Generative engine optimization (GEO) is the practice of making your content easy for AI systems such as ChatGPT, Perplexity, Claude, Copilot and Google's AI Overviews to find, understand and cite in the answers they write. It is no longer a side project: in October 2026, 44.1% of the Google Search Console impressions on our own site that came with a visible query were machine-generated (prompt-shaped queries, rank-tracker probes and scraping queries), and not one of them produced a click.
This guide explains what GEO and answer engine optimization (AEO) are, how they differ from SEO, what the research says actually moves AI citations, what our own data shows, the risks, and the on-site and off-site playbook we apply on technovapartners.com.
In short
- GEO works, but not like SEO: in the original GEO study, adding quotations, statistics, fluent prose or source citations each lifted visibility in generative answers by roughly 28-43%, which the authors summarize as "up to 40%", while keyword stuffing scored below the unoptimized baseline (Aggarwal et al., KDD 2024).
- Fewer clicks reach the top result: Ahrefs measured a 58% lower click-through rate for the top-ranking page when an AI Overview is present, comparing December 2023 with December 2025 (Ahrefs).
- Machines already query search engines, but never click: 44.1% of our attributed impressions in October 2026 (impressions where Google shows the query) came from machine-generated queries, all with zero clicks.
- Off-site signals weigh as much as on-site ones: across 75,000 brands, web mentions correlated with Google AI Overview visibility at 0.664, against 0.218 for backlinks (Ahrefs).
- No special file or markup is required for Google's AI features, according to Google itself (Google Search Central). The fundamentals still decide.
What Is Generative Engine Optimization (GEO)?
Generative engine optimization is the discipline of increasing how often, and how favorably, your brand and pages appear inside answers generated by large language models. The term was coined in a 2023 paper by Pranjal Aggarwal and colleagues, later presented at KDD 2024, which defined generative engines as systems that retrieve sources and then synthesize a single answer with inline citations (arXiv).
The goal shifts from "rank in position one" to "be one of the few sources the model reads, trusts and cites". For a B2B company that means three jobs: structure content so that a passage can stand alone, back it with evidence, and earn mentions on the third-party sites that models also read.
If you are new to the underlying technology, our guide on what an LLM is explains how these models generate text.
GEO vs SEO vs AEO: How AI Search Optimization Differs
GEO, often called AI search optimization, does not replace SEO: it keeps the same foundations (crawlable, indexed, helpful pages) and changes the unit of success from a ranked page to a cited passage or a brand mention. The labels overlap, and the market uses several synonyms for the same work:
| Term | Main focus | Typical surfaces |
|---|---|---|
| SEO | Ranking documents in a list of links | Google and Bing classic results |
| AEO (answer engine optimization) | Being the direct answer to a question | Featured snippets, voice assistants, AI Overviews |
| GEO (generative engine optimization) | Being retrieved, cited and described correctly in synthesized answers | ChatGPT, Perplexity, Claude, Copilot, Gemini, AI Mode |
| LLMO (LLM optimization), AI SEO | Industry synonyms for GEO, often used by tool vendors | Same as GEO |
Definitions as of November 2026. GEO as defined by Aggarwal et al.; the other labels as used across the industry.
What does not change: indexation, crawl access, page quality and links. What changes: you optimize passages rather than pages, you need evidence a model can attribute, brand mentions on other sites matter more, and you measure visibility across many prompts instead of one position. "How to rank in ChatGPT" is, in practice, this whole list, covered step by step below.
Why Does GEO Matter in 2026?
GEO matters because a growing share of searches end inside an AI answer, so a page can rank well and still lose the visit. The scale is already large: OpenAI said in October 2025 that more than 800 million people use ChatGPT every week (TechCrunch).
Pew Research Center tracked 68,879 real Google searches by 900 US adults in March 2025. When an AI summary appeared, users clicked a traditional result in 8% of visits, against 15% when there was no summary, and clicked a link inside the summary in just 1% of visits (Pew Research Center, 2025).
Ahrefs ran the same 300,000-keyword comparison twice, each time measuring the top result's click-through rate before and after AI Overviews. The first study compared March 2024 with March 2025 and found a 34.5% lower rate (Ahrefs, April 2025); the update compared December 2023 with December 2025 and found 58% (Ahrefs, February 2026).
And the platforms now report on it. Microsoft launched AI Performance in Bing Webmaster Tools on 10 February 2026, with citations, cited pages and "grounding queries", calling it "an early step toward Generative Engine Optimization (GEO) tooling" (Bing Webmaster Blog).
For a B2B buyer journey, the shortlist is increasingly written by an assistant. If your company is not in that answer, you are not on the list.
What Our Own Search Console Data Shows About Machine Queries
In October 2026, 44.1% of attributed impressions on technovapartners.com came from machine-generated queries, with zero clicks. Those queries leave a visible trace in Google Search Console: long, conversational queries with impressions and no clicks. We measured it for the 28 days from 3 to 30 October 2026.
The numbers
Over that period the site had 169,597 impressions and 393 clicks. Google showed the query for 100,466 of those impressions. We classified each query by its shape:
| Query type (3-30 Oct 2026) | Impressions | Share of attributed impressions | Clicks |
|---|---|---|---|
| Human-looking queries (short or keyword-style) | 56,181 | 55.9% | 88 |
| AI prompt-shaped queries (10+ words, or 6+ words phrased as a request) | 35,564 | 35.4% | 0 |
Rank-tracker probes (numeric prefix, such as 12602: search engine rank tracker) |
8,226 | 8.2% | 0 |
| Search-operator and scraping queries | 495 | 0.5% | 0 |
Source: Google Search Console, technovapartners.com. The remaining 305 of the 393 clicks came from queries Google anonymizes, which is why the table only shows 88.
That is 44.1% of attributed impressions (44,285 of 100,466) from machine-generated queries, with zero clicks. Measured against all 169,597 impressions, including those Google does not attribute to a query, the machine share is at least 26.1%.
Who is sending these queries?
Most likely a mix of AI assistants grounding their answers and AI-visibility trackers that fire synthetic prompts at search engines; Search Console cannot tell them apart. A typical example: "what is the most accurate rank tracking tool for monitoring keyword positions daily? my location is wilmington, california, united states." Counted at query-and-page level, 28% of the impressions from 10+ word queries contained "location is", a pattern typical of trackers that geolocate their prompts.
The landing pages show what machines look for. Those queries touched 141 URLs, but two pages, our comparisons of rank tracker tools and keyword research tools, took 49% of the long-query impressions, precisely the categories whose buyers run AI-visibility tools. So read these impressions as evidence that machines retrieve our pages for buyer-style questions, not as proof of human demand or of citations.
Why aggregate CTR misleads
Our site-wide CTR was 0.23% (393 / 169,597), but that blends people with machines that never click. On attributed queries the gap is clear: CTR was 0.09% across all of them (88 / 100,466), but 0.17% for queries of five words or fewer that were not tracker probes (75 clicks on 44,399 impressions), almost double. Keep only those short queries before judging titles and snippets.
And remember what this data cannot show: whether a model then cited us. Search Console records retrieval, not citation. That needs the measurement stack described below.
How Do Generative Engines Choose Which Sources to Cite?
Generative engines cite the sources their retrieval step finds and their model judges most useful for the specific sub-question being answered, so citation depends on being indexed, being retrievable at passage level and being clearly attributable.
The pipeline, simplified, looks the same across most engines:
- Query rewriting and fan-out. The engine expands the prompt into several searches. Google states that AI Overviews and AI Mode may use this "query fan-out" technique and can show "a wider and more diverse set of helpful links" than classic results (Google Search Central).
- Retrieval from one or more indexes. Each engine combines its own crawler with partner sources. ChatGPT search uses OAI-SearchBot and also "leverages third-party search providers" (OpenAI); Microsoft Copilot sends "generated search queries" to the Bing search service "to ground responses in web data" (Microsoft Learn); Perplexity runs PerplexityBot; Google uses its own index. None publishes the full mix, so being indexed in both Google and Bing is the safe baseline.
- Passage selection. The model reads chunks, not whole sites. A section that answers one question in its first sentences is easier to quote than a narrative that reaches the point in paragraph four.
- Synthesis and attribution. The model writes the answer and attaches citations to the sources that supplied specific facts.
For Google, there are "no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary" beyond being indexed and snippet-eligible (Google Search Central). That is why our technical SEO service is usually the first step of a GEO engagement.
Which GEO Tactics Actually Work? What the Research Says
The best-evidenced GEO tactics are adding quotations, adding statistics, improving fluency and citing sources, each of which lifted visibility in generative answers by roughly 28-43% over the baseline in the GEO-bench experiments. Keyword stuffing did the opposite.
The study tested nine content rewrites on GEO-bench, a benchmark of 10,000 queries (Aggarwal et al., arXiv 2311.09735). The results that matter for content teams:
| Tactic (GEO-bench, 2024) | Score on position-adjusted word count | Versus baseline (19.5) |
|---|---|---|
| Quotation addition | 27.8 | +42.6% |
| Statistics addition | 25.9 | +32.8% |
| Fluency optimization | 25.1 | +28.7% |
| Cite sources | 24.9 | +27.7% |
| Keyword stuffing | 17.8 | -8.7% |
Scores from the paper; percentage changes are our calculation against its 19.5 baseline.
The most interesting finding for smaller B2B firms: in GEO-bench, on a research engine, "cite sources" raised visibility by 115.1% for websites ranked fifth in the classic results, while the top-ranked site lost 30.3% on average. GEO partly levels the field, because the model rewards the best-evidenced passage, not only the strongest domain.
Two caveats keep this honest. The experiments ran on a research generative engine, not on production ChatGPT or Gemini, and the authors note that efficacy "varies across domains". Treat the numbers as direction, not as a guaranteed uplift. Our SEO copywriting guide covers how to write evidence-led content without keyword density.
The GEO Playbook: 9 On-Page Changes That Make Content Citable
A citable page answers one question per section, proves every claim with a named source, reads clearly, and is technically reachable by AI crawlers. These are the nine changes we apply, in the order we usually work through them on a B2B site.
- Answer first, then explain. Open every H2 with a one or two sentence answer that still makes sense if copied alone.
- One question per section. Keep that direct answer inside a self-contained block of roughly 130-170 words, our working range rather than a published standard. Question-format headings match how assistants decompose prompts.
- Add statistics with a named source and year. "58% lower CTR for the top result (Ahrefs, 2026)" is citable. "Clicks dropped a lot" is not.
- Add attributable quotations. A named, verifiable statement gives the model something to quote and someone to attribute it to. Prefer recognized external experts; label your own team's views as opinion.
- Link to primary sources. Research to the paper, regulations to the official journal, prices to the vendor's page.
- Add a real FAQ section. Question-format H3s with concise answers map to how users prompt, and generate FAQPage structured data on our blog automatically.
- Keep entities consistent. Same company, author and service names everywhere, connected with Organization and Person markup. Our guide to structured data and JSON-LD shows how.
- Write clear, fluent prose. Short sentences, one idea each, defined terms and no filler. Fluency optimization scored +28.7% in GEO-bench, on a par with citing sources.
- Show freshness honestly. Update the content first, then change the
lastModifieddate; never bump the date without a real change.
The authors of the GEO paper put the underlying problem plainly: "Given the black-box and fast-moving nature of generative engines, content creators have little to no control over when and how their content is displayed" (Aggarwal et al.). Evidence and structure are the part you do control.
Our opinion (Alfons Marques, CEO of Technova Partners): GEO is SEO with the bar raised on evidence. The first-hand observation behind that view is the Search Console analysis above: the pages machines retrieve most on our site are the two comparisons with the most verified prices, tables and named sources, not the pages with the most backlinks.
Before and after: one passage rewritten
Before: "AI Overviews are having a big impact on organic traffic, and many sites are seeing fewer clicks than they used to."
After: "An AI Overview correlates with a 58% lower click-through rate for the top-ranking page, according to Ahrefs' comparison of 300,000 keywords between December 2023 and December 2025."
The second version answers first, carries a number, names its source and its sample, and can be quoted on its own without losing meaning.
How to Rank in ChatGPT, Perplexity and AI Overviews: Off-Site GEO
To rank in ChatGPT, Perplexity and AI Overviews, get your brand named on the third-party pages those engines cite for your category, because a large share of their sources are sites you do not own. Profound analysed 680 million citations from August 2024 to June 2025 and found that each platform leans on different sources:
| Platform (Aug 2024-Jun 2025) | Most-cited domain | Second | Third |
|---|---|---|---|
| ChatGPT | Wikipedia (7.8% of citations) | Reddit (1.8%) | Forbes (1.1%) |
| Google AI Overviews | Reddit (2.2%) | YouTube (1.9%) | Quora (1.5%) |
| Perplexity | Reddit (6.6%) | YouTube (2.0%) | Gartner (1.0%) |
Source: Profound, June 2025. Shares of all citations per platform; the mix changes over time and by topic.
For Google's AI Overviews there is also a correlation study: across 75,000 brands, branded web mentions correlated with AI Overview visibility at 0.664, ahead of branded anchors (0.527), branded search volume (0.392), Domain Rating (0.326) and backlinks (0.218) (Ahrefs, May 2025). It covers AI Overviews only and is a correlation, not proof of cause, but it points to where effort pays.
The tactics we use:
- Get into third-party comparisons. For "best X" prompts, engines retrieve listicles and comparison pages; our own data above shows long prompts landing on exactly that kind of page. Pitch editors with facts, not fees disguised as reviews.
- Publish original data others will cite. Digital PR built on your own numbers earns unlinked mentions, which count in the study above.
- Collect reviews where buyers check. G2, Capterra, Google Business Profile or the marketplaces of your industry.
- Take part in communities genuinely. Reddit, LinkedIn and industry forums, answering as a named expert, never with seeded posts.
- Keep your entity data clean. If you are notable, an accurate Wikidata entry, linked through
sameAsto the same profiles your site declares.
GEO Risks: Brand Misrepresentation and Manipulation
The main GEO risk is not invisibility but being described wrongly: an assistant can repeat an outdated price, merge you with a similarly named company or compare you on features you dropped years ago. Because answers vary between runs, one bad answer may be noise; the same error across many runs is a signal.
When a model describes your brand wrongly:
- Record it. Log the exact wording, engine, date and cited URLs in your prompt set.
- Fix the source, not the answer. If a cited page is yours, correct it; if it is a third-party profile, directory or review site, update it or ask its editor to.
- Publish the canonical facts. A clear about page, pricing or scope page and your own comparison pages give engines a better passage to retrieve.
- Align entity data. Organization markup, Wikidata (if you qualify) and
sameAslinks should all state the same name, founding facts and services. - Use the feedback buttons in ChatGPT, Perplexity and Google's AI features, then re-measure over several weeks.
The other risk is manipulation. Hidden text aimed at AI agents (a form of prompt injection), fake reviews and seeded forum posts can look like shortcuts, but hidden text and fake content already breach search engines' spam policies, and a brand caught doing it loses the trust that citations depend on. We do not use them, and we advise clients against them.
Technical GEO: AI Crawlers, robots.txt and llms.txt
Technical GEO means letting the right AI crawlers reach your content, knowing which ones train models and which ones power live answers, and not wasting effort on files that engines do not require.
Training vs search crawlers
Each major provider separates training from search and user-triggered access:
| Provider (as of November 2026) | Search / answer crawler | User-triggered fetcher | Training control |
|---|---|---|---|
| OpenAI | OAI-SearchBot | ChatGPT-User | GPTBot |
| Anthropic | Claude-SearchBot | Claude-User | ClaudeBot |
| Perplexity | PerplexityBot | Perplexity-User | (none for foundation models) |
| Googlebot | Google-Agent | Google-Extended (control token) | |
| Microsoft | Bingbot | (none documented) | No separate bot: NOCACHE / NOARCHIVE meta tags |
OpenAI states that sites that opt out of OAI-SearchBot "will not be shown in ChatGPT search answers" (OpenAI). Anthropic warns that disabling Claude-SearchBot "may reduce your site's visibility and accuracy in user search results" (Anthropic). Perplexity says PerplexityBot is not used for foundation model training (Perplexity). Google-Extended "does not impact a site's inclusion in Google Search" (Google), and Google-Agent serves agents that act "upon user request" (Google). Bing says NOARCHIVE content will not be included in Bing Chat (now Copilot) answers nor used for training (Bing).
So you can block training and still be cited, as long as the search crawlers stay allowed. We allow all of them, because for a consultancy being in training data is also visibility.
Do you need llms.txt?
No, but it is cheap. The llms.txt proposal, published by Jeremy Howard in September 2024, gives agents a curated Markdown map of a site. We publish one, as a convenience rather than a ranking lever: Google says you "don't need to create new machine readable files, AI text files, or markup" for its AI features (Google Search Central).
How Do You Measure GEO and AEO Results?
Measure GEO with visibility percentage across a fixed set of prompts, plus citation data from the platforms that report it, and stop tracking "AI rankings" as if they were stable positions.
AI answers are probabilistic. SparkToro had 600 volunteers run prompts 2,961 times and found "a <1 in 100 chance that ChatGPT or Google's AI, if asked 100X, will give you the same list of brands in any two responses". Rand Fishkin's conclusion is blunt: "any tool that gives a 'ranking position in AI' is full of baloney"; what works is "visibility % across dozens to hundreds of prompts run multiple times" (SparkToro, January 2026).
A GEO measurement stack
| Signal | Where to get it | What it tells you |
|---|---|---|
| Share of answers that mention or cite you | An AI visibility tracker running 50+ prompts daily | Your real GEO trend line |
| Citations and grounding queries | Bing Webmaster Tools, AI Performance (free) | Which pages Copilot and Bing AI cite |
| Prompt-shaped impressions | Google Search Console, queries of 10+ words | Which pages machines are retrieving |
| Human CTR | Google Search Console, queries of 5 words or fewer | Whether titles and snippets still win humans |
| AI-referred sessions | Analytics, referrals from chatgpt.com, perplexity.ai and similar | Whether citations turn into visits and leads |
The stack we run on technovapartners.com, November 2026. Google includes AI Overviews and AI Mode in the "Web" search type, so Search Console cannot isolate them (Google).
The same blind spot exists in analytics: a click from an AI Overview or AI Mode arrives from Google like any other organic visit, so GA4 cannot separate it either. Referrals from chatgpt.com, perplexity.ai or copilot.microsoft.com are visible, but only when the visitor clicks a cited link, which is why citations and AI-referred sessions have to be tracked as two separate signals.
A prompt-set template you can copy
| Prompt | Funnel stage | Origin | Engines | Mentioned? | Cited URL | How we are described |
|---|---|---|---|---|---|---|
| "Best [category] for mid-size companies in Spain" | Shortlist | Sales call | ChatGPT, Perplexity, AI Mode | Yes / No | URL or none | Exact wording |
| "How much does [service] cost?" | Evaluation | GSC long query | ChatGPT, Copilot | Yes / No | URL or none | Exact wording |
| "[Your brand] vs [competitor]" | Decision | RFP question | All tracked | Yes / No | URL or none | Exact wording |
For the tracker itself, we scored ten options with prices verified on vendors' pages in our comparison of the best AI visibility tools. For content optimization software, see our best AI SEO tools.
A 90-Day GEO Plan for B2B Companies
A GEO program needs about 90 days to show a measurable trend: one month to baseline, one to fix, and one to measure. That is our working cadence, not an industry benchmark.
Days 1-30: baseline.
- Build 50-100 prompts from sales calls, RFP questions and long Search Console queries, using the template above.
- Run them daily in a visibility tracker and record share of answers, citations and how the model describes you.
- Verify the site in Bing Webmaster Tools and check robots.txt for the search crawlers listed above.
Days 31-60: fix.
- Rewrite the 10-15 pages that should answer those prompts with the nine-point playbook.
- Close technical gaps: indexation, canonicals, structured data and internal links from hub pages.
- Start the off-site work: comparisons, reviews and original data.
Days 61-90: measure and extend.
- Compare share of answers against the baseline, prompt by prompt.
- Find prompts where competitors are cited and you are not, and create the missing content.
- Report GEO alongside SEO, never as a replacement for it.
If you want a team to run this with you, our SEO consulting service covers the audit, the prompt set and the fixes, within our wider SEO and digital marketing practice.
Write the Passage AI Wants to Quote
Generative engine optimization rewards what good SEO always did, with a higher bar on evidence, structure and reputation:
- Be reachable: allow the search crawlers of OpenAI, Anthropic, Perplexity, Google and Microsoft.
- Be quotable: answer first, one question per section, with statistics, sources and named expertise.
- Be mentioned: earn a place in the third-party pages models read about your category.
- Be measured correctly: visibility percentage across many prompts, not a single "AI rank" or a blended CTR.
44.1% of the attributed impressions our pages received in October 2026 came from machine-generated queries. The question is no longer whether AI engines read your content, but whether they cite it.
Frequently Asked Questions
What is the difference between GEO and AEO?
AEO (answer engine optimization) focuses on being the direct answer to a question, as in featured snippets and voice assistants. GEO (generative engine optimization) is broader: being retrieved, cited and described accurately in answers synthesized by ChatGPT, Perplexity, Claude, Copilot or Google AI Overviews. The tactics overlap heavily.
What is LLMO or AI SEO?
LLMO (LLM optimization) and AI SEO are industry synonyms for generative engine optimization, used mostly by tool vendors and agencies. They describe the same work: making content retrievable and citable by large language models, and measuring how often AI answers mention your brand.
Does GEO replace SEO?
No. Google states that pages only need to be indexed and eligible for a snippet to appear in AI Overviews and AI Mode, with no special optimization required. Indexation, crawlability and helpful content remain the foundation; GEO adds passage structure, evidence, off-site mentions and AI-specific measurement on top.
How do I get my brand mentioned in ChatGPT?
Make sure OAI-SearchBot and Bingbot can crawl your site, publish pages that answer buyer questions directly with sourced facts, and earn mentions in the third-party comparisons, reviews and communities that ChatGPT retrieves for your category. Then track a fixed prompt set to see whether mentions grow.
How do I check if AI cites my website?
Use Bing Webmaster Tools' free AI Performance report for Copilot citations, an AI visibility tracker for ChatGPT, Perplexity and AI Overviews, and analytics referrals from chatgpt.com or perplexity.ai. Single manual checks mislead, because answers change from run to run.
What are examples of generative engine optimization strategies?
Typical GEO strategies are opening each section with a direct answer, adding statistics and quotations with named sources, writing fluent prose, earning mentions in third-party comparisons, Reddit threads and review sites, and keeping entity data consistent. The before-and-after passage in this guide shows the on-page version in a single sentence.
What are the best GEO tools?
GEO tools fall into three groups: AI visibility trackers that run prompt sets across ChatGPT, Perplexity and AI Overviews; free platform reports such as Bing Webmaster Tools' AI Performance; and Search Console plus analytics for retrieval and referrals. We compared ten trackers in our guide to the best AI visibility tools.
Is GEO worth it for small and B2B companies?
Usually yes, because the cost is mostly content work, not media spend, and GEO-bench found the largest gains for sites that were not ranked first. For B2B firms it matters more, as buyers increasingly ask assistants for supplier shortlists. Start with 50 prompts and your ten most important pages.
How much does GEO cost?
The tool part is modest: in our comparison of AI visibility tools, entry plans start at $17-29 a month and wider coverage costs around $300. The larger cost is the work: rewriting key pages, technical fixes and off-site outreach.
Do I need an llms.txt file for generative engine optimization?
It is optional. llms.txt is a 2024 proposal that gives AI agents a curated Markdown map of your site, and it costs little to publish. However, Google says you do not need new machine readable files or AI text files to appear in its AI features, so do not expect it to change citations on its own.
How long does generative engine optimization take to show results?
With our working cadence, about 90 days: 30 to build a prompt baseline, 30 to rewrite and fix the key pages, and 30 to measure the change. Because AI answers vary between runs, judge results on visibility percentage across many prompts, not on single checks in ChatGPT.
If AI assistants are already shortlisting suppliers in your category, talk to us: we will measure where you appear today and build the plan to get cited.
About the author: Alfons Marques is CEO of Technova Partners and leads the SEO and GEO programme of technovapartners.com, a six-language B2B site, including the Search Console analysis published in this guide.
