Generative Engine Optimization (GEO) is the practice of designing content, schema, and measurement so that generative AI systems (ChatGPT, Claude, Perplexity, Google AI Overviews, Copilot, Gemini) consistently surface and cite your content when answering user questions. The term was introduced in the 2023 Princeton paper "GEO: Generative Engine Optimization" by Aggarwal et al.

Generative Engine Optimization (GEO): How Your UTM Taxonomy Becomes an AI Citation Asset

Generative Engine Optimization (GEO) is the practice of designing content, schema, and measurement so that generative AI systems (ChatGPT, Claude, Perplexity, Google AI Overviews, Copilot, Gemini) reliably surface and cite your pages. The term was introduced in the 2023 Princeton paper “GEO: Generative Engine Optimization” by Aggarwal et al. In 2026, GEO has matured into an operational discipline with three legs: content shape that AI systems prefer, schema and freshness signals that drive citation eligibility, and a measurement loop that proves it is working. Your UTM taxonomy is the load-bearing piece of that third leg.

Published by Terminus, the marketing taxonomy governance platform.

TL;DR

GEO was coined in the Princeton paper “GEO: Generative Engine Optimization” (Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan, Deshpande, 2023). Vendor framings (Profound, AthenaHQ, Otterly.AI, BrandRank.AI) build on top.

Citation eligibility is shaped, not random: TL;DR blocks, FAQ density, dated stats, source attribution, and JSON-LD schema all measurably affect inclusion in generative answers.

You cannot improve what you cannot measure. AI-referred traffic shows up in your logs with distinctive hostnames (chatgpt.com, perplexity.ai, claude.ai, copilot.microsoft.com) and benefits from explicit UTM tagging when you control the link.

The robots.txt picture is contested in 2026. The llms.txt spec (llmstxt.org) is a real, proposed standard, but no major AI vendor has publicly committed to honouring it as of mid-2026.

A clean UTM taxonomy is what closes the GEO loop: it lets you separate “AI sent us a click” from “the AI cited us but the user clicked through somewhere else.”

What GEO actually is

The phrase “Generative Engine Optimization” has a primary source: the 2023 paper by a Princeton-led team (Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan, Deshpande). The paper introduced both the term and an empirical benchmark (GEO-bench) for evaluating how content rewriting techniques affect a source’s visibility inside generated answers. The authors tested rewriting strategies (adding citations, quoting authorities, adding statistics, adjusting fluency) and reported double-digit improvements for several techniques under their evaluation setup.

That is the academic anchor. The vendor ecosystem that has formed since then uses overlapping definitions:

Profound frames GEO as “AI Search Optimization” and focuses on prompt-level visibility across ChatGPT, Perplexity, Gemini, and Copilot.
AthenaHQ positions itself around “AI brand visibility” with an emphasis on competitive share-of-voice inside LLM answers.
Otterly.AI uses the term “AI search monitoring” and tracks prompt-level mentions and citations.
BrandRank.AI uses the language of “Answer Engine Optimization” (AEO) and frames it as monitoring how brands are described inside generative answers.

Some practitioners distinguish GEO from AEO (Answer Engine Optimization, used to mean optimising for featured snippets and AI Overviews). In practice the two are converging. In this article, GEO refers to the broader discipline of getting cited by generative AI systems, with the Princeton paper as the canonical reference.

What GEO is not: it is not “SEO for LLMs” rebranded (SEO ranks pages on a SERP; GEO influences whether your content shows up inside a generated answer, often without a user visiting your site); it is not paid placement (no major LLM offers a paid “be cited” slot in 2026); and it is not a one-time technical fix.

How LLMs decide what to cite

The internal ranking signals of major generative engines are not fully public. What is observable, from the Princeton paper’s empirical results and from patterns practitioners see across published GEO experiments, is that citation eligibility consistently correlates with a small set of features.

Content shape. Content structured for retrieval (clear headings, short paragraphs, bullet lists, TL;DR summaries, FAQ blocks) is easier for a retrieval-augmented generation (RAG) pipeline to chunk, embed, and quote. A 2,500-word essay with one continuous wall of prose competes poorly against a 1,200-word piece with 12 H2/H3 anchors and a TL;DR block.

Freshness. Generative systems favour recently-dated content, especially for time-sensitive topics. The signal is not just the HTTP Last-Modified header but a visible, parsable date in the page itself (a datePublished/dateModified JSON-LD pair, plus an in-body “Last updated” line).

Schema markup. Schema.org structured data (Article, FAQPage, HowTo, DefinedTermSet) makes content explicitly machine-parseable. FAQPage schema in particular gets your Q&A blocks chunked cleanly and they show up as standalone answer units.

Source trust. Generative engines preferentially cite sources already trusted by the broader web. Domain authority, inbound link patterns, and corroborating mentions across crawled pages feed into whether your content gets selected from a retrieval candidate set. This is the part of GEO that overlaps with traditional SEO.

Attribution discipline. This is the Princeton paper’s most interesting empirical finding. Adding citations to your own content (linking primary sources, quoting authorities, naming specific studies with dates) measurably improved source visibility in their benchmark. Engines favour content that itself attributes well, presumably because it is easier to chain into a transparent answer.

Tagging discipline (the GEO loop). Citation is the input. Conversion is the output. The middle layer is measurement: knowing which content is being cited, which surface is sending clicks, and which campaigns convert. That is where UTM taxonomy enters the picture.

Measurement: how to know if you are being cited

There are three observable signals that you are being cited by generative engines.

Direct citation tracking

A small market of citation-monitoring tools has formed since 2024. As of mid-2026, the recognisable names include Profound, AthenaHQ, Otterly.AI, BrandRank.AI, Peec AI, Goodie AI, and Daydream. Each works by repeatedly prompting the major generative engines with a defined set of queries and parsing the answers for mentions and citations of target domains. Pricing moves frequently; check the public pages directly.

What they give you: prompt-level visibility, share-of-voice within answers, competitive comparison, and historical trend lines. What they cannot give you: a complete picture of clicks landing on your site from generative answers. For that, you need referrer signals and your own logs.

Referrer signals

When a user clicks a citation in a generative engine’s answer, your server gets a referrer (mostly). Hostnames worth tracking in 2026: chatgpt.com (which replaced chat.openai.com as the canonical user-facing host in 2024), perplexity.ai, claude.ai, copilot.microsoft.com, gemini.google.com, you.com, phind.com, kagi.com, and meta.ai.

Note the asymmetry: browser-based tools send a normal Referer header, but app-based or in-window experiences (the Gemini panel inside Google search, ChatGPT in the macOS desktop app, Copilot in the Edge sidebar) may strip the referrer. Treat the signal as directional.

For a deeper walkthrough of the referrer field, including the regex patterns that catch each engine and the BigQuery-side equivalents for warehouse users, see the companion piece “The ‘AI Traffic’ Channel in GA4: How to Define It Yourself.”

Server log audits

The third signal is the AI bot itself, crawling your pages. Major user agents to track include GPTBot, OAI-SearchBot, ChatGPT-User (OpenAI), ClaudeBot and Claude-Web (Anthropic), PerplexityBot, Google-Extended (Google’s opt-out token for Gemini and Vertex AI training), Applebot-Extended (Apple Intelligence), cohere-ai, Bytespider, and Amazonbot.

A weekly grep over your logs (grep -iE 'gptbot|oai-searchbot|chatgpt-user|claudebot|perplexitybot' access.log) gives a baseline of crawl pressure. Cloudflare, Fastly, and similar CDNs offer pre-built AI-crawler dashboards.

Crawl is not citation. A high GPTBot crawl rate does not mean ChatGPT is citing you. But a near-zero crawl rate strongly suggests you are invisible to that system, and the gap is worth investigating.

The role of UTM taxonomy in the measurement loop

Here is where it gets concrete, and where the UTM taxonomy you already maintain (or should) does load-bearing work.

The GEO measurement loop has three stages: citation (an AI surface includes your URL in its answer), click (a user clicks the citation), and conversion (that user does the thing you wanted). Each stage is measured differently. Citation is measured via the monitoring tools and server logs. Clicks are measured via referrer signals and your analytics platform’s session source/medium attribution. Conversion is measured wherever you measure conversion already.

The problem is that AI-referred clicks arrive on your site without UTM parameters. The AI surface cited the canonical URL. The user clicked. Your analytics platform sees the referrer as chatgpt.com or perplexity.ai and your channel grouping classifies it as “AI Traffic.” But you cannot tell which piece of content was cited, you cannot tell whether the citation came from a specific prompt or competitive context, and you cannot connect that click to a downstream conversion without a clean session-source/medium pairing.

This is where UTM taxonomy steps in, on the surfaces where you do control the link. Consider the categories of URLs AI systems cite: pages you own (no inbound UTM tagging is possible; the AI engine grabs the canonical URL); links you placed elsewhere (guest posts, syndicated content, third-party content marketing, podcast show notes); outbound links from your content to third-party tools (not a measurement opportunity but they affect citation eligibility); and intra-site links (not relevant to inbound GEO measurement).

A worked example (illustrative, framework setup). A B2B SaaS team uses a UTM taxonomy governed in Terminus with allowed values for utm_source, utm_medium, and utm_campaign. Their convention reserves utm_source=ai_referral and utm_medium=ai_search for any link where they anticipate AI citation, with utm_campaign carrying the content surface.

When the team contributes a guest post to a partner site, they place their outbound link with the canonical UTM. Six weeks later, ChatGPT cites the partner article in an answer to a user query, and the user clicks the contextual link inside the partner article. The team’s analytics platform sees:

utm_source = ai_referral
utm_medium = ai_search
utm_campaign = guest_post_utm_governance
referrer    = chatgpt.com

Now they can attribute this conversion not just to “AI traffic” but to the specific outbound placement, content asset, and surface that drove the citation. Multiply that across dozens of placements and you have a measurable GEO program. Without the taxonomy, the same click is one undifferentiated row in the “AI Traffic” bucket.

The taxonomy governance discipline matters because the failure mode is fragmentation: if half the team writes ai_referral and the other half writes AI_Referral, ai-referral, or AI Search, your reports show four different sources where you have one. A governance platform enforces a controlled vocabulary at link-creation time. (For the broader case, see “The 2026 UTM Tagging Guide”.)

GEO without UTM measurement is theatre. You can publish FAQ-dense, schema-rich, beautifully-dated content all year, and your only feedback signal will be “the monitoring tool says we are getting mentioned more this month.” That tells you nothing about which content drove which conversion. UTM taxonomy is what turns GEO from a vanity metric into a marketing channel.

Practical content patterns that get cited

The Princeton paper’s empirical findings, combined with two years of vendor and practitioner reports, point at a consistent set of content patterns that together describe a content shape that maximises citation eligibility.

Lead with the answer

Generative engines retrieve answer chunks. A page that buries its answer in paragraph six is harder to chunk than a page that puts the standalone answer in the first 80 words.

TL;DR blocks and FAQ sections with question-shaped H3s

A blockquoted TL;DR with 4 to 6 bullets gives the retrieval system a pre-summarised chunk. An FAQ section with 6 to 10 question-shaped H3s gives the model 6 to 10 distinct citation candidates from one page. Add FAQPage JSON-LD and the chunks are explicitly machine-parseable. The H3s should be plain questions, matching the natural-language phrasing a user would type. This is the highest-ROI pattern in 2026.

Dated stats and named sources

The Princeton paper’s “quotation_addition” and “statistics_addition” rewriting strategies sit in the 30% to 40% lift range on Position-Adjusted Word Count (the paper’s headline metric). The “cite_sources” strategy is reported separately as reaching up to 40% overall, and as high as 115% for lower-ranked content. When citing a statistic, name the source, the year, the metric, and link the primary document. A specific Forrester/Marketing Evolution 2019 study figure on bad-data waste beats “studies have shown that bad data costs companies.” The latter is unciteable.

Visible dates, schema markup, and comparison tables

A page with a visible datePublished and dateModified in both JSON-LD and human-readable form signals freshness for time-sensitive queries. At minimum, include Article (or TechArticle), Organization (with alternateName if your brand has disambiguation issues), and FAQPage. The schema must mirror the visible content; engines penalise schema that does not match what is on the page.

For comparison queries, engines preferentially cite content containing an actual comparison table. A markdown table with consistent columns and 6 to 12 rows is highly chunkable and answers the comparison question in one structured unit.

The robots.txt and llms.txt picture in 2026

Two competing standards are evolving for site owners to communicate with AI systems about content access.

robots.txt and AI crawlers

The 1994 robots.txt convention is the established mechanism. Major AI vendors now respect named user agents: OpenAI (GPTBot, OAI-SearchBot, ChatGPT-User), Anthropic (ClaudeBot, Claude-Web), Google (Google-Extended, the token controlling Gemini and Vertex AI training, separate from Googlebot), Apple (Applebot-Extended), Perplexity (PerplexityBot, Perplexity-User), Cohere (cohere-ai), ByteDance (Bytespider), and Common Crawl (CCBot).

Verify each agent name against the vendor’s own documentation (platform.openai.com, docs.anthropic.com, Google Search Central) before adding to robots.txt; the agent inventory has churned multiple times.

The complication: blocking these agents forfeits citation eligibility in the corresponding generative engine. Block OAI-SearchBot and ChatGPT cannot retrieve your pages for in-session answers. Most publishers blocking AI crawlers today are doing so for paywalled or proprietary content; most marketing teams want their content crawled and cited.

llms.txt

The llms.txt specification is a proposed standard published at llmstxt.org by Jeremy Howard in late 2024. It proposes a /llms.txt file at the site root, formatted in markdown, that gives LLMs a curated summary of the site’s key documentation.

Be precise about its status: as of mid-2026, no major AI vendor (OpenAI, Anthropic, Google, Microsoft, Perplexity) has publicly committed to honouring llms.txt as an authoritative source. Treat it as a low-cost hedge.

The broader consent and copyright question remains unresolved in mid-2026, both legally and operationally. Verify your organisation’s stance with legal before configuring robots.txt either way.

A 90-day GEO operationalisation plan

GEO is a content discipline, a schema discipline, and a measurement discipline running in parallel. A 90-day plan that gets a marketing team from “we have heard of GEO” to “GEO is a tracked channel”:

Days 1 to 30: audit and baseline

Inventory the 20 to 50 pages that already rank for your priority queries and the 10 to 20 that map to generative answer queries you want to own.

Run a schema audit. For each priority page, check whether Article, Organization, and FAQPage JSON-LD are present and well-formed using Google’s Rich Results Test.

Set a server log baseline: 30 days of logs grepped for the AI crawler user agents listed above, with a per-page crawl-rate view.

Pick an AI-monitoring tool and run a citation baseline. Subscribe to a vendor (Profound, AthenaHQ, Otterly.AI, BrandRank.AI), or build a manual prompt-test framework with 30 to 50 fixed queries run weekly. The manual approach is laborious but free.

Days 31 to 60: structural fixes and tagging discipline

Retrofit schema on every priority page lacking it: Article + Organization + FAQPage JSON-LD, using the existing FAQ H3s as FAQPage source.

Reshape content on the top 20 priority pages: a self-contained answer in the first 80 words, a TL;DR block of 4 to 6 bullets, an FAQ with 6 to 10 question-shaped H3s, and a visible “Last updated” line.

Align UTM taxonomy. Reserve utm_source=ai_referral and utm_medium=ai_search in your governed taxonomy and add utm_campaign conventions for the surfaces you can tag. Update your URL builder and governance platform to recognise these. A governed taxonomy makes the rules enforceable rather than aspirational.

Build a channel grouping. In GA4, create a custom channel group with an “AI Traffic” channel matching the hostnames listed above. Validate in DebugView. Set up equivalent channel definitions in your warehouse.

Days 61 to 90: measurement loop and iteration

Build weekly dashboards showing AI-referred sessions, AI-referred conversions, and (where measurable) prompt-level citation share, cross-referenced with the crawl-rate baseline.

Run a content gap analysis. From the citation tool or manual baseline, identify queries where you are uncited but competitors are, and classify each gap as content (you have nothing), shape (lacks FAQ density or schema), or freshness (dated).

Run a content sprint on the 5 to 10 highest-priority gap pages. Apply the patterns from section 5. Add schema. Republish with the current date.

Repeat the citation baseline and compare. Expect partial movement; large GEO gains generally show up over 90 to 180 days. Codify the playbook from what worked.

A 90-day plan does not finish GEO; it gets you into the discipline. From there it becomes a quarterly content audit, a monthly schema check, a weekly measurement review, and continuous taxonomy hygiene. Without taxonomy hygiene, you cannot tell which interventions worked. With it, GEO joins the rest of your marketing channels as something you can manage by the numbers.

Frequently asked questions

What is GEO?

How does GEO differ from SEO?

SEO optimises for ranking on a search engine results page; the unit of success is a clickable SERP link. GEO optimises for inclusion inside a generated answer; the unit of success is being cited or quoted in the model’s response. SEO and GEO overlap in that both reward trusted, well-structured, schema-marked content, but GEO weights content shape (TL;DR, FAQ density, dated stats) more heavily because generative engines retrieve and quote chunks of text, not whole pages.

Can I track AI citations directly?

Yes, with caveats. A market of citation-monitoring tools (Profound, AthenaHQ, Otterly.AI, BrandRank.AI, Peec AI, Goodie AI, Daydream) repeatedly queries the major engines and logs which sources get mentioned and cited. Pricing changes frequently; check public pricing pages directly. You can also build a manual prompt-test framework with a fixed query set, which is laborious but free. Direct citation tracking does not capture clicks; for clicks you need referrer signals and a clean UTM taxonomy.

Which AI crawlers should I allow or block in robots.txt?

The major ones in 2026 are GPTBot, OAI-SearchBot, and ChatGPT-User (OpenAI); ClaudeBot and Claude-Web (Anthropic); Google-Extended (Google’s Gemini opt-out token); Applebot-Extended (Apple); PerplexityBot; cohere-ai; Bytespider; and CCBot. Blocking them forfeits citation eligibility in the corresponding engine. Most marketing teams allow all of them. Verify current agent names against each vendor’s documentation; the inventory has changed multiple times.

Is llms.txt a real standard?

The llms.txt specification is a real, proposed standard published at llmstxt.org by Jeremy Howard in late 2024. It proposes a markdown-formatted file at the site root that summarises the site’s key content for LLMs. As of mid-2026, no major AI vendor has publicly committed to honouring llms.txt as authoritative. Maintaining one is a low-cost hedge.

How does UTM taxonomy connect to GEO?

GEO has three measurable stages: citation, click, conversion. AI engines cite canonical URLs, so citation is measured via monitoring tools and server logs. Clicks arrive with referrer signals (chatgpt.com, perplexity.ai, claude.ai). Conversion requires connecting that click to a downstream outcome, which is only possible when source/medium attribution is clean. UTM taxonomy keeps it clean across the surfaces where you can tag links (guest posts, syndicated content, third-party placements). See “How to Track AI Referral Traffic” for the broader measurement walkthrough.

Will paying for placement get me cited?

As of mid-2026, no major generative engine offers a paid “be cited” placement. OpenAI, Anthropic, Google, Microsoft, and Perplexity have kept their citation mechanics non-commercial. Perplexity has experimented with sponsored answer formats, but the organic citation list remains separate. The path to citation in 2026 is content shape, schema, freshness, source trust, and attribution discipline, not ad spend.

How long does it take to see GEO results?

Most content interventions show partial movement within 30 days and meaningful movement within 90 to 180 days. Schema additions tend to move the needle fastest because they affect chunk eligibility immediately on re-crawl. Trust signals (domain authority, inbound links) move on longer timescales.

What is the difference between GEO and AEO?

Some practitioners distinguish GEO (optimising for full generative answer engines like ChatGPT and Perplexity) from AEO (Answer Engine Optimization, often used for featured snippets and Google AI Overviews specifically). In practice the two are converging, because the same content shape drives both. GEO is the broader term; AEO is a subset where the answer engine is part of a traditional search experience.

Does GEO replace SEO?

No. SEO and GEO are complementary. Many users still arrive via traditional search results, and many generative engines (including Google AI Overviews and Bing Copilot) draw on the same underlying search index. A well-optimised page is generally well-positioned for both. The shift is that content optimised purely for SERP click-through (clickbait headlines, thin top-of-page content) underperforms in GEO, where the unit of citation is a parseable, self-contained chunk.

Last updated: 2026-07-07