Generative Engine Optimization (GEO): How to Get Cited Inside AI Answers
Search has changed its shape. A growing share of questions never reaches a list of blue links at all: the search engine answers the question itself, in its own words, and tucks a handful of small links into the answer. If your website is one of those links, you are still in the game. If it is not, the visitor may never know you exist — even if you rank first in the old results underneath. Generative Engine Optimization, or GEO, is the work of getting cited inside those AI-written answers — Google AI Overviews and AI Mode, ChatGPT, Perplexity, Gemini and Copilot.
GEO is not a replacement for SEO. It is a layer on top of it. The AI engines ground their answers in the same web that Google crawls, and Google has said plainly that its AI features sit on top of its core ranking and quality systems. What changes is the unit of competition: not the page, but the passage — one clear, quotable, evidence-rich paragraph that an AI can lift into its answer. This post is the full playbook: the evidence, the writing method, the technical steps that matter, and how to measure citation share instead of ranking.
What GEO Actually Is — and What It Is Not
Generative Engine Optimization means optimising your content so AI engines include and cite your business inside their answers, rather than merely listing your page among ten results. In classic SEO you compete for a position on a page; in GEO you compete to be one of the sources an AI names while it answers.
Three things GEO is not. It is not a trick file or a plugin: Google's own Search Central guide says there are no additional requirements and no special optimisations needed to appear in AI Overviews or AI Mode, and no special schema markup to add. It is not a replacement for ranking: the engines retrieve their sources from the same index, so a page Google cannot find or will not show cannot be cited either. And it is not a guarantee: no agency, tool or consultant can promise a citation for a given question.
What GEO really changes is the writing. An AI answer is built from passages, not from pages. It needs a passage that answers one question completely, in forty to eighty words, with a number and a named source, and that still makes sense copied out of your page. Most pages are not written that way: they bury the answer under an introduction and never say anything another site has not said first. GEO is the discipline of writing the quotable version: the answer first, the evidence attached, the original fact included.
The Evidence: Clicks Are Leaving the Results Page
Before spending effort on citations, ask the blunt question: does the click still exist to be won? The most careful independent measurement comes from the Pew Research Center. Pew followed the real browsing of 900 American adults through March 2025, covering 68,879 Google searches. When an AI summary appeared on the results page, people clicked a traditional search result on 8% of visits, against 15% on pages without a summary. They clicked a link inside the summary itself on only 1% of visits, and they ended their browsing session on 26% of visits with a summary, against 16% without one.
Two cautions belong with those numbers. The study is observational: Pew's authors say they cannot prove the summary caused the change. A 2026 field experiment by Wang and colleagues confirmed the pattern with a stronger design: hiding AI Overviews for a random group of Chrome users for seven days raised the click-through rate to other websites by 8.8 percentage points.
SparkToro and Similarweb's clickstream study of 9 June 2026 (January to April 2026 data) found 68% of US Google searches ended without a click, against 58.5% in their 2024 study — different panels, so treat the rise as a direction, not a precise measurement. Roughly two in three searches now end on the results page itself. And yet being cited still beats not being cited: Seer Interactive's 2026 analysis found that brands cited in an AI Overview earned two to five times the organic click-through rate of brands that were not cited. The citation is a smaller door than the old number-one ranking, but for many searches it is now the only door at the top of the page.
Why Ranking First No Longer Guarantees a Citation
For twenty years the rule of search was simple: rank higher, get more clicks. AI answers break that rule. Ahrefs analysed 863,000 keyword results and about 4 million AI Overview URLs in a study published on 2 March 2026, and found that only 37.9% of the pages cited in AI Overviews also ranked in Google's top 10 for the same query. The rest were split almost evenly: 31.2% ranked in positions 11 to 100, and 31.0% ranked beyond position 100. Six out of ten citations now go to pages that are not on page one for the query being asked.
| Where the cited page ranked for the same query | Share of citations (Ahrefs, Mar 2026) |
|---|---|
| Top 10 | 37.9% |
| Positions 11–100 | 31.2% |
| Beyond position 100 | 31.0% |
A year earlier, in Ahrefs' July 2025 study, about 76% of citations came from top-10 pages — but that study counted only the top three citations per overview, and Ahrefs improved its parsing since, so the two are not like-for-like. Part of the fall is better counting. The direction is agreed: a page-one ranking helps, and it is no longer necessary — or sufficient — to be cited.
The reason has a name: query fan-out. Google's AI features run many related sub-searches behind one question and build one answer from all the results, so the links go to the pages that won the sub-searches, not the head query. The 15 November post on this blog, "How to Get Cited in Google AI Overviews & AI Mode", explains fan-out in full; this post takes it as given.
There is a second, uncomfortable fact about where citations land. A Search Engine Land analysis of roughly 9 million AI answers in 2026 found that 82% of citations on commercial-intent queries went to third-party sources, and only about 3% went to the brand's own pages; Muck Rack's October 2026 study of 40 million-plus responses found the same pattern, with owned media at 15.2%. But when a brand's own content is cited, it produces roughly five times the visibility lift, and owned citations persist three to nine times longer.
The Playbook, Part One: Answer-First Passages
If one technique matters more than any other, it is this: put a short, direct answer immediately under each question heading, before any background, story or explanation. The passage should be forty to eighty words long, answer the question completely on its own, and stand alone if it were copied out of your page and pasted somewhere else. The passage that is already written like an answer is the passage that gets used.
Here is the difference in practice. A citable passage, sitting directly under the heading "What is the cheapest way to charge an EV at home in Pakistan?", reads like this: "The cheapest way to charge an EV at home in Pakistan is a 7 kW wallbox on a time-of-use meter, charging off-peak. Off-peak electricity costs about Rs 34.53 per unit before taxes, against about Rs 46.85 at peak hours, and a wallbox wastes only 8–10% of the electricity it draws, against 13–24% lost through an ordinary 3-pin socket." That is 62 words, it answers the question, it names the numbers, and it can be lifted whole.
Apply the same pattern to every question your page answers: the question itself as the heading, the forty-to-eighty-word answer immediately under it, and the detail below for the human reader who clicks through. Three habits make passages quotable: name things precisely — "Rs 34.53 per unit off-peak" is citable, "much cheaper at night" is not; keep each passage about one thing; and keep the facts consistent with everywhere else on your site. Write the headings in the words your customers actually use, and add an FAQ section in the same phrasing.
Evidence Density: One Cited Statistic Every 150–200 Words
The strongest published evidence for what makes AI answers cite a page comes from a research paper, not a marketing blog. "GEO: Generative Engine Optimization", by Aggarwal and colleagues (Princeton and IIT Delhi, presented at KDD 2024), tested different ways of writing the same content and measured how visible each version became in AI-generated answers. The result: adding quotations, statistics and named sources raised a page's visibility in AI answers by up to 40% — 30 to 40% on the paper's main measure. Keyword stuffing, by contrast, gave little or no improvement. What helps a human reader — real numbers, real quotes, real sources — also makes a page visible to an AI.
Turn that finding into a writing rule: one cited statistic every 150 to 200 words, each anchored to a primary source. "Forrester, January 2026" beats "studies show". Publish original information the AI cannot get from ten other sites — your own test, your own price check, your own worked example with the maths shown. An AI that has seen the same paragraph on fifty sites has no reason to cite yours; an AI that finds the only worked example of an EV charging bill in Pakistan has every reason to.
One limit: statistics must be real, current and correctly attributed. Date the figures that move — prices, rates, model versions — and update them when they change. An AI answer that cites your page for a price you changed six months ago is a citation that costs you a customer.
Crawler Access: Let the Bots In
None of the writing matters if the AI's crawlers cannot read your pages. Each assistant has its own crawler, and blocking the wrong one is a common and costly mistake. OpenAI's search crawler is called OAI-SearchBot, and its documentation says plainly that sites that opt out of it "will not be shown in ChatGPT search answers", The training crawler is a different bot, GPTBot, and OpenAI says the two settings are independent: you can allow OAI-SearchBot for ChatGPT search while blocking GPTBot for training. A single robots.txt line blocking all AI bots can remove you from ChatGPT search without you meaning to.
| Crawler | Whose it is | What blocking it costs you |
|---|---|---|
| GPTBot | OpenAI (training) | Your content is not used to train OpenAI models; does not affect ChatGPT search |
| OAI-SearchBot | OpenAI (search) | Your pages will not be shown in ChatGPT search answers |
| ClaudeBot | Anthropic | Your pages are not available to Claude's search and answers |
| PerplexityBot | Perplexity | Your pages are not available to Perplexity's answers |
| Google-Extended | Controls use for AI features; blocking it does not remove you from normal search |
Check your robots.txt once, deliberately: fetch it, read every line that mentions a bot, and ask what each line actually blocks. Muck Rack's October 2026 study gives the stakes: among major publishers that blocked Claude's search crawler, 81% got almost no Claude citations, against 12% of those that allowed it. One more access rule: serve prices, specifications and comparisons in server-rendered HTML — AI crawlers do not execute JavaScript reliably, so a price that only appears after a script runs may never be seen.
Freshness Is Its Own Axis
AI assistants prefer newer content than traditional search does, and the gap is measured. Ahrefs analysed 16.975 million citations across five engines and compared them with organic Google results for the same queries. The average AI-cited page was 1,064 days old; the average organic result was 1,432 days old — a 25.7% gap, whether measured from publication date or last update. ChatGPT leaned newest; Google's own AI Overviews barely differed from plain organic search.
Read that number with its limit attached: the average cited page is still 2.9 years old, against 3.9 years for organic results. Freshness is an advantage, not a gate. What it rewards is genuine maintenance: a page whose prices, versions and figures are updated when the world changes, with a visible update date near the top. For fast-moving topics a quarterly refresh is the minimum; for stable topics, a yearly check is enough.
Two Honest Conflicts: llms.txt and Schema
Two of the most argued questions in GEO have no settled answer, and this post will not pretend otherwise.
The llms.txt question. llms.txt is a proposal, published in September 2024 by Jeremy Howard of Answer.AI, for a Markdown file at your site's root that summarises your site for AI systems. No major AI search engine has said its systems use the file for ranking or citation. Google's Search Central guide is explicit: Google Search ignores llms.txt, and creating one "will neither harm nor help your site's visibility or rankings in Google Search".
The server logs agree. Ahrefs analysed llms.txt requests across 137,210 domains in May 2026 and found that 97% of llms.txt files received zero requests that month — nothing fetched them at all. Of the 3% that were fetched, most requests came from SEO audit tools, general crawlers and site profilers; AI search and retrieval bots made up just 1.1% of requests. SE Ranking, analysing nearly 300,000 domains, found no link between having the file and being cited. Its value today is narrower: for AI coding assistants and other tools that opt in to reading it — not the big search engines.
The schema question. One 2026 vendor playbook claims that adding FAQPage schema gives roughly a 2.7-times citation lift. The claim could not be traced to a named study, and should be treated as a sales claim until its source is shown. Against it stands the only controlled test published so far. Ahrefs, in a study published on 11 May 2026, tracked 1,885 pages that added JSON-LD schema between August 2025 and March 2026, matched them against about 4,000 control pages, and compared citation changes across Google AI Overviews, Google AI Mode and ChatGPT. The result: citations moved by −4.6% on AI Overviews (small but significant; cause unexplained), +2.4% on AI Mode and +2.2% on ChatGPT — both indistinguishable from zero. Adding schema produced no meaningful citation uplift on any of the three platforms.
Two caveats keep the result honest. Every page already had at least 100 AI Overview citations, so the study speaks about pages already inside the consideration set, not pages trying to break in from zero. And it pooled all schema types together rather than testing each separately. A second study points the same way: Chen and Koudas, in a preprint dated 10 October 2026 analysing 158,099 answers and 468,037 citations, found no significant within-website effect for FAQ blocks, structured data, summary boxes, author boxes or page speed.
What to do with this conflict: keep valid structured data that mirrors visible content, but do not buy a promised citation increase because someone will add JSON-LD. Google's FAQ rich result stopped appearing in Search on 7 May 2026, so FAQPage markup is no longer a rich-result play at all. Spend the money on the writing, not on the markup.
How to Measure Citation Share, Not Ranking
GEO cannot be measured with a rank tracker. A single "AI rank" does not exist, because the answer changes every time it is asked. Ahrefs tracked more than 43,000 keywords for a month and found that the average AI Overview kept its content for only 2.15 days — and each time one changed, 45.5% of its cited URLs were replaced. One manual check proves almost nothing.
Start with a fixed list of 50 to 150 real buyer questions — the questions your customers actually ask, in their own words. Ask each one across ChatGPT, Perplexity, AI Overviews and Gemini, repeated many times over weeks, recording how often your brand is mentioned and cited. The number that matters is mention frequency: the share of answers in which you appear, per engine, over time. A single check is noise; a month of checks is a trend. Allow six to eight weeks after a change before judging it.
Google has also given site owners a real measuring tool, covered in full in the 16 November post on this blog, "Search Console Generative AI Report": Search Console's Generative AI performance report (announced 3 June 2026, rolled out worldwide 31 August 2026) shows impressions — how often links to your site appeared in AI Overviews and AI Mode — by page, country, device and date. Know its limits: impressions only, no clicks, no queries, and no split between AI Overviews and AI Mode. Clicks from AI features sit in the main Performance report, under the Web search type. Bing Webmaster Tools has its own AI Performance report, which also shows the grounding queries Google's report does not.
The Pakistan Angle: Bloggers, Shops and AdSense
For Pakistani site owners, the shift arrives with a local twist: Urdu content competes in a smaller pool than English, so a good Urdu article answering a question directly can be the passage an AI quotes, where the same article in English would be one of fifty.
There is also a practical warning for anyone earning from AdSense. The Pew numbers say plainly that fewer people click through when an AI summary is on the page; Ahrefs puts the top result's click loss at about 58% for overview keywords. A citation is therefore not a vanity metric: for many searches it is now the only link at the top of the page. Start where the sub-questions are cheapest to win: prices in rupees, PTA rules, tariff questions, local delivery and warranty terms — the ones global sites answer badly and Pakistani pages can own.
Frequently Asked Questions
What is Generative Engine Optimization (GEO)?
GEO is the work of getting your website cited inside AI-written answers — Google AI Overviews and AI Mode, ChatGPT search, Perplexity, Gemini and Copilot — instead of merely ranking in a list of links. It is a layer on top of SEO, not a replacement.
Can anyone guarantee my page will be cited?
No. The decision is made by each engine's systems, for each query, when it is asked. Anyone offering a guaranteed citation is selling a promise they cannot keep.
Should I add an llms.txt file?
It will not help you get cited by Google: Google Search ignores the file, and Ahrefs found that 97% of llms.txt files received zero requests in May 2026. Keep one only if you want it for AI coding assistants or other tools that opt in to reading it.
Does schema markup help AI citations?
The only controlled test published so far says no. Ahrefs tracked 1,885 pages that added JSON-LD and found no meaningful citation uplift on Google AI Overviews, AI Mode or ChatGPT. Keep valid structured data for normal search, but do not pay for schema as a citation tactic.
How many questions should I track to measure GEO?
Use 50 to 150 real buyer questions, asked across ChatGPT, Perplexity, AI Overviews and Gemini, and repeated many times over weeks. Track mention frequency — the share of answers in which you appear — not a single "AI rank". Allow six to eight weeks after a change before judging it.
Is GEO worth it for a small Pakistani business?
Yes, if your customers research before they buy: Forrester found 94% of business buyers used AI during their buying process in 2026. Start with the sub-questions global sites answer badly — rupee prices, PTA rules, tariffs, delivery and warranty.
How is GEO different from the 15 November AI Overviews post on this blog?
That post explains how to get cited in Google's AI Overviews and AI Mode specifically. This post is the full playbook across all AI engines, and the 16 November post covers the Search Console report you measure it with.
Conclusion
Generative Engine Optimization is not a new trick and not a new file to upload. It is the old discipline of good writing, aimed at a new reader: an AI that builds its answer from the clearest passage it can find, with the strongest evidence attached. Answer the question in the first forty to eighty words, put one cited statistic every 150 to 200 words, let the crawlers in, serve your facts in plain HTML, and keep the page genuinely fresh. Then measure what matters — the share of real buyer questions in which your site is cited, across engines, over weeks — and give it six to eight weeks before you judge.
The click is leaving the results page. That is not a reason to stop publishing; it is a reason to publish the kind of page an AI can quote — the clearest, best-evidenced answer to every question your customers actually ask.
By Umar Farooq, Tech Know.


0 Comments