How to rank in AI search: how ChatGPT, Google, Perplexity and Claude pick their sources.

1. What it means to rank in AI search.

AI search is a group of products, not one engine: ChatGPT search, Google's AI Overviews and AI Mode, Perplexity, Claude, Microsoft Copilot and the Gemini app. None of them shows ten ranked links. Each one writes an answer and attaches a handful of sources. So when people ask how to rank in AI, the honest answer is that there is no position to hold. There is a short list of cited pages per answer, and the job is to be on it for the questions your buyers ask. People call this work by several names, and our page on what SEO for AI is called sorts them out. We call it AI search optimization.

Not every prompt even runs a search. Semrush analyzed more than a billion lines of US clickstream data and found that ChatGPT searched the web on 34.5% of prompts in February 2026, down from 46% in late 2024. Google shows an AI Overview only when its systems decide the overview adds something to the normal results. When no search runs, the answer comes from what the model learned in training and no page gets a link. The prompts that do trigger a search are the ones you can win: specific, current, commercial questions where the model needs fresh evidence.

The engines also disagree with Google and with each other, which is the fact this whole guide rests on. In August 2025 Ahrefs ran 15,000 long-tail queries through Google, Bing and the main assistants. Only about 12% of the links ChatGPT, Gemini and Copilot cited ranked in Google's top 10 for the same prompt, and about 80% did not rank anywhere in Google for it. Perplexity was the outlier, with 28.6% of its citations in Google's top 10. Even Google's own two surfaces split: Ahrefs compared 540,000 query pairs and found AI Overviews and AI Mode cited the same URLs only 13.7% of the time. A Google rank report tells you very little about whether you are cited. Our AI SEO statistics page keeps these numbers current.

2. How each AI engine picks its sources.

Under the differences, every engine runs the same three steps. It turns the prompt into one or more searches. It pulls candidate pages from an index. Then the model writes the answer from passages on those pages and attaches a few of them as links. What changes from engine to engine is the index it searches and the crawler that fills it. That is the table to get right first.

Here is what each row means in practice, with the spoke guide that covers it in depth.

• ChatGPT search. ChatGPT's help pages say results come from outside search providers plus content from partners, name Bing among them, and say a prompt may be rewritten into one or more targeted searches. Its own crawler, OAI-SearchBot, decides whether your site can appear: the crawler documentation says a site that blocks it will not be shown in ChatGPT search answers, apart from navigational links. GPTBot is the training crawler and has nothing to do with it. Our guide to getting cited by ChatGPT covers how it rewrites prompts and which pages it picks.
• Google AI Overviews. The generated block above the normal results. Google's Search Central guidance says a page only has to be indexed and eligible to show a snippet, with no additional requirements. Google runs several related searches from one query (it calls this query fan-out), and Gemini writes the answer from what comes back. In Ahrefs' March 2026 data, 37.9% of cited pages ranked in the top 10 for the query. The 13-point AI Overviews guide walks through eligibility and selection.
• Google AI Mode. The separate conversational tab, which Google said at I/O 2026 had passed one billion monthly users. It reads the same index, but Google says AI Mode and AI Overviews may use different models and techniques, which is why their citations overlap so little. Read how AI Mode in Google Search picks pages before you assume an AI Overview citation carries over.
• Perplexity. Runs its own index. The launch post for its Search API in September 2025 described an index covering hundreds of billions of webpages, with documents split into small units that are scored one by one against the query. PerplexityBot is the crawler that gets you into it. Perplexity-User, which fetches a page when a person asks, generally ignores robots.txt by Perplexity's own description. The Perplexity ranking guide covers freshness and passage structure.
• Claude. Claude searches the web through an outside index, and its maker has not said which. The evidence points to Brave Search: Brave was added to the published list of subprocessors that handle Claude data on March 19, 2025, TechCrunch reported the change two days later, and when developer Simon Willison ran a test search, the ten results Claude cited matched Brave's results for the same query exactly. Claude-SearchBot crawls for search quality, and Claude-User fetches pages when a person asks. Our ClaudeBot guide separates all three bots.
• Microsoft Copilot. Grounds its answers in Bing. In the same Ahrefs study, Copilot had the highest overlap with Bing's top 10, which Ahrefs put down to its use of Bing results. If Bing has not indexed a page, Copilot has nothing to cite.
• Gemini app. Grounds some answers in Google Search and shows a Sources button or inline links when it does. One trap: Google's crawler documentation says the Google-Extended token controls both Gemini training and grounding in the Gemini app. Blocking it to opt out of training also removes you from grounded Gemini answers, while AI Overviews and AI Mode are not affected.

Two consequences follow. First, five different crawlers have to be allowed through for a site to be eligible everywhere, and each has a training sibling with a similar name that you may want to block. Our list of every AI crawler, sorted by job has the user agents and IP lists. Second, at least four separate indexes sit behind these answers (Google's, Bing's, Perplexity's and the one Claude uses), plus ChatGPT's own. Being well indexed in one is no guarantee in the others.

3. AI search ranking: what every engine rewards.

Nobody outside these companies has the ranking weights, and anyone quoting exact percentages for how one engine weighs freshness or authority is guessing. What exists is each company's documentation plus independent studies of what gets cited. Across engines they point the same way, at six properties.

• The page can be read as text. The crawler has to be allowed, and the answer has to be in the HTML. Vercel studied AI crawler traffic across its network in December 2024 and found that none of the major AI crawlers render JavaScript. A price, a spec or an answer that only appears after a script runs is invisible to them.
• The answer comes first. Engines lift passages, not pages. Perplexity scores small units of a document one by one, and Google quotes passages in its overviews. A heading phrased like the question with two or three sentences of answer under it is the easiest thing to lift. That is the core of answer engine optimization.
• The facts are specific and sourced. The GEO benchmark presented at the KDD conference in 2024 tested edits against generative engines and found that adding statistics, quotations and cited sources raised visibility in the answers by up to 40%. The generative engine optimization guide walks through that study.
• The page is current. Ahrefs looked at about 17 million citations in July 2025 and found AI assistants cite pages that are 25.7% fresher than the pages in organic results. The same study warned against changing dates without changing the content.
• The sub-questions are covered. One prompt becomes several searches, so the definition, the how to, the cost, the comparison and the benchmark are each a separate chance to be cited. Give each one its own section, or its own page when it deserves one, and link them together.
• Other sites name you. Ahrefs studied 75,000 brands and found branded web mentions had the strongest correlation with visibility in AI Overviews (0.664), while the number of backlinks came in at 0.218. That is correlation, not proof, but it says that what the rest of the web writes about you matters more than who links to you.

None of this is a trick, and none of it is specific to one engine. It describes a good reference page on a site the web talks about. The broader discipline of writing for machine readers is what we mean by LLM SEO, and it is why GEO and SEO share most of their foundations. The hard part is volume. Every buyer question fans out into five or six sub-questions, each needs its own direct answer with a source behind it, and all of it has to stay true month after month.

4. Where the engines differ, and the fix each needs.

The page work is shared. The access work is not. Each engine has one thing that quietly decides whether your pages can be considered at all, and it is a different thing each time.

• ChatGPT: Bing and OAI-SearchBot. Verify the site in Bing Webmaster Tools and check that your key pages are in Bing's index, since Bing is the provider ChatGPT names. Allow OAI-SearchBot in robots.txt and at the CDN. The crawler documentation says a robots.txt change takes about 24 hours to register. Our breakdown of the four ChatGPT crawlers shows which one to block and which to keep.
• Google AI Overviews and AI Mode: Googlebot. Nothing else. Google-Extended does not affect them, and the opt-out is a snippet control (nosnippet, max-snippet, noindex) or the Search Console setting for generative AI. Run the crawl and indexing checks in our technical audit checklist and you have covered Google.
• Perplexity: PerplexityBot and freshness. Allow the crawler and its published IP ranges at the firewall. Perplexity cites communities heavily: the Tinuiti Q1 2026 AI citation report, built on Profound data, found Reddit was 24% of Perplexity's citations on commercial prompts in January 2026. That is why search everywhere optimization treats forums and video as part of the same job.
• Claude: Claude-SearchBot and Brave. Allow Claude-SearchBot. ClaudeBot is the training crawler, and blocking it does not remove you from Claude's answers. Because the index behind Claude is not Google's or Bing's, search for your key pages in Brave Search and check they show up.
• Copilot: Bing. Same fix as the ChatGPT one, which makes Bing Webmaster Tools the cheapest hour in this whole guide: it covers one engine fully and another in part.
• Gemini app: Google-Extended. Decide it on purpose. Blocking the token keeps your content out of Gemini training and out of grounded answers in the Gemini app. Allowing it does neither harm nor good for AI Overviews.

Put together, a robots.txt that keeps a site in every AI search engine while opting out of the two best-known training crawlers looks like this. Check the CDN as well: Cloudflare started blocking known AI crawlers by default on new domains in July 2025, and its bot settings now treat search, agent and training crawlers separately.

# robots.txt
# AI search crawlers: allow, or you cannot be cited
User-agent: OAI-SearchBot
User-agent: PerplexityBot
User-agent: Claude-SearchBot
User-agent: Googlebot
User-agent: Bingbot
Allow: /

# Training crawlers: a separate decision.
# Blocking these does not remove you from AI search.
User-agent: GPTBot
User-agent: ClaudeBot
Disallow: /

# Google-Extended: opts out of Gemini training AND of
# grounding in the Gemini app. No effect on AI Overviews.
# User-agent: Google-Extended
# Disallow: /

Which engine first? Start with Google, because AI Overviews and AI Mode read the index you already compete in and the only requirement is indexed and snippet-eligible. Do Bing next, because one verification covers Copilot and part of ChatGPT. Then open the crawlers for Perplexity and Claude. After that every hour goes into the pages, which help all of them at once. If you are not sure which engines your buyers use, our comparison of the best AI search engines shows who each one is built for.

5. The AI search checklist: 12 checks for every engine.

Twelve checks, in the order to do them. Access comes first because nothing else counts until the engine can read and index the page. Every check applies to all six engines at once. For the classic search basics underneath them, use our full SEO checklist.

Two items are missing on purpose. There is no llms.txt file on the list, because no engine has said it reads one (our llms.txt guide has the evidence). And there is no special schema, because Google says none is needed for its AI features. Markup that matches the page is still worth having, and our guide to schema markup for AI search covers what it does and does not do.

6. How to measure AI search rankings.

Since 2026, two of the engines report on themselves. The rest you measure from your own analytics and by hand. Use four signals.

• Search Console's generative AI report. Impressions from AI Overviews and AI Mode, grouped by page, country, date and device. Google's help page says it reached all websites worldwide on August 31, 2026. It shows no clicks, and it can be empty for a site with few AI impressions.
• Bing Webmaster Tools' AI Performance report. In public preview since February 2026. It counts how often your pages are cited in Microsoft Copilot, in Bing's AI-generated summaries and in some partner integrations, with the cited pages and the grounding queries behind them. It shows no clicks either, and it does not split out the partners.
• AI referrals in your analytics. Visits from the assistants' domains show up as referral traffic, and ChatGPT adds utm_source=chatgpt.com to the links it shows. This is the only signal of the four that counts real visits.
• A fixed prompt list. Write down 20 prompts your buyers would type, run each one in every engine you care about once a month, signed out, and record whether you are cited. Answers vary from run to run, so follow the trend across months, not one result. The AI visibility tools we reviewed automate this across ChatGPT, Perplexity and Google, and the AI Overview trackers go deeper on Google alone.

Expect the list to move under you. Green Flag Digital tracked the US finance pages ChatGPT cited most and found that 107 of the 215 top pages from October 2025 had no citations at all by September 2026. In Google, the share of AI Overview citations going to top 10 pages roughly halved between two Ahrefs studies eight months apart. A page holds its place only while it stays the best short answer to one of the searches. Plan for a steady supply of current pages, not a one-off rewrite.

A worked example: 3 citations to 27 in 60 days.

Take a fictional company, Tallowmere, that sells invoicing software to freelance designers. The company, its site (on the reserved .example domain) and every number below are invented to show the method. The mechanics match the documentation and studies above.

Tallowmere wrote down 20 prompts its buyers type, such as “best invoicing app for freelance designers” and “how do I invoice a client in another country”, and ran each one in five engines: ChatGPT, Google AI Overviews, Google AI Mode, Perplexity and Copilot. That is 100 checks.

The fixes ran in the same order as the checklist: access first, then indexes, then pages, then mentions.

1. Day 1, access. Allowed the AI search crawlers at the CDN, kept GPTBot and ClaudeBot blocked as a training choice, and confirmed 200 responses in the server logs.
2. Day 2, Bing. Verified the site in Bing Webmaster Tools and submitted the sitemap.
3. Days 3 to 10, existing pages. Rewrote eight pages answer-first, turned the pricing image into an HTML table, and added dated figures from Tallowmere's own invoice data.
4. Days 11 to 40, new pages. Published 12 pages, one per sub-question the 20 prompts fanned out into, each linked from its main page and back.
5. Days 15 to 45, mentions. Sent each roundup author a short note with current pricing and a comparison. Two of the three added Tallowmere.

Look at where the gains came from. The two biggest blockers (a CDN rule and a missing Bing verification) took two days to fix and were invisible in every Google report Tallowmere already watched. The engines that had been locked out moved furthest. Note the shape of the result too: citations and impressions grew far faster than visits, which is normal for AI search. The access fix took two days. The pages took the other 58, and the prompt list keeps growing.

Common mistakes when trying to rank in AI search.

1. Treating AI search as ChatGPT. It gets the headlines, so it gets the whole budget. But Google's two surfaces read a different index, and Perplexity, Claude and Copilot each read another. Map where your buyers ask before you decide where to spend.
2. Blocking the search crawler while meaning to block training. The names are close (GPTBot and OAI-SearchBot, ClaudeBot and Claude-SearchBot), and a blanket AI-bot rule catches both. Block training by name and keep the search crawlers allowed.
3. Ignoring Bing. It has a small share of classic search, so teams skip it. But Copilot runs on it and ChatGPT names it as a provider. One verification covers both.
4. Buying a file or a schema fix. It is quick to ship and easy to invoice. Google says no AI text file or special schema is needed for its AI features, and no other engine has said it reads llms.txt. Spend the time on the pages.
5. Reading a Google rank as an AI rank. It feels safe because the page already wins on Google. But only about 12% of the links ChatGPT, Gemini and Copilot cite rank in Google's top 10. Check the answers themselves.
6. Reporting AI impressions as traffic. Both the Search Console and Bing reports count appearances, not clicks. Put them next to referral visits and signups before you call it a win.