The whole GEO industry has spent eighteen months treating AI Search citation rates as a signal to be optimized, the way it once treated rankings. New Similarweb data says something less comfortable. The citation rate is not a signal at all. It is an inventory level, set by the platform, and it moved by a factor of five in a year.

  • ChatGPT carried a citation in 6.8% of US desktop answers in May 2026, up from roughly 1.3% in June 2025, with a drop to about 4.5% in February along the way.
  • The rate varies almost fivefold by topic: 22.6% of travel answers versus 4.8% of education answers.
  • Citation density is highest where commercial intent is highest, which is the same gradient that set CPC in paid search for twenty years.
  • In the commercial verticals, sponsored placement is already denser than organic citation, so the paid lane is bigger than the earned one where the money is.
  • News and publisher sites supply 26% of ChatGPT's citations and hold no position in that layer's rate card.

What a citation rate actually measures

An AI citation rate is the share of generated answers that include at least one external source link. It measures the engine's willingness to route a user outward, not the quality of any given publisher or brand. That distinction matters commercially, because a rate that the platform controls is inventory, and inventory that nobody has contracted for is an allowance.

Similarweb's 2026 Generative AI Landscape report, covered by Search Engine Journal on 27 July 2026, puts that allowance at 6.8% of US desktop ChatGPT answers in May 2026. Nine sectors were ranked: travel led at 22.6%, retail at 13.5%, sports at 10.7%, finance at 8.0%, with technology, health, media and entertainment, and education all below the average and education lowest at 4.8%.

Ten stable organic slots versus a 6.8% discretionary allowance

The open-web search economy ran on a structure that was gameable but stable. Ten organic positions, holding steady for two decades, plus a parallel paid market with explicit inventory, an auction, a rate card, and per-click measurement. Marketers complained about Google's control, but they could plan against it.

The agentic web inverts the risk profile. The earned layer is a discretionary percentage that rose more than fivefold in under a year, dipped by roughly a third over one winter, and recovered by spring. The paid layer, by contrast, has a fixed surface: OpenAI began placing sponsored units below ChatGPT answers on 9 February 2026. Brands and publishers are treating the earned layer as the reliable one and the paid layer as the experiment. It is the other way around.

What Search Engine Journal reported, and which numbers survive scrutiny

Matt G. Southern's report is careful, and the caveats deserve repeating rather than burying. Similarweb sells AI visibility tracking, the report sits behind a download form, the figures are described by Similarweb itself as estimates and extrapolations, coverage is US desktop only, it excludes API traffic, desktop applications, and AI features embedded in other apps, and the topic breakdown reflects a single month.

There is also a competing measurement worth flagging. Resoneo, reported by SEJ in April 2026, found ChatGPT citing roughly 20% fewer websites per response after the GPT-5.3 Instant update. The two datasets count different things: one asks whether an answer links out at all, the other counts domains inside a single answer. Read together they describe breadth rising while depth thins, which is worse for publishers than either number alone.

Why a rising citation rate is still not a channel

The strongest case for patience on earned citation

The optimistic reading is credible. A fivefold rise in twelve months is a trend, not noise. Grounding answers in sources improves accuracy and reduces liability, litigation and licensing pressure both push toward attribution, and a vendor selling citation-tracking software has an obvious interest in making the metric look decisive. If citation rates keep climbing toward double digits across all topics, the firms that built citable content early will own a compounding asset that no competitor can buy outright.

Why an allowance you cannot contract for is not plannable inventory

No media director would accept a channel where the counterparty can cut available impressions by a third with a model checkpoint and owes you no notice. That is what the February dip was. The deeper tell is the topic gradient: citation density peaks in travel and retail, the categories closest to a transaction, which is precisely the gradient that determined price in paid search. Platforms do not route the most value outward where it is least valuable. They do it where an ad market is easiest to build on top, and independent ad-tracking research reported in 2026 by OtterlyAI put sponsored placement on roughly three quarters of shopping-related ChatGPT questions. The denominators are not directly comparable to Similarweb's topic rates, so treat that as directional, but the direction is unambiguous: in the money verticals the paid layer is already the larger surface.

What this means for media buyers and for the publishers being cited

For CMOs, media buyers, and agencies: budget against inventory that exists

Stop setting citation share targets before checking whether the answers exist to be cited in. A citation programme in education or health is competing for a pool where fewer than one answer in twenty carries a link, and no amount of content changes that denominator. Pull the topic-level citation rate for your category first, size the earned pool honestly, and route the remainder into paid placement inside the answer and into the source types that actually get cited in your vertical. In beauty that is retail and e-commerce properties at over half of citations; in finance it is finance-specific sites and news outlets. There is no single playbook, and treating one as if there were is how budget gets burned.

For publishers: 26% of the answer and no line on the rate card

News and publisher sites accounted for 26.0% of ChatGPT's citations in Similarweb's May sample, second only to reviews and user-generated content at 28.9%. Professional publishing is therefore the second-largest supplier of the answer layer while holding no position in how that layer is monetized, and in travel, the citation-richest category of all, reviews and UGC took 54.1% of citations. The lesson is not to publish more. It is to price the citation itself, as disclosed sponsored inventory inside the answers your content already grounds, because the alternative is having your supply volume set annually by a dial you cannot see.

Three signals to watch through 2027

Three things will tell you when the answer layer becomes a real market. First, whether any engine commits to a citation floor or a usage-based licensing term, which would convert an allowance into an obligation. Second, whether the gap between sponsored density and organic citation density in commercial categories keeps widening, because if it does, the answer layer prices before it opens. Third, whether the topic gradient hardens: if travel and retail get citation-rich answers while education and health stay under 5%, the agentic web splits into a cited economy and an uncited one, and entire publishing sectors are excluded by architecture rather than by quality.

Conclusion

Hold on to this: AI Search citation rates are an inventory level controlled by the engine, not a ranking signal you can move, and a market built on an unpriced allowance will eventually be priced by whoever gets there first. Smalk AI exists to price it properly, as generative engine advertising that places native ads for AI agents and pays the media sources whose content grounds those answers. Watch for the first engine or coalition to publish a citation commitment with a number attached, because that is the day the allowance becomes inventory and everyone starts bidding for it.