In mid-2025, Claude was the fifth-largest AI referral source, at 1.35%. By April 2026, it was second, at 18.5%.
ChatGPT went the other way over the same window, from 89% of measurable AI referrals down to 62.6%.
Those are from Goodie's 2026 AI search traffic report, brand-averaged across March and April 2026 against a May-August 2025 baseline of 41 brand sites and 2.8 million referral sessions. Disclosure while it's early: Goodie AI is a company I lead, and every dataset I cite here is published with its methodology so you can check it.
The practical version: optimizing only for ChatGPT reaches roughly a third less of the AI landscape than it did in mid-2025. Claude stopped being a rounding error.
So the question clients actually bring me has changed. It used to be "how do I get into ChatGPT?" Now it's "do I need a second playbook for Claude?"
The short answer. Mostly no, and that surprised me. Getting cited by ChatGPT and Claude, and by AI search engines generally, takes the same three things: a crawler that can reach you, third parties describing you rather than you describing yourself, and content with something in it worth lifting. On the published evidence, these two models behave more alike than any other pair in the market. They pull from social at nearly identical rates, they lean on the same platforms, and they both honor your robots.txt. Where they diverge is how selective they are and how much traffic they send. Here's what that changes.
What Does Getting Cited Actually Mean?
Three states get collapsed into one word, and the levers for each are different.
- Mentioned. The model names your brand in the answer text. No link required. This is the one that moves purchase consideration.
- Cited. A clickable source appears, pointing at a specific page. This is what citation tracking counts.
- Retrieved. The model fetched your page. It may have read it and used nothing. Only your server logs see this state.
None of the three is a rank. Nothing is ordered, and nothing is scored, so "how to rank in ChatGPT" is the wrong question to bring to this. You're competing to be selected, not placed.
And the selection is brutal. A search results page had ten blue links and infinite scroll, so a page-two site still picked up something. An AI answer has three to eight citation slots and nothing after that. That's from Goodie's 31-million-citation publishers study, and it's the single structural fact that makes AI citations concentrate harder than search rankings ever did.
Pick which of the three you're managing before you pick tactics.
I laid out the broader factors that decide whether AI names your brand separately, and if you need the ground floor, the guide to answer engine optimization covers it. I'm not redefining AEO or generative engine optimization here. This is what happens after that work, when one model still won't name you.
How Does ChatGPT Decide What to Cite?
ChatGPT can reach the web through more than one path, and which path is in play changes what matters.
Free Tier vs. Paid Reasoning: Different Retrieval, Different Rules
OpenAI's help documentation names third-party search partners including Bing. But RESONEO captured raw ChatGPT response streams across plan tiers and countries (1,200 answers and 26,900 distinct pages) and published the teardown through Search Engine Land in August 2026. In their sample, the free-tier default ran through an OpenAI-operated retrieval path, with only 1.5% of its URLs appearing in Bing's top 20 for the same queries. Paid reasoning mode leaned on scraped Google results bought from data vendors.
Both things can be true. Say so rather than picking the tidier story.
The practical read: on the free tier, where the default experience lives, your Bing rank is not the lever the conversation turns on.
ChatGPT Rewrites Your Prompt Before It Searches
There's a second thing in OpenAI's own docs that most teams ignore. ChatGPT rewrites the prompt before it searches. One question becomes several targeted queries. So you're competing for whatever the model decided the user meant, which is usually more specific and more comparative than what they typed. Ask it which tool to pick, and the fan-out reaches pricing, integrations, alternatives, complaints. Cover the sub-questions, not just the head one.
Retrieval Shortlists You. Opening Cites You.
The same teardown surfaced the step almost nobody optimizes for. ChatGPT retrieves a set of pages, then opens some of them. In that sample, a page it opened got cited 74% of the time, compared with 7% for a page retrieved and never opened. Retrieval shortlists you. Opening cites you. One team, one sample, so hold it as a mechanism rather than a constant.
Two Fixable Technical Limits
Pages over 4 MB were rejected outright with an HTTP 400 rather than truncated. And the robot that opens pages doesn't execute JavaScript, so client-side content is invisible to it. That last one has the strongest evidence of anything in this article. Vercel's network-level testing found the same across every major AI crawler in December 2024, Goodie's agent-readiness audit of flagship publishers found the same pattern in July 2026, and nothing published since has contradicted either.
Server-render the part of the page that answers the question. Keep the page under 4 MB. That's most of the technical work for ChatGPT.
How Does Claude Decide What to Cite?
Claude is the model I get asked about most and the one almost nobody writes about with data behind it. How to optimize for Claude visibility comes down to four things: what triggers a search, how often that happens, which index it hits, and the one path that skips search entirely.
What Triggers a Claude Search
Anthropic's documentation says Claude searches when a request depends on current information, changing, or outside its training data, and answers directly on facts, math and science fundamentals, creative writing, and analysis of content you supplied. Two details most write-ups skip: that triggering behavior is steerable by system prompt, and there's a cap on searches per turn.
Why Published Retrieval Rates Range From 2% to 98%
That cap explains part of it. But the bigger reason is that a retrieval rate is a property of the test harness as much as the model. Measure through a platform hitting a search-enabled API endpoint, and you'll approach 100%, because the harness is asking for search. Measure real consumer sessions on defaults, and you'll get a fraction of that. Both numbers are right. They're answering different questions. Ask what the harness was before you believe any of them.
One number circulating doesn't hold up at all: 36.6% of prompts triggering a search against roughly 90% for ChatGPT. That traces to one practitioner's conference session relayed by Search Engine Land, with no sample size or method published. Directional at best. Nobody has a defensible precise number for the gap, and anyone quoting you one is quoting a slide.
Comparison Prompts Are Where Claude Visibility Is Winnable
Here's what does hold up, and it's more useful. The docs note that simple factual questions use one to three searches, while comparative or multi-entity research can use ten or more. If you want to gain visibility in Claude, comparison prompts are where it's winnable. If you publish one new asset for Claude this quarter, make it a real comparison page in your category.
Citations Aren't Optional Once Claude Searches
Anthropic's docs state citations are always enabled for web search and that developers displaying Claude's output must show them. There's no cited-but-hidden state on that path. If you made it into a retrieved Claude answer, the user can see you.
The Brave Search Question
Anthropic has never publicly documented a search backend. TechCrunch reported Brave Search appearing on Anthropic's subprocessor list in March 2025, when Claude web search launched. Third-party testing has since found strong overlap between Claude's citations and Brave results, including one test where 13 of 15 cited results matched. Thirteen of fifteen across a handful of queries is a lead worth chasing, not confirmation of anything. The practical move is unglamorous, and almost nobody does it: check whether your pages are in Brave's index at all, then see whether it moves your Claude numbers.
The Non-Search Path: Model Context Protocol
Anthropic publishes a connector directory built on the Model Context Protocol, where tools plug in and answer from live data. That's distribution, not SEO. I've written separately about marketing to agents, which is where this goes next.
ChatGPT vs. Claude: What Gets Cited
Here's the comparison, drawn from three published Goodie AI studies. Different windows and different samples, so read each row against its own source rather than across the table.
Sources: the AI search traffic report (rows 1-2), the publishers study, 31M citations across 11 surfaces (rows 3 and 7), and the social citations study, 6.1M citations across 10 platforms (rows 4-6).
Four Things To Act On
The Closest Pair in the Market on Source Behavior
1.1% against 1.0% on social. Reddit first for both, LinkedIn second for both. I expected a gap and there isn't one. Most AEO content sells you two playbooks here. The published evidence says one playbook covers most of it.
The Real Splits Are Elsewhere
Grok pulls 14.5% of its citations from social, and 87.4% of that is X, because xAI owns X. Instagram citations come from AI Overviews 98.98% of the time. YouTube concentrates 82.47% in Google surfaces. Those are structural couplings created by ownership and licensing, and they're where "each model is a different source market" actually bites. ChatGPT and Claude are the exception, not the illustration.
Claude Is More Selective With Citations
Claude cites news at 5.1% against ChatGPT's 11.9%, the lowest share among major assistants. Combined with three-to-eight citation slots per answer, that means mid-tier sources get squeezed out of Claude first. Being the sixth-best source is worth less on Claude than on ChatGPT.
Robots.txt Is the Real Lever for Both
The New York Times blocks all thirteen major AI user agents. Across 31 million citations, ChatGPT cited it zero times and Claude once. Grok cited it 37,642 times, AI Overviews 8,007, Perplexity 4,303. For this specific pair, the crawler layer is the highest-leverage thing you control, because these are the two models that will actually respect what you put there. That cuts both ways, which is the next section.
The Limits
Three studies, three windows, three samples. The referral figures are brand-averaged shares, not slices of one panel. The news-citation shares describe a narrow slice of the citation universe, since news is only 1.3% to 1.9% of everything AI cites. Full methodology sits in each linked study.
Where ChatGPT and Claude Agree
More than the market assumes, and the overlap is where your cheapest work lives.
Earned sources carry both. Across 6.1 million citations, earned made up 72.2% of everything cited. Owned content isn't competing for the slot.
Social behavior is near-identical, Reddit-led and LinkedIn-second for both. Reddit and LinkedIn were the only two platforms cited by all ten models in that study, which makes them the broad-coverage play rather than a model-specific one.
And the technical layer transfers cleanly: crawler access, server-side rendering, and page weight matter the same to each. Both honor robots.txt.
Do the technical work once. It counts twice.
Where ChatGPT and Claude Differ
Scale, selectivity, and trajectory.
Claude sends far less traffic in absolute terms: 18.5% of measurable AI referrals against 62.6%, but it's the one moving. From 1.35% to 18.5% across those two windows, while ChatGPT gave up 26 points.
Claude is stingier with citations. 5.1% of news citations against 11.9%, the lowest of the major assistants, against a ceiling of three to eight slots per answer.
And Claude has a non-search path that ChatGPT handles differently: Anthropic runs an integration catalog; OpenAI runs a merchant feed spec.
What doesn't differ is the asset list. Reddit, LinkedIn, third-party coverage, and crawler access serve both. What differs is how much room there is once you're in.
Why Owned Content Alone Isn't Enough
Across the 6.1-million-citation study, social citations ran 2.31 times owned citations over the full period, and 4.17 times by November. Earned sat at 72.2% of everything.
Your own domain is not the citation graph. It's the thing the citation graph describes.
Traffic doesn't rescue it either. In the publishers study, The New York Times ranks first in visits and sixth in citations. Fox News is top three in visits and holds 0.12% of citation share. Axios, at a fraction of their size, out-cites publishers twenty times bigger. Being visited and being cited are separate outcomes with separate rules, and most teams still only measure the first one.
That reframes the content budget. Publishing is how you become describable. Getting described is a different job on a different clock, usually owned by PR and review operations rather than the content calendar. I'd rather say that plainly than sell you a publishing cadence the data doesn't support.
The 7-Step AI Citation Playbook
Ordered by cost. Cheap diagnostics first.
1. Search your server logs for the seven user-agents. Dev, thirty minutes.
If OAI-SearchBot and Claude-SearchBot aren't in your logs, stop and fix robots.txt before anything else. The row people get wrong is OAI-SearchBot: blocking GPTBot opts you out of training and leaves ChatGPT Search untouched, while blocking OAI-SearchBot removes your content from search summaries. Very different decisions, and a lot of 2024-era robots.txt files made the second one by accident.
2. Confirm the answer is in the initial HTML. And that pages are under 4 MB.
3. Stand up a prompt panel. Twenty prompts, five runs each, both models, logged out. Tooling here.
4. Build real third-party review presence. In Seer Interactive's 804,491-response study, brands without a Trustpilot profile had a 1% median citation rate against 53.5% for brands in the minimal-profile tier. Correlation, not proof that a profile causes a 52-point lift, and Claude wasn't in that study. Strong enough to audit this week anyway.
5. Check whether you're in Brave's index. The Claude-specific step nobody takes.
6. Put one checkable statistic in every section you want cited. The Princeton and Georgia Tech team ran 10,000 queries through generative engines and tested nine optimization methods: adding quotations lifted visibility 27 to 28%, adding statistics 25 to 26%, keyword stuffing did close to nothing. That paper predates every retrieval system in this article, and measures answer share rather than clickable citations. Caveat it. But line it up against the 74% open-to-cite rate and the mechanism makes sense: a page with a number gives the model something to lift.
7. Build on Reddit and LinkedIn, then diversify. Those two were the only platforms cited by all ten models. But treat any single platform as fragile: when Reddit sued Perplexity in October 2025, Perplexity's Reddit citation share fell from 19.51% to 2.67% in a few days, and YouTube absorbed the difference. Access rules move citations faster than content quality does. Re-measure against baseline, not against last week.
What Not to Do
Don't Trust the Multipliers
"Schema gets you cited 2.8x more." "Pages under 0.4 seconds are 3x more likely to be cited." "Citations arrive in 4 to 8 weeks." Every one of those circulates without a published method, a sample size, or a control group. Ask for the denominator. If nobody can produce one, it's marketing.
Don't Expect Schema to Move Citations
The one controlled test on this question tracked 1,885 pages that added JSON-LD against 4,000 matched controls: -4.6% on AI Overviews, +2.4% on AI Mode, +2.2% on ChatGPT, with the positives statistically indistinguishable from zero. Don't overclaim the negative either. Both treated and control pages were already declining, so this isn't evidence that schema hurts. The honest position: schema is real hygiene for Google rich results and entity disambiguation, with no demonstrated causal lift for AI citations. Claude wasn't tested at all.
Don't Treat LLMs.txt as a Citation Lever
In Goodie's audit of flagship publishers, adoption was zero, and no major AI lab has confirmed using it for retrieval. Cheap to publish, no demonstrated lift. I've written the full assessment separately.
Don't Blanket-Block AI Crawlers
Two reasons. It works on exactly the models you probably want, since ChatGPT and Claude honor it and Grok, AI Overviews, and DeepSeek largely don't. And Wharton and Rutgers researchers tracking the 2023 wave of robots.txt blocking found roughly a 7% decline in weekly visits within six weeks, showing up in human browsing data rather than bot metrics.
Don't Assume a Licensing Deal Replaces the Crawl Question, or Vice Versa
AP blocks every OpenAI crawler and still draws 80% of its citations from ChatGPT, because licensed content travels through the contract rather than the crawl.
How to Measure Citation Performance
Three things will mislead you.
GA4's AI Assistant channel keys on source and medium together, so one source fragments across three channels. chatgpt.com / ai-assistant lands in AI Assistant, chatgpt.com / referral in Referral, chatgpt.com / (not set) in Unassigned. Build a source-only custom channel group alongside it.
Referral traffic is the wrong headline metric, but not for the reason people say. Cited-and-not-clicked is the majority outcome, so sessions understate visibility. And the mix moves fast: ChatGPT, Claude, Gemini, and Perplexity together held about 99% of measurable AI referrals as of April 2026, but the distribution inside that four shifted hard between the two measurement windows. If you built your reporting around ChatGPT in 2025, it's measuring a third less of the landscape than it was.
And prompt tracking without sampling discipline is noise. Answers are non-deterministic. Three to five runs per prompt, logged out, browsing state constant, model version recorded. Otherwise you'll read variance as a trend and rebuild your content strategy around it.
One caveat on logs, since I told you to read them. ChatGPT holds a short-lived cache of pages it has opened, so a repeat answer can cite you without generating a new request. Logs are a diagnostic, not a tally.
I've laid out the fuller framework for measuring AI search visibility separately.
Two models, one playbook, and a second one that's growing faster than anything else in the mix. Go read your logs before you write another page.
How to Get Your Brand Cited in ChatGPT & Claude: FAQs
