1 What broke — from ten blue links to a generated answer
The visible part of the change is the page. As of February 2026, BrightEdge's tracking found AI Overviews appearing on 48% of tracked Google queries, up from roughly 30% twelve months earlier — a 58% year-over-year expansion. AI Mode, the conversational interface Google launched in 2025 as a separate surface, is rolling out into the main results experience. The ten blue links still exist; they are now positioned beneath a generated answer that occupies most of the above-the-fold area on roughly half of informational queries.
The interesting part of the change is what happens to traffic. In September 2025, Seer Interactive published the largest publicly disclosed measurement on this question — 3,119 queries, 42 organizations, 25.1 million organic impressions:
- Organic click-through rate fell 61% on SERPs where an AI Overview appeared (1.76% → 0.61%).
- Paid click-through rate fell 68% on the same SERPs.
- But brands cited inside an AI Overview earned 35% more organic clicks and 91% more paid clicks than uncited brands on the same SERPs.
The implication is sharp. The middle position — visible on the SERP, not cited in the AI answer above it — is the worst place to be. The traditional rank-one organic listing is no longer a rank-one outcome. A page that ranks first organically but is uncited by the AI Overview that appears above it earns roughly a third of the clicks it used to.
This is not a near-future projection. It is current measurement of current behaviour by an industry-standard tooling vendor (BrightEdge) and a named agency analytics team (Seer). Both are published. Both quote sample sizes. Neither is contested at the direction-of-travel level: AI Overviews are common, they suppress clicks below them, and being inside the answer now matters more than being directly beneath it. Verified. The only quibble worth raising is that the Seer dataset represents a particular slice of B2B-leaning queries from 42 organizations; the magnitude of the CTR drop may vary by sector. The direction does not.
The hard-to-accept part is that the discipline most small-business sites are still optimized for — keyword-targeted pages competing for ten-blue-link ranking — is a discipline that now competes for the less valuable half of the page.
Chapter 1, in one line. AI answers now sit above the links on roughly half of queries, they cut the clicks on the links beneath them by sixty-one percent, and being inside the answer is where the traffic now lives.
2 Why keywords stopped working
The keyword-targeting playbook assumed a one-to-one relationship: a search for "best running shoes for flat feet" ran a single query against the index, retrieved ten ranked URLs, and displayed them. The page that out-optimized that exact phrase won the click. That is no longer how Google search works on queries that surface an AI Overview or AI Mode response.
Google's own current documentation describes the mechanism it now uses. In the AI Features documentation maintained by Search Central, Google states:
AI Overviews and AI Mode may use a "query fan-out" technique — issuing multiple related searches across subtopics and data sources.
One user query becomes ten to fifteen sub-queries issued in parallel. Passages are pulled from across the index — different URLs, different sites, different keyword profiles — and assembled into the answer. The page that earned its position by hyper-targeting the original phrase is competing against pages that never targeted it at all, but happen to answer one of the sub-questions Google fanned out into. Citation no longer tracks one-keyword targeting. It tracks topic coverage.
This is the structural reason the old playbook hurts. And there is a measured version of "hurts" — not an opinion, a result.
The Princeton finding
In 2024, a team from Princeton, IIT Delhi, Georgia Tech, and the Allen Institute for AI published "GEO: Generative Engine Optimization" at KDD '24 (Aggarwal et al., arXiv:2311.09735). It is the most cited public study on what content patterns lift AI citation. The team built GEO-bench — 10,000 queries across 9 source datasets and 25 domains — and tested nine optimization tactics on a generative retrieval engine.
The results on the paper's headline metric (Position-Adjusted Word Count), per Table 6:
| Tactic | Effect |
|---|---|
| Adding direct quotations | +41% |
| Adding statistics | +31% |
| Citing sources (author + institution + URL) | +28% |
| Improving fluency | +28% |
| Keyword stuffing | −8% to −10% |
Across the nine tactics tested, keyword stuffing is the only one that performs worse than baseline. The old discipline — keyword density, exact-match anchors, semantic packing of the same term across the body — does not merely produce no lift in generative retrieval. It produces a measurable penalty. Verified at primary source (paper v3, Table 6). Tested on 2024-era engines (GPT-3.5-turbo + a 200-query Perplexity validation set). Whether the same lifts hold on Gemini 3.5 / GPT-5 / Claude 4 is direction-only.
The decoupling
A second piece of evidence corroborates the first. AI citation has decoupled sharply from Google organic rank in 2026:
- Ahrefs Q1 2026 (863K keywords, ~4M AI Overview URLs): only 38% of AI Overview citations come from Google's top-10 organic — down from 76% in July 2025.
- eMarketer / Profound (Dec 2025): only ~8% of ChatGPT citations and ~8.6% of Gemini citations come from Google's top-10 organic.
- BrightEdge (Feb 2026): organic / AI overlap fell to ~17%.
Two channels, not one. A page that ranks well organically is not guaranteed to be cited by AI; a page cited by AI is not guaranteed to rank well organically. The competent strategy is to ship for both — and the substrate that serves both is structured, sourced content, not single-keyword landing pages.
Chapter 2, in one line. Google now fans one query into many, so citation tracks topic coverage rather than exact-keyword targeting — and the only intervention the foundational study measured that performs worse than doing nothing is the old keyword-stuffing discipline.
"From Google Search's perspective, optimizing for generative AI search is optimizing for the search experience, and thus still SEO."
Google Search Central · 15 May 2026
Optimizing your website for generative AI features on Google Search.
3 What works now — the new substrate
On 15 May 2026 Google published
Optimizing your website for generative AI features on
Google Search
on the
Search Central developer site.
It is the most consequential piece of search documentation
Google has shipped this year, and the single sentence above is
the one the rest of the document is built around.
Optimizing for AI is still SEO.
Not a parallel discipline. Not a separate set of files (Google
explicitly says llms.txt is unnecessary). Not a
separate schema (Google explicitly says structured data is not
required for AI Overviews). The same playbook, applied to a
search experience that now generates answers as well as
retrieving them.
The same document is unusually direct about what content earns the citation. Google's own example, on what does not:
Be sure that you're writing non-commodity content that your readers will find helpful and reliable. Commodity content (for example, something like "7 Tips for First-Time Homebuyers") is often based on common knowledge, which could originate from anyone, and typically adds little unique insight for readers.
The Princeton paper, the Google guidance, and the recency data all point at the same thing. Citation eligibility is driven by a small number of properties that are visible in the rendered HTML:
- Direct quotations from named sources. +41% in the Princeton study. Wikipedia is the #1 cited domain in Google AI Mode (11.22%) precisely because it operationalizes verifiability — every claim attached to an inline citation. (Surfer SEO · Ahrefs Q1 2026)
- Concrete statistics with dates. +31%. Numbers, percentages, currency, counts, sample sizes. Vague hedging — "many businesses," "most of the time" — is what the paper measured as low-lift content.
- Named source citations (author + institution + URL). +28%. The signal that turns a passage from opinion into an extractable fact.
- Recency. Seer Interactive's October 2025 study of 5,000+ AI-cited URLs found 65% of AI-bot hits target content published in the past year, 79% under two years, 89% under three, 94% under five. Only 6% of hits go to content older than six years. A brochure site published once and never updated falls out of the 65% bucket within twelve months.
- Comprehensive topic coverage. The consequence of query fan-out is that pages which answer a topic and its adjacent sub-questions get cited across queries they never directly targeted. Depth and cross-linking beat thin single-keyword pages.
A second independent dataset corroborates: Ahrefs analyzed 17 million AI citations in 2025 and found AI-cited content averaged 1,064 days since publication versus 1,432 days for traditional organic results — a 25.7% freshness advantage on publish date. Two independent datasets, same direction, same order of magnitude. The recency bias is not a vendor talking point. Verified.
What is not the lever
Two interventions popular in 2026 SEO writeups do not, on current evidence, move the needle on AI citation. Naming both is part of the discipline.
Schema markup. In April 2026, Ahrefs published a controlled study on 1,885 pages adding JSON-LD between August 2025 and March 2026, matched against 4,000 control pages. Their finding: "Adding schema produced no major uplift in citations on any platform." AI Overviews on treated pages declined 4.6% more than controls — statistically significant. Google's own May 2026 documentation says the same thing in different words: "Structured data isn't required for generative AI search." Schema is necessary infrastructure for entity disambiguation, rich results, and the Knowledge Graph. It is not a citation hack. The Ahrefs study is a single dataset and is contested by Suganthan Mohanadasan and SchemaApp on grounds that it only measures one of three schema effects; we treat schema as hygiene — required — and decline to sell it as a growth lever. Single-source.
Core Web Vitals as an AI-visibility growth lever. Dan Taylor's January 2026 Search Engine Land analysis of 107,352 pages found a small negative correlation between LCP and AI citation (r = −0.12 to −0.18) and a smaller one for CLS (r = −0.05 to −0.09). His conclusion is the clearest available framing: Core Web Vitals are a gate, not a signal of excellence. Severe failure suppresses AI citation; going from "good" to "great" does not lift it. The dataset is single-contributor and undisclosed in its methodology, so treat the magnitudes as direction-only. The direction matches the rest of the picture. Directional.
A handful of 2026 SEO blogs (Digital Applied first, then several downstream) claim Google tightened the LCP "Good" threshold from 2.5s to 2.0s with the March 2026 core update, citing a "Search Central blog post" that does not exist. Google's official Core Web Vitals documentation, last updated 10 December 2025, still states LCP ≤ 2.5s. Treat the 2.0s claim as unconfirmed industry rumour until Google publishes otherwise.
Chapter 3, in one line. Citation eligibility runs on five properties — direct quotations, concrete dated statistics, named source attribution, recent publication, and comprehensive topic coverage — none of which is a schema-markup or speed-score hack.
4 What this means for a small business
The small-business web is, for the most part, still optimized for the 2019 playbook. Five service pages. A list of locations served in the footer. A blog updated quarterly with the kind of commodity content Google's May 2026 documentation names by example. A title tag tuned for the local keyword. The site is not bad. It is competing in a category that has been reorganized.
In late 2025, Whitespark's 2026 Local Search Ranking Factors report — the practitioner survey that has set the local-search agenda since 2008 — added "AI Search Visibility" as a formal local ranking category for the first time. Structured data, consistent citations, and curated list mentions were named as direct inputs. "Dedicated page for each service" was the top-ranked factor in Local Organic and the second-ranked factor in AI Visibility. Local search is not a separate channel from AI search any more. They have begun to share the same scorecard.
The asymmetric opportunity follows from that. A small business that ships structured, sourced, date-stamped content competes against a field of competitors whose sites still treat content as long-form prose stapled to a template. The bar to clear right now is not "publish more." It is "publish things AI engines can extract" — direct quotes from named sources, regional or sector-specific statistics with explicit dates, a dedicated page per service instead of a single bundled list, and a visible published / updated date on every page. None of this requires a new platform. It requires editorial discipline that most small-business sites are not currently practising.
The pessimistic version of the same point: Ahrefs' 14-billion-page study, published 2025, found that 96.55% of pages get zero traffic from Google. Most of the long tail is not in the game. The long-tail pages that are in the game share a small number of properties: they are well-maintained, they are structurally clean, their freshness is current, and they answer questions in extractable form. Those properties have not changed; only the engines reading them have. A small business with one or two of those properties is closer to the front of the field than the market structure suggests.
None of this implies a rebuild. The audit-then-edit version is cheaper and faster and clears the field-measurable problems in most cases. The question for a small business in 2026 is not whether its site is fast enough — measured against the gate, it probably is. The question is whether the content on that site is extractable. Most of it, on most sites, is not.
Chapter 4, in one line. The bar to clear is not "publish more" but "publish things AI engines can extract" — and the field of small-business sites that have done so is small enough that one or two disciplined pages put a business closer to the front than the market structure suggests.