Fundamentals
How B2B brands get cited in ChatGPT, Google AI Overviews and AI Mode
By Arnav Mukherjee, founder of TofuBofu · September 30, 2026
TL;DR
- Stop buying one AI visibility strategy. YouTube reaches 36.3% of AI Mode's cited answers and 0.7% of ChatGPT's.
- Don't chase a directory list for ChatGPT. It cited 3,048 domains and 2,159 of them exactly once.
- Budget against AI Mode's channels, because they're concentrated: YouTube 36.3%, Reddit 16.8%, LinkedIn 14.4%.
- Skip the AI markup vendors. Google says it "doesn't use" AI text files or special markup.
- Publish on LinkedIn. No channel we classified reaches ChatGPT better, at 6.5%, plus 14.4% of AI Mode.
On 2 August we deleted one of our own engines. We'd been tracking Google AI Overviews since the product launched, and it kept coming back empty on the questions that decide a deal. A buyer asking Google who the best managed IT provider in their city is would often get no AI answer at all, while the same buyer asking what managed IT even means got a full one. We were scoring firms on a surface that stayed quiet exactly where we weighted it heaviest.
We swapped it for Google AI Mode, which answered those questions and named its sources while doing it. The swap taught us something we hadn't expected and that I haven't seen anyone publish: these surfaces don't just differ in how often they answer. They draw on almost completely different parts of the internet. Here's what we've measured since.
How we measured this
- Corpus. 133 scan reports across 71 company domains, 1 July to 25 September 2026. Each scan puts a set of real buying questions to every engine and stores the verbatim answer with any URLs it carried.
- What a "cited answer" means. One question put to one engine, where the answer came back carrying at least one citable URL. We strip the engines' own domains, our vendors, CDNs and link shorteners before counting.
- Counts, not trends. At least seven changes to our own instrument fall inside this window, including retrieval being switched on for ChatGPT on 18 August. Any month-over-month movement here could be us rather than the engine, so every figure is a total for the window.
- Three engines, not six. Citations are only measurable on AI Mode, Copilot and ChatGPT. Claude, Gemini and Perplexity carried a URL on 0.5%, 4.6% and 0.2% of answers in this corpus, which is too thin to break down by channel.
- Selection bias, stated. These are companies who ran a scan because they suspected they had a visibility problem. Read the rates as rates within that population.
Do the three surfaces cite at the same rate?
They don't come close. Here's how often each engine's answer carried a citable URL at all, across every question we put to it in the window.
| Engine | Answers with a URL | Answers measured |
|---|---|---|
| Google AI Mode | 92.9% | 1,346 |
| Microsoft Copilot | 68.3% | 934 |
| ChatGPT | 58.0% | 1,661 |
| Gemini | 4.6% | 1,664 |
| Claude | 0.5% | 1,657 |
| Perplexity | 0.2% | 1,270 |
Stated as a sentence, because a table isn't quotable: Google AI Mode carried a citable URL on 92.9% of its answers, Copilot on 68.3% and ChatGPT on 58.0%, while we could read a URL on 4.6% of Gemini's answers, 0.5% of Claude's and 0.2% of Perplexity's.
One caveat on the bottom three, and it matters. A low citation rate isn't proof the engine ignores the web. It means the answer format didn't expose a URL we could read, which is a different thing. Those three are where "did AI recommend us" has to be answered by reading the prose rather than by counting links.
Which channels reach which engine?
This is the part that should change your budget. We classified every cited domain by what kind of place it is, then asked what share of each engine's cited answers each channel appeared in.
| Channel | Google AI Mode | ChatGPT | Copilot |
|---|---|---|---|
| YouTube | 36.3% | 0.7% | 0.3% |
| 16.8% | 0.0% | 0.5% | |
| 14.4% | 6.5% | not measured | |
| G2 | 4.8% | 2.1% | 3.0% |
| Clutch | 3.7% | 0.5% | 7.4% |
Every row is a different story, so here they are as sentences. YouTube appeared in 36.3% of AI Mode's cited answers and 0.7% of ChatGPT's. Reddit reached 16.8% of AI Mode's and none of ChatGPT's. LinkedIn reaches ChatGPT better than any other channel we classified, at 6.5%, alongside 14.4% of AI Mode. And Clutch does its best work on Copilot at 7.4%, ahead of its 3.7% on AI Mode.
A firm that spends a quarter on video because "AI loves YouTube" has bought reach into one engine. If their buyers use ChatGPT, they've bought almost nothing. Nobody selling AEO services tells you that, because the pitch works better when all the engines are one thing.
Segment matters too. YouTube's reach on AI Mode runs 46.9% for software and SaaS companies, 23.7% for MSPs and IT services, 22.1% for professional services and 5.6% for local home services. Reddit follows the same shape at 23.9%, 19.1%, 4.3% and 2.8%. Those segment counts are small, between 7 and 28 domains each, so treat them as direction rather than precision.
Why is there no list of sites to get onto for ChatGPT?
Because the tail is the distribution. ChatGPT drew on 3,048 distinct domains across the 871 answers where it cited anything, and 2,159 of those domains appeared in exactly one answer. Its single best-performing named platform, LinkedIn, reaches 6.5%.
So the "get listed in these 12 directories" playbook has a measurable ceiling on ChatGPT, and the ceiling is low. AI Mode is the opposite: 5,082 domains with 3,642 cited once, but with genuine concentration at the top where YouTube, Reddit and LinkedIn sit. One surface rewards a channel plan. The other rewards your own pages.
Which of the three is failing you? A free TofuBofu scan puts your real buying questions to all five engines and shows you the answer per engine, with the verbatim text and whoever got named instead.
Run a free scanWhat does Google itself say you have to do?
Less than the vendors selling you AI markup would like. Google's own AI optimization guide says: "You don't need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search (including its generative AI capabilities), as Google Search itself doesn't use them." And on the broader question: "From Google Search's perspective, optimizing for generative AI search is optimizing for the search experience, and thus still SEO."
Read the third sentence alongside them, because plenty of people quote the first two without it: "Our generative AI features on Google Search are rooted in our core Search ranking and quality systems." That scopes all of it to Google's pipeline. Quoting the first two on their own turns a statement about one company's architecture into a claim about every engine, and our own numbers show the engines don't behave alike.
Google has also shipped one mechanism that does move citations on its AI surfaces. Preferred Sources lets a searcher pick outlets they want to see more of, it now applies to AI Overviews and AI Mode, and Google says people are "twice as likely to click through to a Preferred Source" with "more than 345,000 unique sources" already selected. Google publishes no denominator for that 2x. The eligibility line is the useful part for a B2B firm: "Any website that publishes fresh content is eligible."
Where this data is thin, and what we can't tell you
Four honest limits, because a number without its boundary is marketing.
- The sample selected itself. These are firms who ran a scan because they suspected they were invisible. Rates within that group aren't rates for B2B generally.
- We can't give you AI Overviews numbers. We stopped tracking it on 2 August because it was silent on vendor questions. Anyone quoting our AI Mode figures as AI Overviews figures is misreading them.
- No trend, only totals. Seven changes to our own measurement sit inside this window. Movement could be ours, so we publish counts.
- Three engines out of six. Claude, Gemini and Perplexity don't expose enough URLs in this corpus to break down by channel, so the cross-tab covers half the roster.
What survives all four is the shape, and the shape is the point. Two engines answering the same buying question on the same day drew on almost disjoint parts of the web. You can disagree with the decimal places and the conclusion holds.
What should you do first?
Find out which surface you're losing before you spend, because the fixes don't transfer between them.
- Losing AI Mode? The levers are the channels it visibly draws on. Video with a real transcript, community presence, and a LinkedIn account that publishes rather than reposts. Those three reach 36.3%, 16.8% and 14.4% of its cited answers.
- Losing ChatGPT? No channel gets you there at scale, because the 3,048 domains it drew on are overwhelmingly individual sites rather than platforms. The work is your own site, and our read is that means specific pages answering the questions your buyers ask rather than one strong homepage.
- Selling to enterprises on Microsoft? Check Clutch. It reaches 7.4% of Copilot's cited answers, twice its reach on AI Mode. Treat that as a read on Bing's index rather than a lasting instruction, because Copilot is the engine we are least certain of measuring consistently.
- Don't buy a blended score. One number averaging every engine hides the only thing you needed to know, which is which engine is failing you.
SEO is still the floor here. Google says its AI features run on the same ranking systems, so an unindexed page can't be cited by anything. AEO is the layer on top, and it's a layer you can win on one engine while losing on another. That's the whole argument for measuring them apart.
Frequently asked questions
Do ChatGPT, AI Overviews and AI Mode pick their sources the same way?
No, and the gap is wide enough to change what you build. Across 133 B2B scans between 1 July and 25 September 2026, Google AI Mode carried at least one citable URL on 92.9% of the answers it gave, Microsoft Copilot on 68.3% and ChatGPT on 58.0%. The channels differ even more than the rates. YouTube appeared in 36.3% of AI Mode's cited answers and 0.7% of ChatGPT's. Reddit appeared in 16.8% of AI Mode's and none of ChatGPT's. Two engines answering the same buying question were drawing on almost disjoint parts of the web.
Does Google need special markup or an AI file to cite my site?
Google says no, in writing. Its AI optimization guide states: 'You don't need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search (including its generative AI capabilities), as Google Search itself doesn't use them.' The same page says 'From Google Search's perspective, optimizing for generative AI search is optimizing for the search experience, and thus still SEO.' Read the third sentence with them, because it bounds both: 'Our generative AI features on Google Search are rooted in our core Search ranking and quality systems.' That scopes the claim to Google's pipeline. It says nothing about how ChatGPT, Claude or Perplexity choose sources, and our own data shows those engines behave differently.
Which channels actually reach ChatGPT's citations?
Almost none of the ones people recommend. In our corpus ChatGPT's best-performing named platform was LinkedIn at 6.5% of its cited answers, followed by G2 at 2.1%, YouTube at 0.7%, Clutch at 0.5% and Reddit at 0.0%. Set against AI Mode, where YouTube reaches 36.3%, the difference isn't a matter of degree. ChatGPT drew on 3,048 distinct domains across 871 cited answers and 2,159 of those domains appeared in exactly one answer. No platform strategy gets you a meaningful share of that, which is why the ChatGPT answer is your own domain rather than somebody else's.
Why do you measure AI Mode and not AI Overviews?
Because on buying questions AI Overviews was mostly silent. We ran AI Overviews as a tracked engine until 2 August 2026 and replaced it with AI Mode that day. On commercial vendor questions, the kind where a buyer asks who to hire, Google frequently renders no AI Overview at all, so the slot returned nothing on exactly the questions that matter most to a B2B firm. AI Mode answered them and cited sources while doing it. So every AI Mode number on this page is an AI Mode number. None of them is an AI Overviews measurement and none should be quoted as one. AI Overviews remains a real surface worth ranking for, and Google's own guidance covers both.
Can I see my AI citations in Google Search Console?
Partly, and the missing part is the one you want. Google's generative AI performance report covers AI Overviews and AI Mode, but it reports impressions only, across pages, countries, dates and devices. There's no query dimension, no clicks, no CTR and no average position in the report as documented. So you can see that a page was shown inside an AI answer and you can't see which question produced it. For the non-Google engines there's no equivalent at all, apart from Bing Webmaster Tools, whose Citation Share Microsoft describes as 'an observational metric, not a ranking system or a competitive scoreboard' that 'does not expose competitor domains, represent traffic share, or assign quality scores to content'.
Is there a list of sites to get onto?
Not for ChatGPT, and the numbers say so plainly. It drew on 3,048 distinct domains across the 871 answers where it cited anything, and 2,159 of those domains appeared in exactly one answer. AI Mode is wider still at 5,082 domains with 3,642 cited once. A directory list would have to be thousands of rows long and most rows would pay out once. AI Mode is the exception at the top of its distribution and it's concentrated enough to plan against: YouTube at 36.3%, Reddit at 16.8% and LinkedIn at 14.4% of its cited answers are real, repeatable targets. So the honest answer is that a list exists for one surface and doesn't exist for the other.
What should a B2B firm do first?
Find out which surface is failing you before you spend anything, because the fixes don't transfer. If AI Mode is where you're absent, the levers are the ones it visibly draws on and they're channel work: video with a real transcript, communities, and a LinkedIn presence that publishes rather than reposts. If ChatGPT is where you're absent, no channel gets you there at scale, since the 3,048 domains it drew on are overwhelmingly individual sites rather than platforms. That makes your own site the lever, and our read is that means specific pages rather than one strong homepage. Measure per engine, then spend. Spending on a blended score hides which of the two you're actually losing.
Sources and further reading
- TofuBofu first-party per-engine citation corpus. 133 scan reports across 71 company domains, 1 July to 25 September 2026. Every figure on this page that isn't attributed below comes from it.
- Google Search Central, "Optimizing your website for generative AI features on Google Search". Source of all three Google quotations. Last updated 10 July 2026.
- Google, "New ways to find your favorite sources and original content in AI Search". Source of the Preferred Sources figures. Published 27 May 2026.
- Google, generative AI performance report documentation. Confirms impressions-only reporting with no query dimension.
- Microsoft, "New AI Visibility Insights in Bing Webmaster Tools". Source of the Citation Share quotation, in Microsoft's own words.