SEO for AI answers: getting cited by ChatGPT, Gemini and Perplexity

In short

• There is no "ranking in ChatGPT" you can buy or force. AI assistants run a web search, read a few pages and cite the ones that answer best. Being findable and quotable is still classic SEO, done with more rigour.
• Google says it in black and white: no special optimisation is needed for AI Overviews and AI Mode. A page must be indexed and eligible for a snippet. That's it.
• AI crawlers don't all do the same job. Blocking GPTBot (training) does not remove you from ChatGPT search; blocking OAI-SearchBot does. Most sites that "vanish" from AI answers blocked themselves without knowing it.
• llms.txt is a useful but marginal proposal: Google has said it doesn't use it, and no major answer engine has publicly committed to using it to pick sources. It costs little and works no miracles.
• What makes the difference: clear, verifiable answers, data nobody else has, an identifiable author, a consistent brand across the web, and pages crawlers can read without running JavaScript.
• Arabic content is a real opportunity: fewer quality sources means better odds of being the page that gets cited.

Why this guide, and why now

For two years my clients have all asked the same question, in French, English or Arabic: "How do I get ChatGPT to mention me?" And for two years they've been sold magic answers: "GEO audits" costing thousands, miracle files, "AI visibility scores" down to the decimal.

The reality is both simpler and more demanding. Answer engines (ChatGPT search, Google AI Overviews and AI Mode, Gemini, Perplexity, Claude, Copilot) nearly all rely on real-time web search. They query an index, fetch a handful of pages, extract passages and write an answer that cites its sources. If your page isn't in the index, if it's blocked, or if it has no clear passage that answers the question, you won't be cited. If it is, you're in the running.

GEO (Generative Engine Optimization), AEO (Answer Engine Optimization), LLMO: the acronyms change, the substance doesn't. This guide sorts it out, with a level of evidence for every claim.

Proven, likely and hype

Before any tactic, here is the grid I use with clients. It stops you spending time and money on rumours.

My rule: I prioritise only what is proven or likely and also improves classic SEO. That way nothing is wasted, whatever happens to answer engines.

How each engine picks its sources

Google: AI Overviews and AI Mode

Google uses its own search index. The official documentation is clear: there are no additional requirements to appear in AI Overviews or AI Mode, no special file to create, no particular markup. The page must be indexed and able to show with a snippet. Clicks from these features are counted in Search Console's Performance report under the "Web" search type; they are not broken out into a separate report.

Practical consequence: your Google SEO is your SEO for Google's AI answers. A slow, poorly indexed or thin site won't be cited, whatever its "GEO strategy".

Gemini

Gemini can rely on Google Search to ground its answers (what Google calls grounding). The Google-Extended token lets you refuse that your content be used to train future Gemini models and for grounding in some Gemini systems, without touching your presence in Google Search. It's an editorial decision, not an SEO one.

ChatGPT search

ChatGPT has its own search crawler, OAI-SearchBot, which is used to surface sites in its answers. OpenAI also says it works with third-party search providers, without publishing the list. In practice, being well indexed in Bing remains a sensible precaution, but OpenAI hasn't officially confirmed it: I rate it "likely". When a user explicitly asks ChatGPT to read a page, ChatGPT-User visits it.

Perplexity

Perplexity uses PerplexityBot to surface and link websites in its results; the company states it is not used to train foundation models. Perplexity-User visits a page at a user's request and, according to Perplexity's documentation, generally ignores robots.txt because a person asked for the fetch. In 2025 Cloudflare accused Perplexity of getting around some blocks, which the company disputed: if this matters to you, watch your logs rather than rely on a single statement.

Claude

Anthropic separates three bots: ClaudeBot (collection that may be used for training), Claude-SearchBot (indexing to improve search result quality) and Claude-User (fetching a page when a user asks a question that needs it). Blocking the last two reduces your visibility in Claude's answers.

Copilot and the rest

Microsoft Copilot relies on Bing. Apple uses Applebot for Spotlight, Siri and Safari; the Applebot-Extended token doesn't visit any page, it only lets you refuse that your content train Apple's models, and pages that refuse it stay in search results.

AI crawlers: who does what

This is the most misunderstood part, and the number one cause of invisible sites. Separate three families:

1. Training: collecting content to train future models. Blocking has no effect on being cited in answers (unless stated otherwise).
2. Search and citation: indexing to surface your site in answers. Block it = disappear from that engine's answers.
3. User request: a one-off visit when someone asks the assistant to read a page. These bots may not follow robots.txt, since a human triggered the action.

Careful with Googlebot: it feeds Google Search and its AI features. There is no separate Google crawler you could block to leave AI Overviews while staying in Google. To limit what gets reused, Google offers nosnippet, data-nosnippet and max-snippet, at a cost to your classic listings.

Which robots.txt strategy for which goal

There's no universal right answer. There's a decision to make, knowingly.

Example for the most common goal among my clients, "be cited without feeding training":

# Search and citation: allowed
User-agent: OAI-SearchBot
Allow: /

User-agent: Claude-SearchBot
Allow: /

User-agent: PerplexityBot
Allow: /

# Training: refused
User-agent: GPTBot
Disallow: /

User-agent: ClaudeBot
Disallow: /

User-agent: CCBot
Disallow: /

User-agent: Google-Extended
Disallow: /

User-agent: Applebot-Extended
Disallow: /

# Everyone else, including Googlebot and Bingbot
User-agent: *
Allow: /

Sitemap: https://www.example.com/sitemap.xml

Three precautions:

• Check your firewall and CDN too. Cloudflare, for example, offers one-click blocking of AI crawlers, and turned it on by default for new domains in 2025. A perfect robots.txt is useless if the CDN returns a 403 first.
• Never block Googlebot or Bingbot "to protect yourself from AI": you would leave classic search engines, and therefore most AI answers.
• robots.txt is an instruction, not a lock. Serious crawlers respect it; truly confidential content belongs behind a login.

To generate this file without mistakes, with the right crawler names and a matching llms.txt, use my llms.txt and AI robots.txt generator.

llms.txt: what it is, and what it's really worth

The proposal

llms.txt was proposed by Jeremy Howard in September 2024 (official site: llmstxt.org). It's a Markdown file at the site root, /llms.txt, that gives language models a clear map of your important content. Its structure:

• an # heading with the site name (the only required element);
• a > quote with a short summary;
• optional context paragraphs;
• ## sections containing link lists, with an optional note after each link;
• by convention, an ## Optional section for secondary links.

The proposal also suggests publishing a clean Markdown version of important pages (for example page.html.md).

# OUFRID

> Web developer, SEO expert and Android apps. Services in French, English and Arabic.

## Services
- [SEO](https://oufrid.com/services/seo.en.html): audits, technical SEO, multilingual content
- [Android apps](https://oufrid.com/services/android.en.html): design and Google Play publishing

## Guides
- [SEO for AI answers](https://oufrid.com/ia/seo-ai-answers.en.html): what's proven, likely or speculative

## Optional
- [Glossary](https://oufrid.com/glossary.en.html)

Who actually reads it

• Google: no. Gary Illyes said Google doesn't support llms.txt and isn't planning to; John Mueller compared it to the keywords meta tag. Google's documentation also states that no "AI text file" is needed for its AI features.
• Other answer engines: no major player has publicly announced using it to choose sources. Some site owners see certain crawlers fetch the file in their logs, but a fetch is not proof it's used for ranking or citation.
• Developer tools and agents: this is where the file makes the most sense today. For technical documentation, an API or a software product, an llms.txt helps coding assistants and agents find the right page without crawling the whole site.

My cost/benefit verdict

• Cost: thirty minutes, then an update when your structure changes.
• Proven benefit: low for SEO and citations; real for technical docs read by agents.
• Risk: none, as long as you don't sell it to a client as a visibility lever.

So I install it on documentation and technical product sites, and offer it as an option elsewhere. I never bill it as "optimisation for ChatGPT".

What really helps you get cited

An answer engine doesn't cite a site, it cites a passage. Your goal: every important page should contain passages that can be extracted, understood on their own and verified.

1. Clear answers at the top of each section

• Open every section with a sentence that directly answers the question, then expand.
• A definition should fit in one or two sentences and stand on its own.
• Phrase headings like people's real questions: "How much does… cost", "What's the difference between…", "Should I…".

Before: "In a world where digital is taking an ever greater place, the question of a website's price arises…"

After: "A professional showcase website usually costs between X and Y in Morocco in 2026, depending on the number of pages, languages and features. Here's the breakdown."

2. Sourced figures and original data

Assistants prefer verifiable claims. Cite your sources, date your figures. Better still: publish data nobody else has: real prices observed on your projects, the results of a test you ran, a survey of your clients, speed measurements on a sample of sites in your sector. Original data is the best reason to be cited, because it can't be found anywhere else.

3. Tables, lists and FAQs

Comparison tables, numbered steps and FAQs produce exactly the kind of short, structured passages that systems extract easily. It's not a trick for robots: it's also what busy readers prefer.

4. An identifiable author and demonstrated expertise

Sign your content, link the author to a real "About" page, show your work, certifications, years of experience. This is the heart of what Google calls E-E-A-T (experience, expertise, authoritativeness, trustworthiness). Answer engines have every reason to cite reliable sources: make their job easy.

5. A clear, consistent entity everywhere

Your name, activity, city, contact details and services must be identical on your site, Google Business Profile, LinkedIn, directories and social networks. A consistent entity is easier to recognise and recommend. Mentions of your brand on third-party sites (articles, interviews, customer reviews, forum discussions) strengthen that recognition.

6. Freshness

For fast-moving topics (prices, regulation, tools, technologies), show a real update date and genuinely update the content. Changing the date without changing the text doesn't fool anyone for long.

7. Clean structured data

Google requires no special markup for its AI features, but valid structured data (Organization, Person, Article, Product, LocalBusiness, BreadcrumbList) helps every engine understand who you are and what the page is about. It must match the visible content. Don't add FAQ markup the user can't see.

8. Pages readable without JavaScript

Many AI crawlers fetch raw HTML and don't run JavaScript, or run it poorly. If your text only appears after a client-side React or Vue app loads, some crawlers will see an empty page. Prefer server-side rendering or static generation for important content. Simple test: view the page source (not the inspector) and look for your text.

9. Speed and availability

A crawler that hits a slow page, a 5xx error or a timeout moves on to the next source. Speed won't get you cited, but slowness can get you skipped.

10. Bing Webmaster Tools and IndexNow

Register your site in Bing Webmaster Tools, submit your sitemap and enable IndexNow, the protocol that instantly notifies participating engines (Bing, Yandex, Naver, Seznam, Yep, Amazon) when a page is published or changed. Google isn't among the listed participants, but for Bing, and therefore Copilot, it's a concrete freshness gain for minimal effort: a key file at the root and a call on every publish.

Measuring: what you can see, and what you can't

Let's be honest: measurement is imperfect. Part of the traffic from assistants arrives with no referrer and lands in "direct". But you can still track a trend.

In your analytics tool

Create a segment or channel group for these referrers:

• chatgpt.com (and the old chat.openai.com)
• perplexity.ai
• gemini.google.com
• copilot.microsoft.com
• claude.ai

ChatGPT often adds the utm_source=chatgpt.com parameter to links it cites, which helps attribution. This tagging isn't systematic and its behaviour has varied over the months: treat it as a hint, not an exact counter.

In Search Console

Clicks from AI Overviews and AI Mode are included in the Performance report, "Web" search type. They aren't separated. Watch especially how impressions and click-through rate evolve on your informational queries: rising impressions with falling clicks often means the answer is given directly on the results page.

In your server logs

Your logs tell you which crawlers actually visit, how often, and whether they get errors. It's the best way to check that a CDN or firewall isn't blocking what your robots.txt allows.

The structured manual test

Once a month, ask ChatGPT, Perplexity, Gemini, Claude and Copilot a fixed list of 10 to 20 questions your customers really ask, in each of your languages. Note whether you're cited, which page, and who is cited instead. It's not statistically perfect (answers vary between sessions), but it's the most telling indicator for a business owner.

The Arabic and multilingual opportunity

This is the point I see least exploited, and the one that gets my clients in Morocco, the Middle East and the diaspora the best results.

• Less quality competition: on many professional topics there are few complete, structured, up-to-date pages in Arabic. When a user asks in Arabic, the assistant has fewer good sources to choose from.
• A real translation, not a raw one: have it reviewed by a native speaker, adapt examples, prices and references to the market. Unreviewed machine translation shows, and it doesn't inspire trust.
• Correct hreflang: each language version declares its equivalents (fr, en, ar, and ar-MA for Darija if you have it), plus an x-default. It helps engines serve the right language and stops your versions competing with each other.
• One URL per language, indexable, with its own title and description. No translation injected with JavaScript after load.
• Real vocabulary: in Arabic, people sometimes search in Modern Standard Arabic, sometimes in Darija, sometimes mixed with French. Study the real queries before writing.

30-day action plan

Week 1: stop being invisible

• Read your robots.txt line by line and check every AI search crawler.
• Check the firewall and CDN (AI crawler blocking settings, bot rules).
• Check indexing in Google Search Console and Bing Webmaster Tools; submit sitemaps.
• View the source of your 10 key pages and check the text is there.
• Decide your training policy (allow or refuse) and write it down.

Week 2: make your pages quotable

• Pick your 10 most business-critical pages.
• Rewrite the opening of each section as a direct answer.
• Add a table or list wherever the text compares or enumerates.
• Add a visible FAQ based on real customer questions.
• Show the author, the real update date and the sources of figures.

Week 3: strengthen entity and trust

• Align name, address, phone, description and services across the site, Google Business Profile, LinkedIn and directories.
• Complete Organization or LocalBusiness and Person structured data.
• Publish one piece of original data: a mini-barometer, a quantified case study, a tested comparison.
• Enable IndexNow.
• Optional: publish an llms.txt if you have documentation or a technical product.

Week 4: measure and iterate

• Create the "AI assistants" segment in your analytics.
• Run the monthly manual test (fixed question list, all your languages).
• Analyse server logs: which crawlers visit, with which response code.
• Identify questions where a competitor is cited instead of you, and improve the matching page.
• Plan the Arabic or English version of your best pages if it doesn't exist.

The checklist

• robots.txt: OAI-SearchBot, Claude-SearchBot, PerplexityBot allowed if you want to be cited.
• Googlebot and Bingbot never blocked.
• CDN and firewall checked (no 403 for allowed crawlers).
• Training policy decided and applied (GPTBot, ClaudeBot, CCBot, Google-Extended, Applebot-Extended).
• Site verified in Google Search Console and Bing Webmaster Tools, sitemaps submitted.
• IndexNow active.
• Main content present in the HTML source, not dependent on JavaScript.
• Every key page opens with a direct answer.
• Figures sourced and dated; at least one piece of original data published.
• Tables, lists and FAQs visible where useful.
• Identifiable author, real "About" page.
• Consistent entity across the web.
• Valid structured data matching visible content.
• Correct hreflang for each language, one URL per language.
• "AI assistants" segment and monthly manual test in place.

For the technical foundation (speed, security, indexing, Arabic SEO), add the complete website launch checklist.

The mistakes I see most often

• Blocking search crawlers while thinking you're blocking "AI". A Disallow: / on OAI-SearchBot or PerplexityBot, or a "block AI bots" toggle switched on in the CDN, and you vanish from answers. Decide crawler by crawler.
• Blocking Googlebot to leave AI Overviews. You leave Google entirely.
• Generating hundreds of AI pages with no review or added value. They all look alike, bring no new data and have no reason to be cited. They can also drag down the perceived quality of the whole site.
• Believing promises of "ranking in ChatGPT". There is no fixed position: the answer depends on the question, language, location, conversation and timing. Beware anyone selling a guaranteed spot.
• Changing the date without changing the content. Systems compare versions; so do readers.
• Hiding content behind JavaScript, unrendered tabs or pop-ups.
• Machine translating without review, especially into Arabic.
• Measuring nothing, and so never knowing what worked.

FAQ

Do I need an llms.txt file to appear in ChatGPT?

No. There's no public sign that a major answer engine uses it to choose sources, and Google has said it doesn't support it. It can be useful for technical docs read by agents. To be cited, rely first on indexing, robots.txt and the quality of your passages.

If I block GPTBot, do I disappear from ChatGPT?

No. GPTBot is for training. For ChatGPT search, OAI-SearchBot is what matters: OpenAI states that sites blocking it are not shown in ChatGPT search answers.

Does blocking Google-Extended cost me rankings in Google?

No. Google states that Google-Extended has no effect on inclusion or ranking in Google Search. It concerns Gemini training and some grounding uses.

Can I leave AI Overviews while staying in Google?

Not with a dedicated crawler: AI Overviews relies on Googlebot. You can limit what's reused with nosnippet, data-nosnippet or max-snippet, but these directives also reduce your snippets in classic results.

How do I know if ChatGPT or Perplexity send me visitors?

Filter the referrers chatgpt.com, perplexity.ai, gemini.google.com, copilot.microsoft.com and claude.ai, and the utm_source=chatgpt.com parameter. Part of the traffic will stay invisible (counted as "direct"), hence the value of the monthly manual test.

Does GEO replace SEO?

No. Answer engines rely on search indexes. GEO is SEO applied with higher standards on passage clarity, sources and entity. A poorly optimised site won't be cited.

Is writing in Arabic really worth the effort?

On most professional topics, yes: there are fewer complete, up-to-date Arabic sources, so better odds of being the page chosen, provided the writing or review is native and hreflang is correct.

---

Want to know where your site stands? Start with the free SEO audit, generate a suitable robots.txt with the llms.txt and AI robots tool, or hand me the whole project through my SEO service, in English, French or Arabic.