GEO in 2026: Evidence, Limits, and a Practical Workflow

The Short Answer

Generative engine optimization (GEO) is a useful name for the work of making content eligible, useful, well-supported, and attributable when an AI-assisted search system builds an answer. It is not one ranking algorithm, and it is not a bag of markup tricks.

For Google, the official position is deliberately unexciting: AI Overviews and AI Mode are rooted in the same core Search ranking and quality systems, and there are no additional technical requirements or special optimizations for inclusion. Other products have different retrieval stacks and crawler controls, so a responsible GEO practice begins by naming the system rather than speaking about “the AI” as if it were one machine.

I use GEO as an operating label for five questions:

1. Can the relevant system access the page?
2. Does the page contain something distinct and useful?
3. Can its claims be checked?
4. Can a passage be attributed without losing its meaning?
5. Can I measure what happened without inventing causality?

That is less exciting than “rewrite three headings and win AI search.” It is also much closer to what the evidence can carry.

This is chapter 1 of 6 in my GEO series. It provides the evidence boundary and operating model; later chapters can go deeper without turning early findings into universal laws.

What Changed in Search—and What Did Not

Clicks are under pressure, but the numbers need a denominator and a date.

A Pew Research Center study from March 2025 examined 68,879 Google searches from 900 consenting US adults. About 18% of those searches produced an AI summary. Users clicked a traditional result after 8% of visits with a summary, compared with 15% without one. They clicked a link inside the summary in only 1% of visits.

Those results support a narrow statement: in that sample and period, the presence of an AI summary was associated with fewer outbound clicks. They do not tell us that every query class, country, or later version of Google behaves the same way.

Bain’s February 2025 analysis reported a different survey: about 80% of respondents relied on zero-click results for at least 40% of their searches, and Bain estimated a 15%–25% reduction in organic traffic across affected contexts. The underlying Bain–Dynata survey was conducted in December 2024 with 1,117 respondents. Again, it is evidence of a broad change in user behavior, not a conversion guarantee for sites that receive an AI citation.

What did not change is equally important. Search still needs to discover, index, and assess pages. Google explicitly says its generative features use core Search systems, including retrieval-augmented generation and query fan-out. A page must be indexed and eligible to show a snippet. Helpful, original, people-first content remains more important than a special “AI format.”

The useful conclusion is not that links have stopped mattering. It is that a publisher should distinguish at least three outcomes:

• visibility inside a generated answer;
• an attributed link or citation;
• a visit or conversion after that attribution.

They are related, but they are not interchangeable.

What the Original GEO Paper Actually Demonstrated

The term was formalized in the KDD 2024 paper GEO: Generative Engine Optimization by Aggarwal and colleagues. The paper introduced GEO-bench, with 10,000 queries across multiple domains, and evaluated nine ways of modifying source text.

The headline result is real: some strategies improved source visibility by up to 40% in the authors’ experimental generative engine. Adding citations, quotations, or statistics often helped, and the effectiveness varied substantially by domain.

The qualifications matter:

• Visibility was measured with constructs such as position-adjusted word count and a model-graded subjective impression. It was not simply “the probability that a current product cites this page.”
• The main engine retrieved sources and generated answers with gpt-3.5-turbo; production systems in 2026 do not share one fixed pipeline.
• The live Perplexity experiment used 200 examples and supplied source text through file uploads because the researchers could not choose the retrieved URLs. Quotation addition improved one visibility metric by 22%; statistics improved another by as much as 37%.
• The combination experiment also used a 200-example subset. Its best pair exceeded the best individual strategy by about 5.5%, while combinations involving cited sources averaged a 31.4% improvement in the selected metric.
• The authors did not evaluate effects on ordinary search rankings and warned that methods would need to adapt as engines and query distributions changed.

This is valuable early evidence. It is not permission to promise that adding three quotations will raise a page’s ChatGPT, Gemini, or Perplexity citation rate by 40%.

My practical reading is modest: verifiable evidence and relevant quotations can make a source more useful during synthesis, but the effect must be tested by domain and product. The paper gives us hypotheses and an evaluation pattern, not a timeless recipe.

Crawler Controls Are Product Controls, Not Ranking Magic

The earlier version of this article put GPTBot, OAI-SearchBot, ChatGPT-User, PerplexityBot, and Google-Extended in one list and called them trucks carrying content into AI answers. That metaphor concealed important policy choices.

Google

Google Search uses Googlebot. Google-Extended is not a separate HTTP crawler; it is a robots token controlling whether already-crawled content may be used for some Gemini model training and grounding in Gemini Apps and Vertex AI. Google’s crawler documentation states that it has no effect on Google Search and is not a Search ranking signal.

OpenAI

OpenAI’s publisher guidance distinguishes search discovery from potential model training. A publisher that wants content included in ChatGPT search summaries and snippets should not block OAI-SearchBot. GPTBot is the control for pages a publisher wants to exclude from potential training. Allowing one does not imply consenting to the other.

Perplexity

Perplexity’s crawler documentation describes PerplexityBot as the crawler used to surface and link websites in search results. Perplexity-User supports fetches initiated by user requests and generally ignores robots rules because the user requested the access. Their purposes and enforcement behavior differ.

A correct robots.txt can preserve eligibility or express a policy. It cannot make an unhelpful page authoritative, and allowing a training crawler does not buy a citation.

What Google Says You Do Not Need

Google’s current guide to generative AI features in Search is unusually direct about several popular GEO claims:

• Google Search ignores llms.txt; it neither helps nor hurts visibility or ranking there.
• There is no special schema.org type required for AI Overviews or AI Mode.
• Structured data is not required for generative search.
• Pages do not need to be split into tiny “AI-readable chunks.”
• Publishers do not need to rewrite text into a special style for AI systems.
• Inauthentic off-site mentions and scaled pages designed to manipulate generative results remain spam, not a durable shortcut.

Structured data still has ordinary uses. Valid Article, Person, or breadcrumb markup can support Search features and make site semantics easier to maintain. It just should not be sold as a direct AI-citation switch.

The same restraint applies to FAQ and HowTo markup. Google deprecated HowTo rich results and restricted FAQ rich results to well-known government and health sites in 2023. A real FAQ can still help a reader. Adding FAQPage to an ordinary personal blog does not create a documented advantage in AI Overviews.

I keep an llms.txt only when I can name a consumer that documents using it. “It is cheap” is not enough if the file becomes another artifact I must keep accurate.

A Five-Layer Working Model

The following model is my engineering checklist, not a list of proven ranking factors:

flowchart TB
    L1["1. Accessible"] --> L2["2. Distinct and useful"]
    L2 --> L3["3. Evidenced"]
    L3 --> L4["4. Extractable without distortion"]
    L4 --> L5["5. Measured and corrected"]

1. Accessible

Use normal technical SEO: return a successful response, expose the main content without authentication, avoid accidental noindex, publish a sitemap, use stable canonical URLs, and make the page work for users on real devices. Then choose crawler permissions separately for each product and purpose.

2. Distinct and useful

Google’s strongest current recommendation is to publish non-commodity, people-first work: original analysis, first-hand testing, or a point of view that a generic summary cannot reproduce.

A clear opening and descriptive headings often help readers. I still use them. I no longer call them universal citation factors or promise that question-form headings are “cited far more often.” That claim needs a controlled test, not an agency blog.

3. Evidenced

Separate three kinds of material:

• externally verifiable claims, linked to the closest primary source;
• measurements with collection date, scope, and method;
• experience or interpretation, labeled as such.

A statistic without its population and date is decoration. A citation that does not support the adjacent sentence is worse than no citation because it borrows authority without transferring evidence.

4. Extractable Without Distortion

Tables, lists, definitions, and short summaries can make complex material easier for both people and software to navigate. The KDD study gives limited experimental support to making evidence visible during synthesis.

Extractability should not become fragmentation. A passage is useful only if it retains the conditions that make it true. “Up to 40% in one experimental visibility metric” must not be shortened to “40% more citations.”

5. Measured and Corrected

Keep discovery, attribution, traffic, and business outcomes separate. Record a baseline, make one class of change, and preserve enough context to explain a later comparison. If an engine changes, treat the series as a new regime rather than splicing it silently into the old one.

The fifth layer protects the first four from becoming ritual.

Revisiting My Own Blog Without Inventing Causality

The original draft used a private Google Search Console snapshot from the old domain: a rolling three-month view with 852 clicks, 878,000 impressions, 0.1% average click-through, and average position 13.2. High-impression queries included unrelated MBTI, medical, and local-history questions; more relevant clicks came from technical searches around Hugo, LangGraph, GPT Researcher, and Go directives.

That is useful first-party evidence about this site’s query mix. It is not evidence that an AI system copied those pages. Search Console impressions and click-through do not reveal whether a model retrieved or cited a page. The proper conclusion is narrower: the old site’s visibility was broad, but much of it did not match the audience I wanted to serve.

A one-run Lighthouse audit also returned SEO 100 and Best Practices 100. That is a configuration smoke test, not a GEO score. It can catch technical mistakes; it cannot prove authority, originality, citation selection, or user value.

The stronger insight from the blog is editorial. A small set of technical pages has repeatedly attracted relevant readers:

• the Hugo build and advanced Hugo notes ;
• MarkItDown , mem0 , LangGraph , GPT Researcher , and NotebookLM ;
• automation directives and TDD .

My next step is not to manufacture dozens of query variants. It is to improve the few clusters where I have direct experience, correct stale claims, link the work coherently, and see whether readers and external systems find it useful.

How to Measure GEO Without Fooling Yourself

1. Use Platform-Owned Reports

Google’s 2026 guidance points publishers to the Generative AI performance report in Search Console for Google Search and Discover. Use that report for Google’s own generative surfaces rather than inferring them from ordinary high-impression pages.

For incoming visits, record referrers such as ChatGPT or Perplexity in analytics and measure downstream behavior separately. A referral proves a visit, not the exact answer or citation that caused it.

2. Audit Answers with Repeated Prompts

A fixed prompt set is useful, but one run is not a measurement. Store:

• the exact prompt and target intent;
• product, model or mode when visible, locale, account state, and date;
• several repetitions;
• whether the site was retrieved, cited, and accurately represented;
• the cited passage and whether it actually supports the generated claim.

Model and index changes create discontinuities. Report them instead of hiding them inside an average.

3. Test Changes as Hypotheses

“Add a direct summary,” “replace secondary sources,” or “publish a first-hand benchmark” are testable interventions. Change a bounded set of pages, preserve the before state, and decide in advance which observation would count against the hypothesis.

Third-party visibility tools can automate collection, but they do not have access to proprietary ranking systems. Evaluate their sampling and normalization before treating a share-of-voice chart as ground truth.

A 30 / 60 / 90-Day Workflow

Days 1–30: eligibility and evidence inventory

• Verify indexability, canonical URLs, sitemap coverage, and page rendering.
• Review crawler policy by purpose: Search, user fetch, or training.
• List the ten pages that already serve the intended audience.
• Mark unsupported statistics, stale product claims, and secondary citations.
• Capture a dated baseline from first-party platform reports.

Days 31–60: improve the work itself

• Replace important secondary claims with primary sources.
• Add first-hand examples that a commodity summary cannot supply.
• Rewrite ambiguous passages so their conditions travel with the conclusion.
• Improve internal links among two or three genuine topic clusters.
• Create a small, versioned prompt audit with repeated runs.

Days 61–90: expose and learn

• Publish the improved pages and record the release date.
• Share them where a real community can challenge the work—not to seed fake mentions.
• Review Search Console’s generative report, verified citations, referrals, and reader behavior as separate series.
• Correct misquotations and stale claims.
• Stop any tactic that produces no interpretable evidence.

The schedule is not a promise of ranking. It is a way to make the experiment auditable.

Ethics: Make the Source Worth Citing

Fabricated reviews, synthetic consensus, and mass-produced pages are not a clever GEO tier. They pollute the evidence that retrieval systems and people depend on. They also violate the spirit—and often the letter—of platform spam and advertising rules.

The durable alternative is slower: publish experience that happened, state the limits, link the primary evidence, correct mistakes, and earn independent discussion without manufacturing it.

My standard is now simple: if a passage would become misleading when lifted into an answer, the passage is not ready. Optimization begins by making the source true enough to survive extraction.

FAQ

Does GEO Replace SEO?

No. For Google Search, official guidance treats generative search optimization as SEO because AI features rely on core Search systems. Across the wider market, GEO remains a useful umbrella for product-specific discovery, attribution, and measurement work.

Should Every Article Start with a 40–100 Word Answer?

Not as a ranking ritual. A concise opening can help a reader and make the article’s scope explicit. Use it when the question has a direct answer; do not flatten an essay or an uncertain investigation merely to create an extractable block.

Should a Personal Blog Publish llms.txt?

Only for a known consumer whose documentation says it reads the file. Google Search ignores it. Keep it accurate if you publish it, and do not count it as ranking work.

Does Structured Data Improve AI Citation Rate?

There is no general evidence for that claim. Use supported structured data for ordinary Search features and maintainable semantics. Google says no special schema is required for its generative Search features.

What Is the Fastest Useful Change?

Correct the most important unsupported claim on a page and replace it with first-party evidence or an honest statement of uncertainty. That improves the work even if no engine changes its answer.

Closing

GEO is young enough that confidence often travels faster than evidence. The original version of this guide made that mistake: it combined one promising paper, several marketing summaries, private analytics, and product folklore into a certainty the sources did not justify.

The correction is not to abandon the subject. It is to practice the discipline the subject claims to reward: make pages accessible, publish something distinct, show the evidence, preserve the limits, and measure each outcome without pretending it proves the next one.

Search interfaces will keep changing. A source that remains useful after the interface changes is the only optimization I trust to compound.

Primary Sources

1. Google: optimizing for generative AI features in Search
2. Google: AI features and your website
3. Google: common crawlers and Google-Extended
4. OpenAI: publishers and developers FAQ
5. Perplexity crawler documentation
6. Pew Research Center: clicks when Google AI summaries appear
7. Bain: zero-click search and AI summaries
8. Aggarwal et al.: GEO, KDD 2024
9. Google: changes to FAQ and HowTo rich results