ChatGPT vs Gemini: how each one finds your content, and what that means for your business

The question is not which tool, it is how each one finds a page

When this article first went up, the advice of the day was to pick a side. ChatGPT for the technical crowd, Gemini for the Google crowd, and a content strategy split down the middle. It sounded tidy. Most of it was guesswork about who uses what, and I have cut that guesswork from this version because I could not find a source for any of it.

What can be sourced is more useful anyway. OpenAI and Google both publish how their tools go and fetch pages from the web, which crawler does the fetching, what a page needs to be eligible, and how the sources show up in the answer. Once you know the two paths, the question "where should I optimise" mostly answers itself. You make one good page, you make sure both paths are allowed to reach it, and you check, on a fixed schedule, whether it is being cited.

This article does three things. It traces how ChatGPT search and Google's AI features (AI Overviews, AI Mode and the Gemini app) find and cite a page, with a diagram of each path. It sets out what the published figures for Ireland actually say, and where there is no Irish figure at all, from Coimisiún na Meán and the Reuters Institute rather than from a vendor's blog. And it shows a simple, honest way to measure your own visibility in both, with an invented Cork family law practice and a month by month count.

Two things it does not do, because sister articles already do them well. The vocabulary of SEO, AEO, GEO and AIO is in the complete guide to AI search optimisation. How to write a passage that an engine can lift is in winning visibility in the age of AEO and GEO. If ChatGPT is your only concern, how to rank in ChatGPT is the shorter read.

Worked example: the practice in this article

Shandon Family Law is an invented practice of six solicitors in Cork. It handles divorce and judicial separation, guardianship, custody and access, and financial and property settlements, and it sells a fixed fee first consultation at €195. The principal is Aoife. Nobody real, every number made up.

Aoife's problem arrived in a sentence from a new client: "I asked ChatGPT how long a divorce takes and it sent me to a firm in Limerick." She did not know whether that was bad luck or a pattern. She also did not know whether Gemini would have done the same, or whether the answer was even the same twice. By the end of this article she will have a way to find out, for €0 and about two hours a month.

How ChatGPT search finds and cites a page

ChatGPT can answer in two ways. It can answer from what the model learnt in training, which carries no live sources. Or it can search the web first and answer from what it finds, with links. OpenAI launched the second mode for everyone as ChatGPT search on 31 October 2024. In its announcement it says that ChatGPT can choose to respond with information from the web, that answers include links to sources such as news articles and blog posts, and that a Sources button under the reply opens a sidebar with the references.¹

Where do the results come from? OpenAI's own wording is that ChatGPT search draws on third party search providers, as well as content provided directly by its partners.¹ The announcement does not name the providers. The old version of this article said flatly that "ChatGPT uses Bing". I would not write that today on the strength of OpenAI's own page, which does not say it, so I have corrected it to what OpenAI does say.

Then there is OpenAI's own crawler. OpenAI runs three different bots, and the difference between them matters more than anything else in this section.²

• OAI-SearchBot is for search. OpenAI states that sites opted out of it will not be shown in ChatGPT search answers, though they can still appear as plain navigational links.
• GPTBot gathers content that may be used to train OpenAI's foundation models. Disallowing it says your content should not be used for training. It does not, on OpenAI's description, remove you from search.
• ChatGPT-User visits a page when a person in ChatGPT asks for it, for example by pasting a link. Because a person started it, OpenAI says robots.txt rules may not apply, and it is not used to decide what appears in search.

So there are two separate switches. One says "you may show me in answers". The other says "you may train on me". A good few sites have flipped both off by accident, usually with a robots.txt line written in a hurry in 2023 that blocked anything with "GPT" in its name.

Worked example: Aoife's robots.txt

Aoife asks her web developer to send over the practice's robots.txt. It contains this:

User-agent: GPTBot
Disallow: /

User-agent: OAI-SearchBot
Disallow: /

Step one: read the first block. It refuses the training crawler. That is a perfectly reasonable business choice. A solicitors' practice may not want its guides used to train a model, and nothing in OpenAI's documentation says this costs it a place in search answers.²

Step two: read the second block. It refuses the search crawler. On OpenAI's own description, that means Shandon's pages will not be shown in ChatGPT search answers.² This is very likely why the client was sent to Limerick. The Limerick firm was not better. It was simply allowed in.

Step three: decide on purpose. Aoife keeps the training refusal and removes the search refusal. The file becomes:

User-agent: GPTBot
Disallow: /

User-agent: OAI-SearchBot
Allow: /

Step four: write the decision down, with the date. Next year, when someone asks why GPTBot is blocked, the answer is a sentence, not an archaeology project.

How Google's AI features and Gemini find and cite a page

Google has three products here that people lump together as "Gemini". It helps to pull them apart.

AI Overviews are the AI written summary at the top of an ordinary Google results page. AI Mode is a separate tab for longer, conversational questions. Google extended it on 7 October 2025 to over 200 countries and territories, and its help page for AI Mode lists Ireland among them.³ Google says both run on a custom version of Gemini 2.5.⁴ The third product is the Gemini app itself, the chatbot, which is a different place again.

For AI Overviews and AI Mode, Google's documentation is unusually direct. Both may use a "query fan out" technique, issuing multiple related searches across subtopics and data sources before writing the answer.⁵ To be shown as a supporting link, a page must be indexed and eligible to be shown in Google Search with a snippet. Google adds that there are no additional requirements, no special optimisations, and no special schema.org structured data you need to add.⁵ The controls are the ordinary search ones: nosnippet, data-nosnippet, max-snippet and noindex. And visits from these features are counted inside your normal search traffic in Search Console, under the Web search type, not in a separate line.⁵

The Gemini app and the Gemini models that developers use work in a similar way, which Google documents for its developer interface as grounding with Google Search. The steps it lists are: the prompt arrives, the model decides whether a Google Search would improve the answer, if so it writes one or more queries and runs them, it reads and synthesises the results, and it returns an answer with inline citations that link parts of the answer to their sources.⁶

One switch here behaves differently from what most people expect. Google's crawler documentation describes a control token called Google-Extended. It covers whether your content may be used to train future Gemini models, and whether it may be used for grounding in the Gemini app and in grounding with Google Search on Vertex AI. Google is explicit that it does not affect a site's inclusion in Google Search and is not a ranking signal there.⁷ So blocking Google-Extended can keep you out of the Gemini app's grounded answers while leaving you in AI Overviews and AI Mode, which are part of Search. That is the opposite of the intuition "block the AI bot, vanish from the AI".

Worked example: Aoife's fees page

Aoife's developer checks the Google side. The practice's pages are indexed, so the first condition is met. Then he finds this on the fees page, wrapped around the paragraph that gives the €195 consultation price and the typical cost of an agreed divorce:

<div data-nosnippet>

Step one: what does it do? It tells Google not to use that text in a snippet.⁵ Step two: what does that mean now? The only passage on the site that answers "how much does a divorce solicitor cost in Cork" cannot be quoted in a snippet, and a page needs to be eligible for a snippet to be a supporting link in AI Overviews and AI Mode.⁵ Step three: the fix. Remove the attribute from that paragraph, and keep it only on the one block that genuinely should never be quoted, a disclaimer about individual circumstances.

Step four: check Google-Extended. The robots.txt does not mention it. Aoife decides that the practice is grand with being used for grounding in the Gemini app, so it stays unmentioned. That is a decision now, not an accident.

The two paths, top to bottom

Here are both paths, each from the question to the cited answer. They are drawn separately because the switches are separate.

Read each flow from the top. Both start with a question and both end with an answer carrying links. In between, each has one gate that you control: for ChatGPT it is whether OAI-SearchBot may visit, for Google it is whether the page is indexed and may show a snippet, with Google-Extended as an extra switch that only touches the Gemini app. The takeaway is that the two paths look alike from the outside and are opened by different keys. Neither path asks for special markup or a special kind of page.

Worked example: one question through both paths

A prospective client in Douglas types "how long does a divorce take in Ireland" into both tools.

In ChatGPT, step one, the tool decides the question benefits from a search, because it is about current law and timescales. Step two, it searches through its providers. Step three, it reaches the gate. Before the robots.txt fix, Shandon's guide fails here and cannot be shown. After the fix, it passes. Step four, the model writes the answer from the pages it read and lists them under Sources. Whether Shandon is one of them now depends on the page itself, which is the subject of the sister article on passages.

In Google, step one, the question is split into searches such as "how long do you have to live apart before a divorce in Ireland" and "how long does the Circuit Court take to grant a divorce". Step two, those run against the index. Step three, the gate: Shandon's guide is indexed and, since the fees page fix, nothing on it blocks a snippet. It passes. Step four, the answer is written and supporting links are attached.

Two different gates. Two different fixes. One page.

What the Irish figures say, and what they do not

The old version of this article had a section called "who uses what". It said ChatGPT dominates developers and B2B SaaS buyers and Gemini dominates local businesses and enterprise teams on Google Workspace. I could not find a primary source for any of those splits, so the section is gone. Here is what the published figures do say for Ireland, and it is less than you might hope.

I could not find a published Irish figure for how many adults use ChatGPT or Gemini at all, for how many use Google Search in a month, or for how often Irish searches carry an AI Overview. Rather than borrow another country's numbers and pretend they are ours, I am leaving those gaps as gaps.

What does exist is about news. The Digital News Report Ireland 2026, published by Coimisiún na Meán with the analysis by DCU FuJo, surveyed 2,053 adults in January and February 2026.⁸ It found that 7% use AI chatbots for news each week, doubling to 14% among 18 to 24 year olds, against 5% of over 35s.⁸ Only 1% name an AI chatbot as their main source of news.⁸ In a repolled sample of 1,002 adults, 32% had come across news through search engines in the last week.⁸ The Reuters Institute's country page for Ireland reports the same 7%.⁹ The report's authors add a caution I agree with: people may not count the AI written summaries in a search engine as a "chatbot", so engagement with AI generated news may be much higher than 7% suggests.⁸

For the wider trend, the best I can offer is not Irish. The Reuters Institute's Generative AI and News Report 2026 surveyed nine countries, and Ireland is not one of them.¹⁰ Across the six countries it has tracked since 2024, weekly use of ChatGPT rose from 22% to 34% and of Gemini from 11% to 24%, and 57% of people across all nine said they had seen or engaged with AI generated answers in a search engine in the last week.¹⁰ That tells you the direction of travel. It does not tell you the Irish level.

Notice what these numbers measure, because they do not measure the same thing. The Irish 7% is the share who say, in a survey, that they used an AI chatbot for news in the last week. It is news only, and it leaves out anyone who met AI in a Google results page without thinking of it as a chatbot. The cross country figures are about any use, in countries that are not Ireland. A careful reader keeps the measure next to the number.

Worked example: what the figures mean for Aoife

Let me turn the Irish shares into people, so the sizes are easy to hold. This is arithmetic for scale, not a forecast for Cork, and it is about news, the only thing the Irish figures measure.

Take 1,000 Irish adults. In a given week, $1{,}000 \times 0.32 = 320$ of them come across news through a search engine and $1{,}000 \times 0.07 = 70$ use an AI chatbot for news.⁸ Among 1,000 people aged 18 to 24, the chatbot figure is $1{,}000 \times 0.14 = 140$.⁸

Now the caution from the report. Some of the 320 who meet news through search will have read an AI Overview at the top of the page and not counted it as using a chatbot.⁸ So the AI written answers that most people in Ireland see are very likely the ones already sitting in Google, met without ever opening a chatbot. There is no Irish figure that puts a number on that, and I will not invent one.

What Aoife takes from this is not "ignore ChatGPT". It is that the largest AI audience she has is the one already searching on Google, and that the chatbot audience is smaller but clearly growing, younger, and actively choosing to ask. Both are worth being open to. Neither is worth a separate content strategy.

What actually differs for your content, by kind of business

Here is the honest core of it. On the published documentation, both paths want the same thing from a page: that it can be fetched, that it is allowed to be quoted, and that it answers the question plainly. Google says no special optimisation is needed for its AI features.⁵ OpenAI's only stated requirement for search answers is not blocking its search crawler.² Everything past that is judgement, and below it is mine, labelled as mine.

Local services, such as a solicitor, a dentist or a plumber. In my view this is where the Google path matters most, simply because the Irish figures put far more people finding news through search engines than through AI chatbots,⁸ and local questions are typed into Google out of habit. What I check first: the pages are indexed, nothing blocks snippets, and every service page states the town, the price range and the timescale in plain sentences. I do not see a separate Gemini tactic here.

eCommerce shops. In my experience the product questions people take to a chatbot are comparisons ("which is better for a small kitchen") rather than "buy this". The pages that get cited in the shops I have worked on are buying guides and comparison pages written by a person, not the product pages. That is true on both paths, and it is my observation, not a documented rule.

SaaS and technical products. The old article claimed ChatGPT "dominates" technical queries. I have no source for that, so I will not repeat it. What I see in my own work is that technical buyers ask longer, more specific questions in both tools, and the pages that get cited answer one specific question each: a setup guide, a pricing explanation, an honest comparison with a named alternative.

Professional services, such as accountants, consultants and solicitors. In my experience the thing that most often keeps them out of AI answers is not the writing. It is a robots.txt or a snippet setting nobody remembers adding, exactly like Aoife's. Fix the gates first, then the pages.

The old article also suggested posting on Medium, Quora and Reddit so the models would "train" on you, and putting content in Google Drive to surface in Gemini. I have cut both. Training is not the same as being cited in a live answer, and neither OpenAI nor Google documents Drive content as a route into public answers.

Worked example: Aoife's priority list

Aoife has limited time, so she ranks the work by which gate it opens.

Step one: the ChatGPT search gate. Allow OAI-SearchBot. Ten minutes for her developer.

Step two: the Google snippet gate. Remove data-nosnippet from the fees paragraph. Ten minutes.

Step three: the pages themselves. Shandon has 14 guides. Aoife picks the three that answer the questions clients ask most often (how long, how much, what happens to the house) and asks one solicitor to rewrite each so that the question is the heading and the first sentence gives the answer, with the detail after it. The method is in the sister article on being the answer in AI search. Three guides at about two hours each is six hours of a solicitor's time.

Step four: measure, which is the next section. Without it, she will never know whether steps one to three did anything.

Measuring it: the same 30 questions, both tools, every month

Neither OpenAI nor Google gives you a report of how often your site is cited in AI answers. Google counts clicks from its AI features inside your ordinary Search Console figures, without separating them.⁵ In the analytics I look after, visits from ChatGPT show up as referrals from chatgpt.com, but a referral only tells you about the people who clicked, not about the answers where you were cited and nobody clicked.

So the honest way to measure is the plain one. Write a fixed list of questions your clients actually ask. Ask every one of them in each tool, once a month, in a clean session. Record whether your site is cited in the answer. Then count.

The number you track is a rate:

$$r = \frac{c}{q}$$

Here $r$ is the citation rate for one tool in one month, $c$ is the number of questions where the answer cited your site at least once, and $q$ is the number of questions asked, which stays fixed. Foundations: ratios and rates.

Keeping $q$ fixed is the whole trick. If you change the questions every month, you are measuring your choice of questions, not your visibility.

And because 30 questions is a small sample, the rate will wobble from month to month even if nothing has changed. The standard error tells you how much wobble to expect:

$$SE = \sqrt{\frac{r(1 - r)}{q}}$$

Here $SE$ is the standard error of the citation rate, $r$ is the rate you measured, and $q$ is the number of questions asked. A change smaller than about two standard errors is the kind of change chance alone produces often. Foundations: sampling.

Worked example: Shandon's first three months

Aoife writes 30 questions from the practice's own enquiry emails, such as "how long does a divorce take in Ireland", "do I need a solicitor to apply for custody and access" and "fixed fee divorce solicitor Cork". In the first week of each month, a trainee solicitor asks all 30 in ChatGPT with search on and in the Gemini app, each in a fresh chat, and ticks a spreadsheet column when Shandon's site appears among the sources. It takes about an hour per tool.

These are the invented counts. August is before any changes. The robots.txt and snippet fixes went live in the second week of August, and the three rewritten guides in September.

Along the bottom are the two tools and the three months. Up the side is the number of the 30 questions whose answer cited Shandon's site at least once. The takeaway: both tools cite the practice more often in October than in August, and the rise is steady rather than a single jump. Whether that is a real change or a lucky month is what the arithmetic below settles.

Step one: turn the counts into rates with $r = c / q$. ChatGPT in August is $6 / 30 = 0.200$ and in October $11 / 30 = 0.367$. The Gemini app in August is $9 / 30 = 0.300$ and in October $15 / 30 = 0.500$.

So ChatGPT went from 20.0% to 36.7%, a rise of 16.7 percentage points. The Gemini app went from 30.0% to 50.0%, a rise of 20.0 points.

Step two: the standard error of the October ChatGPT rate is $\sqrt{0.367 \times 0.633 / 30} = \sqrt{0.00774} = 0.088$, which is 8.8 percentage points. For the Gemini October rate it is $\sqrt{0.5 \times 0.5 / 30} = \sqrt{0.00833} = 0.091$, which is 9.1 points.

Step three: compare the rise with the wobble. The standard error of a difference between two months combines both months' errors. For ChatGPT, August to October, it is $\sqrt{0.200 \times 0.800 / 30 + 0.00774} = \sqrt{0.00533 + 0.00774} = \sqrt{0.01307} = 0.114$.

A rise of 16.7 points against a standard error of 11.4 points is about 1.5 standard errors. On its own, that is not enough to call it a real change. One more question cited moves the rate by 3.3 points, and asking the same question twice can bring back different sources. With 30 questions, the rule I use is that one month's move of under about 20 points proves nothing by itself.

Step four: read the pattern, not one pair of months. Both tools rose in both months after the fixes, the Gemini rise is 20.0 points against a difference error of about 12 points, and the timing matches the work. That is a reasonable basis for Aoife to keep going. It is not proof, and I would tell her so. Three more months on the same 30 questions will settle it either way.

Step five: decide what to do with the gap between the tools. In October the Gemini app cites Shandon in 15 answers and ChatGPT in 11. Four questions apart, out of 30, is well inside the wobble, so Aoife does not start a separate ChatGPT project. She keeps one set of pages and one list of questions, and lets the monthly count tell her if a real gap opens up.

What I would do on Monday

If I sat down with your site tomorrow, this is the order.

1. Read the robots.txt. Decide, on purpose, for OAI-SearchBot (search answers), GPTBot (training) and Google-Extended (Gemini training and grounding). Write the decisions down.
2. Check that the important pages are indexed and that nothing stops them showing a snippet. Look for nosnippet, data-nosnippet and max-snippet that nobody remembers adding.
3. Pick the three questions your clients ask most, and make sure one page answers each, plainly, near the top.
4. Write 30 fixed questions and count citations in both tools once a month. Treat any one month's move of under about 20 points as noise.
5. Do not build a separate strategy per tool until the counts show a gap that survives three months.

Less exciting than two strategies side by side. Also what the documentation supports, and it costs less to run than to argue about.

Footnotes

1. OpenAI, Introducing ChatGPT search, OpenAI, 31 October 2024, https://openai.com/index/introducing-chatgpt-search/

2. OpenAI, Overview of OpenAI crawlers, OpenAI developer documentation, https://developers.openai.com/api/docs/bots

3. Hema Budaraju, AI Mode is now available in more languages and locations around the world, Google, 7 October 2025, https://blog.google/products-and-platforms/products/search/ai-mode-expands-languages-locations/ (Ireland is listed on Google's help page, Get AI powered responses with AI Mode in Google Search, https://support.google.com/websearch/answer/16011537)

4. Elizabeth Reid, AI in Search: going beyond information to intelligence, Google, 20 May 2025, https://blog.google/products/search/google-search-ai-mode-update/

5. Google, AI features and your website, Google Search Central documentation, https://developers.google.com/search/docs/appearance/ai-features

6. Google, Grounding with Google Search, Gemini API documentation, https://aistudio.google.com/docs/google-search

7. Google, Google's common crawlers, Google Search Central documentation, https://developers.google.com/search/docs/crawling-indexing/google-common-crawlers

8. Dan Lloyd, Jane Suiter, Eileen Culloty, Callum Craig and Eirini Psychari, Digital News Report Ireland 2026, Coimisiún na Meán and DCU FuJo, June 2026 (online survey by YouGov and Dynata of 2,053 adults, January to February 2026, news discovery question repolled with 1,002 adults in March 2026), https://www.cnam.ie/app/uploads/2026/06/Digital_News_Report_2026.pdf

9. Jane Suiter, Ireland, Digital News Report 2026, Reuters Institute for the Study of Journalism, 16 June 2026, https://reutersinstitute.politics.ox.ac.uk/digital-news-report/2026/ireland

10. Felix Simon, Rasmus Kleis Nielsen and Richard Fletcher, Generative AI and News Report 2026: from standalone tools to information infrastructure, Reuters Institute for the Study of Journalism, 29 October 2026 (countries surveyed: Argentina, Australia, Denmark, Finland, France, Japan, Norway, the UK and the USA), https://reutersinstitute.politics.ox.ac.uk/generative-ai-and-news-report-2026-standalone-tools-information-infrastructure