The answer layer
Zero-click, version two
A featured snippet was one slot with one winner. An AI Overview cites about five sources at once. That sounds like a small difference and it is not: across six domains I pulled, the citable surface roughly doubled while the number of slots a brand can actually win went up about twenty-fold.
Version one: the snippet
Rand Fishkin named the zero-click search around 2019, and the featured snippet was its face. Google lifted a paragraph out of a page, put it above the results, and the searcher got their answer without a click. If you owned the snippet you got the brand impression and some of the traffic. If you did not, you got neither.
The share of searches ending this way has only grown. SparkToro's clickstream analysis put US zero-click Google searches at 68.01% in the first four months of 2026, up from 60.45% in 2024 and roughly 45% a decade ago. Rand Fishkin, who named the pattern, draws a blunt conclusion from it:
Invest in marketing on platforms you don't own or control. Free yourself from the goal of directly driving traffic back to your website. Promote your brand (subtly). Mention your product (when relevant). But don't obsess over link inclusion. The people who are truly interested will seek you out.
The important structural fact about that era: one query, one snippet, one winner. Being the second-best answer on a page was worth approximately nothing. That is why snippet optimization looked like a lottery, and why so few brands ever held many of them. Elastic ranks for over sixty thousand keywords in the US and at its peak held 114 featured snippets.
Version two: the answer
AI Overviews replaced the quote with a synthesis, and synthesis needs more than one source. The winner-take-all slot became a set of seats. Surfer's analysis of 405,576 searches puts the average at five sources per Overview, with 90% citing eight or fewer, and found that 52% of cited sources also rank in the organic top ten for the same query. That second number is the one to hold onto: the answer layer is not a separate game from search, it is a second prize awarded on the same board.
Two clarifications worth making, because the category gets sloppy about both.
AI Overviews are Google's instance of this, not the whole of it. The broader shift is to answer engines generally: ChatGPT, Claude, Perplexity, Copilot. AI Overviews matter disproportionately for measurement because they sit inside the SERP, which means the SEO tooling that was already crawling SERPs has years of history on them. The assistants mostly do not, which is why almost every "AI visibility" product starts its clock the day you subscribe.
Presence and citation are different metrics and are routinely reported as if they were the same. "An AI Overview appeared on 48,632 of your keywords" tells you the weather. "You were named inside 3,616 of them" tells you whether you are winning. Any tool that reports only the first is measuring the surface, not your position on it.
Google disagrees, and that is worth reading
Everything above cites people whose business is telling you search changed. The obvious counter-argument comes from Google itself:
Overall, total organic click volume from Google Search to websites has been relatively stable year-over-year. Additionally, average click quality has increased and we're actually sending slightly more quality clicks to websites than a year ago... This data is in contrast to third-party reports that inaccurately suggest dramatic declines in aggregate traffic, often based on flawed methodologies, isolated examples, or traffic changes that occurred prior to the roll out of AI features in Search.
Two things can be true. Aggregate click volume across all of Google can hold steady while the click-through rate on any individual result with an Overview above it falls, because Overviews appear on only about a fifth of all SERPs and Google keeps adding queries. Ahrefs measured the per-result effect across 300,000 keywords and found that the presence of an AI Overview correlates with a 58% lower click-through rate for the top-ranking page. Note the word correlates; they were careful and so should anyone quoting it be.
This is exactly why the numbers in the next section are counts of citations rather than counts of clicks. Whatever is happening to aggregate traffic, whether an answer engine names you is observable, attributable to you, and not a matter of interpretation.
The size of the difference, measured
Six domains, all pulled from the same Semrush series. For each, the peak featured-snippet footprint against where AI Overviews sit today.
| Domain | Keywords carrying the feature | Slots the brand actually won |
|---|---|---|
| elastic.co | 22,550 → 48,632 (2.2x) | 114 → 3,616 (32x) |
| seon.io | 5,857 → 8,563 (1.5x) | 105 → 111 (1.1x) |
| flex.one | 2,149 → 2,816 (1.3x) | 49 → 166 (3.4x) |
| absencesoft.com | 1,368 → 3,837 (2.8x) | 14 → 290 (21x) |
| tines.com | 526 → 880 (1.7x) | 1 → 53 (53x) |
| revpartners.io | 221 → 1,575 (7.1x) | 2 → 49 (25x) |
Median on the left, about 2x. Median on the right, about 22x. The surface got somewhat bigger. The number of ways onto it got an order of magnitude bigger, because a synthesis has room for you even when a snippet did not.
One caveat on the arithmetic: Tines' 53x rests on a denominator of one, so read it as "effectively nothing to something" rather than as a precise multiple.
And one caveat on the metric itself, which matters more. SEON's row looks like a failure and is not. Its citations peaked at 544 in June 2025 and sit at 111 now, share down from 5.6% to 1.3%. But a count cannot tell you where those citations sit, and SEON's are concentrated in the informational layer: what is transaction fraud, how does device fingerprinting work. It is thin exactly where the buyer chooses. That is a different problem from being uncited, it needs a different fix, and no citation total will ever surface it.
One conversation now spans the whole funnel
The reason a citation count is a weak metric on its own is that the queries it counts used to be separated by weeks, and no longer are.
A buyer evaluating fraud tooling used to run three searches at three different times. "What is transaction fraud" in week one. "Best fraud prevention platforms" in week three. "SEON vs Sift pricing" when a shortlist existed. Three SERPs, three moments, three separate chances to be present, and a funnel stage you could infer from the query itself.
That is now one conversation. The prospect opens with the problem, gets educated, asks who does this well, gets a shortlist, asks how two of them compare, and arrives at a preference, without starting a new session. Problem discovery, brand awareness and high intent happen in a single thread, often in a few minutes.
Which means being cited is necessary and nowhere near sufficient. If you are named while the model explains the problem and absent when it names vendors, you paid for the education and someone else got the shortlist. The citation count records a win. The conversation records a loss. This is what SEON's row in the table actually shows, and it is invisible in the total.
Two consequences worth taking seriously.
Segment citations by stage, or the number will mislead you. Split the prompt set into problem-stage, category-stage and decision-stage questions and track the three separately. A brand with 60% informational citation and 5% decision-stage citation has a completely different problem from one sitting at 20% across all three, and an aggregate figure reports them as the same.
The decision stage is being taken from review sites. The comparison layer used to belong to G2, Capterra and TrustRadius, and to a lesser extent to analyst grids. It is moving to AI answers and to Reddit threads that those answers cite heavily. AI Overviews and Reddit are to review sites what Craigslist was to newspaper classifieds: not a competitor within the format, a replacement of the format, and the incumbents kept reporting healthy numbers well into it. If your decision-stage presence is a review-site profile you refresh quarterly, that is the asset being disintermediated.
Training data and retrieval are not the same lever
This is where the category oversells, so it is worth being exact.
Retrieval is the near-term, measurable one. When ChatGPT searches, when Perplexity answers, when an AI Overview renders, the model is fetching documents at query time and citing what it used. That is retrieval-augmented generation, and it has three useful properties: it responds to changes you make in weeks rather than years, it names its sources, and those sources can be counted. Everything in the table above is retrieval. So is anything a citation tracker can legitimately sell you.
Training data is real, slower, and not measurable from outside. What a model knows without searching came from its training corpus, and being well represented there genuinely helps. But the effect is gated by training cycles measured in months or years, you cannot observe your own share of any corpus, and you cannot verify a change took effect. If a vendor offers you a dashboard for "training data optimization," ask what it is counting. In my experience the answer is either retrieval wearing a different label, or nothing.
The practical consequence is a sequencing one rather than a choice between them. Structure your material to be retrievable and quotable now, because that is the part you can measure and move this quarter. The same work compounds into the training corpus later, since the documents that get cited and syndicated are the documents that get scraped. You just should not promise anyone a number for the second part.
What this changes about the work
If a snippet was a lottery, the answer layer is a market you can take share in. Three things follow.
Structure beats volume. Answer-first sections, explicit question headings, and schema on the pages that already rank. The edit that makes a page quotable is usually the same edit that makes it convert, which is why it tends to be the cheapest thing on the list. The mechanics are written up separately, across every engine in answer engine optimization and for Google specifically in AI Overview optimization. This page is the argument for why the work is worth doing at all.
Measure citation, not presence. One number, tracked on a fixed prompt set, weighted by the engines your buyers actually use. If the prompt set changes, the trend has to restart, or you are comparing different questions to each other and calling it progress.
Give the answer away. The instinct to gate the good part survives from an era when the click was the conversion. Amanda Natividad's framing is the one I keep coming back to:
Zero-Click content is content that offers valuable, standalone insights (or simply engaging material), with no need to click. Clicking might be additive, but it's not required.
That was written about social feeds and it applies exactly to answer engines. A model cannot cite what sits behind a form.
Do not wait for a vendor. Every engine worth measuring publishes its own citations through a documented API. I wrote the harness for this and published it here, along with the audience side that decides which engines are worth metering for a given buyer.
And a caution about whose advice you follow, including mine:
Google's guidance on AI Search is one opinion. It is the opinion of the company with the most to lose from a multi-platform world. Read it. Take what is useful. Apply it where it applies. Don't mistake it for the truth.
The same applies to this page. Everything on it is reproducible from named sources and a Semrush seat, which is the only reason to believe any of it.
Common questions
What is answer engine optimization (AEO)?
AEO is the practice of getting your brand named inside AI-generated answers rather than ranked in a list of links. It covers the same raw material as SEO, your published pages, but optimises for a different outcome: being quotable and retrievable by a model that is synthesising an answer, instead of being clickable by a person scanning results.
What is the difference between AEO and SEO?
SEO competes for a position; AEO competes for a citation. The practical difference is that ranking is exclusive and citation is not. Ten blue links have ten places, one featured snippet had one, but an AI Overview cites several sources at once, so being the third-best answer can still get you named. That is why the citation numbers in the table above are so much larger than the snippet numbers they replaced.
Is AEO the same as generative engine optimization (GEO)?
In practice yes, and the distinction is mostly vendor positioning. AEO tends to be used for anything that produces a direct answer, including AI Overviews and featured snippets before them. GEO tends to be used specifically for generative assistants. Both describe the same work: making material a model can retrieve, parse and safely quote.
What is a zero-click search?
A search that ends without the searcher visiting any website. The answer arrives in the results page itself, historically via a featured snippet or knowledge panel, and now via an AI Overview. The searcher's need is met; the publisher who supplied the answer gets an impression rather than a session.
Why do zero-click searches happen?
Because the search engine can answer the question itself, and answering is better for the searcher than sending them away. This is not an accident or a temporary policy: a result page that resolves the query keeps the user in the ecosystem. Treating it as something that will be reversed is the most expensive assumption in the category.
How do you rank in AI Overviews?
Rank first in organic for the query, then make the page liftable. AI Overviews draw heavily on pages that already rank, so conventional SEO remains the entry ticket. On top of that, the things that get quoted share a shape: a question as a heading, the answer in the first sentence beneath it, claims that survive being read alone without the surrounding paragraph, and schema that removes ambiguity about what is a question and what is an answer.
How do you track AI Overviews?
Track three things, because most tools report one. Presence is whether an Overview appeared on
your keyword. Citation is whether you were named inside it; in Semrush those are
serp_ai_overview_keywords and serp_ai_overview_positions respectively,
both monthly and both available back to September 2024. Stage is the one nobody reports: split the
prompt set into problem, category and decision questions and track them separately, because an
aggregate rate hides whether you are present at the moment someone chooses. For the assistants
beyond Google, each publishes its own citations through its API, which is what the
open harness reads.
How do you measure ROI from zero-click searches?
Stop trying to attribute them individually and measure them as a share of surface. A citation you cannot click cannot be attributed with last-touch tooling, and pretending otherwise produces numbers nobody believes. What does work: citation share on a fixed prompt set tracked over time, branded search volume, direct traffic, and self-reported attribution on forms. Those four together move when answer-layer presence moves.
Is answer engine optimization worth it?
It depends entirely on whether an AI Overview sits on the queries you care about, and that is measurable before you spend anything. For one domain in the table above, Overviews now appear on 76% of its ranking keywords; for another, on far fewer. Pull the presence number for your own domain first. If the surface is not there, this is not your priority this quarter. If it is on three quarters of your keywords, it already is, whether or not you are measuring it.
The blueprint
The method behind this page is written up separately: how the claim was chosen, where the evidence came from, why the counterexample stayed in the table, and the structural checklist at the end. It is deliberately reusable.