How AI Overviews and Answer Engines Are Changing What Content Even Ranks

In July 2025, Ahrefs found that about three in four pages cited in Google's AI Overviews also ranked in the organic top ten for the same query. When it re-ran the study in March 2026 on a much larger sample of 863,000 keywords and four million cited URLs, the figure was about 38%. More than six in ten citations were going to pages that did not sit on page one.

That shift captures the whole story. For two decades, ranking was the currency of content strategy: a position on a list, a click-through rate attached to that position, a traffic forecast built on both. AI Overviews, AI Mode, ChatGPT search and Perplexity break that chain in the middle. The interface now retrieves several pages, writes one answer, and attaches a few links, and the rules for which pages get read, which get quoted, and which get clicked are not the rules that set position.

It would be easy to stop at "SEO is dead" or at "nothing has changed." Neither survives the data. Google's own documentation says its generative features are built on its core ranking and quality systems, while independent studies show citations increasingly coming from pages that do not rank on page one. Click data shows real losses, but concentrated in particular kinds of content and partly predating AI altogether. This piece separates what is well supported from what is contested and what is mostly vendor marketing, then turns that into a plan.

Ranking Used to Mean One Thing. Now It Means Three.

Google has now documented how its generative features work, and the description is more useful than most of the commentary about it. Two techniques do the heavy lifting. The first is retrieval-augmented generation: the system relies on Google's core ranking systems to pull relevant, current pages from the search index, reads what those pages say, and writes a response with links. The second is query fan-out: the model generates a set of related queries at the same time and fetches results for each. Google's own example is a question about fixing a lawn full of weeds, which might fan out into searches about the best herbicides, removing weeds without chemicals, and preventing them from returning.

Two consequences follow. First, there is no separate "AI index" to get into. The pool of candidates is the ordinary index, so being indexed and eligible to appear with a snippet is a hard requirement. Second, the page that ranks first for the query someone typed is no longer the only page in contention, because the model may be drawing on results for several sub-queries you never see. Google says this approach lets it show a wider and more diverse set of links than a classic search.

The practical way to think about it is as three gates rather than one position:

One more detail matters for interpretation. Ahrefs has argued that since Google made Gemini 3 the default model for AI Overviews in late January 2026, the Overviews appear to lean less on results for the typed query and more on results for the fan-out queries. Google has not confirmed any change in fan-out behavior, so treat that as an informed reading of the data rather than an established fact.

This Is No Longer a Niche Feature

At Google I/O in May 2026, Sundar Pichai said AI Overviews now have more than 2.5 billion monthly users and that AI Mode, launched a year earlier, has passed 1 billion. Google says AI Mode queries have more than doubled every quarter since launch and that overall Search queries reached an all-time high last quarter. It also unveiled what it called the biggest upgrade to the Search box in over 25 years. People are not searching less. They are searching more, and more of that searching ends in a generated answer.

How often an AI Overview actually appears is where the numbers start to disagree, and the disagreement is instructive.

The spread is not noise. It is the finding. AI Overviews cluster on informational, question-shaped and comparison queries, and they are largely absent elsewhere. Your exposure depends on your own query mix, not on an industry average. Publishers interviewed by Press Gazette also described the growth in Overview frequency as flattening, which suggests the first land-grab has slowed even as AI Mode grows.

What Happened to Clicks: Fewer, Unevenly, and Not Only Because of AI

Every serious study finds that clicks fall when an AI answer appears. They disagree about how far, and most of the disagreement comes from what is being measured.

Read together, three things stand out.

Being cited beats being ignored, but it does not erase the loss. In Seer's data, brands cited inside an Overview earned roughly twice the organic click-through of those that were not, yet both sat below the rate for searches with no Overview at all. Google has said that links inside AI Overviews earn more clicks than the same links would as traditional listings. The independent data above does not test that claim directly, and a page's presence in the answer is not the same as a click.

Clicks are being redistributed, not only destroyed. Seer found that click-through on queries without an Overview rose from 2.8% in early 2025 to 3.8% by February 2026. One plausible reading is that AI answers absorb the quick, factual questions, so the people who still click are the ones who want more. Google's May 2026 changes point the same way: inline links placed beside the text they support, hover previews, a "Subscribed" label on links from publications a reader pays for, and first-hand perspectives from forums and social posts credited by name. The clicks that remain go to sources the answer highlights.

The damage is uneven, and headline averages hide it. Chartbeat data reported by Press Gazette shows Google search referrals to publishers down about a third worldwide in the year to November 2025 (down 38% in the US, 17% in Europe). Chartbeat's breakdown by size found referral traffic down 60% for small publishers, 47% for mid-sized ones and 22% for the largest over two years. Yet Similarweb data published by Graphite found organic Google traffic to the top 40,000 US websites down only 2.5% in 2025, and Bauer Media reported a single-digit decline. The Daily Mail said most of its search traffic is branded and more than half its readers arrive directly.

Zero-click behavior also predates AI. SparkToro's panel already showed 60% of US searches ending without a click two years before the 68% figure. The Daily Mail's head of SEO pointed to years of results pages crowded with featured snippets, forums and knowledge panels, and to journeys that now begin on Reddit or TikTok. AI Overviews accelerated a trend; they did not start it.

Rank and Citation Have Come Apart (By an Amount Nobody Agrees On)

This is the statistic people quote most and get wrong most often. Different studies measure different Google products, different markets, and different definitions of "citation," and they use different parsers to detect one. The table lays the main ones side by side.

Add up the caveats and three lessons survive.

Rank helps, but it is neither necessary nor sufficient. One review of published estimates found top-10 overlap figures ranging from 12% to 93% depending on surface, market and definition. The safest reading is that page-one rank raises the odds of being retrieved (Gate 1) but is not what makes a page get chosen (Gate 2).

The mix of sources is changing shape. Among AI Overview citations that did not rank in the top 100 for the keyword, over 18% were YouTube URLs, and Ahrefs reports YouTube as the most-cited domain in Overviews in its tracking. AI Mode, in Ahrefs' December comparison, leaned more on Wikipedia and Quora. Google's May 2026 update adds forum and social voices to the mix. The answer can come from a video, a discussion thread or a site you never considered a competitor.

Vertical matters. BrightEdge's data, as summarized by trade press, reportedly shows healthcare, insurance and finance with the highest overlap between citations and top-10 rankings, at roughly 68–75%. A plausible explanation is that trust signals drive both ranking and citation in high-stakes categories. If you publish in those areas, rank remains a better predictor than the overall averages suggest.

What the Evidence Says Earns a Citation

Most advice about "GEO" mixes three kinds of claim: things Google says outright, things independent studies support, and things that are plausible but unproven. The table separates them. The confidence labels are a judgment about the strength of the evidence, not a measurement.

Notice what sits at the top of the list: the least glamorous items, and the ones Google itself emphasizes. The tactics that sell the most consulting hours, such as special files, rewriting for machines and manufactured mentions, sit at the bottom. That does not prove they never work on some engine. It does mean the burden of proof is on the people selling them. Google's own guide advises caution with third-party tools that claim to use internal Google metrics, since no outside tool has access to Google's ranking or AI systems.

Every Answer Engine Has Its Own Source Pool

"Optimize for AI search" hides the fact that the engines behave differently. A page can be prominent in one and invisible in another.

Two vendor-run studies, one covering 680 million citations and another 118,000 responses, both found that only about 11% of domains cited by ChatGPT are also cited by Perplexity. The studies are commercial and their methods are not fully open, but they agree with each other. The implication is that visibility on one engine tells you little about the others.

The traffic reality is more modest than the excitement suggests. Conductor's study of 13,770 domains, reported in trade coverage, found AI referrals averaged about 1% of sessions versus about 25% for organic search, with ChatGPT supplying roughly 87% of them. Chartbeat found chatbots still accounted for under 1% of publisher page-view referrals, even after ChatGPT referrals more than doubled in a year. Yet Similarweb reports that brands recommended by ChatGPT drew about 2.5 times more visits than competitors that were not, and that many of those visits show up in analytics as branded organic search rather than as referrals. So answer engines are small as a traffic source and potentially large as an influence on demand. Claims that AI-referred visitors convert better range from about 1.5x the rate of other digital channels (early Criteo data) to about 9x the rate of Google organic (Seer's 2025 figures). Most rest on small, self-selected samples, and the volumes are low.

Beyond Google, the category is also fragmenting. Similarweb data shows ChatGPT's share of worldwide generative AI web traffic falling from roughly 76% in June 2025 to about 53% in May 2026, as Gemini passed a quarter of the total and Claude grew fastest from a small base. A strategy built for a single engine covers a shrinking part of the audience.

Which Content Is Exposed and Which Is Resilient

The best evidence on who is hurt comes from people running the sites. Press Gazette's interviews with audience leaders at Bauer, the Daily Mail, the Telegraph and Candr Media describe a consistent pattern.

• Facts anyone can supply get absorbed. Bauer said car specification data hit hard because the facts are not in dispute and are not proprietary to Bauer. A reader who only wants a number has no reason to click.
• Real-time and high-consequence content holds up. TV listings rarely trigger Overviews, likely because the data is too close to real time to synthesize. Breaking news still rarely triggers them at the Daily Mail. Where a decision is costly, such as choosing a car, readers still click through to a source they trust.
• Evergreen explainers and how-tos are the most exposed. Candr's chief executive described them as being eaten up, and reported staff cuts since Overviews arrived, while opinion-based content was less affected.
• Debate protects clicks. Where there is no single settled answer, Bauer said click-through stays healthy. Subscription publishers such as the Telegraph are considered safer because their readers are less likely to be satisfied by a summary.
• Brand demand is a moat. Amsive found branded queries with an Overview saw click-through rise by 18%, and the Daily Mail's mostly branded and direct audience shielded its total traffic.

The map is a triage tool, not a verdict. Most sites have pages in all four quadrants. The useful exercise is to sort your own library and decide, page by page, whether to rebuild, consolidate, or leave alone. Bauer's audience director put it bluntly: be unemotional about the content types where the decline is unlikely to reverse, and put the effort where readers still want a trusted source.

Measuring a Ranking That Changes Every Time You Look

Two developments this year change how you can measure this. First, Google launched generative AI performance reports in Search Console on June 3, 2026, starting with a subset of UK sites, and says they had reached all websites worldwide by August 31. The reports show impressions from AI Overviews, AI Mode and Discover's generative features by page, country and date, with a device breakdown in the Search report. They do not show clicks or queries, and sites with too few AI impressions may not see a report at all. Alongside them came a property-level control that lets a site exclude itself from those surfaces. Google says the setting is not used as a ranking signal elsewhere in Search and does not affect AI training, which is governed separately by Google-Extended. The UK's Competition and Markets Authority issued a conduct requirement the same day. Google has until March 2027 to offer page-level controls, and the regulator expects click data too.

Second, AI answers are not repeatable. SparkToro and Gumshoe.ai had 600 volunteers run 2,961 prompts across ChatGPT, Claude and Google's AI features in late 2025. The chance of receiving the same list of brands twice for one prompt was under one in 100, and the chance of the same list in the same order was closer to one in 1,000. Yet in tight categories the leading brands reappeared in most responses. That points to the right metric: how often you appear across many runs is stable enough to track, while your "position" in any single answer is not. The authors suggest 60–100 runs before patterns become meaningful.

No single row is enough. Together they give an honest picture, and our upcoming post on measuring content success in a zero-click world goes deeper on building the dashboard.

A Practical Playbook, in Priority Order

1. Confirm eligibility before anything else. Check that key pages are indexed, carry no snippet restrictions, and are not caught by robots.txt rules written for something else. Review robots.txt against each vendor's crawler documentation. OpenAI, for example, treats its search crawler and its training crawler separately, so a blanket AI blocklist can quietly remove you from answers you wanted to be in. Confirm your Search generative AI setting while you are there.
2. Sort your library by "answerable without you." For each high-traffic page, ask whether a competent summary would fully satisfy the reader. Use Fig. 3. Pages in the exposed quadrant need rebuilding, consolidating, or a decision to stop investing in them.
3. Rebuild commodity pages around something only you have. First-hand testing, proprietary data, original photos and screenshots, named experts with a point of view, calculators and tools. Google's guide contrasts a generic list of tips with an account of what one person actually did and learned. That gap is the opportunity.
4. Answer the neighboring questions on the page. Fan-out means the system asks related questions you can predict: cost, alternatives, risks, who it suits. Mine support tickets, sales calls and on-site search for what people ask next, and answer under clear headings. Do not create a near-duplicate page for every variant; Google treats that as scaled content abuse.
5. Build presence where the engines look. Earned coverage, YouTube, credible communities and review profiles. The evidence is correlational, so treat this as brand-building that also helps AI visibility, not a shortcut. Manufactured mentions are exactly what Google says not to bother with. See our upcoming post on E-E-A-T for the AI era.
6. Refresh substantively, on a schedule. Prioritize pages where facts change: prices, regulations, tools, benchmarks. A refresh means new data, corrected claims and reworked sections. Change the date only when the content changes.
7. Show your evidence. Name sources, date your statistics, attribute quotes. It is what a careful reader wants anyway and it is cheap. Do not expect a guaranteed visibility lift, because the research is mixed.
8. Win the click you still get. Visibility is not traffic. Bauer reported that concentrating on how content is packaged across discovery surfaces, including headlines and descriptions, helped turn visibility into visits. If you run a paywall, look at Google's subscription linking, which powers the "Subscribed" label.
9. Measure with the new tools and report rates. Set up the Search Console report, track branded search and direct traffic, segment chatbot referrals, and run a prompt panel. Report visibility as a percentage with the number of runs behind it, never as a position.

One note on paid. Ads now sit alongside AI answers, including on Google and, since February, in ChatGPT. OpenAI says ads do not influence what ChatGPT answers. That gives brands two routes to visibility, earned and bought, and the two should be reported separately so that neither flatters the other.

What Is Still Unsettled

• Whether the click-through rebound holds. Seer's move from 1.3% to 2.4% is one quarter of data across one sample, and Seer's own language treats it as stabilization rather than recovery.
• How much of the fall in top-10 overlap is behavior and how much is measurement. Ahrefs itself says parsing improvements account for part of the change, and the studies disagree on direction depending on how "overlap" is defined.
• What AI Mode does to clicks. In SparkToro's January–April window, only 0.34% of searches moved into AI Mode, so its effect barely registered. Early vendor analyses from Semrush and SE Ranking put the zero-click share of AI Mode sessions above 90%. With more than a billion monthly users and doubling query volume, this is the biggest open question.
• Whether Google's link changes move traffic. Google says labeled subscription links drew significantly more clicks in testing but has not published numbers, and Search Console does not report clicks from AI features.
• What agents do to all of it. Google's guide already includes a section on agentic experiences, and I/O introduced background "information agents." If software rather than a person reads your page, a click may stop being the goal.
• How to reconcile Google's story with publishers' story. Google says people search more when they use AI features, and queries are at a record high. Publishers say referrals are down. Both can be true: more searching, fewer clicks per search.

The Bottom Line

Ranking has not been replaced. It has been demoted from the outcome to the entry ticket. Retrieval still runs through the index, and Google is clear that the same foundations apply. What changed is everything after retrieval: a model chooses among the candidates, writes something new, and credits only a few of them, on criteria that include how distinctive your material is and what the rest of the web says about you.

Rank used to stand in for being the best page on a list. Now the interface picks pieces, and the pieces it trusts are distinctive, corroborated and easy to attribute.

The read here is that the unit of competition is shifting from the page to the source: your name, your evidence, and your reputation beyond your own domain. That is uncomfortable for content built to be interchangeable, and encouraging for anyone with something real to say. Google's guide, the citation studies and publishers' own experience converge on the same advice. Be the source someone would want to cite, make sure both machines and people can reach you, and measure what you can with honest error bars.

A Practical Audit Checklist

• Key pages are indexed, deliver their content in the HTML crawlers receive, and carry no accidental snippet restrictions such as nosnippet or a noindex tag
• robots.txt is reviewed against each vendor's current crawler list, with search and answer crawlers allowed where you want to appear and training crawlers set by deliberate policy
• The Search generative AI setting is confirmed and the decision to stay in or opt out is documented
• Top pages are sorted by how answerable they are without a click, and commodity pages are flagged to rebuild, consolidate or retire
• Each priority page contains something only you can provide: first-hand testing, proprietary data, a named expert's view, or a tool
• The questions a reader would ask next are answered on the page under clear headings, without near-duplicate pages
• Statistics, quotes and claims carry named, dated, linked sources
• A substantive refresh schedule exists for pages where facts change, and dates change only when content does
• An off-site presence plan covers earned coverage, video and community participation, and none of it is manufactured
• Measurement is in place with a named owner: the Search Console generative AI report, branded and direct trends, chatbot referrals, and a 60–100-run prompt panel reported as visibility rate

Frequently Asked Questions