Citations reached 6.8 percent of ChatGPT prompts, and the visit landed on the homepage

How often does a ChatGPT answer carry a citation?

Similarweb's AI search statistics page, which carries a date of 29 July 2026, states that its data shows the presence of citations in US ChatGPT prompts rising from about 1.6 percent in June 2025 to roughly 6.8 percent by May 2026, more than quadrupling over 11 months.

There are two honest readings of that sentence and they point in opposite directions. The first is the trend: a rate that more than quadruples in under a year is a platform changing its behaviour quickly, and anyone planning for 2027 should assume the line keeps moving. The second is the level: if citations were present in roughly 6.8 percent of US prompts in May 2026, then something close to 93 in every 100 carried none. A citation count is a measurement taken inside a small and unevenly distributed subset of what the platform does. Growth and scarcity are the same number here, and a vendor quoting only one of them is choosing which half to show you.

This site has already set out the test to apply to a figure like that. Writing about the difference between a Gemini citation tool and an AI Overview tool, it argued that when you see a citation rate from anybody, the questions to ask are what the denominator was, how many engines it covered, and whether opaque URIs were counted or quietly removed, because those three choices move a headline percentage further than the underlying reality does. That standard applies to a number this post is reporting as much as to anyone else's, so here is the answer as far as the page gives it. The denominator is US ChatGPT prompts observed in Similarweb's panel. It is one engine, not a cross-platform average. The page describes its inputs as Similarweb data and a worldwide traffic panel, and it does not publish a panel size, a prompt sampling method or a definition of what counts as a citation for this purpose. That is a real gap, and it is the reason this post reports the figure as Similarweb's rather than adopting it.

The sector breakdown on the same page is the more actionable half, and it is rarely quoted. Travel and Hospitality prompts carry a citation about 23 percent of the time, the highest rate the page reports, followed by Automotive at roughly 20 percent. Professional Services sits under 4 percent. The spread between those two ends is close to six to one, and none of it is about the quality of any individual site. It is about what kind of question people bring to the assistant in that category and whether the answer needs a source to be useful. Anyone doing generative engine optimization in a low citation category is competing for a much smaller pool of opportunities than the headline rate suggests, and no amount of on-page work changes the size of that pool.

One scale figure is worth keeping separate rather than multiplying into the others. The same page reports average monthly web visits across generative AI platforms worldwide growing 70 percent year over year to 9.5 billion between June 2025 and May 2026, with unique visitors growing more slowly, by 57 percent, to 655 million. That is worldwide traffic across platforms. The citation rate is US ChatGPT prompts. The two are measured on different populations and combining them into an estimate of citation events would be arithmetic dressed as evidence.

What ChatGPT changed on 7 May 2026, and what it did to referral traffic

The second figure comes from a separate Similarweb analysis, ChatGPT referral traffic near triples overnight, published on 25 May 2026 and written by Adelle Kehoe. It describes a product change rather than a gradual trend. On 7 May 2026, the post states, ChatGPT began surfacing more prominent links to brands directly within its answers, so that rather than burying source links in footnotes or follow-up suggestions, brand names became clickable callouts placed in the body of the response.

The measured effect is large and it is uneven. Week on week, the post reports total ChatGPT referrals increasing by 157.7 percent, and homepage referrals, which it defines as traffic landing directly on brand root domains, increasing by 354.7 percent. The share moved with the volume: before 7 May, roughly 26 to 32 percent of ChatGPT referrals arrived at a brand's homepage, and after 7 May that figure jumped immediately to around 60 percent and has stayed there consistently.

The method is stated and it is worth repeating because it bounds every one of those numbers. The analysis uses Similarweb clickstream panel data across tracked websites from 30 April to 20 May 2026, desktop panel only. Desktop only is the constraint that matters most, because assistant use skews heavily to mobile and to native apps, and neither is in this measurement. A twenty-one day window either side of a single product change is also a short base for the word consistently. What the analysis supports is that the shift happened and held for roughly two weeks on desktop, which is a narrower claim than a permanent change in how the platform sends traffic.

The two Similarweb pages do not state the pre-change homepage share identically. The May analysis gives roughly 26 to 32 percent; the July statistics page gives roughly 26 to 29 percent before the update and about 62 to 63 percent by late May, described as based on weekly referral data. The gap is small and both bands overlap, but they are not the same numbers, and this post is not going to silently pick the tidier one. Where the two disagree, the figures above are from the May analysis, which is the page that describes the underlying study.

A third Similarweb page, published on 23 June 2026, measures what happens after the visit rather than the visit itself. It followed thousands of real user journeys across Finance, Travel and Beauty, tracking users who asked ChatGPT an industry question, received a brand recommendation, and were then observed for seven days, with users excluded if they had previously visited the brand or named it in the prompt. It reports that those users were 2.5 times more likely to visit that brand's website in the seven days that followed, that 56 percent of those arrivals came through branded search against roughly 40 percent for standard visits, and that they viewed nearly twice as many pages and spent twice as long on site. The exclusion rule is the part that makes it interesting rather than circular. The branded search finding also implies that a referral count is a floor rather than a total, since a visit that arrives as a branded search a day later carries no ChatGPT referrer at all. Similarweb states the consequence plainly, that none of this shows up properly in your analytics yet, which is an unusually direct admission from a company selling the measurement.

A citation, a click and a crawl are three separate events

Put the two findings together and the shape of the problem changes. A citation is awarded to a specific URL, usually a deep page that answers a specific question. A visit, after 7 May 2026, lands on the root domain roughly six times in ten. Those are different pages, reached by different clients, and they fail for different reasons.

Start with the fetch, because it comes first and is the only one of the three that a site owner directly controls. OpenAI's crawler documentation names three agents and gives each a distinct job. GPTBot is used to make its generative AI foundation models more useful and safe. OAI-SearchBot is used to surface websites in search results in ChatGPT's search features. ChatGPT-User is used for certain user actions in ChatGPT and Custom GPTs, and the page states that when users ask ChatGPT or a CustomGPT a question, it may visit a web page with a ChatGPT-User agent.

One sentence on that page deserves more attention than it usually gets. Because those actions are initiated by a user, OpenAI states, robots.txt rules may not apply. The fetch that supports a live answer is therefore not necessarily governed by the file most people think of as the control surface. Resolving your file against a named token, which is what the robots.txt tester does, tells you what a compliant crawler is asked to do. It does not tell you what a user-initiated fetch will do, and the population of clients that never announce a name you would recognise is a separate problem covered in the crawler tokens that never appear in your logs.

The second event is selection, and this site has written about it before from a research angle rather than a panel one. A measurement framework covered in more AI citations did not mean more of your page in the answer splits the problem into citation selection, where a platform picks sources, and citation absorption, where a cited page actually contributes text to the answer, and reports that the platform citing most broadly had the lowest mean influence per cited page. That work makes clear that being cited and being used are already two things. The Similarweb referral data adds a third: being visited. A page can be fetched and not cited, cited and not read, cited and read and never visited, or visited without ever having been the cited page at all. Only the last of those is new, and it is the one the 7 May change made common.

The practical consequence is that a single number cannot carry the diagnosis. If your citations are flat, the question is whether you were fetched. If your citations are up and your traffic is not, the question is whether the answer was complete enough that nobody needed to click, which is the ordinary condition of an assistant answer rather than a fault. And if traffic is up on the homepage while the cited pages are unchanged, nothing about your content changed at all: the interface did. Platform-specific behaviour of this kind is why the guidance for getting cited in ChatGPT is kept separate from the guidance for every other engine.

Why a homepage landing puts rendering on the critical path

If roughly six in ten ChatGPT referrals now arrive at the root domain, then the homepage is doing work it was not previously asked to do. It is receiving a visitor who has already been told what your company does by a third party, who arrives with an expectation formed elsewhere, and who has never seen the page that earned the citation.

That matters for readability in a way a deep page does not, because homepages are where the marketing stack concentrates. They carry the animation, the personalisation, the carousels and the interactive elements, and on several popular frameworks they are the page most likely to ship as an application shell that assembles its text in the browser. A client that does not run JavaScript sees whatever the server returned, and if that is a shell, it sees nothing worth reading. The gap between the text a browser renders and the text a crawler receives is the condition prose parity is named for, and it is checkable in a single request: what GPTBot sees fetches a page the way a crawler would and shows the text that survives. On a rendering-heavy stack the specific remedies differ, which is why the Next.js guidance is written separately from the rest.

This is the failure that scanning exists to catch, and it is worth being exact about the evidence for it rather than implying more. A capture of real storefronts published here in what a crawler actually meets on a real storefront reports what was measured, on which pages and on what date. This post is not extending those measurements to homepages in general, and Lantad has published no figure for how often a homepage fails a crawler fetch. The claim here is narrower and is about consequence rather than prevalence: whatever your homepage's rendering behaviour is, the 7 May change raised the cost of getting it wrong, because a page that used to be a starting point for human visitors is now the landing point for a majority of assistant referrals.

There is a second reason a homepage landing changes the requirement. A visitor arriving on a deep page has already been given the specific answer and needs the specific page. A visitor arriving at the root has been given a recommendation and needs to confirm, quickly, that this is the company that was recommended. That is an identity question rather than a content question, and it is the one entity confidence describes: whether a machine, and now a person arriving with a machine's summary in their head, can tell who this is, what they sell and where they operate from the page in front of them. The structural signals that carry that information are set out in five structural signals that tell an AI who you are, and they are cheap to get right and easy to leave implicit on a homepage designed for people who already know the brand.

Both requirements sit on top of the older one, which has not gone away. Access and rendering are separate layers and a site can pass one and fail the other, the argument made in two layers decide whether AI can read your site. Admitting every crawler in robots.txt is worth nothing if the response body is empty, and a perfectly rendered homepage is worth nothing to a crawler that was refused at the edge.

What a clickstream panel cannot tell you about your own domain

Everything above is a population measurement, and the useful response to a population measurement is to stop generalising from it. None of these figures describe your domain, and several of them cannot in principle.

The referral analysis is desktop panel data over a twenty-one day window on tracked websites. It cannot tell you your own homepage share, because your category, your brand strength and your site's structure all move that number, and a panel average absorbs all three. The citation figures are US prompts on one engine, so they say nothing about Gemini, Perplexity or Claude, and nothing about any market outside the United States. The seven day journey study excluded users who had already visited the brand, which is the right call for isolating the effect and also means the result does not describe your existing customers. And Similarweb's own caveat, that none of this shows up properly in analytics, applies with full force to any attempt to verify these numbers in your own reporting: a visit that arrives through branded search after an assistant recommendation is indistinguishable in most analytics from a visit that would have happened anyway.

This site holds itself to the same limit rather than filling the gap. Lantad runs no prompt panel and publishes no citation rate, and the research page reports scan composite and sub-score aggregates and grade bands rather than citation statistics, because the sample that would support a citation statistic does not exist here. What a scan does measure, and the boundaries of it, are written down in the methodology: it resolves a real robots.txt against a registry of crawler tokens, fetches pages the way a crawler would, and reports what was readable and what was not. That is one input to an AI visibility assessment and it is not a prediction of citations. When a measurement cannot be taken, the honest output is to say so and withhold the grade, the position argued in why we will not grade a page we could not measure.

Three checks follow from these findings and each is about your own domain rather than the population. First, fetch your homepage the way a crawler would and read what comes back, because the 7 May change made that page the landing point for the majority of ChatGPT referrals in the panel and rendering is the failure mode most likely to affect it. Second, check which named agents your site actually admits, since the fetch behind a live answer may come from a client OpenAI documents as possibly not bound by robots.txt, and the AI crawler reference enumerates the tokens worth resolving. Third, separate your own numbers the way the events separate: fetches in your server logs, citations in whatever tool you use to observe answers, and landings in your analytics. If those three move together you have a story. If they move independently, which after May 2026 is the more likely outcome, then a single citation count was never going to tell you which one changed.