Over the summer, the companies that sell AI visibility tracking started publishing their own data on where AI engines get their sources. They measured different engines, over different months, with different samples. On the headline question, they landed in almost the same place.
Profound's index, built on 11.84 billion citations across eight AI systems between April 16 and July 16, 2026, put brand sites at 57% of citations. OtterlyAI's AI Search Index, published on September 8 and built on 6,095,520 citations collected in July and August, puts company-owned pages at 53.7%. When two independent datasets agree on "more than half", that part of the story is probably stable.
The more useful part of Otterly's report isn't the average. It's how badly the average describes any single industry.
What the AI Search Index actually measured
Otterly describes its method in one sentence: "We asked all seven AI Search engines the same thousands of questions every day through July and August 2026, in each of 17 industries (US only), written the way buyers actually ask: ranked lists, comparisons, pricing, how to questions and trust checks." The seven engines were ChatGPT, Gemini, Perplexity, Microsoft Copilot, Google AI Mode, Google AI Overviews and Claude. Besides the 6,095,520 citations, the dataset includes 107,633 ChatGPT ad units and 15,363 shopping cards, each one classified into one of ten source categories.
Two things are worth keeping in mind before using any of these numbers. The first is who published them. OtterlyAI sells an AI search monitoring tool, starting at $29 a month, that tracks those same seven engines. Vendor research isn't wrong by default, but it is also marketing, and it's fair to read it that way. The second is what isn't published. The report doesn't detail how each domain gets assigned to a source category, and the sample is US-only buyer questions that Otterly chose. A different set of questions would produce different shares.
Neither point cancels the data. They define how far it travels.
The average hides a spread of more than two and a half times
"Company owned pages take 53.7% of every citation the seven engines make. News and media takes 16.9%, government and NGO sources 8.1%, and nothing else clears 5%." That's the headline, and it's consistent with the argument about owned visibility I made when ChatGPT stopped citing Reddit.
Now the breakdown. "Industrial/Manufacturing draws 77.4% of its citations from company websites. Nonprofit draws 29.6%." Same seven engines, same method, same two months. In one sector, company pages are more than three quarters of what gets cited. In another, they're under a third.
Journalism moves even more. "Journalism dependence swings 7.3x by sector, from 32.2% in Technology to 4.4% in AI Data Governance." For a technology company, coverage in the trade and tech press is a real citation channel. For a company in AI data governance, it barely registers.
And some sectors run on sources marketing doesn't produce at all. Nonprofit draws 42.4% of its citations from government and NGO sources, and Pharma 26.2%. Educational Platforms draws 30.7% of its citations from education domains, "against a median sector of 1.1%."
| Sector | Source that stands out | Share of citations | What it points to |
|---|---|---|---|
| Industrial/Manufacturing | Company websites | 77.4% | Product and technical pages on your own domain carry most of the weight |
| Nonprofit | Government and NGO sources | 42.4% (company sites: 29.6%) | Being referenced by institutions matters more than your own pages |
| Pharma | Government and NGO sources | 26.2% | Regulatory and public health sources sit between the brand and the answer |
| Technology | News and media | 32.2% | Earned coverage in trade and tech press works as a citation channel |
| AI Data Governance | News and media | 4.4% | Press coverage barely registers as a source |
| Educational Platforms | Education domains | 30.7% (median sector: 1.1%) | Universities and education sites are part of the competitive set |
If you took only the 53.7% and built a GEO plan from it, you'd overinvest in owned pages in Nonprofit and underinvest in them in Manufacturing. The average is accurate. It just doesn't describe anyone in particular.
A three-axis read before any GEO decision
The way I'd use this report is as a reason to stop asking whether AI search cites brands, and to start asking three narrower questions. I think of it as a three-axis read: engine, sector and source type. Which engines matter for this audience. What the source mix looks like in this category. And which of those source types a company can realistically influence.
Source type is where it gets practical. Owned sources are the part you control directly: product pages, documentation, comparison content, pricing, the middle-of-funnel material most sites never publish. Earned sources, news and trade media, respond to PR and to having something worth reporting. Institutional sources, government bodies, NGOs, universities, respond to neither in the short term. In a regulated sector, the way into those answers runs through being referenced where regulators and researchers already publish, which is slower work and usually not owned by marketing. Otterly's own recommendation for that case is blunt: "Look outside marketing if you are in a regulated sector."
This is also where a topical content map earns its keep, with one adjustment. Mapping coverage tells you what you've written about a topic. The sector's source mix tells you how much of the answer your own writing can realistically reach.
August 8 was a ChatGPT change, not an AI search change
The second finding matters for anyone who watched their numbers move in August. "On 8 August, ChatGPT changed what it cites. No other engine made the same move." According to Otterly, ChatGPT halved its journalism citations that day, from 21.3% to 9.8%, while its citation volume roughly doubled, from a median of 19,801 a day to 39,446. Community and forum sources fell 67% across the same break, then fell again on August 14 to end 80% down. Reddit alone went from 4.42% of all ChatGPT citations before August to 0.39% after August 14, a fall of 91%.
The control group is what makes it credible: "Five other engines held within 7%, so this is ChatGPT, not AI Search."
It also lines up with a separate dataset. When I wrote about the Reddit drop, the source was Promptwatch, which measured Reddit at 3.83% of ChatGPT citations between July 18 and August 7, and 0.52% between August 14 and 17, an 86.4% relative drop. Promptwatch traced the break to August 8, when the share of ChatGPT's fanout queries using the site: operator went from about 0.37% to 16.8% in a single day, and average searches per response went from about 1.08 to 1.83.
Two trackers, different samples, different decimals: 86.4% in one, 91% in the other. Same date. Same direction. Promptwatch put the right frame on it at the time, telling readers to "treat the direction and the timing as the reliable part, not any single tool's decimal point." That advice applies to both reports, including the one I'm writing about now.
The doubling in citation volume and the jump in site: queries point to the same mechanism: ChatGPT running more searches per answer, more of them scoped to specific domains. Some sources gained in the process. Otterly measured youtube.com going from 0.150% of all ChatGPT citations before August to 0.510% after August 14.
Seven engines, seven readings of the web
The engines don't only differ in when they change. They differ in how much of the web they're willing to cite at all. Google AI Mode accounts for 21.02% of all citations recorded in the study and cites 69,390 distinct domains, yet reaches half of its citations with just 436 of them. Claude cites 19,732 distinct domains in total. Microsoft Copilot accounts for 7.78% of all citations.
Platform preferences diverge just as sharply. In two months, ChatGPT cited TikTok 163 times. Google AI Mode cited it 1,650 times, and Google AI Overviews 1,004.
| Measure | Engine | Value |
|---|---|---|
| Share of all citations recorded | Google AI Mode | 21.02% |
| Share of all citations recorded | Microsoft Copilot | 7.78% |
| Distinct domains cited | Google AI Mode | 69,390 |
| Domains that make up half its citations | Google AI Mode | 436 |
| Distinct domains cited | Claude | 19,732 |
| TikTok citations in two months | ChatGPT / Google AI Mode / Google AI Overviews | 163 / 1,650 / 1,004 |
| Reddit share of citations, before August vs after August 14 | ChatGPT | 4.42% → 0.39% |
This is the practical reason behind another of Otterly's recommendations: "Never diagnose a drop from one engine." A brand whose ChatGPT citations fell in August and concluded that its content had gotten worse would have been diagnosing a retrieval change it had nothing to do with. It's a close relative of what I wrote about Cloudflare's new crawler categories: each company reaches the web through its own crawlers and its own retrieval logic, and treating all of them as one channel hides exactly the movements that matter.
Some categories are already decided
One line in the report deserves more attention than it will probably get. "In 7 of the 17 industries, one brand leads on all seven engines: Coursera in Educational Platforms, Tesla in Energy, YouTube in Media/Entertainment, Pfizer in Pharma, Zillow in Real Estate, Nike in Sports Brands and Uber in Transportation."
When seven engines with different crawlers, different domain appetites and different retrieval logic all put the same brand first, the broad question in that category has an answer the engines agree on. For a challenger, competing head-on for that question is probably the least efficient use of effort. The open ground is in narrower questions the leader doesn't cover as completely: comparisons for a specific use case, pricing for a specific segment, the trust checks buyers run before committing. Those are the question types Otterly built its sample from, and they're where a smaller site with complete coverage of its niche has something the category leader doesn't.
Otterly's recommendation here is a check rather than a tactic: "Check whether your category is already decided." It's a cheap check. It changes what you should be trying to win.
The commercial layer is still only on ChatGPT
One more finding that's easy to miss. "ChatGPT is the only engine in the study running commercial units, and its two commercial surfaces behave in opposite ways." Shopping cards concentrate where there's something to buy: "Consumer Goods draws 27.3% and Sports Brands 18.3%. Eight sectors return none at all." Ads don't follow that logic: "Fifteen of the 17 sectors sit inside a narrow band, and commercial and non commercial categories score alike."
That fits what I found when I looked at advertising in ChatGPT: ads are served based on conversational context rather than keywords, so they don't need a shoppable category to show up. For GEO planning, it means that in ChatGPT the organic citation and the paid unit can sit in the same answer. In the other six engines studied, the citation is still the only way in.
How to read a benchmark you didn't collect
Otterly closes with seven recommendations. Two of them are the ones I'd keep even if the rest of the report didn't exist: "Judge your share against your category, not the study." And: "Measure category mix, not just your own citations."
The first protects against the average. A share of owned citations that looks weak next to 53.7% can be entirely normal in a sector where company websites draw 29.6%. The second protects against only looking at yourself. If government and NGO sources account for 42.4% of citations in your category, your share of the remainder tells you less than knowing which institutions are being cited in your place.
There's a third check that applies to the report itself. Otterly, Profound and Promptwatch all sell tools that measure this, and each has a commercial reason to publish this kind of data. That doesn't make any of them wrong. It's why the agreement between them matters more than any single figure: 57% in one study and 53.7% in another, with different engines, different months and different samples, pointing to the same conclusion. The same goes for Reddit on ChatGPT, where two separate trackers found the same break on the same date with different decimals.
"AI search" is a reporting convenience
It's convenient to talk about AI search as a channel, the same way we talk about social or paid. The data from this summer keeps undoing that. The share of citations going to company pages changes by more than two and a half times depending on the industry. The weight of journalism changes by seven. One engine rewrote its sourcing in a day while five others held still.
The category you're in decides which kinds of sources get to answer. The engine decides how wide it looks. Your own pages decide the part that's left.