against OpenAI’s documentation.
What not to do
Several popular GEO tactics are either unsupported by the platforms’ documentation or contradict it.
- Changing dates without changing content. Google’s helpful content guidance lists updating dates to look fresh, without substantial changes, as a warning sign. Refresh a page when facts change, and say what changed.
- Adding schema in the hope of AI citations. Google says no special structured data is needed for AI features. Structured data still helps for the rich results it supports, but only when it matches visible content. If you added FAQ markup for this reason, see our keep, rewrite or delete checklist for FAQ schema.
- Bolting on a generic FAQ. An FAQ of six invented questions with 50-word answers adds repetition, not coverage. Add questions only when they come from real users and the body does not answer them.
- Generating many near-identical pages for long-tail questions. Using automation to produce many pages without added value can fall under Google’s scaled content abuse policy.
- Inventing statistics or expert quotes. They may be quoted back by an AI answer with your name on them, which is worse than not being cited.
- Hiding text for bots. Content shown to crawlers but not to people is cloaking, and violates Google’s spam policies.
How to tell whether it is working
Measuring AI visibility is possible but imperfect, because neither Google nor OpenAI gives site owners a full citation report. Combine three signals and compare them over weeks, not days.
Google: Search Console plus engagement
Clicks from AI Overviews and AI Mode are counted in the Search Console Performance report under the Web search type, together with other search traffic; there is no separate AI filter described in Google’s documentation. Watch the pages you reworked for changes in impressions and clicks on longer, question-style queries, and use your analytics tool to check engagement and conversions from Google organic traffic to those pages.
ChatGPT: the referral parameter
OpenAI adds utm_source=chatgpt.com to links from ChatGPT search, so in Google Analytics 4 you can filter sessions by that source. Create a simple exploration showing landing pages for that source, and note which pages receive visits. If the number is zero across the whole site for weeks, check robots.txt for OAI-SearchBot before anything else.
Manual spot checks
Keep a short list of 10 to 20 questions your pages should answer, and check them monthly in Google Search and ChatGPT. Record whether your site is cited and which page. Results vary by location, account and phrasing, so treat this as a sanity check, not a metric. Third-party AI visibility tools sample prompts at larger scale, but their numbers are estimates from their own prompt sets, not platform data.
A short checklist per page
- Googlebot and OAI-SearchBot are allowed; the page is indexed and not set to nosnippet.
- The first 2-3 sentences answer the main question on their own.
- Each heading says what its section answers.
- Each section opens with a standalone sentence and avoids references to “above”.
- Claims have a source, a number or a concrete example, and none is invented.
- Time-sensitive facts were checked against current official pages.
- Important content is in text, not only in images or scripts.
- The page links to and from related pages on your site.
- You have a baseline in Search Console and GA4 before you change anything.
