Being cited by AI won't save your traffic (and the numbers say so)

Someone is looking for exactly what you sell. They ask their question, they get a written answer at the top of the page, they read it and they close the tab. Your site may well have been ranking first. They will never know, and neither will you, because a visitor who never arrived leaves no trace in your analytics.

A small industry has grown around this, with its own acronym (GEO, for Generative Engine Optimization), its tools and its recipes. The reasoning fits in one line: since the AI cites sources, arrange to be the cited source. It is intuitive, and it is what everyone repeats.

The trouble is that the available data tells a distinctly less comfortable story. Here is what it actually says, starting from nothing. You need no SEO background to follow what comes next.

For twenty years a search worked the same way every time: you typed a question, you got a list of links and you clicked. The engine pointed, the sites answered. Everyone got something out of it, the engine had its audience and you had your visitors.

That contract has changed. On a growing share of searches, Google now shows a written answer above the links, the AI Overview, with a few sources cited off to the side. It also offers a fully conversational mode, AI Mode. And in parallel, some people no longer go through a search engine at all: they put the question straight to Claude, to Perplexity or to another assistant.

In every one of these cases the answer is assembled from existing content, potentially yours. You supply the raw material. The question is what you get back in exchange.

On the scale of the phenomenon, the most honest measurement comes from the Pew Research Center (opens in a new tab). Its method is what makes it interesting: rather than asking people what they do, the researchers observed the real browsing of 900 US panellists running a tracking tool, across 68,879 Google searches in March 2025. This measures behaviour, not self-reporting.

These figures date from March 2025 and the phenomenon has grown since. They are enough to set the scene: when a written answer appears, a large share of the traffic that used to reach you no longer does.

So far, nothing very surprising. What follows is a great deal more so.

In April 2026, the consultancy Seer Interactive published an analysis (opens in a new tab) that makes exactly the comparison nobody makes. Instead of asking "does an AI Overview hurt?", it asks the real question: when an AI Overview is there, what difference does it make to be cited or not? Here is the result, on informational queries, scaled to one million organic impressions.

Read the second row slowly. Being cited means losing 38% of your clicks compared with a world where the AI Overview simply wasn't there. The citation is not a win, it is a softened loss. You don't take the race, you lose it less badly than the next site.

That is the unspoken part of the whole conversation. Optimising for the citation gets sold as the new grail, when the best outcome available inside that frame still sits below the old status quo. Being cited is genuinely far better than not being cited, the gap is roughly double. But the match is now played between a bad outcome and a less bad one.

One honest caveat on this number: it comes from data observed across a portfolio of client sites, not from a controlled experiment. Seer says as much itself, this is a correlation, it does not prove that the citation causes the extra clicks. The direction of the gap is solid, the decimal precision is not. Don't treat these as physical constants, treat them as an order of magnitude pointing somewhere very clear.

Another marker points the same way. Ahrefs compared (opens in a new tab), across 300,000 keywords, the real click-through rate of position one against what would have been expected without AI Overviews. In April 2025 the gap was 34.5%. In the February 2026 update it reaches 58%. The trend is not turning around.

The natural reflex: "fine, just work on the SEO and the citation follows". Except the link between the two is far looser than people assume.

A team of researchers (Grossman and co-authors (opens in a new tab), work accepted at SIGIR 2026) measured the overlap between the sources cited by generative engines and the classic organic results for the same questions. The score lands between 0.11 and 0.18 on an index that reads 1 when the two lists are identical and 0 when they share nothing. Translation: the two worlds barely overlap. Ranking first does not buy you the citation, and being cited does not require ranking first.

Where it gets genuinely interesting is that Google says the opposite. Its official documentation (opens in a new tab), updated on 10 July 2026, states plainly that its generative AI features are built on the same ranking and quality systems as classic search. Meaning: do good SEO, there is no separate discipline.

The two statements aren't necessarily contradictory. The same ranking foundation can feed an additional selection step that keeps only a handful of sources on criteria of its own, and it takes very little for the final lists to diverge. But for you the practical consequence is identical: your ranking is not a reliable indicator of your presence in AI answers. The two have to be looked at separately.

A vacuum always creates a market. Here are the most heavily sold approaches and what the data says about them.

The llms.txt file. The idea: drop a file at the root of your site explaining to AI systems what your site contains. Appealing on paper. In practice, a server-log analysis counted 84 requests to that file out of 62,100 AI bot requests, or 0.1%. Nobody goes looking for it. Google states in its documentation (opens in a new tab) that it ignores the file and that creating one will "neither harm nor help" your visibility. The delicious detail: Google itself published an llms.txt on its own developer documentation on 3 December 2025, then pulled it the same day.

Schema.org markup. Useful for other reasons, but Google explicitly writes (opens in a new tab) that structured data is not required to appear in its generative features and that no AI-specific markup exists. If someone is selling you schema on a promise of citations, the promise rests on nothing.

Content written for machines. The academic work on this (opens in a new tab) shows that stuffing a text with keywords performs worse than leaving it readable. Which ought to be reassuring: content a human enjoys reading is not penalised, it tends to be favoured. Writing for the AI at the reader's expense loses on both counts.

This is where it gets concrete, and it is the part the marketing conversation ignores entirely.

AI crawlers do not execute JavaScript. This is the finding with the heaviest consequences. The bots of the main players (opens in a new tab) fetch raw HTML without running the page's JavaScript. They even download the JS files without executing them, which achieves nothing. Concretely: if your site builds its content in the browser, which is what a standard React or Vue application does without server-side rendering, those bots see an empty page. You are not ranking poorly in AI answers, you are absent from the corpus. Google, for its part, has rendered JavaScript for years, which is why a site can be in perfectly good health on the SEO side while being invisible on the AI side.

That measurement dates from late 2024 and crawlers evolve, so treat it as a signal worth checking rather than an eternal truth. Checking is easy: view your page's raw source (Ctrl+U) and look for your text in it. If it isn't there, JavaScript is building it. On what that implies and how to fix it without rebuilding everything, there is a whole article: progressive enhancement, when JavaScript fails (opens in a new tab).

The "block AI" button trap. Cloudflare, which sits in front of more than 20% of web domains, changes its defaults (opens in a new tab) on 15 September 2026: on new domains, bots classified as "training" and "agent" get blocked by default on pages that display ads, while search bots stay allowed. The trap is elsewhere: switching on the block for the "training" category also blocks Googlebot, Applebot and BingBot, which are multi-purpose crawlers. You tick a box to stop AI from hoovering up your content and you drop out of Google's index. If anyone touches these settings on your side, check that specific point.

Citations often point at nothing. A telling detail: the bots of two of the main assistants were spending 34.8% and 34.2% of their requests (opens in a new tab) on 404 pages, against 8.2% for Googlebot. These systems work partly from stale or guessed URLs. One direct consequence for you: never break your URLs without a redirect, or you risk being cited towards a dead page.

A fair question: how does anyone know where they stand? Short answer: very poorly, and it's better to know it.

On 3 June 2026 Google launched a dedicated generative AI report (opens in a new tab) in Search Console. It is real progress, after long saying it wouldn't build one. But look at what it gives and above all at what it doesn't.

Add to that a subtlety that breaks naive readings: when one of your URLs appears both in the AI Overview and in the classic links of the same search, it counts as a single impression. So the report cannot isolate what the AI adds on top.

On the analytics side it is worse. A visitor arriving from an AI Overview lands tagged as an ordinary Google organic visit, strictly indistinguishable from the rest. And a good share of assistant traffic shows up nowhere at all: people copy the address and paste it into a new tab, which reaches you with no referrer and gets filed under direct traffic. Mobile apps that open links in an embedded window strip the referrer too.

The conclusion: anyone quoting you a precise "AI traffic" figure is selling an estimate dressed up as a measurement. The one place you see something tangible is your server logs, where the bots pass under their real names: GPTBot, ClaudeBot, PerplexityBot. That doesn't tell you whether you are cited, only whether anyone came to read you.

No miracle recipe, and that is precisely the message. But a few conclusions hold up.

And one last thing, which matters more than the rest. Everything just described, from the crawler that can't read your JavaScript to the bot citing a broken URL, describes machines coming to fetch something from you. They manufacture nothing. They rephrase what somebody took the trouble to write and publish somewhere.

That somewhere is your site. It remains the only place where you decide what gets said, how it gets said and what happens when someone arrives. Generated answers summarise you without asking, and hand the reader a flattened version of what you are. The visitor who walks through the door anyway has become rarer, and therefore more valuable. This is no moment to greet them with a lukewarm page.

So the conclusion is not "the site matters less". It is the opposite. When discovery happens elsewhere, what happens once someone is with you carries the whole weight of the relationship. Traffic becomes a variable you only partly control. What the visitor experiences once they arrive is still entirely yours.