Protecting traffic from AI Overviews and zero-click searches

Your rankings can hold position one and still lose more than half their clicks. Ahrefs found a 58% lower CTR for the top-ranking page when an Overview is present, and 68.01% of US Google searches ended without a click between January and April 2026. The rank report looks healthy while the traffic report does not.

To protect traffic from AI Overviews, run a five-step procedure: diagnose the loss by isolating stable-rankings, falling-clicks pages in Search Console, triage which queries carry zero-click risk, earn citations inside the Overview with extractable 40–60 word answer blocks, build assets an Overview cannot summarize, and diversify distribution so Google is one channel among several.

How do you protect traffic from AI Overviews?

The five steps run in order because each one narrows the next. You cannot triage pages you have not diagnosed, and you cannot decide what to rewrite until you know which queries the Overview answers completely.

• Diagnose the loss: Pull pages where average position and impressions held while clicks fell, then confirm the drop in GA4 landing-page sessions.
• Triage vulnerable pages: Sort affected queries by whether the Overview satisfies the intent. Identify pages to protect, then decide which remaining pages to summarize or rewrite.
• Earn citations: Restructure the pages you keep into answer-first 40–60 word chunks that retrieval systems can lift, with a byline and a machine-readable date.
• Build un-summarizable assets: Shift production budget toward calculators and proprietary data. Also prioritize bottom-of-funnel pages where a 60-word summary cannot complete the job.
• Diversify: Grow branded search, earned mentions, newsletter, and community so a Google-side shift moves one line of the dashboard instead of the whole pipeline.

Most teams skip the first step and jump to rewriting, which is why they spend a quarter restructuring pages that were never losing clicks to an Overview in the first place.

How do you confirm AI Overviews are causing the drop?

The signature of an AI Overview loss is a page whose position and impressions stayed flat while clicks fell, and that pattern is visible in Search Console before you touch GA4. Across 53 brands and 5.47 million queries, Seer Interactive measured organic CTR of 2.36% on queries with an AI Overview against 3.82% without one in February 2026. A query can lose a third of its clicks with no ranking movement at all.

A governance software team sent a rank-decline list and asked for a recovery plan. Search Console showed the opposite of a ranking collapse. Average position had improved, impressions were down, and clicks on the refreshed pages were roughly flat. Question queries lost about twice the clicks of head terms. The Overview was taking the click. Starting from the rank report would have sent the refresh queue at ranking recovery instead of citation-ready answers.

• Export the comparison: Start in the Performance report with the Web search type, dimension by page, and set the comparison to a window before Overviews reached your market. Export clicks and impressions for both windows. Include average position.
• Identify candidates: Keep the pages where position moved less than a spot and impressions held or rose while clicks fell. Those are your candidates. A page that lost impressions along with clicks has a ranking or demand problem, which is a different fix.

Position data understates the damage because of how Google counts it. An AI Overview occupies a single position in results, every link inside it inherits that same position, and a click on one of those links counts as an ordinary click. Your page can report a flat position-two average while sitting inside a block that sends a fraction of the clicks it used to.

Then open the Generative AI performance report, which Google rolled out to all properties worldwide on August 31, 2026.

It covers AI Overviews and AI Mode together. Impressions only, broken down by page, country, date, and device. No clicks, no queries. Discover’s generative features sit in a separate report.

GA4 closes the loop. Filter sessions to google / organic by landing page over the same two windows and check whether conversions fell in proportion to sessions. If sessions dropped and conversions held, the page belongs lower on your triage list.

Scale the Seer rates onto a page holding position three with 50,000 monthly impressions. The no-Overview CTR of 3.82% gives about 1,910 clicks. The with-Overview CTR of 2.36% gives about 1,180.

That calculation represents a 38% loss with zero assumed movement in the rank report. Seer labels its own data as correlation rather than causation, so treat the figures as a benchmark for scale rather than a position-three prediction for any single page.

Which queries are most at risk of zero-click losses?

Informational queries carry the risk, and query length makes it worse. Across 146 million desktop SERPs in September 2025, Ahrefs found an AI Overview on 21.4% of informational SERPs, 4.3% of commercial, 2.1% of transactional, and 0.9% of navigational.

The same study put the trigger rate at 9.5% for one-word queries and 46.4% for queries of seven or more words. Long-tail is a trap if you read it as safe. The seven-word informational question is the single most exposed query shape you own. Seer’s client data points the same direction on format: comparison queries triggered an Overview 95.4% of the time and price, cost, or buy queries 83.4%, against 27.3% for single-word queries.

Trigger rate is half the risk score. The other half is whether the Overview completes the job or only starts it.

A query like “how to calculate net revenue retention” ends inside the Overview, because a formula and two sentences satisfy it. A query like “migrate from [competitor] to [your product] without losing historical data” triggers an Overview too, but the searcher still has to land on a page that owns the migration. Score each candidate page on both axes and sort into three buckets.

• Protect: Bottom-of-funnel and commercial pages where the action requires your product. Keep them open and fast. Make sure the Overview can cite them.
• Summarize, or rewrite or consolidate: Summarize informational pages where you can realistically become the cited source. Restructure them into answer blocks and accept a smaller, higher-intent click volume. Rewrite or consolidate generic informational long-tail where the Overview answers completely and you are not cited. Merge thin pages into one deep asset, or redirect them toward a tool or dataset from step four.

The judgment call most teams get wrong is bucket three. Before deleting pages that still earn clicks, compare their conversion value with the value of consolidating them into one asset that can compete for the Overview citation.

How do you get cited inside AI Overviews?

A citation retains roughly 62% of the clicks a query sent before the Overview existed, and no citation retains under a third. Per one million informational impressions in 2025, Seer counted about 33,500 clicks with no Overview, 20,743 when the brand was cited, and 9,445 when it was not. Seer flags the authority confound (brands that get cited tend to be brands people already click), but the gap is the reason the goal shifts from ranking to being the source.

Ranking does not get you there on its own. In March 2026, Ahrefs found only 37.9% of URLs cited in AI Overviews ranked in the organic top ten for the query, across 863,000 SERPs and four million cited URLs. The other 62% came from outside the first page. Citation is a separate target with its own mechanics.

The mechanic is RAG. The engine retrieves 40–60 word chunks from across the web, synthesizes them, and generates an answer with sources attached. Your content gets cited when one of its chunks is the cleanest standalone answer to the retrieved sub-question.

The four AEO pillars map onto that: content clarity so the chunk reads as complete on its own, E-E-A-T signals so a byline and credentials travel with it, technical accessibility so the crawler reaches it, and citation surface area so you have a chunk for every sub-question the engine fans out to.

Page-level features matter less than the chunk. In a controlled study of 3,096 pages from 486 US Overview searches, Underneath found a machine-readable date was the only feature of 14 tested that stayed significant after correction, at +7.9 points. Question-style headings and HTML tables did not clear the bar. Put dateModified in your schema and keep it honest.

A rewrite of the kind that works, from a SaaS metrics page:

Before: “Net revenue retention is one of the most important metrics for any SaaS business, and something every founder should understand, because investors look at it closely and it tells you a lot about how healthy your customer base is over time. There are many ways to think about it, so let’s start from the beginning.”

After: “Net revenue retention (NRR) measures how much recurring revenue you keep from an existing customer cohort over 12 months, including expansion and net of contraction and churn. Calculate it as starting ARR plus expansion, minus contraction and churn, divided by starting ARR. An NRR above 100% means the cohort grew without any new logos.”

The before version says the metric matters. The after version gives the definition and formula. It also states the threshold, completing the sub-question a retrieval system would ask. The rule: the first chunk under any H2 must answer the heading completely, with the definition and the number in the same block.

How is GEO different from traditional SEO?

GEO optimizes how a passage presents itself to a generative engine, where SEO optimizes how a page competes for a ranking. The term comes from the Princeton-led paper published at KDD ‘24, which defines it as a black-box optimization framework that tailors a page’s presentation and text style to increase visibility inside engines that retrieve documents and generate a grounded, cited response. The framework also adjusts the content.

The paper’s results show why the two disciplines diverge. Adding cited sources or quotations produced a 30–40% relative improvement on the Position-Adjusted Word Count metric. Adding statistics produced improvements in the same range. Keyword stuffing, the oldest SEO lever, performed 10% worse than no optimization on Perplexity. And the gains were not evenly distributed: adding citations lifted visibility 115.1% for the site ranked fifth in the SERP while the top-ranked site lost 30.3% on average. A generative engine redistributes attention away from the incumbent toward whichever source packaged the evidence best.

The mechanical differences affect what you optimize and measure.

• Passage-level competition: Put the answer where a retrieval system can extract it. Underneath researchers found word count stopped mattering once position was held constant.
• Citation measurement and signals: The reward is a citation rather than a click, so you measure share of answers rather than share of positions. Presentation and provenance correlate with winning: dates and bylines establish provenance, while sourced numbers support the passage. These differ from the link graph that SEO teams spent years building.

Your GEO work still depends on SEO. A page has to be indexed and eligible for a snippet before any of this applies, and the retrieval step in Google’s pipeline draws on the same index. Treat GEO as a layer you add to pages that already clear the technical bar, run by the same team, with a different definition of done.

How do you build content an Overview cannot replace?

The test for an un-summarizable asset is whether a 60-word summary of it satisfies the query. If it does, the Overview will deliver that summary and keep the click. If it cannot, the searcher has to come to you. Five asset types pass the test.

• Original research: Other authors can cite a dataset you collected and published, putting your statistics inside the chunks the engine retrieves. The GEO paper’s statistics-addition result is the citation mechanism, and buyers visit your page to access the data itself. Proprietary data is how you earn the citation without hoping the Overview paraphrases a generic explainer.
• Proprietary data: Publish benchmarks and pricing indices only you can produce. Include your own usage numbers. The Overview can quote the headline figure but not the breakdown, and the breakdown is what a buyer needs.
• Interactive tools and calculators: The output depends on the user’s input, so there is no static answer to summarize. An NRR calculator survives where an NRR explainer does not.
• Gated resources: Put templates and teardowns behind an email form. Gate playbooks the same way. The engine can describe them and link to them, but buyers get the resources on your page.
• Bottom-of-funnel content: Publish migration guides and integration setup pages. Build comparison pages that lead to a trial. Buyers complete the job on your site.

Be honest about what the evidence supports. The controlled evidence that originality earns more citations is thin and mixed, so treat these assets as a click-retention play first and a citation play second. Reallocate from the “what is X” explainer backlog toward one calculator and one dataset per quarter, and measure retained sessions and conversions on those pages rather than citation counts.

Should you opt out with nosnippet or block AI crawlers?

For most B2B SaaS teams the answer is no for sitewide opt-outs and yes for surgical ones. The controls differ in what they remove and what they cost in snippets. Google’s robots meta tag documentation and the Search generative AI control set the Google-side trade-offs. OpenAI’s crawlers are a separate file.

• nosnippet: Removes the text snippet and video preview in web search, Images, Discover, AI Overviews, and AI Mode. Google states it also stops the content being used as a direct input for those AI features. It blocks featured and regular snippets. Deploy as a robots meta tag or an X-Robots-Tag header.
• data-nosnippet: Excludes only the marked span, div, or section from snippets. If nosnippet is also present, nosnippet wins. HTML attribute only. Google says not to add or remove it with JavaScript on existing nodes.
• max-snippet:[n]: Caps snippet length in characters and limits how much content can be used as a direct input for AI Overviews and AI Mode. 0 equals nosnippet. A lower value makes featured snippets less likely without guaranteeing they stop.
• robots.txt Disallow for GPTBot: Stops OpenAI’s training crawler. None in Google. Separate from OAI-SearchBot, which controls appearance in ChatGPT search answers. OpenAI treats each setting independently. Changes take about 24 hours on the search side.
• Search generative AI control: Removes the whole site from AI Overviews, AI Mode, and Discover generative features. Opted-out sites get no traffic or impressions from them. Google says it is not a ranking signal elsewhere. Search Console toggle, taking effect in one to two days.

Two precedence rules shape how you combine them. When directives conflict, the more restrictive one applies, so max-snippet:50 plus nosnippet resolves to nosnippet. And none of the meta or header directives are read if robots.txt disallows the page, because Google never fetches them.

The common mistake is blocking Google-Extended and expecting it to remove you from Overviews. Google’s own control page says the appearance toggle does not affect AI training and points you to Google-Extended for that, which means the reverse holds too: Google-Extended governs Gemini training and grounding, not Search features.

The pattern that holds is data-nosnippet around the proprietary table or the full dataset you want to gate, with everything else left open. You keep the citation and the definition-level chunks that earn it, and you keep the one thing the Overview would otherwise lift wholesale.

What metrics should you track beyond organic sessions?

Organic sessions alone cannot distinguish a visibility loss from a click loss, and the pages you worked on in step three will often show falling sessions and rising presence at the same time. Four measures separate the two.

• Generative impressions: The Search Console Generative AI report from step one, trended weekly by page.
• AI citation appearances: A prompt-level tracker that records which of your pages appear across AI Overviews, AI Mode, ChatGPT, and Perplexity for a fixed prompt set. Track the page, not only the brand, because the citation-driving page is what you replicate.
• Branded search: Branded query impressions in Search Console, trended month over month.
• Conversions by landing page: Trial starts and demo requests attributed to the pages you protected and rebuilt, not sitewide. This is the number the board reads.

Attribution is broken for the AI side of this. In one operator write-up, 80–90% of leads from AI platforms were tagged organic or direct in standard analytics, so a sessions-only dashboard undercounts the channel. Pipeline attribution for AI search has to sit beside the click report.

Add a self-reported “how did you hear about us” field to the trial form and read it next to the referral data.

For your monthly gap diagnosis, compare the first metric with the last.

With rankings flat, generative impressions zero, clicks down, and conversions down, move that page to the top of the summarize bucket.

The reporting surfaces here keep changing. Google launched the Generative AI report with impressions only and says more metrics are coming.

The reporting surfaces keep changing. Google launched the Generative AI report with impressions only and says more metrics are coming. Do not treat this month’s dashboard as the last word on the channel.

How do you reduce dependence on Google?

Brand mentions predict AI Overview presence better than links do, which makes brand defense and diversification the same workstream. Across 75,000 brands, Ahrefs found brand web mentions correlated with AI Overview mentions at 0.664, branded search volume at 0.392, and backlinks at 0.218.

Top-quartile brands for web mentions averaged 169 Overview mentions against 14 for the next quartile. Ahrefs labels this as correlation. Treat the ordering as a place to put effort, not as a proof that mentions cause citations.

• Digital PR: Digital PR moves the top of that list. Every data story you place with a trade publication is a third-party mention a retrieval system can read, and the same proprietary dataset from step four is the asset you pitch. Run it as a quarterly cadence tied to each research release rather than a one-off campaign.
• Entity optimization: Entity optimization makes the mentions cohere. Use a consistent company name and maintain your Organization schema. Keep an unambiguous description on your site and review platforms, then carry it into your directory listings to reduce the chance an engine attributes your category to someone else. The research I have seen offers no controlled test that a Knowledge Panel changes an AI answer, so treat entity work as hygiene that supports mentions rather than a lever on its own.

Branded search is the outcome you compound toward. When buyers type your name instead of the category, the Overview has nothing to intercept, and the branded impression trend from the previous section becomes your leading indicator for the whole program.

Newsletter and community are the channels that do not route through a search engine at all. A weekly email reaches the same buyer the Overview now answers, and a community gives them a reason to come back without a query. For a one-person growth function, start with the newsletter and treat the community as the thing you build once the list has given you the first few hundred people who reply.

Pick the three affected pages with the highest conversion value from your triage list and rebuild their first chunk this week. Set the generative impressions report to a weekly export so you can see whether the citation lands. The weekly notes on how teams actually run that scorecard land in The Messy Middle newsletter.