Beyond SEO: winning visibility in the age of AEO and GEO

Search now answers before it links, and the rules for being the answer are mostly public

For twenty years the deal was simple. You wrote a page, Google ranked it, and if you ranked well enough people clicked. Search Engine Optimisation (SEO) was the craft of earning that position. The deal has changed shape. Google now often writes an answer at the top of the results and cites a handful of pages inside it. ChatGPT answers questions with a short essay and a row of source links. Bing does the same through Copilot. The page that gets quoted in the answer is visible. The page that merely ranks seventh is not.

Two new names have grown up around this. Answer Engine Optimisation (AEO) is the craft of being the answer the engine lifts out and shows, in a featured snippet, a voice reply or an AI summary. Generative Engine Optimisation (GEO) is the craft of being the source a language model draws on and cites when it writes a reply. The four terms side by side, including the fourth, are in the complete guide to SEO, AEO, GEO and AIO. This article is about the mechanics underneath: how a question turns into an answer with citations, which three things decide whether you are in it, and what I actually do on the sites I build.

First, the size of the shift, with sources rather than conference slides. Pew Research Center watched the real browsing of 900 adults in the United States in March 2025. On Google result pages that carried an AI 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 AI summary itself in 1% of visits. About 18% of all the searches in the study produced a summary, and people ended their browsing session more often after a summary page (26%) than after a plain results page (16%).¹ The message is plain: when the engine answers, roughly half the clicks disappear, and the sources inside the answer get almost none of the rest.

The old version of this article said that "over 60% of Google searches now result in zero clicks". I cannot stand that up as written. The closest serious figure is from SparkToro's analysis of a clickstream panel supplied by Datos, covering September 2022 to May 2024. It found that 59.7% of Google searches in the European Union, Ireland included, ended without any click, against 58.5% in the United States. Of every 1,000 searches in the EU, about 374 clicks went to the open web, and in the United States about 360 reached a site that was neither owned by Google nor paying Google for an ad.² It is a secondary source, a vendor's panel with the caveats its author lists (the panel is not the whole population, mobile data only begins in January 2024, Apple devices are thin), so I quote it as such. Nearly six in ten, here in the EU, is the honest phrase. "Over 60%" was a rounding up the data did not earn.

Worked example: the question Róisín is not in

Calder Ledger is an invented accountancy software firm in Limerick. It sells bookkeeping software to small practices, about 400 of them, and its managing director is Róisín. Nobody real, every number made up. Róisín's best page, by her own reckoning, is the one on moving a practice onto Calder Ledger partway through a tax year. It ranks fourth on Google for the obvious search and it has done for two years.

One morning she searches the question herself, "can you switch accounting software in the middle of a tax year", and the page is still fourth. Above it sits an AI Overview that answers the question in five sentences and cites three pages. None of them is hers. Two are from software firms larger than hers and one is from a practice in Galway that wrote a plain, step by step page with dates and a named author. Her ranking has not moved. Her share of the attention has.

How a question becomes an AI answer with citations

The engines publish more about this than people assume. Google's own documentation says that both AI Overviews and AI Mode may use what it calls a query fan out technique, which it describes as "issuing multiple related searches across subtopics and data sources" before assembling a response from what comes back.³ So the model does not read the web live. It runs several ordinary searches against the ordinary index, picks passages from the results, writes, and attaches links to the passages it used.

That has a consequence worth sitting with. To be in the answer you first have to be in the index and allowed to show a snippet. Google states the eligibility rule in one line: "a page must be indexed and eligible to be shown in Google Search with a snippet", and adds that there are no additional requirements to appear in AI Overviews or AI Mode, no special optimisation, and "no special schema.org structured data that you need to add".³ OpenAI describes the equivalent on its side: a crawler called OAI-SearchBot "is used to surface websites in search results in ChatGPT's search features", separately from the crawler that gathers training data.⁴ Same shape: fetched pages, a writing step, citations.

The diagram runs top to bottom through nine steps. The three places where you can act are the index (be crawled and indexed), the gate (be open to the bot and allowed to show a snippet) and the passage selection (have a passage that answers a sub question plainly enough to be lifted). Everything that follows in this article is about those three.

Worked example: Róisín's question, step by step

Take Róisín's search through the diagram. The question "can you switch accounting software in the middle of a tax year" fans out into something like "what happens to opening balances when you switch", "is it better to switch at year end", "how long does an accounting software migration take" and "what does Revenue need when you change software". Each of those runs as a search. The Galway practice's page has a heading that is nearly word for word the third sub question and a first sentence that answers it with a number of weeks. Róisín's page has a heading that says "Why Calder Ledger?" and a first paragraph about the firm's values. Both pages are indexed. Both are open to Googlebot. Only one has a passage a system can lift and attribute. The gate was open for both. The passage selection chose one, and fair play to Galway, it chose the page that earned it.

Be the answer: a clear question, a direct first sentence, and structure the machine can lift

Google has been honest for years about how featured snippets work, and the same logic carries into AI answers. Can you mark a page as a featured snippet? The documentation answers in two words: "You can't."⁵ Google's systems "determine whether a page would make a good featured snippet for a user's search request", and the one condition they publish is that "featured snippets will only appear if enough text can be shown to generate a useful featured snippet".⁵ There is no minimum length. There is a passage, or there is not.

So the craft of being the answer is the craft of writing passages. A heading that is the question in the words a person would type. A first sentence under it that answers the question directly, with the number or the yes or no in it. Then the detail, the caveats and the context for the reader who stays. Someone who reads only the heading and the first line gets the answer. That is good for a reader in a hurry, and it is exactly the unit an engine extracts. I write that way because it is how people read. The machines happen to reward it.

Structured data belongs here too, with two corrections to how it is usually sold. First, the markup has to describe what is on the page. Google's guidelines say "your structured data must be a true representation of the page content" and "don't mark up content that is not visible to readers of the page", and a page whose markup describes hidden or misleading content can lose its eligibility for rich results altogether.⁶ Second, markup is not a ticket to a feature. From August 2023, Google showed FAQ rich results only for what it called well known, authoritative government and health websites, and the HowTo rich result was retired entirely.⁷ Since 7 May 2026, FAQ rich results no longer show at all.⁸ So the FAQPage type on schema.org, "a WebPage presenting one or more 'Frequently asked questions'",⁹ will not get a Limerick software firm a drop down of questions in the results. What structured data still does is tell every machine, in a vocabulary they all read, what kind of thing the page is and who stands behind it. The Article type with a named author and a date, and the Organization type with a legal name, an address and sameAs links to the firm's other profiles,¹⁰ do that quietly and cost nothing. Written as JSON-LD, which Google recommends,⁶ they sit in the head of the page.

Worked example: one section, before and after

Here is one section of Róisín's page, before. The heading is "Why Calder Ledger?" The first paragraph reads: "At Calder Ledger we believe switching software should never hold a practice back. Our dedicated onboarding specialists have helped hundreds of practices transform the way they work." Nothing in that paragraph answers a question anyone typed. There is no number, no date, no step.

After. The heading is "Can you switch accounting software in the middle of a tax year?" The first sentence is: "Yes. A practice can move to Calder Ledger at any month end, because the opening balances are imported as at that date and the prior months stay in the old software for reference." The second paragraph gives the three steps and how long each has taken across the firm's last 60 migrations: export (one working day), import and reconciliation (three to five working days), parallel running (one month end). The third paragraph says when not to do it: the fortnight before a VAT return is due. At the foot, a byline with Róisín's head of onboarding named, a date, and a line saying where the timing figures come from (the firm's own migration log, which is first party data and needs no outside source).

Same product, same page, same position in the rankings. The difference is that a system now has a heading that matches a sub question, a first sentence that answers it, and a passage with figures and a named person to attribute. That is all AEO is.

Be citable: facts with sources, numbers, quotes, and a clear sense of who you are

The term GEO comes from a paper by Pranjal Aggarwal and colleagues, presented at the KDD conference in 2024, which asked an unusually practical question: if you can change your page but not the engine, what changes to the page make a generative engine cite you more?¹¹ They built a benchmark of 10,000 queries across many domains, tried nine kinds of edit, and measured how much of the generated answer each source accounted for, adjusted for where in the answer it appeared.

The results are the useful part. Adding quotations from relevant sources, adding statistics, and citing sources each lifted that measure of visibility by roughly 30% to 40% over the untouched page. Making the text more fluent helped nearly as much. Making it sound more authoritative, on its own, did not reach significance. And the one classic tactic everybody has heard of, keyword stuffing, made things worse: the stuffed pages scored 17.8 on the position adjusted measure against 19.5 for the untouched baseline, about 9% below it.¹¹ The authors also found that lower ranked sources gained the most. For a source sitting fifth in the traditional results, citing sources raised its visibility by 115.1%, while the top ranked source lost visibility on average.¹¹ Fourth place, like Róisín's, is where this craft pays off most.

Two caveats. The engine in the study was a research setup the authors built from a search backend and a language model, not Google's or OpenAI's production system, and their measure counts attributed words, not clicks or enquiries. The direction of the findings matches what I see on the sites I work on, and it matches what the engines themselves say they want, which is why I lean on it. The precise percentages are theirs and belong to their setup.

The second half of being citable is being identifiable. A model composing an answer about accounting software for small practices has to decide which firm this page belongs to, whether that firm is the same one named in a review on another site, and whether it is in Ireland at all. The clearer the entity, the more confident the attribution. That means one spelling of the name everywhere, the same legal name and address on the page as on the firm's record at the Companies Registration Office (CRO), an Organization block that points at the firm's LinkedIn page and its listing in the professional directories through sameAs, and an author who exists and can be found. The entity side is covered in the article on semantic SEO and entity based optimisation, and it is the foundation the citation sits on.

To talk about this sensibly you need one measurement, and it is a rate:

$\text{citation share} = \frac{\text{answers that cite the page}}{\text{questions asked}}$

Here $\text{answers that cite the page}$ is the count of generated answers, across whichever engines you are checking, that include a link to the page, and $\text{questions asked}$ is the number of questions in a fixed list you ask every time you check. It is a plain rate, events over opportunities, and the usual warnings about small denominators apply (Foundations: ratios and rates, Foundations: sampling). Twenty questions asked once tell you very little. The same 40 questions asked every month for six months tell you a great deal.

Worked example: Róisín makes the page citable

Róisín's after page has the structure. Now it needs the substance a model can quote. Three additions, each of a kind the paper found to matter.

A statistic: "Of the 60 practices that moved to Calder Ledger in the last two years, 38 did so between months four and eight of the tax year, and none reported a reconciliation difference above €50 at the first month end." Her own data, stated as hers, dated.

A quotation: two sentences from her head of onboarding, named, about the one mistake practices make (exporting before the final bank reconciliation is posted). A person's words in a person's name are something a model can attribute.

Sources: where the page states something outside the firm's own knowledge, such as what a VAT return requires, it links to the guidance page on Revenue's own site rather than paraphrasing it unsupported. Not because the link brings traffic, it rarely will, but because a passage with a source attached reads as a passage worth repeating.

Then she writes down 40 questions a practice manager might ask about switching software, asks them of Google, ChatGPT and Copilot on the first Monday of each month, and counts the answers that cite her page. In March, before the work, the count was one out of 40. That is her baseline, and it is honest to call it a small number from a small sample.

Be crawlable by the AI bots, or deliberately not. Either way, decide.

Everything above assumes the bots can read the page. That is now a choice with several switches, and most sites have never looked at them. The useful thing is that the switches are documented and they are separate from each other.

On Google's side, Googlebot is the one that matters for Search, and Google's documentation says crawling preferences addressed to it "affect Google Search (including Discover and all Google Search features)",¹² which is where AI Overviews and AI Mode live. There is a second token, Google-Extended, which controls whether content "may be used for training future generations of Gemini models" and for grounding in Gemini apps and Vertex AI. Google is explicit that this token "does not impact a site's inclusion in Google Search nor is it used as a ranking signal in Google Search".¹² So blocking Google-Extended keeps you out of Gemini's training while leaving you fully present in Search and its AI answers. Blocking Googlebot removes you from both.

On OpenAI's side there are three agents. GPTBot gathers content that may be used to train models, and "disallowing GPTBot indicates a site's content should not be used in training generative AI foundation models". OAI-SearchBot fetches pages to show them in ChatGPT's search answers, and disallowing GPTBot does not affect it, so a site can allow search and refuse training at once. ChatGPT-User fetches a page when a person asks ChatGPT to look at it, "is not used for crawling the web in an automatic fashion", and because a person triggered it, robots.txt rules may not apply.⁴

On Bing's side the control sits in the page rather than in robots.txt. In September 2023 Bing announced that the existing NOCACHE and NOARCHIVE robots meta tags govern Bing Chat, now Copilot. With NOCACHE, "content with the NOCACHE tag may be included in Bing Chat answers" but only the URL, title and snippet are shown. With NOARCHIVE, "content tagged NOARCHIVE will not be included in Bing Chat answers, not be linked to in the answers". Either way the page "will still appear in our search results".¹³

And there is a finer dial on Google than the bot itself. The nosnippet, data-nosnippet and max-snippet controls limit how much of a page can be shown as a snippet, and John Mueller's post on succeeding in AI search states the consequence plainly: "more restrictive permissions will limit how your content is featured in our AI experiences".¹⁴ A page with nosnippet is in the index and out of the answer. That is a legitimate choice, for a paywalled report, say. It is a poor choice made by accident, which is how I usually find it.

Worked example: Róisín's robots.txt

Róisín wants to be cited in ChatGPT and Google's answers, because that is where her prospects are asking. She does not want her documentation used to train models that compete with her support team. Those two wishes are compatible, and the file is short:

User-agent: Googlebot
Allow: /

User-agent: Google-Extended
Disallow: /

User-agent: OAI-SearchBot
Allow: /

User-agent: GPTBot
Disallow: /

In the head of every public page she leaves the snippet controls alone, so the full passage is available to be lifted. On the two pages that hold her pricing for resellers, which she does not want quoted anywhere, she adds nosnippet for Google and NOARCHIVE for Bing, and accepts that those pages will rank but never appear in an answer. Each line is a decision she can explain. Before this exercise she had a robots.txt file generated by a plugin in 2019 that nobody had read since, and it blocked nothing and allowed everything, including training.

Worked example: the whole page, with before and after figures

Put the three parts together on Róisín's migration page and watch for six months. The work happened in the last week of April: the question headings and direct first sentences, the statistic and the quotation and the sources, the Article and Organization markup, the named author, and the robots.txt above. Nothing else on the site changed, and she did not buy a single link.

Across the bottom are the months, January to August, and the restructure sits between April and May. Up the side is the count of answers, out of the same 40 questions each month, that cited the page. Before, one or two. After, nine in the first month and 16 by August, which is a citation share of 16 over 40, or 40%. The honest reading is that the jump is far too large and too well timed to be noise on a sample of 40, and that the exact level is still uncertain by a few questions either way. The other honest reading is that the page's Google position moved from fourth to third over the same period, which is grand and not the point.

Same months along the bottom. Up the side, enquiries from practices whose first visit landed on the migration page, whatever brought them there. The average of the four months before was 6.75 a month, the average of the four after 15.25. Using the plain percentage change,

$\Delta\% = \frac{v_{\text{new}} - v_{\text{old}}}{v_{\text{old}}} \times 100 = \frac{15.25 - 6.75}{6.75} \times 100 = 125.9\%$

where $v_{\text{old}}$ is the monthly average before the change, $v_{\text{new}}$ the monthly average after, and $\Delta\%$ the percentage change between them (Foundations: percentages). Clicks from Google to the page, for what it is worth, rose by about a fifth, because the AI answers that cite the page send a trickle and the third position sends a little more. The enquiries more than doubled. Several of them, when asked how they found Calder Ledger, said "ChatGPT suggested you" or "it came up in the Google summary", and the visit that followed was often a direct one, typed in. That is the shape of visibility now: seen in the answer, visited later, counted under "direct" unless you ask.

Everything in this example is invented, and the figures are small on purpose, because they are the size of figures a 400 customer firm actually has. What is not invented is the mechanism, which the engines document, and the direction, which the paper measured.

What I do, and what I would not bother with

My own practice on this is short, and it is the same on a Rails application as on a Solidus shop.

I write pages as questions and answers because that is how my readers think, and I let the engines benefit as a side effect. Every page that states a fact from outside my own work cites it, which you can see at the foot of this article. Every site I build ships with Organization markup that matches the registered company exactly, Article markup with a real author on every article, and a robots.txt file whose every line the owner has read and agreed to. I decide the training question with each client rather than for them, because it is their content. And I measure with a fixed list of questions asked monthly, alongside Search Console, where Google counts AI Overview and AI Mode traffic inside the ordinary "Web" search type rather than separately,³ so the question list is the only instrument that isolates the effect.

Things I would not bother with. FAQPage markup in the hope of a rich result, for the reason above. Rewriting a whole site "for AI" in one sweep, when the paper and the engines both say the gain comes from passages, which you can improve one page at a time starting with the pages that matter. Any tool that scores your "GEO readiness", because the inputs are the three things in this article and an afternoon. Keyword stuffing, which lost in 2012 and lost again in the 2024 paper by about 9%. And hiding content behind nosnippet by accident, which I check for on every site before I do anything else.

What this article deliberately does not cover, because its siblings do. The specifics of ChatGPT and the other chat engines, including how far they still lean on Bing's index, which nobody outside OpenAI can say for sure, are in how to rank in ChatGPT and AI search engines. The strategy of giving the whole answer away on the platform and expecting no click at all is in zero click content. This one is the mechanism: be in the index and through the gate, write a passage worth lifting, give it facts and a name worth citing. Róisín's page did not move much in the rankings. It moved a great deal in the answers, and the answers are where her prospects now read.

Footnotes

1. 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/

2. Rand Fishkin, 2024 Zero Click Search Study: For every 1,000 US Google searches, only 374 clicks go to the open web. In the EU, it's 360, SparkToro, 2 July 2024 (the post's own text gives 374 for the EU and 360 for the US, the headline swaps them; secondary source: a Datos clickstream panel, September 2022 to May 2024, with the author's caveats on representativeness), https://sparktoro.com/blog/2024-zero-click-search-study-for-every-1000-us-google-searches-only-374-clicks-go-to-the-open-web-in-the-eu-its-360/

3. Google, AI features and your website, Google Search Central documentation, https://developers.google.com/search/docs/appearance/ai-features

4. OpenAI, Overview of OpenAI crawlers (GPTBot, OAI-SearchBot, ChatGPT-User), OpenAI developer documentation, https://developers.openai.com/api/docs/bots

5. Google, Featured snippets and your website, Google Search Central documentation, https://developers.google.com/search/docs/appearance/featured-snippets

6. Google, General structured data guidelines, Google Search Central documentation, https://developers.google.com/search/docs/appearance/structured-data/sd-policies

7. Google Search Central, Changes to HowTo and FAQ rich results, 8 August 2023, updated 14 September 2023, https://developers.google.com/search/blog/2023/08/howto-faq-changes

8. Google Search Central, Latest Google Search documentation updates, entries of 8 May 2026 (Deprecating the FAQ rich result feature) and 15 June 2026 (Removing documentation for the FAQ rich result feature), https://developers.google.com/search/updates#faq-deprecation

9. schema.org, FAQPage, https://schema.org/FAQPage

10. schema.org, Organization (properties legalName, address, sameAs, url), https://schema.org/Organization

11. Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande, GEO: Generative Engine Optimization, Proceedings of KDD 2024, arXiv 2311.09735, 2024, https://arxiv.org/abs/2311.09735

12. Google, Google's common crawlers (Googlebot and Google-Extended), Google Search Central documentation, https://developers.google.com/search/docs/crawling-indexing/google-common-crawlers

13. Bing Webmaster Blog, Announcing new options for webmasters to control usage of their content in Bing Chat, Microsoft, 22 September 2023, https://blogs.bing.com/webmaster/september-2023/Announcing-new-options-for-webmasters-to-control-usage-of-their-content-in-Bing-Chat

14. John Mueller, Top ways to ensure your content performs well in Google's AI experiences on Search, Google Search Central Blog, 21 May 2025, https://developers.google.com/search/blog/2025/05/succeeding-in-ai-search