Google and ChatGPT just moved citations to the bottom of the answer

Two competing AI search products ran citation experiments in the same week, and both moved in the same direction: away from the top of the answer.

Barry Schwartz reported on Search Engine Roundtable on July 22 that both Google AI Mode and ChatGPT are running live tests that relocate citation cards away from inline placement, in the same week, independently.

What changed

Sachin Patel spotted the Google side of it first: opening a free-listing product in AI Mode now sometimes anchors the user straight down to a block of citation cards sitting at the bottom of the page, instead of surfacing sources inline as the answer is generated.

Brodie Clark, writing on SERP Alert, caught OpenAI running a different mechanic toward a similar end. ChatGPT is testing an overlay: instead of anchoring the page down, a click on a citation now sometimes pops a box over the answer rather than jumping straight to a visible source list.

Two different mechanics, one direction

Google is testing an anchor-and-scroll. OpenAI is testing an overlay-on-click. The interaction patterns do not match, but the effect is the same: the source stops living next to the specific claim it backs, and moves somewhere the user has to go looking for it.

Schwartz raised the obvious question in his write-up: will this move mean more clicks or fewer? Nobody outside Google and OpenAI has that answer yet, and neither company is saying.

This isn't a one-off

Read alongside Google's other May move, this is not an isolated interface experiment. In May, Google expanded Preferred Sources into AI Mode and AI Overviews, and reported that users had already selected over 345,000 unique sources, with a Preferred Source getting clicked twice as often as an unpicked one. Google also rolled out a Perspectives carousel and a wider Highly Cited label. Every one of these changes reshuffles where and how a source gets seen. None of them puts a plain citation link back at the top of the answer, next to the claim.

Taken together, the pattern across 2026 has been consistent: Google and OpenAI keep adding structure around citations, curation layers, trust signals, carousels, without giving up any of the real estate a plain inline link used to occupy. Each addition is defensible on its own. A Perspectives carousel surfaces genuinely useful firsthand context. Preferred Sources gives users real control. But stacked together, they are all pulling attention away from the one thing GEO practitioners have spent two years optimizing for: a clean, inline, attributable citation next to the claim it backs.

What placement does to answer engine optimization

Most answer engine optimization work still treats a citation as binary: you're in the answer, or you're not. Placement testing breaks that model. A citation buried behind a scroll-and-anchor, or hidden behind an overlay click, still counts as a citation in almost every GEO tracking tool on the market. It does not behave like one for the buyer who never scrolls that far or never clicks the overlay.

This compounds a problem we've written about before: production citation counts already under-report a brand's real retrieval footprint, sometimes by a factor of three to ten, because so much of what an agentic system reads never makes it into the visible answer. Now add a second layer, where even the citations that do surface can sit somewhere the user has to work to find. Two different failure modes, both invisible to a dashboard that only checks whether you're cited at all.

There's a reason engines are willing to run this kind of interface test in the first place: they have room to experiment because trust in the underlying answer is already established through other signals. Ahrefs' 75,000-brand study from August 2025 found brand mentions correlating with AI visibility at 0.664, against just 0.218 for raw backlinks, and the top quartile of brands by mention volume pulled in roughly 10 times more AI citations than the next quartile down. Engines are increasingly confident about which brands deserve to be in the answer at all. What they're now iterating on is how much friction to put between the reader and the proof.

What to do about it in your answer engine optimization plan

Track citation placement, not just citation presence, for your highest-value queries. Manually check whether your source link is inline, anchored below the fold, or hidden behind a click, and log it alongside your usual mention-rate tracking, because a blended visibility score will not show you this shift. Push for Preferred Source status where your brand already has trust: getting picked once by an engaged user roughly doubles the odds of a click every time your content is cited afterward, per Google's own numbers.

None of this changes the underlying GEO fundamentals: machine access, extractable structure, and third-party authority still decide whether you're in the answer at all. It just means presence is no longer the finish line. Where the citation sits, and how much friction stands between the buyer and your link, is quickly becoming its own metric, one we're already building into the reporting we run for clients alongside rank and crawl data from search console.

The stakes for click behavior were already high before this test. Similarweb data from July 2025 found zero-click searches had grown from 56% to 69% within a single year of AI Overviews launching broadly, with mobile searches ending without a click at 77.2%. A Botify and Demandsphere study around the same period found AI Overviews and Featured Snippets together consuming roughly 67.1% of desktop screens and 75.7% of mobile screens before a user even reaches a traditional result. Against that backdrop, a citation that requires an extra scroll or an extra click isn't a minor UX tweak. It's competing for attention that was already scarce and getting scarcer.

Growth Memo's own May research on user behavior backs this up from the demand side: a clickstream analysis of 846,000 sessions found AI Overviews functioning as a comparison environment, where users cross-check sources before committing, while AI Mode drives faster, more direct acceptance of whatever the model surfaces first. If AI Mode is where users accept the first thing they see, moving the citation away from the point of acceptance is not a neutral interface change. It's a test of whether trust in the summary is now strong enough to survive without an easily visible source right next to it.

Practically, that means a manual weekly check on your top ten buyer prompts is no longer optional busywork. Screenshot the answer, note whether the citation is inline, anchored below a scroll, or hidden behind a click, and compare week over week. Two engines moved in the same direction in the same week. There's no reason to assume this is the last placement test either of them runs before the end of the year.

It's also worth watching whether the two engines converge on one pattern or keep diverging. Right now Google is testing anchor-and-scroll and OpenAI is testing an overlay, two genuinely different mental models for the user. If one pattern proves to convert better, the industry will likely see the other engine adopt something similar within a quarter or two; that's been the pattern with nearly every other AI search UI experiment so far, from citation cards themselves to Preferred Sources-style curation layers. Don't optimize your tracking process around one specific mechanic. Optimize it around the general shift toward more friction between the claim and the source.

Build the habit now rather than after a quarter of confusing attribution data. Teams that wait until a placement change shows up as an unexplained dip in referral traffic lose weeks figuring out whether the problem is their content, their crawl access, or, as turned out to be the case this week, an interface test they never knew was running. A five-minute weekly screenshot habit is cheap. A quarter spent debugging the wrong variable is not.

See where you are cited today

A free snapshot audit of your rankings and AI citations before we ever talk.