Search Generative Experience and the changing landscape of organic traffic

Search Generative Experience was an experiment, and it grew into two products

When this article first went up, everybody called Google's AI answers "SGE". Half the advice still doing the rounds was written for a thing that no longer exists under that name, so the names come first.

Search Generative Experience (SGE) was an experiment. Google announced it on 10 May 2023 as something people could sign up for in Search Labs, its opt in testing area, on Chrome on a computer and in the Google app. It was available in the United States, in English, and nowhere else. For a complex question it showed a generated snapshot of the key information at the top of the page, with links to dig deeper, and a conversational mode for follow up questions.¹

A year later the experiment graduated. On 14 May 2024 Google began rolling out AI Overviews to everyone in the United States, and said people had already used them "billions of times" through the Labs experiment.² Google told the press at the time that "AI Overviews in Search are replacing Search Generative Experience", and that the SGE name would stay behind in Labs for testing new features.³ Ireland waited longer than most English speaking markets. On 25 March 2025 Google brought AI Overviews to Ireland, in English, as part of a wider European rollout, for signed in users aged 18 and over.⁴

The conversational half of SGE took longer again. AI Mode, a separate tab where the whole page is a generated answer with links and follow up questions, opened to everyone in the United States on 20 May 2025 without a Labs sign up.⁵ It reached Ireland on 8 October 2025, as a tab on the results page and in the Google app for Android and iOS.⁶

So when someone says "SGE" today, they mean one of two things. An AI Overview is a generated summary placed on top of the ordinary results. AI Mode is a separate place where there are no ordinary results at all, only an answer with links in it. Both affect your organic traffic, differently.

The timeline reads from the top. The first five steps are Google's launches. The last two are the regulator and the tools, and they change what a business can question, measure and decide. Each gets its own section further down.

Worked example: one question, asked in Cork in four different years

Lee Bakehouse Supplies is an invented shop in Cork. It sells sourdough kit online: proving baskets, flour, starter kits, a few books. Siobhán runs it, and the shop's guides at leebakehouse.example bring in most of its new customers. Nobody real, every number made up.

Take the question a new baker types at ten at night: "why is my sourdough flat".

In May 2023 her customers in Cork saw ten blue links. SGE existed only for Labs users in the United States.

In September 2024 they still saw ten blue links. AI Overviews were live in the United States and a handful of other countries, but not yet in Ireland.

In November 2025 the same search could show an AI Overview above the links: five sentences on underproofing, weak starters and shaping. Her guide still ranked third, lower on the screen. And there was the AI Mode tab, where her guide appears only if the answer links to it.

In September 2026 Siobhán can, for the first time, see in Search Console how often her pages were shown inside those features, and switch them off if she wants to. Same question, same guide, four different shop windows.

Two kinds of generated answer, and both start with ordinary searches

Here is the mechanism in short. The full version, with diagrams, is in the article on AEO and GEO.

Google's documentation says that AI Overviews and AI Mode may use a technique it calls query fan out, "issuing multiple related searches across subtopics and data sources" before writing a response.⁷ The model does not read the web live. It runs several ordinary searches against the ordinary index, takes passages from what comes back, writes an answer and links to some of the pages it used.

Two consequences follow. To be in either kind of answer, a page must first be in the index and allowed to show a snippet. Google states that rule in one line: "a page must be indexed and eligible to be shown in Google Search with a snippet".⁷ And because the answer is built from several searches, a page can be cited for a question it never ranked for, simply because it answered one of the sub questions well.

Worked example: where Siobhán's holiday guide comes in

A customer asks "how do I keep my sourdough starter alive while I'm on holiday for two weeks". The fan out might produce three searches: "can you freeze sourdough starter", "how long can a starter go without feeding" and "how to dry a sourdough starter".

Siobhán's guide has a section headed "Can you freeze a sourdough starter?" with a plain answer in the first sentence. It has a section on feeding gaps. It says nothing about drying. So two of the three sub questions can be answered from her page, and the third will be answered from someone else's. Her guide never ranked for the long holiday question at all. It can still be cited in the answer to it.

Why sections like hers are understood that way is in the article on passage indexing.

What the click studies actually measured

Every owner asks me how much traffic an AI Overview costs. There are good studies now. They measure different things, and the differences matter more than the headlines. One thing they share: none of them is Irish. Pew's is US data, and neither Seer nor Ahrefs says which country its data comes from, so read all three as a guide to the direction, not as Irish figures.

Start with the rate they all talk about. The click through rate (CTR) is a plain ratio:

$\text{CTR} = \frac{\text{clicks}}{\text{impressions}}$

Here $\text{clicks}$ is the number of times someone clicked through to a page from the results, and $\text{impressions}$ is the number of times the page was shown in the results. It is events over opportunities (Foundations: ratios and rates).

Pew Research Center: what real people did in the United States

Pew watched the actual browsing of 900 adults in the United States during March 2025, 68,879 unique Google searches in all. About 18% of those searches produced an AI summary. On pages with a summary, people clicked a traditional result link in 8% of visits. On pages without one, they clicked in 15% of visits. They clicked a link inside the summary itself in 1% of visits. And they ended their browsing session more often after a page with a summary (26%) than after one without (16%).⁸

Along the bottom are three situations: a results page with no summary, a results page with one, and the summary's own links. Up the side is the share of visits in which someone clicked. The takeaway is the gap between the first two bars, and how short the third one is. Being cited inside the summary is not, on these numbers, a replacement for the click you lost below it.

From 15% to 8% is a fall of seven percentage points. As a relative change it is:

$\Delta\% = \frac{v_{\text{new}} - v_{\text{old}}}{v_{\text{old}}} \times 100 = \frac{8 - 15}{15} \times 100 = -46.7\%$

where $v_{\text{old}}$ is the click share without a summary, $v_{\text{new}}$ the share with one, and $\Delta\%$ the relative change (Foundations: percentages). "Seven points" and "nearly half" describe the same finding, so say which one you mean. Pew measured real people over one month in the United States, not your industry, Ireland, or a particular ranking position.

Seer Interactive: what happened to a set of client queries

Seer Interactive, a US marketing firm, tracked 3,119 informational queries across 42 of its client organisations from June 2024 to September 2025, using their Search Console and Google Ads data. For queries that showed an AI Overview, organic CTR went from 1.76% to 0.61%. For queries without one, it went from 2.74% to 1.62%. Brands cited inside the AI Overview had a higher organic CTR than those not cited, 0.70% against 0.52% in the third quarter of 2025.⁹

Three things to notice. Seer's headline calls the first fall 61%, but from its two published monthly figures I get 65.3%, so recompute headlines when you can. Queries without an AI Overview lost about 41% too, so not all of the fall is the Overview's doing. And Seer says it cannot prove citation causes the higher CTR, since strong brands may be both cited and clicked more, and that the results cover informational queries only.⁹

Ahrefs: position one, with a modelled comparison

Ahrefs compared 300,000 keywords, half with an AI Overview and half informational keywords without one, using aggregated desktop Search Console data for December 2023 and December 2025. For the page in position one, CTR on the AI Overview keywords fell from 0.073 to 0.016. On the informational keywords without an Overview it fell from 0.076 to 0.039.¹⁰

Because both groups fell, Ahrefs asked what the AI Overview keywords would have done had they fallen only as much as the others:

$\widehat{\text{CTR}}_{2025} = \text{CTR}^{\text{AIO}}_{2023} \times \frac{\text{CTR}^{\text{other}}_{2025}}{\text{CTR}^{\text{other}}_{2023}} = 0.073 \times \frac{0.039}{0.076} \approx 0.037$

Here $\text{CTR}^{\text{AIO}}_{2023}$ is the position one CTR of the AI Overview keywords in December 2023, $\text{CTR}^{\text{other}}_{2023}$ and $\text{CTR}^{\text{other}}_{2025}$ are the position one CTRs of the comparison keywords in December 2023 and December 2025, and $\widehat{\text{CTR}}_{2025}$ (the hat means "estimated") is what the AI Overview keywords would have shown in December 2025 had they followed the comparison group. The fraction is an index: it scales the old figure by how much the comparison group moved (Foundations: indexing and base rates).

The reduction against that forecast is then:

$\text{reduction} = 1 - \frac{\text{CTR}_{\text{actual}}}{\widehat{\text{CTR}}_{2025}} = 1 - \frac{0.0157}{0.0373} \approx 0.58$

where $\text{CTR}_{\text{actual}}$ is the measured December 2025 figure and the result is read as a share, so 0.58 is 58% (Foundations: percentages). That is Ahrefs' headline: the presence of an AI Overview "correlates with a 58% lower average clickthrough rate for the top ranking page".¹⁰ It is desktop only, it rests on a modelled comparison rather than a controlled experiment, and Ahrefs says "correlates", not "causes".

And what Google says

In May 2024 Google said that "the links included in AI Overviews get more clicks than if the page had appeared as a traditional web listing for that query".² Its documentation says that clicks from result pages with AI Overviews are "higher quality (meaning, users are more likely to spend more time on the site)".⁷ Neither statement comes with published data, so I cannot check them, and neither contradicts the studies above: a link inside an Overview can do well while the page as a whole is clicked less.

Worked example: reading the studies for Siobhán

Siobhán's question is narrower: what happens to her guides? Pew says roughly half the clicks on ordinary links go when a summary appears. Seer says informational queries are where it bites. Ahrefs says that at position one, on desktop, it is closer to 58%. Google says the clicks that remain are better ones.

So for her product searches, "banneton basket 1kg" and the like, I would expect little change. For her guides, which answer questions, I would plan for a fall of somewhere around half in the clicks on the queries that show an Overview, and I would not trust any one of these figures to more than one significant figure for her shop, least of all figures measured somewhere other than Ireland. The next section turns that into a number she can plan with.

The questions people type decide how exposed you are

Not every search gets a generated answer. Pew found that searches containing a question word ("how", "why", "what") produced a summary 60% of the time, searches written as a full sentence 36% of the time, and searches of 10 words or more 53% of the time. Searches of one or two words produced a summary only 8% of the time.⁸

The Irish figures are thinner. The Central Statistics Office (CSO) found in its 2025 household survey that 42% of internet users in Ireland had used generative AI in the three months before they answered.¹¹ The survey does not ask about summaries in search results, so it tells you the habit is spreading here, not how often an Irish searcher sees an Overview. For that, your own Search Console data is the better source, and I come to it below.

That gives a simple way to estimate exposure, group by group. For one group of queries:

$E[\text{clicks}] = I \times c \times (1 - a \times d)$

Here $E[\text{clicks}]$ is the expected number of clicks after AI answers arrive, $I$ the monthly impressions of the group, $c$ its CTR before, $a$ the share of those searches that show an AI Overview, and $d$ the relative drop in clicks when one appears. The bracket is a weighted average of two outcomes: a share $a$ of searches lose a share $d$ of their clicks, the rest lose nothing (Foundations: expected value, Foundations: ratios and rates).

Worked example: Siobhán's two query groups

Siobhán exports a year of Search Console data from before March 2025, when the Overviews reached Ireland, and splits her queries into two groups.

Guide questions: 20,000 impressions a month, CTR 4.0%, so $20{,}000 \times 0.04 = 800$ clicks. Most of them start with "how" or "why", so she takes $a = 0.6$ from Pew. For $d$ she uses Pew's 46.7%, rounded to 0.47:

$E[\text{clicks}] = 800 \times (1 - 0.6 \times 0.47) = 800 \times (1 - 0.282) = 800 \times 0.718 \approx 574$

The symbols are as above, with $I \times c = 800$ (Foundations: expected value).

Product searches: 15,000 impressions a month, CTR 3.0%, so 450 clicks. They are one or two words, so $a = 0.08$:

$E[\text{clicks}] = 450 \times (1 - 0.08 \times 0.47) = 450 \times (1 - 0.0376) = 450 \times 0.9624 \approx 433$

Here $I \times c = 450$ and $a = 0.08$, the rest as before (Foundations: expected value).

Together: 1,250 clicks a month before, about 1,007 after. That is a fall of $\frac{1{,}007 - 1{,}250}{1{,}250} \times 100 \approx -19.4\%$ for the shop, and almost all of it in the guides (Foundations: percentages). If she uses Ahrefs' 0.58 for $d$ instead, the guides come out at $800 \times (1 - 0.348) \approx 522$ clicks, and the shop's total fall is about 23%.

The point is not the decimal. She now knows where the loss sits (the guides), roughly how big it is (about a fifth of her organic clicks), and which assumption moves it most ($a$, which she can measure, rather than $d$, which she borrows).

Measuring the change on your own site

Borrowed studies get you a plan. Your own data tells you what happened. Until recently that was awkward, because Google puts AI feature traffic inside the ordinary numbers. Its documentation says sites appearing in AI Overviews and AI Mode are "included in the overall search traffic in Search Console", reported in the Performance report under the "Web" search type.⁷ There was no separate line.

That changed in 2026. Search Console now has a Generative AI performance report, rolled out to all websites worldwide on 31 August 2026. It shows impressions for your site in AI Overviews and AI Mode, defined as "how many times links to your site were shown to a user in a generative AI feature on Google Search", by page, device and country, over time. Its data comes from the same Web search type. It excludes Search Labs experiments, and a site with too few impressions may not see it at all.¹² The help page I read describes impressions. It does not describe a click figure for these features, so the clicks still have to be read from the ordinary Performance report. The country breakdown is handy if you sell into both Ireland and the UK: you can see the Irish impressions on their own.

So the method I use has three parts. Group the queries, as Siobhán did. Watch CTR by group in the Performance report, quarter by quarter, against the same quarter a year earlier, because seasonality would otherwise fool you (Foundations: seasonality). And use the Generative AI report to see which pages are being shown in the answers at all.

Worked example: what Siobhán's numbers did

Here are Siobhán's two groups, quarter by quarter, from the summer of 2024 to the summer of 2026. The CTRs are invented, but they follow the shape the studies describe.

Along the bottom are nine quarters, from the third quarter of 2024 to the third quarter of 2026. Up the side is the organic CTR of each group. The guide questions slide from 4.0% to 2.5%, a relative fall of 37.5%. The product searches barely move. Note that the slide starts before the Overviews reached Ireland in March 2025, which is what Seer saw too: question searches lose clicks even where no Overview shows. The takeaway is that the shop's organic traffic did not fall evenly. One kind of page lost a third of its clicks and the other lost nothing, which is what the exposure estimate predicted, a little milder than feared.

Then the new report. In September 2026 it shows 41,000 impressions for her guides inside AI Overviews and AI Mode. Her starter guide is shown most. Her flat loaf guide barely appears, because the answers lean on two larger baking sites. A list of pages shown and not shown is far more useful than a total.

In my experience, on the shops I have worked on, the visits that still come from guide questions convert better than before, because people who click past a summary want more than a summary. It does not make up for the lost volume.

The controls: snippets, the new opt out, and what each one costs

There are three levels of control, and each has a price.

The snippet controls are the oldest. The nosnippet rule stops Google showing a text snippet for a page, and Google's documentation now states that this applies to "AI Overviews, AI Mode" and "will also prevent the content from being used as a direct input for AI Overviews and AI Mode". The max-snippet rule caps the length of the snippet and also limits "how much of the content may be used as a direct input" for those features. The data-nosnippet attribute does the same for chosen parts of a page, and noindex removes the page from Search altogether.¹³ The cost: a page with nosnippet also loses its ordinary text snippet in the normal results, which tends to cost clicks there too.

The new control is the Search generative AI setting in Search Console, rolled out to all websites worldwide on 31 August 2026. A site is included by default. Set it to exclude, and Google says your content "is prevented from being visible to users in Search generative AI features", and that your site "won't receive any traffic or impressions from these features". Google also says the setting "isn't used as a ranking or inclusion signal affecting other parts of Search", and that it does not affect AI training, which is still governed by the separate Google-Extended robots token.¹⁴ Its optimisation guide makes the other side explicit: a site must be included in that setting to be eligible for these features at all.¹⁵

Ireland has no rule of its own that requires any of this. The question sits with Brussels. On 9 December 2025 the European Commission opened a formal antitrust investigation into whether Google used web publishers' content for AI Overviews and AI Mode "without appropriate compensation to publishers and without offering them the possibility to refuse such use of their content", given that many publishers depend on Google Search for traffic and do not want to risk losing access to it.¹⁶ Opening an investigation does not decide it, and no ruling has come out of it yet. So for an Irish business the Search Console setting is a tool Google offers worldwide, not a right an Irish or EU regulator has granted.

The training question is separate and is covered in the article on AEO and GEO, with the robots.txt lines for Google's and OpenAI's crawlers. Here I only need the distinction: excluding a site from Google's AI answers is not the same as refusing to let models train on it, and neither one removes the site from the ordinary results.

The flow runs from the top, from the bluntest control (out of Search entirely) to the finest (one paragraph). Most sites end at the bottom, rightly.

Worked example: should Siobhán opt out?

Siobhán has been giving out about it for months. She writes the guides, the summaries reuse them, and the clicks have fallen. Should she switch the new control to exclude?

What she would lose: every impression and click from AI Overviews and AI Mode. Her September report showed 41,000 impressions inside those features. If Pew's 1% click rate on summary links held for her, that is about $41{,}000 \times 0.01 = 410$ visits a month, and 410 visits she cannot easily replace (Foundations: ratios and rates).

What she would gain: nothing from Google directly. Excluding her site does not bring the ordinary links back above the summary. The summary still appears, built from other sites' pages, and her guide sits below it as before. Google says the exclusion is not a ranking signal, so it should not cost her the ordinary positions either.¹⁴

What she actually objects to is one page: a printable feeding schedule she sells as part of a starter kit, which the summaries had started paraphrasing. The answer is the finest control. She wraps the schedule table in a section with data-nosnippet, leaves the site included, and keeps the 410 visits. One decision, written down, with the reason next to it.

What to change in the content, and what to ignore

The old version of this article recommended schema markup, definition lists and bullet points so that content could be "lifted" into summaries. Google now says the opposite. Its guide for site owners says that, from Google Search's point of view, optimising for generative AI search is simply optimising for search, which is still SEO. It tells site owners they can ignore tactics like "chunking" content and AI text files such as llms.txt, says there is no special schema for these features, and warns that mass producing pages to manipulate AI answers may breach its scaled content abuse policy.¹⁵

What is left is the ordinary craft, done well. A heading that is the reader's question. A first sentence that answers it. Facts with sources and a named author. I have written that up in full, with before and after figures, in the article on AEO and GEO, and for ChatGPT and the other chat engines in how to rank in ChatGPT. The strategy of giving the whole answer away where the reader already is has its own piece too, on zero click content.

The part specific to this article is a business decision, not a writing one. If a summary can replace a page, the page will lose clicks whatever you do to it. So the question for each guide is what it offers that a five sentence summary cannot: your own data, a tool, a product, a person to ask. Those are the pages that keep earning a click.

Worked example: Siobhán's flat loaf guide, rebuilt around what a summary cannot do

The flat loaf guide answered a common question with common knowledge: underproofing, a weak starter, poor shaping. Any summary can say that, and does.

Siobhán keeps the plain answer at the top, because the reader still deserves it. Then she adds three things only her shop has. First, her own numbers: of 212 customers who emailed her a photo of a flat loaf last year, 131 had a starter that had not doubled within eight hours of feeding. That is 62% ($131 \div 212 = 0.618$), stated as hers (Foundations: percentages). Second, a short tool: enter the time your starter takes to double and your kitchen temperature, and it suggests a proving time. Third, a named person: Siobhán, with a photo of her own failed loaf from 2019.

The summary can still answer "why is my sourdough flat". It cannot run the tool or show her 212 photos. Three months later the guide's CTR has not recovered to its old level, and I would not expect it to. But the people who click now use the tool, and one in nine looks at a starter kit before leaving. That is the trade on offer: fewer visits, from people who came for something a summary could not give them.

What I do on the sites I build

I split the organic queries into groups before anything changes, so that there is a baseline per group, not one total that hides everything. I estimate exposure with the formula above, using the shop's own share of question searches. I read the Generative AI report monthly, page by page, filtered to the countries the shop sells in. I leave sites included in Google's AI features unless there is a specific reason not to, and when there is, I use the finest control that solves it and write the reason down. I decide the training question separately, with the owner, because it is their content.

And I stopped calling it SGE in 2024, as Google did. The experiment is over. What it became is the first screen most of your customers see, and the useful response is to know which pages it affects, by how much, and what each offers that a summary cannot.

Footnotes

1. Elizabeth Reid, Supercharging Search with generative AI, Google, 10 May 2023, https://blog.google/products/search/generative-ai-search/

2. Elizabeth Reid, Generative AI in Search: Let Google do the searching for you, Google, 14 May 2024, https://blog.google/products/search/generative-ai-google-search-may-2024/

3. Barry Schwartz, Google AI Overviews Launch In US Search Results, Search Engine Roundtable, 15 May 2024, quoting Google, https://www.seroundtable.com/google-ai-overviews-launch-37389.html

4. Hema Budaraju, We're bringing the helpfulness of AI Overviews to more countries in Europe, Google, 25 March 2025, https://blog.google/feed/were-bringing-the-helpfulness-of-ai-overviews-to-more-countries-in-europe/

5. Elizabeth Reid, AI in Search: Going beyond information to intelligence, Google, 20 May 2025, https://blog.google/products-and-platforms/products/search/google-search-ai-mode-update/

6. Brian O'Donovan, Google launches new AI mode search tool in Ireland, RTÉ News, 8 October 2025, https://www.rte.ie/news/ireland/2025/1008/1537357-google-ai-ireland/

7. Google, AI features and your website, Google Search Central documentation, last updated 10 December 2025, accessed 8 November 2026, https://developers.google.com/search/docs/appearance/ai-features

8. Athena Chapekis and Anna Lieb, Google users are less likely to click on links when an AI summary appears in the results, Pew Research Center, 22 July 2025, https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/

9. Tracy McDonald, AIO Impact on Google CTR: September 2025 Update, Seer Interactive, 4 November 2025, https://www.seerinteractive.com/insights/aio-impact-on-google-ctr-september-2025-update

10. Ryan Law, Update: AI Overviews Reduce Clicks by 58%, Ahrefs, 4 February 2026, https://ahrefs.com/blog/ai-overviews-reduce-clicks-update/

11. Central Statistics Office, Household Digital Consumer Behaviour 2025, chapter Generative AI, 5 December 2025, https://www.cso.ie/en/releasesandpublications/ep/p-isshdcb/householddigitalconsumerbehaviour2025/generativeai/

12. Google, Generative AI performance report (Search), Search Console Help, accessed 8 November 2026, https://support.google.com/webmasters/answer/16984139?hl=en

13. Google, Robots meta tags specifications, Google Search Central documentation, last updated 24 March 2026, accessed 8 November 2026, https://developers.google.com/search/docs/crawling-indexing/robots-meta-tag

14. Google, Search generative AI control, Search Console Help, accessed 8 November 2026, https://support.google.com/webmasters/answer/16908024?hl=en

15. Google, guide for site owners on generative AI features in Google Search (page title in American spelling), Google Search Central documentation, last updated 10 July 2026, accessed 8 November 2026, https://developers.google.com/search/docs/fundamentals/ai-optimization-guide

16. European Commission, Commission opens investigation into possible anticompetitive conduct by Google in the use of online content for AI purposes, press release IP/25/2964, 9 December 2025, https://ec.europa.eu/commission/presscorner/detail/en/ip_25_2964