How many sources does ChatGPT cite per answer?

Semrush's 2026 AI Visibility Index, covering 126 million prompts, measured an average of about 15 sources per answer in ChatGPT against roughly 3 in Gemini. ChatGPT is the engine with the most room for a citation.

Which sites does ChatGPT cite most?

According to Profound, across more than 680 million citations, Wikipedia is the single most cited domain in ChatGPT at around 7.8%. Reddit dominates in Perplexity instead, where it accounts for close to a quarter of citations, and is almost absent in Gemini at about 0.1%.

Does traffic from ChatGPT convert?

Seer Interactive measured conversion rates of 15.9% for ChatGPT referrals against 1.76% for Google organic, roughly nine times better. Volume is much lower, but the intent is far higher because the user arrives after the answer has already filtered the options.

What kind of content does ChatGPT actually link out to?

Content that supports a decision rather than closing a fact. Pages answering what something is, where it is or its opening hours get resolved inside the answer with no click. Pages comparing real options, explaining total cost or covering logistics from a specific origin get cited with a link, because the user still needs to verify and buy.

What ChatGPT actually cites, and what it sends back to you

3 September 2026 · Joan Sanz · Research note · 7 min read

Short answer

ChatGPT cites far more sources per answer than any other engine, which makes it the easiest one to get into. But it only links out when your page helps someone decide. If your page closes a fact, the answer keeps the visit.

Five findings

ChatGPT cites ~15 sources per answer. Gemini cites ~3. There is more room at the table than people assume.
Each engine leans on a different surface. Wikipedia rules in ChatGPT, Reddit rules in Perplexity, and Gemini barely touches either.
It is volatile. Reddit went from appearing in ~60% of ChatGPT answers to ~10% in six weeks.
The traffic converts about 9x better than Google organic, on much lower volume.
Content type beats technical setup. In my own network, two structurally identical sites sit at 43.4% and 2.0% of AI-referred traffic.

1. ChatGPT has the most room for citations

Semrush's 2026 AI Visibility Index analysed 126 million prompts and found ChatGPT surfacing around 15 sources per answer, against roughly 3 in Gemini. That single number reframes the whole game: being cited by ChatGPT is not a lottery for one slot, it is a shortlist with room for a dozen. Gemini is where the squeeze really is.

2. Every engine trusts a different neighbourhood

Profound analysed more than 680 million citations. The pattern is not subtle:

Surface	ChatGPT	Perplexity	Gemini
Wikipedia	Most cited domain (~7.8%)	Present	Present
Reddit	>5% of citations	~24% of citations	~0.1%
YouTube	4.4% of citations	38.7%	Present

The practical consequence is that "optimising for AI" is not a thing. If your buyers use Perplexity, a Reddit and YouTube strategy is rational. If they use ChatGPT, the same strategy is close to a waste of money, and you are better off being the encyclopaedic, verifiable source that Wikipedia-shaped citations reward.

3. The ground moves fast

The same Semrush work tracked Reddit falling from appearing in roughly 60% of ChatGPT answers to about 10% in six weeks. Not a slow decline, a cliff. Any agency selling you a strategy anchored to a single surface is selling a screenshot of a moving picture.

Read these numbers carefully. "Reddit is 5% of ChatGPT citations" and "Reddit appears in 60% of ChatGPT answers" look contradictory and are not: one is share of all citations, the other is frequency of appearance. A lot of the noise written about GEO comes from mixing the two, sometimes on purpose.

4. Low volume, high intent

Seer Interactive measured conversion at 15.9% for ChatGPT referrals against 1.76% for Google organic, roughly nine times better. That makes sense: by the time someone clicks through from an answer, the model has already filtered the options and they are checking a shortlist, not browsing. You should expect far fewer sessions and value each one much more.

5. What I measured myself

I run a network of more than 60 travel sites on the same generator, the same hosting and the same technical rules. The spread between them was wide enough that I first assumed a measurement bug:

Site	What it answers	AI-referred traffic
Site A	A full-day island trip: which operator, from where, real price	43.4%
Site B	Trips to a set of waterfalls: options and logistics	41.7%
Site C	What a specific cave is and when it opens	2.0%

Site C is technically flawless and commercially irrelevant to an assistant, because everything it answers fits inside the answer. Site A cannot be closed that way: twenty operators, prices that double between them, three departure cities, and a real difference between a catamaran and a speedboat. The model has to lean on a source and send the user to verify.

AI competes with you on facts and collaborates with you on decisions. If your content is an encyclopaedia, you have built free training data. If it is a criterion, you have built something that needs citing.

How this was measured: first-party analytics on each site, classifying sessions by referrer from assistant domains, September 2026. These are people who landed on the site after an AI answer, not crawler hits. AI crawlers are counted separately at the edge, because they do not execute JavaScript and folding them in would inflate the number.

What to do with this

Before writing the next page, ask whether an assistant can answer it completely without sending anyone anywhere. If yes, that page will not bring you traffic no matter how much schema you bolt on.

The angles that do work are the ones that force a choice: comparisons between real options, origin (how do I get there from my city, the most neglected angle of all), true cost including the catch, logistics, and the plain is it worth it answered with an actual opinion instead of "it depends".

Then let the crawlers in. Explicit allows for GPTBot, OAI-SearchBot and ChatGPT-User in your robots.txt cost nothing and are the difference between being readable and being invisible. That part is necessary and not sufficient: it earns you the right to be considered, not the citation.

Questions

Is 43% a realistic target?

No. It is the best case in my network and that is exactly why it is worth studying. Most sites sit far closer to 2% than to 43%. What matters is that two sites identical on the inside produce such different outcomes, because that isolates the cause.

So schema and llms.txt are useless?

They are the second layer, not the first. They make you readable and quotable. If what you write does not deserve a citation, perfect markup only means you get read more efficiently before being ignored.

How long before a change shows up?

Weeks, not days, and you need patience with the measurement. At low volumes noise looks exactly like a trend, so give it at least 30 days before concluding anything.

Sources

Semrush AI Visibility Index 2026 · Profound, AI platform citation patterns · Semrush, most cited domains · Tinuiti via CMSWire · Otterly, YouTube AI citation study 2026. Third-party figures are as published by each study; first-party figures are my own analytics.

Joan Sanz. Fifteen years in travel and hotel tech. I run a network of 60+ sites where I test this before recommending it to anyone. More at /about.

Does your content answer facts or decisions?

If you want to look at it with your own numbers in front of us, write to me.

Message me on LinkedIn