LLM SEO: How to Get Mentioned in ChatGPT, Gemini and Claude

Decide first which path you can act on: training data, live search, or both. Live search is the path you can act on most directly today, because the vendors publish how their bots and search tools work.

For the wider difference between GEO and classic SEO, read the GEO vs SEO guide.

Yoast describes the two paths this way. Without retrieval, the model answers only from what it learned in training. With retrieval, the system gets content from live sources and passes it to the model with the query.

Ahrefs frames LLM search the same way: the model fetches new information instead of answering only from what it already knows.

OpenAI says its foundation models learn from public internet information, information from partners, and information from users, human trainers and researchers. Its models do not store copies of that data, and the same question can give different answers.

Mozilla Foundation reviewed 47 text models published from 2019 to October 2023. At least 64% of them, 30 models, used at least one filtered version of Common Crawl for pre-training. That study predates the current models, so use it to show how common web crawls are in training, not to predict today's models.

Anthropic's models overview lists a June 2026 training data cutoff and a June 2026 reliable knowledge cutoff for its current Claude models.

OpenAI's models page lists knowledge cutoffs that differ by model, from April 30, 2026 to June 1, 2026. Check the cutoff of the model you test, because the answer you see may come from a model with an older cutoff.

A model cannot draw on training data from after its cutoff, so a brand that launched after that date can reach an answer only through live search. Anthropic's web search tool gives Claude real-time web content beyond its knowledge cutoff, with citations for the sources it draws on.

ChatGPT searches the web when a question benefits from current information, and a user can start a search by hand. People who are not signed in can use web search too.

ChatGPT search rewrites the question into one or more targeted queries and sends them to search providers. It can send more specific queries after it reviews the first results. If memory is enabled, saved memories can change the rewritten query, so two users can get different sources for the same question.

OpenAI does not say which index ranks ChatGPT search results. Its help page says ChatGPT search sometimes partners with other search providers, and it links the Microsoft and Shopify privacy statements.

Microsoft says its grounding powers nearly every major AI assistant on the market. That is Microsoft's own claim about its position, so do not treat it as an independent measure of ChatGPT.

Claude searches the web when a topic benefits from current information, and every response with search includes citations. In the new Claude experience there is no web search toggle, so Claude decides when to search.

Anthropic's help page says Bing powers the images Claude selects from web results. It does not name the provider for text search, so do not assume one.

Google's Gemini API documentation says grounding with Google Search connects the model to real-time web content. The model decides whether a search can improve the answer, and it writes the search queries itself. That page covers the developer API, not the consumer Gemini app or AI Mode.

Google says AI Overviews and AI Mode may use query fan-out, which issues several related searches across subtopics and data sources.

Ahrefs reports that 14.4% of 66.6 million free, logged-out ChatGPT responses contained fan-out queries. That sample covers free logged-out use only, so it does not describe paid or signed-in use.

No vendor publishes which sites are in its training data or how often a brand appears in answers. The vendor pages describe source types and crawler controls only, so treat any claim that a specific site sits in a training set as unsupported.

Choose the path by the question type. Each assistant decides when to search, so treat these paths as likely rather than fixed.

For how other search ecosystems find and cite content, read the search ecosystems page. If you want prompts for SEO tasks rather than brand mentions, use the AI prompt library.

Set each bot's access on its own. Blocking a training bot and blocking a search bot have different effects, and each vendor says so in its crawler documentation.

OpenAI runs separate robots for separate jobs. You can allow OAI-SearchBot to appear in search while you disallow GPTBot to opt out of training.

OpenAI says a robots.txt update can take about 24 hours to take effect in its search systems. Wait that long before you judge a change.

To appear in ChatGPT search results, allow OAI-SearchBot and confirm that your host or CDN allows traffic from OpenAI's published searchbot IP addresses. Placement is not guaranteed.

OpenAI publishes a JSON file of IP ranges for each bot, such as searchbot.json, for allowlisting at a CDN or web application firewall.

Anthropic uses three robots, one for each job: model development, web search, and retrieval at a user's request.

Anthropic honors robots.txt and supports Crawl-delay. It also says blocking by IP address may not work, because that stops Anthropic from reading your robots.txt file.

A robots.txt block must be repeated on every subdomain you want to opt out.

Google-Extended is a control token, not a crawler. It has no HTTP user agent of its own, and Google crawls with its existing user agents.

Google says user-triggered fetchers generally ignore robots.txt rules, because a user requested the fetch.

Google also intends to support the AIPREF training-preference signal, with the user agent token __usage_ai-training.

Microsoft says Copilot and other experiences are powered by Bing's search index. Crawlability, metadata, internal linking and backlinks remain essential, but Microsoft calls them only the starting point.

Bing says it respects robots.txt and other content owner controls, and its AI Performance report shows your citations in Microsoft Copilot, AI-generated summaries in Bing and select partner integrations.

Common Crawl's CCBot can be blocked with a robots.txt rule. Common Crawl also warns that some crawlers falsely identify themselves as CCBot, and it publishes its IP ranges as JSON, so check the IP before you trust a log line.

For most sites, allow the search and fetch bots you want in answers, then set training bots by a separate policy. Confirm that your CDN lets the published IP ranges through, or the robots.txt allow rule never reaches the bot.

Build mentions across many independent sites before you add more pages to your own domain. The correlation data points that way, and it also shows where the evidence stops.

Ahrefs compared 75,000 brands with millions of AI responses from its Brand Radar tool, using a Spearman correlation.

YouTube mentions correlate most strongly with AI visibility, at about 0.737, across ChatGPT, AI Mode and AI Overviews.

In Ahrefs' brand visibility correlations study, branded web mentions correlate with AI visibility at 0.709 for AI Mode, 0.664 for ChatGPT and 0.656 for AI Overviews.

Your own page count matters less. Site pages correlate at about 0.194 with AI visibility, and Ahrefs says that is almost no relationship. Ahrefs concludes that what matters most is getting mentioned across a broad scope of sites.

ChatGPT shows weaker correlations with classic authority signals than AI Mode does. For ChatGPT, branded search volume correlates at 0.352 and domain rating at 0.266. For AI Mode, branded anchors reach 0.628.

Ahrefs says correlation is not causation. The study cannot show that raising these metrics raises AI visibility.

Mention gaps are common. Ahrefs' earlier study of AI Overviews found that roughly 26% of brands had zero AI Overview mentions.

Add Organization structured data to your home page, with a sameAs array that links your official profiles. Google says this helps it understand your administrative details and tell your organization apart in search results.

Google does not say that AI assistants read this markup. Treat it as disambiguation, not as a way to earn mentions.

Microsoft says assistants act more as retrievers, drawn to structured, verifiable and applicable content. It also advises that text, images and video describe the same entities, products and concepts.

If you run physical locations, Bing Places for Business helps keep address, hours and contact details current for AI answers on local queries.

Use the same brand name, product names and category labels across your site, profiles and listings. Mismatched names make it harder to connect mentions to one entity.

Pick the source types that your target platforms cite for your category, then earn mentions on those pages. Start from the source types the answers already cite.

Muck Rack, a PR software vendor, analyzed links from ChatGPT, Claude and Gemini responses across many industries. Earned media makes up most of the citations in its study, and paid and advertorial content makes up very little.

Muck Rack reports that ChatGPT leans on Wikipedia, Claude leans on PubMed Central and Gemini leans on Reddit. Query type changes what gets cited, too: industry trend questions drive journalism citations at more than double the rate of how-to questions.

Semrush's 2026 AI Visibility Index analyzed 126 million U.S. AI search prompts from January to April 2026. It counts an average of 15 sources per ChatGPT response and 3 per Gemini response. Counts differ between studies, so compare them only inside one study.

Semrush says AI platforms also rely on customer reviews, community discussions, independent publishers, retailers and industry sources to understand and recommend brands. Its example, Patagonia, has consistent descriptions across OutdoorGearLab, REI, Switchback Travel, GearJunkie and Reddit.

Category concentration differs. Semrush reports that the top three brands hold 82.9% of visibility in News and Media and 41.4% in Finance. In a category where three brands hold 41.4%, a newcomer has more room than in one where they hold 82.9%.

Peec AI's study of cited domains across ChatGPT, Google AI Mode, Gemini, Perplexity and AI Overviews lists Reddit, YouTube, LinkedIn and Wikipedia among the most cited domains.

AirOps, which sells off-site AI visibility software, reports that most brand mentions in early commercial discovery come from external domains.

Nearly 90% of third-party mentions come from listicles, comparison pages and review roundups.

Earn mentions where your category already gets cited: listicles, comparison pages, review roundups and community threads. Ask the owners of those pages to describe your product in the same terms your site uses. The off-page SEO guide covers outreach for those placements.

Measure mentions across repeated runs, track mentions and citations as separate numbers, and use first-party data for traffic. A single prompt check tells you almost nothing.

Ahrefs Brand Radar defines a mention as your brand name appearing in AI answer text, and a citation as an AI link to your pages or domains as a source.

OpenAI says the same question can yield different answers, and saved memories can change how ChatGPT rewrites a search query.

Run each tracked prompt this way:

1. Pick 20 to 50 prompts that a buyer would type, and group them by topic. Keep the same list for the whole measurement period.
2. Run each prompt on each platform 60 to 100 times. The SparkToro study used this range across ChatGPT, Claude and Google AI.
3. Record a baseline before you change anything, then score the percent of runs that name your brand.
4. Do not score the position of your brand in a list. The order is so random that two lists appearing in the same order takes about 1 in 1,000 runs.
5. Log how many options each answer draws from, because spaces with limited options repeat more often than spaces with many.
6. Read a sample of answers by hand, because the study did not collect how AI describes each brand or its sentiment.

Bing Webmaster Tools AI Performance reports total citations, average cited pages, grounding queries and page-level citation activity. Microsoft says the grounding data is a sample of overall citation activity, and that Total Citations does not show placement in an answer.

Citation Share is the percentage of citations attributed to your site out of all citations for the same grounding query. Microsoft calls it an observational metric, not a ranking.

Microsoft's Topics Portfolio, in beta, groups queries into topic clusters. Microsoft says to treat its numbers as directional signals for planning, not as a precise ledger.

Search Console counts AI Overviews and AI Mode in the Web search type, without a separate line.

OpenAI documents that ChatGPT referral traffic carries the utm_source=chatgpt.com parameter, and publishers who allow OAI-SearchBot can track it in analytics tools.

Semrush's survey of marketing leaders reports that 45% cannot accurately measure brand visibility in AI answers, and only 9% have tools to track all relevant metrics across platforms. Semrush does not state the survey's sample size.

A companion Semrush survey reports that 81% of organizations that fully integrate SEO and AI visibility saw more traffic or leads from AI platforms, against 36% of those that manage the two separately. The answers are self-reported.

A summary of Semrush's second-half 2026 index by Relevant Audience reports that fewer than half of the brands AI answers mention most often are also among the sites those answers cite most. The summary says the study excludes Claude, Perplexity and Copilot.

Use one tracker for trend lines, keep the prompt list fixed for the whole period, and treat vendor survey numbers as context, not as benchmarks.

Check access before you change any copy. Each failure stops a bot before it reads your page, so better writing does not fix any of them. Work the list in order.

1. Blanket AI bot blocks. Blocking only the training bot does not remove you from search. Blocking the search or user-fetch bots can remove you from answers.
2. CDN and WAF defaults. Cloudflare changed its default to block AI crawlers unless they pay creators. Check your own Cloudflare settings, because a default block can stop a search bot without any robots.txt change.
3. IP blocks on Anthropic bots. Do not block Anthropic bots by IP address, because that can stop Anthropic from reading your robots.txt file.
4. Subdomains. A robots.txt block applies only to the host where it sits. Repeat it on every subdomain you want to opt out.
5. Client-side rendering. Vercel's December 2024 analysis found that none of the major AI crawlers it tested rendered JavaScript. Gemini is the exception, because it uses Googlebot infrastructure that renders pages. ChatGPT and Claude crawlers fetch JavaScript files but cannot read client-side rendered content.
6. Hidden answers. Microsoft advises that AI systems may not render hidden content, so key details can be skipped. Keep key answers out of tabs and expandable menus.
7. Core facts in PDFs or images. Microsoft's guidance also says not to rely on PDFs or images for core facts. That is guidance, not a documented rule, so test it on your own pages.
8. Snippet eligibility. A page must be indexed and eligible to show in Google Search with a snippet to be a supporting link in AI Overviews or AI Mode. Google adds no other technical requirements.

Fix access and naming first, then measure, then earn mentions. The first 30 days decide whether the later measurements mean anything.

1. Days 1 to 14: audit access. List every AI bot rule in robots.txt and at your CDN. Compare each rule with the bot's job in the tables above, then allow about 24 hours after each change before you judge it.
2. Days 15 to 30: fix naming. Use one brand name, product names and category labels across your site, profiles and listings. Add Organization markup with a sameAs array to the home page.
3. Days 15 to 30: set the prompt list. Choose 20 to 50 buyer prompts, group them by topic, and record a baseline for each platform. Keep the list fixed for the full 90 days. At 60 to 100 runs per prompt that is 1,200 to 5,000 runs per platform, so use an API script or a tracker, not manual checks.
4. Days 31 to 60: earn third-party mentions. Target the listicles, comparison pages, review roundups and community threads that your category already cites.
5. Day 60: measure again. Run every tracked prompt the number of times set in the measurement steps, and compare the percent of runs that name your brand with the baseline.
6. Day 90: decide. Keep the prompts and sources where the mention rate moved, and drop the ones that did not. Then repeat the cycle on the prompts you still lack.

Skip new machine-readable files for this plan. Google says you do not need to create new machine readable files, AI text files or markup to appear in its AI features.

Put the work into pages that answer the prompts. The SEO content strategy page covers how to group topics into clusters. The SEO guide groups the classic SEO work by goal, from traffic drops to AI search.