- AI Overviews now appear on roughly 48% of tracked queries (February 2026) and cut organic CTR on those queries by 61%, yet cited brands earn about 120% more organic clicks per impression than uncited ones.
- The retrieval pipeline selects passages, not pages, through query fan-out: one query expands into 8 to 12 parallel sub-queries you compete for separately.
- Visibility equals the retrieval surface: the breadth of sub-queries you cover times the liftability of each passage, gated by accessibility and trust.
- The strongest citation correlates are URL accessibility (9.5), search rank (9.4), fan-out rank (9.3), preview controls (9.2), and query-answer match (9.2), per Cyrus Shepard's Zyppy Signal analysis.
- Engineer extractable passages: definition-first openings, 134 to 167 word answers, self-contained blocks, 15 or more named entities, and headings phrased as fan-out questions.
- Measure impressions and cross-engine citations, not CTR, then convert the smaller, higher-intent click stream that survives the overview.
Table of Contents
Google’s AI Overviews now appear on roughly 48% of tracked search queries as of February 2026, up from 31% a year earlier, according to BrightEdge. Seer Interactive’s September 2025 study measured the cost: organic click-through rate on queries that trigger an AI Overview fell 61%, from 1.76% to 0.61%. Semrush, tracking more than 10 million keywords across 2025, watched AI Overview prevalence climb from 6.49% in January to a 24.61% peak in July before settling near 15.69% in November, a reminder that measured incidence depends heavily on the keyword set and the methodology.
That decline is not the whole story, and treating it as one leads marketing teams to the wrong response. Seer Interactive’s 2026 update, covering 53 brands and 5.47 million queries, found organic CTR on AI Overview queries rebounded from a 1.3% floor in December 2025 to 2.4% by February 2026. More important for budget decisions: brands cited inside an AI Overview earn roughly 120% more organic clicks per impression and 41% more paid clicks than uncited brands on the same queries. Being the answer behind the answer is now a measurable commercial position, not a vanity metric.
What most teams misunderstand is the unit of competition. Classic SEO competes for a ranking position on a single query. AI Overviews select passages across a fan of synthetic sub-queries, then assemble them. This article lays out the mechanism, introduces the retrieval surface as the model that replaces keyword rank, and walks through the implementation, measurement, and failure modes. It covers what AI Overview optimization is, why it matters now, how the retrieval pipeline works, when to invest in it, and when it actively wastes resources.
How AI Overviews actually select sources
Google’s AI Overviews use retrieval-augmented generation: rather than answering from the model’s training weights, the system retrieves passages from the live index and constructs the answer around them, citing what it drew from. The retrieval step is where selection happens, and it does not run a single search. A user query is decomposed into multiple sub-queries that execute in parallel, and your content competes separately for each one.
This is query fan-out. Google’s own description frames it as expanding one query into several sub-queries that capture different intents and pull from different sources, including the live web, the Knowledge Graph, and structured data. AI Mode runs the aggressive version, issuing 8 to 12 sub-queries for a standard prompt and hundreds for Deep Search. AI Overviews use a lighter variant for quicker answers. Seer Interactive measured Gemini generating an average of 10.7 fan-out queries per prompt. A search for “best project management tools for remote teams” does not look up that phrase. It fires sub-queries like “top project management software 2026,” “remote team collaboration features,” and “project management pricing comparison,” retrieves passages for each, then merges and deduplicates by semantic relevance.
Two consequences follow directly from this architecture:
- Retrieval is passage-level. The question is whether any single block of your content is the best available answer to a specific sub-query, not whether your page is broadly relevant.
- Rank is not the gate. Because the pipeline scores passage quality and trust rather than organic rank alone, AI Overviews frequently cite pages from positions 4 to 20 and beyond. A page with strong, specific content that has not yet earned the link authority to rank top three can still be selected.
The mechanism also explains the citation distribution. Surfer SEO’s analysis of 46 million citations found YouTube supplying 23.3% and Wikipedia 18.4%, sources that combine broad topical coverage with clean, extractable structure. Both win across many fan-out queries at once.
AI Overviews are not featured snippets, and not AI Mode
These three surfaces behave differently, and conflating them produces bad strategy.
| Surface | Source model | Linking behavior | Optimization implication |
|---|---|---|---|
| Featured snippet | Extracts from one source | Drives clicks to that source | Win one query, capture the click |
| AI Overview | Synthesizes multiple sources | Cites several, fewer clicks | Win sub-queries, earn citation |
| AI Mode | Heavy fan-out, deep synthesis | Designed to reduce external linking | Cover the topic, accept zero-click |
A featured snippet rewards a single best answer and still sends traffic. An AI Overview rewards being one of several cited passages and sends less. AI Mode, powered by a custom Gemini model since 2025 and improved alongside Gemini 3 in late 2025, fans out hardest and links least. Optimizing as if all three reward the same thing is the most common strategic error in this space.
The retrieval surface: the model that replaces keyword rank
If retrieval is the gate, then the right planning unit is the retrieval surface: the total set of fan-out sub-queries your content is eligible to be retrieved for, multiplied by how cleanly any single passage can be lifted into an answer. Breadth times liftability, gated by accessibility and trust. Visibility in AI Overviews is a function of how many sub-queries you can be retrieved for and how extractable your answer is at each one, not where you rank for a head term.
Breadth comes from topical coverage. A single page targeting one keyword gives the system one passage to evaluate. An interlinked cluster covering a topic from multiple angles gives the system many high-quality passages spread across many sub-queries. Clusters expand the surface area available for citation. Ekamoira’s research framed the inverse as a topical coverage gap: when content targets a single keyword instead of the full set of sub-questions a fan-out generates, AI systems skip it even when it ranks well in classic results.
Liftability comes from passage construction. Wellows analyzed 15,847 AI Overview results and found passages of 134 to 167 words achieving the highest citation rates, with 62% of featured content landing between 100 and 300 words. The same analysis scored semantic completeness as the strongest correlate of citation, and content with vector embedding alignment above a 0.88 cosine similarity threshold showed 7.3 times higher selection than poorly aligned content. The takeaway is operational: extractability is an editable property of a passage, not a quality you either have or lack.
The retrieval surface reframes the work. Instead of asking “what do I want to rank for,” you ask “what sub-queries will a fan-out generate for my topic, do I have a clean passage for each, and can the crawler reach all of them.” That question maps directly to a content plan.
Why ranking number one no longer guarantees inclusion
The instinct that fixing your rank fixes your AI visibility is half right and dangerous in the half it gets wrong. Cyrus Shepard’s Zyppy Signal analysis, published in May 2026, scored 23 factors correlated with AI citations across 54 experiments, patents, and case studies. Search rank scored 9.4 out of 10, second only to URL accessibility at 9.5. Ranking well remains one of the strongest predictors of getting cited, because Google’s AI Overviews lean on the existing index. But search rank shares the top of the list with fan-out rank (9.3) and query-answer match (9.2), and those two can override raw position.
Two of those factors deserve a closer look:
- Fan-out rank measures how well you rank for the sub-queries, not the head term. You can hold position 1 for “best CRM software” and still be invisible if the model generates sub-queries around “CRM with email automation” or “HubSpot vs Salesforce for SMBs” and your content does not address them.
- Query-answer match measures whether a specific passage directly resolves a specific sub-query. A page at position 8 with a tight, self-contained answer to “how does retrieval-augmented generation affect AI Overview source selection” beats a vague position 1 page that covers AI search at a high level.
This is the structural opening for challenger sites. In classic search, outranking an entrenched competitor for a commercial head term can take years of link building. In AI Overviews, a page with a better passage for an underserved sub-query can be cited next week. The cost is specificity: you have to commit to answering narrow questions completely rather than producing broad explainers that summarize cleanly and leave the user with no reason to click.
Where this approach fails: on queries with no meaningful fan-out, like navigational or pure brand searches, passage quality gives you nothing, because there is one obvious answer and the model returns it. Spending passage-engineering effort on those queries is wasted.
Engineering passages for extraction
Passage engineering is the highest-leverage editable lever, because it is the one factor under full editorial control. The target is a block of text that an answer engine can lift without surrounding context and present as a complete answer.
Build each answerable section to this spec:
- Open with the answer. Definition-first openings get cited several times more often than narrative introductions, per Megrisoft’s synthesis of citation studies. Lead the section with a direct, declarative resolution to the question, then expand.
- Hold the core answer to 134 to 167 words. This is the band Wellows found most cited. Longer passages dilute the semantic signal; shorter ones fail to fully resolve the query.
- Make it self-contained. The passage should answer the sub-query without depending on the paragraph above it. If understanding the block requires the reader to have read the previous section, the model cannot lift it cleanly.
- Raise entity density. Pages with 15 or more recognized entities showed 4.8 times higher selection probability in Wellows’ data. Name the tools, standards, people, and concepts explicitly rather than using pronouns and vague references.
- Match headings to fan-out queries. Replace generic labels like “Benefits” with the specific question a sub-query would phrase, for example “what to do when CTR drops after an AI Overview appears.” The heading tells both the reader and the retrieval system exactly what the section resolves.
Structured data supports this. For Google’s AI Overviews specifically, traditional E-E-A-T signals weigh more heavily than on other AI platforms, and 96% of AI Overview citations come from sources with strong E-E-A-T signals according to citation-factor analysis. FAQ, HowTo, and Article schema help the system parse which passage answers which question. Schema does not force a citation, and Google has deprecated rich-result display for some FAQ markup, but the structured signal still aids parsing.
What this costs: passage engineering is slower than writing prose, because every answerable section becomes a small standalone artifact with its own word budget and entity checklist. It also constrains voice. Teams that prize a distinctive narrative style will find the definition-first, self-contained format clinical. The trade is real, and on informational content built for retrieval, extractability wins.
Accessibility and the preview-control trade-off
The single highest-scoring citation factor is also the one teams most often sabotage by accident. URL accessibility scored 9.5 in Shepard’s framework, the top spot. If an AI crawler cannot fetch your URL, return a clean 200 status, and read it without an authentication wall, nothing else on the list matters. In 2026, accessible means accessible to a dozen distinct user agents, each operated by a different AI company, so server-log audits to confirm those crawlers actually reach your priority pages are now part of technical SEO hygiene.
The harder problem is preview controls, which scored 9.2. The nosnippet and data-nosnippet directives, along with broad AI-crawler blocking deployed through services like Cloudflare, reduce citation probability. Many publishers deployed those blocks after watching their content get scraped with no traffic return. The evidence Shepard reviewed is consistent: blocking carries a direct cost in AI visibility.
That creates a genuine business decision with two defensible sides:
- Block AI access. Protects content from being absorbed and summarized without attribution. Costs you presence in the answers that now mediate a growing share of discovery, and forfeits the 120% click-per-impression lift cited brands see.
- Allow AI access with permissive previews. Maximizes citation probability and the measurable click lift. Accepts that some queries will be fully resolved in the overview with no click, and that your content trains and grounds systems you do not control.
There is no universally correct answer. The decision depends on whether your content’s value is in driving site visits or in being the recognized source. For most commercial B2B sites where a citation builds buying-stage authority, the cost of blocking outweighs the protection. Make the call deliberately, because the default Cloudflare and robots configurations many teams inherited were set before AI citations carried measurable revenue.
Measuring AI Overview performance when clicks lie
The metric that defined SEO for two decades is now actively misleading. Organic CTR drops when an AI Overview appears, so a team that reads falling CTR as failing content will kill pages that are succeeding as cited sources. Click-through rate stopped being a clean proxy for visibility the moment attention started getting spent inside the overview before any link is reached.
Re-instrument around impressions, citations, and assisted outcomes:
- Impressions over CTR. Google Search Console surfaces AI Overview activity within the Performance report; filter by search type to isolate URLs appearing in AI Overviews and read impression and click data for cited pages. Track whether impressions hold even as CTR on those queries falls, which indicates you are being shown rather than dropped.
- Citation tracking across engines. AI Overviews are one surface. ChatGPT, Gemini, Perplexity, and Copilot all run their own fan-out and citation behavior. Tools that monitor brand mentions across these engines give a fuller visibility picture than GSC alone, because a single passage often gets cited across several platforms.
- Cited-query click lift. Segment queries where you are cited from queries where you are not. Seer’s 120% organic and 41% paid lift figures are population averages; your own delta is the number that justifies budget.
- Assisted conversions and brand-query growth. Citations that drive no immediate click still build recognition. Watch for downstream brand-query volume and assisted conversions from users who saw your name in an overview, then returned directly.
A caution on measurement maturity: more citations do not automatically mean more clicks, influence, or revenue. The factors that correlate with citation are correlations drawn from public studies, not a platform formula, and a tactic that works in Google’s AI Overviews may have a weaker effect in another engine. Use the data to improve page architecture, not to promise certainty the evidence cannot support.
Where AI Overview optimization fails or backfires
Chasing AI Overview visibility is the wrong call for several query and content types, and treating it as a universal mandate burns budget.
Pure informational content built only to answer simple questions is the clearest trap. If your strategy is surface-level explainers, the overview absorbs them and the user never visits. The sites with the worst traffic declines built their content around answering simple, surface-level questions, the exact material the overview summarizes and replaces. The defensive move is depth: original data, named methodology, proof the model has to cite because it cannot generate it from elsewhere.
Transactional and navigational queries see lighter AI Overview impact, and brand queries see almost none. Commercial-investigation queries trigger overviews roughly 35 to 45% of the time, less than informational ones, and the CTR damage there is more nuanced because buyers still click to evaluate. Spending passage-engineering effort on bottom-funnel and brand terms returns little.
The costs that get ignored:
- Commoditization risk. Content optimized purely for clean extraction reads like every other extractable source, which weakens differentiation and gives buyers no reason to choose you once they do click.
- Maintenance load. AI engines favor recently updated content, so the cluster you build needs continuous refresh, not one-time publication. That is an ongoing editorial cost most teams underestimate.
- Zero-click reality. Roughly 26% of users end their session after reading an overview, and most scan only about 30% of it. Some share of your best informational work will satisfy intent without ever sending a visitor.
When not to use this approach: if your business depends on informational top-of-funnel traffic that an overview can fully satisfy, optimizing harder for citation accelerates your own disintermediation. The sounder move there is to shift the content mix toward material that requires a visit, such as interactive tools, proprietary data, and decision-stage depth, and to convert the surviving clicks far harder than before.
How Pathmonk helps you convert the shrinking AI-era click stream
AI Overviews change the shape of the traffic that reaches your site more than they change the total. Informational top-of-funnel clicks get absorbed into the overview, so a larger share of the visitors who do click through are past the research stage and closer to a decision. The strategic response to fewer, higher-intent clicks is to convert them harder, because each surviving visit now carries more weight in the pipeline.
Pathmonk runs as one AI engine across acquisition, conversion, and analytics, fed by the first-party, cookieless behavioral data your website produces. Three mechanisms map directly to the AI Overview problem.
Its intent engine classifies each visitor’s buying-journey stage in real time from on-site behavior, without cookies or consent banners, then serves microexperiences, small adaptive interactions that surface the right proof, answer, or offer at the moment a visitor needs it. A decision-stage visitor who arrived from an AI Overview citation gets social proof, a dynamic FAQ, or a one-click path to book, rather than the generic homepage that loses 80% of decision-stage visitors who leave. Because the classification is behavioral rather than segment-based, it adapts per visitor instead of bucketing traffic into broad rules, which matters when your incoming mix skews later-stage and varies by the query that cited you, the same problem as visitors who are not yet ready to convert.
On the visibility side, Pathmonk’s LEO agent works on presence inside AI search engines, and its Website MCP makes your brand and conversion actions transactable inside AI tools like ChatGPT, Claude, and Perplexity. As AI Mode is built to discourage external linking, the ability to let a transaction happen inside the AI conversation is a hedge against the click that never comes.
Measurement holds up where cookie-based analytics degrade. Pathmonk’s cookieless fingerprinting maps full buyer journeys and filters bot traffic, so you can read the behavior of AI-referred visitors even when third-party data is thinning. To prove the lift is real, Pathmonk runs a 50/50 A/B test against a control; once you are confident in the results, you can manually scale personalization toward 95% of traffic while keeping a 5% control group. The system does not declare a winner and scale on its own, which keeps the uplift figure honest. The business math is direct: if overviews cut your informational clicks while raising the intent of the rest, a conversion-rate lift on that traffic compounds against a smaller but more valuable base.
How Doctoralia lifted conversions across three markets in two weeks
Doctoralia, one of the largest healthcare booking platforms, runs across multiple countries and languages and pulls heavy organic search traffic to a site serving several distinct audiences. AI Overviews reshape exactly this profile: informational health queries increasingly resolve in the overview, so the visitors who still click through skew toward active booking and evaluation.
The constraints that made this hard:
- A single generic experience could not serve a research-stage visitor in one market and a decision-stage visitor in another.
- Manual segmentation across three markets and languages would not scale or keep pace with shifting intent.
- Rebuilding the site per market was off the table on time and cost.
The biggest available gain sat in conversion of the traffic Doctoralia already earned.
Doctoralia deployed Pathmonk’s intent-based personalization to read each visitor’s behavior in real time and serve microexperiences matched to journey stage and market. One engine ran across all three markets, classifying and adapting per visitor rather than per segment, with no developer work and no redesign.
The results:
- +82% average conversion lift across the three markets.
- Delivered within two weeks of going live.
- Achieved with zero redesign and no added acquisition spend.
That speed is the point: when AI Overviews compress the top of the funnel, the gain comes from converting existing traffic better, not from waiting on a ranking cycle.
FAQs on Google’s AI overview optimization
Do AI Overviews use the same ranking signals as classic search?
Partially. Google’s AI Overviews draw heavily from the existing index, so search rank is the second-strongest citation correlate after URL accessibility. The pipeline adds passage-level selection through query fan-out, so a clean, self-contained answer at position 8 can be cited ahead of a vague position 1 page. Optimize for both classic rank and passage extractability.
What is query fan-out and why does it matter for content planning?
Query fan-out is the retrieval technique where one user query is expanded into multiple sub-queries that run in parallel, each retrieving its own passages before the answer is synthesized. AI Mode generates 8 to 12 sub-queries for standard prompts; Gemini averages around 10.7. It matters because you compete separately for each sub-query, so topical clusters that cover the full fan beat single keyword-targeted pages.
How long should a passage be to get cited?
Analysis of 15,847 AI Overview results found the highest citation rates for passages of 134 to 167 words, with 62% of featured content falling between 100 and 300 words. Hold the core answer to that band, open with the direct resolution, and make the block understandable without surrounding context.
Does blocking AI crawlers hurt my visibility?
Yes, based on consistent evidence. Preview controls like nosnippet and data-nosnippet, and broad AI-crawler blocking through services like Cloudflare, reduce citation probability; preview control scored 9.2 in Shepard’s framework. Blocking protects content from uncompensated scraping but forfeits the roughly 120% click-per-impression lift cited brands see. Treat it as a deliberate business trade-off.
How is optimizing for AI Overviews different from optimizing for ChatGPT or Perplexity?
The mechanics are similar, all use query fan-out and passage retrieval, but citation behavior differs by engine. Google’s AI Overviews weight traditional E-E-A-T and structured data more heavily and lean on Google’s index. ChatGPT tends to add commercial modifiers like “best” and “2026” to its sub-queries, and Perplexity favors aggressive, fresh citation. A passage that wins in one may underperform in another.
Can I still get cited if I rank outside the top three?
Yes. Because the retrieval pipeline scores passage quality and query-answer match rather than position alone, AI Overviews frequently cite pages from positions 4 to 20 and beyond. This is the main opening for sites with strong content that have not yet built the link authority to rank at the top.
Why did my CTR drop but my impressions stay flat?
That pattern usually means you are being shown inside or alongside an AI Overview, so attention is spent before the click. Falling CTR with steady impressions on AI Overview queries is a signal of changed user behavior, not failing content. Segment those queries in Search Console and track citation and assisted conversions instead of CTR alone.
Does structured data guarantee a citation?
No. Schema such as FAQ, HowTo, and Article helps the system parse which passage answers which question and supports E-E-A-T signaling, but it does not force selection. Google has also reduced rich-result display for some FAQ markup. Treat schema as a parsing aid layered on top of strong passages, not a standalone tactic.
When is AI Overview optimization a waste of effort?
On navigational and pure brand queries, where fan-out is minimal and the answer is obvious, and on bottom-funnel transactional terms where overviews trigger less often. It also backfires on surface-level informational content, which the overview absorbs and replaces, accelerating your own disintermediation rather than building visibility.
Key takeaways
- AI Overviews appear on roughly 48% of tracked queries as of February 2026 and cut organic CTR on those queries by 61%, but cited brands earn about 120% more organic clicks per impression than uncited ones.
- The retrieval pipeline selects passages, not pages, through query fan-out, so you compete separately for each sub-query a query generates.
- The retrieval surface model: visibility equals the breadth of sub-queries you can be retrieved for, times the liftability of each passage, gated by accessibility and trust.
- The top citation correlates are URL accessibility (9.5), search rank (9.4), fan-out rank (9.3), preview controls (9.2), and query-answer match (9.2), per Cyrus Shepard’s Zyppy Signal analysis.
- Build extractable passages: definition-first openings, 134 to 167 word answers, self-contained blocks, 15 or more named entities, headings phrased as fan-out questions.
- Accessibility is the highest-scoring factor, and blocking AI crawlers carries a direct, measurable visibility cost; decide deliberately.
- Re-instrument measurement around impressions, cross-engine citations, and assisted outcomes, because CTR no longer proxies visibility.
- Optimization backfires on surface-level informational content and brand or navigational queries; shift toward content that requires a visit and convert the surviving, higher-intent clicks harder.