The AI citation trap: when being cited costs you more

Google published an optimisation guide this week telling site owners that AEO and GEO are "still SEO." Four days earlier, Ahrefs published controlled-test data showing schema markup didn't move AI citations. Three days before that, Google killed FAQ rich results. Yesterday, GA4 quietly added an "AI Assistant" default channel group so you can finally see how much traffic ChatGPT, Gemini and Claude are actually sending you.

Read those four updates in the order they happened and the industry conversation looks one way. Read them as a single package and they say something different.

The thing they say, taken together, is that we've been arguing about the wrong question.

Every "should you optimise for AI citations" debate of the last eighteen months has assumed that being cited is the goal. Being cited is good. Not being cited is bad. The whole GEO discipline organises itself around increasing your citation share. But nobody has stopped to ask the prior question: what is a citation actually worth to you?

And now that GA4 will tell us — properly, by default, as a line item next to Organic Search — we're about to find out. The early signs are not great.

What we're actually measuring now

Until this month, AI assistant traffic in GA4 was a regex problem. You built a custom channel group, you guessed at the referrer patterns, you compared notes with three other people on LinkedIn who'd built slightly different regex, and you accepted that your numbers were directionally correct at best.

That changes now. Google has published a default channel group with a reserved campaign label. ChatGPT, Gemini and Claude are explicitly recognised. Sessions get classified automatically. Whatever AI assistant traffic actually looks like — volume, session quality, conversion behaviour, bounce — it's about to become legible to every business with GA4 installed, which is most of them.

This is the bit nobody in the GEO industry seems prepared for. The pitch for the last year has been "you need to be cited in AI Overviews, AI Mode and ChatGPT because that's where discovery is moving." The pitch has been allowed to float on the absence of measurement. Once measurement arrives, the pitch needs a referent. You need to be cited because… what, exactly, happens when you are?

The pitch has been allowed to float on the absence of measurement.

For some businesses the answer will be: a meaningful trickle of high-intent traffic that converts well. Great. For others it will be: a citation count that climbs while the traffic line stays flat or trends down. The second pattern is the one I expect to see most often, and it's the one nobody has a strategy for.

The Ahrefs data is more useful than people are treating it

Ahrefs tracked 1,885 pages that added JSON-LD schema and matched them against controls. Schema added precisely nothing to AI citation rates across Overviews, AI Mode and ChatGPT. I wrote about this earlier in the week and won't rehearse the whole argument, but there's a second-order finding inside the report that I think got missed.

If the citation game is the SEO game, then the question isn't "how do I get cited more.

Every page in the dataset already had more than 100 AI Overview citations before any intervention was added. These are pages that AI systems already know about, already crawl, already cite. The test isn't really "does schema help you get cited" — it's "does schema increase your citation share among pages already in the consideration set."

The answer is no. And that's a more interesting finding than the headline, because it tells us something about how AI systems decide who to cite. They aren't doing it on markup. They're doing it on whatever signals already put those pages in the consideration set: authority, links, topical coverage, brand recognition, the same things that put a page in the top ten organic results.

Which means the citation game is mostly the SEO game with extra steps. Google's new documentation now says this out loud — "from Google Search's perspective, optimizing for generative AI search is optimizing for the search experience, and thus still SEO." Gianluca Fiorelli's comparison was apt: schema is a label on a bottle already on the shelf. The label doesn't get the bottle on the shelf. The label doesn't even meaningfully change pick-up rate.

If the citation game is the SEO game, then the question isn't "how do I get cited more." The question is "given that I'm already competing on the same surface area I always was, what does a citation actually return to me that an organic listing didn't?"

The two scenarios, and only one of them is good

There are basically two ways AI citation can play out for a publisher or business.

Scenario one: citation is incremental. You get cited in an AI response, the user reads the summary, the summary isn't sufficient for their need, they click through to your site. Your AI assistant traffic in GA4 grows, your conversion behaviour from that channel is reasonable, and the citation is doing roughly what a top organic position used to do — funnelling considered intent.

Scenario two: citation is substitutive. You get cited in an AI response, the user reads the summary, the summary is sufficient, they don't click. Your content has been mined to satisfy the query. You appear in a footnote nobody clicks. The "value" of the citation is brand impression at best, content donation at worst. Meanwhile the AI Overview is occupying the SERP real estate where your organic listing used to sit, so your direct organic clicks are also down.

Both scenarios will show up in the same GA4 dashboard. The first looks like a healthy new channel. The second looks like a citation count that grows while your traffic line declines. The shape of the second pattern — strong impressions, weak clicks, weaker still revenue — is going to be familiar to anyone who watched featured snippets play out a decade ago, except this time the answer box doesn't even reliably surface the source.

The honest reading of the available data is that scenario two is more common than the industry has admitted. Ahrefs' own separate work on AI chatbot traffic, which I've referenced before, suggests that even sites with high citation counts get only modest click-through. Most queries answered in AI are queries that no longer need a click.

What the cloaking conversation actually reveals

Mike King's piece this week on JavaScript cloaking for LLMs is worth reading in this context, even if — as I argued in detail — I think the tactic is wrong for most businesses that will deploy it. The reason it's worth reading is what it implicitly assumes.

King's argument only makes sense if you've concluded that AI crawlers extract more value from your content than they return. He's not wrong about that calculation for some publishers — proprietary research, original datasets, unique frameworks. For those publishers, being indexed by an AI that doesn't drive traffic and does enable competitors to republish your work via a single prompt is a net loss. Cloaking is rational.

But notice what he's done. He's looked at the same data the GEO industry has been celebrating — schema doesn't move citations, JavaScript isn't rendered, citations don't drive traffic — and drawn the opposite conclusion. Not "how do we get cited more" but "how do we stop being mined." That inversion is the part the rest of the industry needs to engage with, even if you don't agree with the specific tactic.

Because somewhere in the middle of the King position and the GEO consultancy position is the actual question every business needs to answer: is AI citation, for my business specifically, a traffic source or a content liability?

Most agencies haven't asked that question. They've defaulted to "of course it's a source, we need to optimise for it" without checking whether their client's particular content shape — informational, transactional, branded, proprietary — actually benefits from being cited.

How to actually answer the question

Now that GA4 has a default AI Assistant channel, this becomes tractable. Not easy, but tractable.

The diagnostic isn't complicated. Look at your AI Assistant channel over the next three to six months as the data populates. Compare session quality, time on site, pages per session, conversion rate, and assisted conversion paths against Organic Search. Then look at the question that GA4 alone won't answer: are you being cited in queries you used to rank for, and what happened to the organic clicks for those queries?

That second piece needs Search Console data overlaid with manual citation monitoring. It's tedious. There aren't great tools for it yet, and the tools that exist tend to be expensive log analysis with markup. Most businesses will need to do it semi-manually for a representative sample of queries and extrapolate.

What you're looking for is the shape of the trade. If AI citations are coming with healthy click-through and good conversion, you have an incremental channel and you should keep building for it. If AI citations are coming with vanishing clicks and your organic traffic for the same queries is dropping, you have a substitution problem and the strategy needs to change.

The strategy when you have a substitution problem isn't to cloak. It's usually to move further up the funnel — to content and brand activities that AI summaries can't satisfy. Comparisons that need lived experience. Original research with proprietary data. Strong points of view from named experts. Tools, calculators, configurators. The things that require a click because the summary isn't the answer.

This is, incidentally, exactly the kind of content Google's new documentation calls "non-commodity." The example they give is illustrative: not "7 Tips for First-Time Homebuyers" but "Why We Waived the Inspection and Saved Money." One can be summarised away. The other can't, because the value is in the specific lived account.

Where this leaves GEO as a discipline

There's a version of the GEO industry that survives this transition by becoming useful. It looks like measurement, attribution, content portfolio analysis, and honest advice about which businesses benefit from citation optimisation and which don't. It looks like helping clients understand whether their content shape is incremental or substitutive in AI contexts, and adjusting strategy accordingly.

There's another version that doesn't survive, which is the version currently selling schema templates, llms.txt files, and "AI optimisation audits" that are mostly old SEO audits with a few new acronyms stapled on. That version has been running on the absence of measurement. The measurement is arriving.

Google's documentation this week — "AEO and GEO are still SEO" — was helpful in the way that an awkward family dinner is helpful. It said the quiet thing out loud. The schema markup pitch has been weakening for two years; the citation pitch survived because nobody could check the numbers. Now we can.

The honest limits

A few things this analysis doesn't account for, in fairness.

The recognised AI assistant list in GA4 is small — three platforms. Perplexity, Copilot, You.com and everything else still need custom channel groups. The default group is a start, not a complete picture. If you've already got a custom group running, keep it running alongside the default.

The data is also early. The first three to six months of AI Assistant channel data will be noisy. Some of what looks like substitution will be measurement lag — AI Overviews showing your content to users who then search for your brand and arrive via Organic. The full attribution picture takes time to settle.

And the substitution-vs-incremental framing isn't binary. Most businesses will find that some content categories work well as AI citations and others get mined. The job is portfolio-level: knowing which is which and allocating effort accordingly.

The point isn't that AI citation is bad. The point is that "more citations" is not a strategy. It's a metric, and like all metrics it can be optimised without producing business value.

What I'd actually do this quarter

If I were briefing a marketing team on this right now, the work would be:

Set up the GA4 AI Assistant channel reporting properly, alongside any existing custom channel groups, and start building a baseline. Cross-reference AI Assistant sessions against conversion behaviour and assisted-conversion paths. Don't just count sessions — assess whether they're worth anything.

Then take the top 100 queries you used to rank for in organic, check which now trigger AI Overviews or AI Mode responses, and look at what happened to your organic clicks for those queries over the last six months. If clicks are down on queries with AI presence, you have your answer about substitution.

Then segment your content. The pages that get cited frequently in AI but don't drive AI Assistant traffic are donating to a system that isn't paying you back. Decide what to do with them — gate them, restructure them, move the unique value behind a click, or accept the donation as a brand impression and stop investing in expanding the page.

The pages that get cited and drive traffic are working. Build more of those.

The pages that don't get cited at all and you wanted to be cited — that's where the genuine GEO work lives, and it's overwhelmingly the same work as SEO: authority, links, useful content, brand recognition. None of it involves llms.txt.

This is unglamorous compared to a deck about "the GEO playbook." It's also the work that survives contact with the data GA4 is about to put on everyone's dashboard. The industry has been operating on vibes for eighteen months. The vibes era is closing.