ChatGPT doesn't cite sources when it decides not to search the web, and when it skips the search, whatever it says about your brand comes out of what the model already learned. A study Goodie published last week found the share of ChatGPT answers with no sources at all went from 1.5% to 12.2% starting October 31, so the piece of your AI visibility that runs on memory got a good deal bigger, and most of the reports I see don't measure that piece at all (Goodie).
I've been writing about citations a lot on this blog, so I figured I owed you a post on what happens when there aren't any.
The short version
ChatGPT decides on its own whether to search. When it doesn't, you get an answer with no links, built from training data.
How often it searches swings a lot by topic and has changed several times this year, sometimes overnight.
In my AI Verdict Study, ChatGPT gave the same verdict on a brand with search turned off in 29 of 36 conversations, so the opinion it holds about you before it searches carries a lot of weight.
If your tracking tool forces ChatGPT to search, it's measuring a version of ChatGPT that a lot of your buyers never see.
What Goodie found
Goodie tracks ChatGPT answers for its clients, about 3,000 a day pulled from the ChatGPT web interface, and it looked at 94,851 of them between October 9 and November 9 (Goodie). From October 9 to 30, 1.5% of answers came back with no sources. From October 31 on, it was 12.2%, and the later period never dropped below 10%. Citations per answer fell 17%, from 5.44 to 4.53, and about 63% of the lost citations came from answers that cited nothing at all. The rest came from answers that still searched but pulled a few fewer sources.
Two days before the jump, OpenAI updated GPT-6 in ChatGPT to give shorter answers to Free, Go and logged-out users, with the stated goal to "reduce overly flattering language" (MIXED News). Goodie points out that nothing in that release note mentions search or citations, so the link is timing plus the idea that a shorter answer needs fewer sources. OpenAI hasn't said anything to confirm it.
A couple of caveats. Goodie sells AI search analytics, so it has a stake in people caring about this, and the write-up doesn't describe the prompt set, which matters a lot here (more on that in a second). I'd treat the exact numbers as one company's tracked prompts and pay more attention to the direction they point.
Why ChatGPT doesn't cite sources on some answers
The plain mechanical answer is that ChatGPT only cites when it searches, and it decides for itself whether a question needs a search. In Graphite's words, "By default, ChatGPT decides whether to search" (Graphite). If the model thinks it already knows enough to answer, it answers from what it learned in training, and there's nothing to link to.
How often it decides to search depends a lot on what you ask. Similarweb's numbers for May, which cover U.S. desktop use of ChatGPT overall (homework, recipes, everything), found that only 6.8% of answers included a citation. Travel answers were at 22.6% and education answers were at 4.8% (Search Engine Journal on Similarweb). Graphite looked only at prompts comparing brands and products, the kind a buyer asks, and found evidence of search in 58% of logged-out answers. Those two numbers measure different sets of questions, so I wouldn't line them up against each other, though together they tell you that commercial questions search a lot more often than ChatGPT does on average and still don't search every time.
The rate also moves. SISTRIX saw the share of ChatGPT answers that searched drop from above 15% to below 2.5% over two weeks in September 2025, at least in sessions with no account (Search Engine Journal on SISTRIX). seoClarity found the U.S. zero-citation rate went from 28% to 48% in March of this year, with citation volume down 86 to 94% across markets before it recovered in May (seoClarity). So October 31 is the latest swing, and I'd bet a fair amount it won't be the last.
How often ChatGPT shows its sources
Answers with no sources are growing
What ChatGPT says about your brand when it doesn't search
My own data comes in here.
In the AI Verdict Study, I ran a scripted buyer through a seven-turn conversation with four engines and then asked each one whether the buyer should choose a specific brand. Most runs had web search on. I also ran 108 conversations with search turned off to see what changed. ChatGPT gave the same verdict with search off in 29 of 36 conversations. Gemini matched in 30 of 36. Claude was the outlier, matching in only 17, and with search off it refused to give a verdict 11 times (with search on, it never refused once in 108 conversations).
So for ChatGPT and Gemini, the opinion about a brand was mostly already there before any search happened. The search added sources and links, and in most of those conversations it didn't change the answer. I want to be careful here, because that study is small, it ran through the API without system prompts, and I built the tool that ran it. I'd call this a strong hunch backed by a small sample. Still, it lines up with what Goodie says in its write-up, that a no-source answer "comes from what the model already believes about your category and your brand."
AI Verdict Study · search on vs. search off
Turn search off and the verdict mostly holds
Source: Jarred Smith, AI Verdict Study v1, Sept 2026. 36 search-off conversations per engine, compared with the same engine's search-on verdicts. API runs, no system prompts, small sample.
Say you're a brand that's done a lot of good work getting cited. If a buyer's question doesn't trigger a search, none of those citations show up. ChatGPT just tells them what it thinks of you, and what it thinks was set months ago by whatever was written about you across the web when the model was trained. I wrote a whole post on where AI engines go to decide about your brand, and that list of sources matters even more now, because it's also roughly the list of places the model learned from.
What this does to your AI visibility reports
Most AI visibility reporting is built around citations, which makes sense, because citations are the part you can see and the only part that can send anyone to your site. I made the case a few weeks ago that mentions and citations are two different numbers and should be tracked separately. Now there's a third number underneath both of them, which is how often the engine searches at all.
Graphite found that some prompt tracking tools appear to force ChatGPT to search on every run. One tool showed six prompts searching in all 18 of its responses, while those same prompts showed no search across 600 normal responses (Graphite). For prompts that never search on their own, forced and unforced visibility differed by about 20 points on average. If your tool does that, your report is describing the version of ChatGPT that searches, which may be the minority case for your category.
I'd ask your vendor two questions. Does the tool force search, and can it show you the share of runs where ChatGPT searched on its own? Graphite calls that the grounding rate, and I think it should sit at the top of every ChatGPT report, right next to the date of October 31 marked on the chart so nobody panics about a citation drop that had nothing to do with your content.
What to do about it
Number one, find out what ChatGPT says about you when it doesn't search. Ask it about your category and your brand in a logged-out window, and check whether the answer has sources. If it doesn't, read it closely, because that's the version built from memory. Is it current? Does it get your pricing, your positioning or your product line wrong? Do it a few times, since answers vary from run to run.
Two, work on what the model remembers. Goodie calls this "be remembered," which means consistent facts about your brand repeated in a lot of places the model learns from. That's your own site, sure, but it's also Wikipedia if you qualify, trade press, review sites, comparison articles, and the forums your buyers read. When those places disagree with each other about what you do or who you're for, the model has to average them, and you probably won't like the average. I spend a good part of Explainable on why some brands become the default answer, and most of it comes back to this kind of consistency.
Three, keep doing the retrieval work. Graphite's 58% says commercial questions still search more often than not, and when ChatGPT searches it's reading pages in real time, which I covered in how ChatGPT searches your website. Both layers matter, and the memory layer is the one most teams have stopped looking at.
Then last, claim your listing in OpenAI's Business Hub if you haven't, so you have at least one source of facts about your business that OpenAI gets straight from you.
If you'd like a quick read on where you stand, my free AI Visibility Quick Check walks you through the questions to ask ChatGPT and what to look for in the answers. The self-check below will give you a rough idea first.
Self-check · 6 questions
Is your brand ready for answers with no sources?
Answer for your own brand. Nothing is saved or sent anywhere.
Where this leaves the click question
I've argued for a while that AI visibility is the metric that matters now, and that a lot of the value happens in answers where nobody clicks (The Click That Didn't Happen). Answers with no sources are the far end of that. There's no link to click and no citation to count, only a recommendation that either includes your brand or doesn't. You can't see those in your analytics, and most tools can't see them either unless they look for them on purpose.
I'd love to hear what you're seeing in your own tracking since October 31, especially if your citation counts dropped and you're trying to figure out whether it was you or ChatGPT. Send me a note or find me on LinkedIn.
Jarred Smith is the author of Explainable: Why AI Recommends Some Brands & Ignores Others, an Amazon bestseller on AEO, GEO, and SEO. He's a marketing leader with nearly 20 years of experience across healthcare, public media, retail, and environmental services. Find him at jarredsmith.com.