Generative engine optimization: what GEO is and when to pay for it

In Ahrefs' US database, the query generative engine optimization carries 8,000 searches a month, a keyword difficulty of 72 and a cost per click of around 11 USD, consulted on 17 August 2026. Eleven dollars is what advertisers now pay to reach someone who is asking what the letters mean. That price says more about the selling than about the discipline, because the study that named the field is public, free and specific about which techniques moved visibility and which made it worse. The first page of results for the query, read the same day, sells GEO in nearly every position and quotes the study's failed methods in none.

What is generative engine optimization?

Generative engine optimization (GEO) is the work of getting a page cited inside answers written by ChatGPT, Perplexity, Google AI Overviews and AI Mode. On a new site it waits behind indexing; on a mid-authority site it is the highest-leverage edit available; on a consolidated site it is mostly defending citations already won.

Quick answer

• GEO optimizes for the citation, not the click. The result is your page named inside a generated answer, and the mention often arrives without a visit attached.
• The evidence base is one public study, and it is narrower than the pitch. Nine techniques were tested; three moved visibility clearly, and keyword stuffing performed below baseline.
• Eligibility is nothing new. Google documents no additional requirements beyond being indexed and eligible to appear with a snippet.
• The measured gains concentrate in the middle of the results page. Sites ranked around fifth gained the most; the top result lost visibility in the same experiment.

What kind of problem this is

The pitch arrived before the evidence.

The model most buyers bring to this topic in the United States was installed by headlines, and the loudest of them declared SEO dead with GEO as its successor. From there the conclusion feels natural: a new discipline, a new budget line, a new vendor. The market behaves accordingly, which is what an 11 USD cost per click on a definitional query means in practice.

The correct model is smaller and cheaper. Google's own documentation, checked on 17 August 2026, states that there are no additional requirements to appear in AI Overviews or AI Mode, no special optimizations, no new machine-readable files and no special schema.org markup: a page must be indexed and eligible to appear with a snippet, and the controls that keep one out are the old snippet controls. Whatever GEO is, it is not an infrastructure purchase.

What remains once the infrastructure claim falls away is a writing layer over the same index, plus a naming problem: GEO, AEO, AIO and LLMO circulate as labels for it, with no consensus definition separating them. A buyer who understands that spends on editing and measurement. A buyer who does not, pays for the acronym. Which of those labels names a different job and which is the same invoice with another sticker is the whole of AI search optimization.

What the study that named the field measured

The tables are more useful than the abstract.

The term comes from Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande (arXiv 2311.09735, presented at KDD 2024), who tested nine optimization methods on GEO-bench, a benchmark of 10,000 queries drawn from nine datasets. The abstract's headline, a visibility boost of up to 40 %, is the number that circulates. The tables underneath it are the part a buyer can actually use.

What worked. Adding quotations, adding statistics and citing sources were the strongest methods. The best of them improved on baseline by 41 % on Position-Adjusted Word Count, the metric that measures how much of the generated answer comes from your page and where, and by 28 % on the study's Subjective Impression metric. Statistics Addition alone achieved a relative improvement of 30 to 40 % on the first metric.

What failed. Keyword stuffing, the reflex imported from a decade of SEO habit, showed little to no improvement and performed below baseline. The one tactic most sites already know how to do is the one the measurement argues against.

Where the gains landed. The effect was not evenly distributed. With the Cite Sources method, sites ranked fifth in the classic results gained 115.1 % in visibility while the top-ranked site lost 30.3 % on average. Sites already winning the list have little to gain here and something to lose; sites stuck in the middle have the largest measured upside available to them. The effect also varied by domain, with statistics strongest in law and government topics and an authoritative tone strongest in debate and history.

The full breakdown of the paper, its metrics and its limits is in the Spanish guide on the Princeton GEO study.

The numbers a sales deck will show you

The spread matters more than any single figure.

How often AI Overviews appear depends entirely on who measured. Conductor reports 25.11 % of queries (21.9 million queries, first quarter of 2026), Similarweb reports 43 % (Generative AI Landscape 2026), BrightEdge reports 48 % in commercial verticals (March 2026) and Xponent21 reports 60.32 % in the United States (April 2026). Four answers to the same question, from 25 to 60 %, because each study uses its own keyword sample. A deck that shows one of these without the others is choosing the one that supports its price.

The click-through data comes as a pair, and the pair is the point. Seer Interactive, measuring 3,119 informational queries across 42 organizations between June 2024 and September 2025, found organic click-through where an AI Overview appears fell from 1.76 % to 0.61 %, which Seer reports as a 61 % decline. Where no AI Overview appears it also fell, from 2.72 % to 1.62 %, a 41 % decline. Only the difference between the two declines is attributable to the summary. A pitch citing the first number without the second is counting half the story as the whole of it.

And the click is disappearing as the unit of account. Zero-click searches went from 56 % to 69 % between May 2024 and May 2025 according to Similarweb. In Google's AI Mode, 93 % of queries end without a single click to an external site, a figure with two independent measurements behind it: Semrush, over 69 million US desktop sessions between 1 May and 5 July 2025, and Seer Interactive, over 25.1 million impressions. Where that is the environment, the honest metric is the mention, not the visit, and a report built on sessions is measuring the wrong thing.

The pattern under all three blocks is the same. The drop concentrates in informational queries while transactional ones hold up better, which decides who should care about any of this and who should not.

The work, piece by piece

A pillar earns its name by pointing somewhere.

Each layer of this work has its own article on this blog. What follows is the map: what each piece does, and where the detail lives.

Being in the candidate pool

Every generative system that cites pages selects them from an index it can reach. A page that is not indexed is not a candidate for anything, which makes indexing the gate in front of all of this: the diagnosis order is in why is my page not indexed.

Measurement of that gate comes free and is the second check: a verified property shows which pages Google has stored and which queries they surface for, and the afternoon it takes to configure is in how to set up Search Console.

For a site nobody has audited before, the order of operations starts earlier than either of those, with crawlability, titles and the dozen mechanical checks that decide whether a model has anything to find, collected in the basic SEO checklist.

Writing the passage that survives extraction

The unit that competes in a generated answer is not the page but the block of forty to sixty words that can be lifted out and still make sense. The mechanical test: any fifty-word passage, pasted somewhere with no context, either still says something or it does not.

The difference is visible in a pair. "Our content is optimized for the new era of AI search" cannot be cited, because it asserts nothing checkable. "Of nine optimization methods tested on 10,000 queries in the GEO study, keyword stuffing was the only one that fell below baseline" can travel on its own. How this unit of work differs from classic ranking work, and where each one pays, is in answer engine optimization vs SEO.

Getting named by ChatGPT

ChatGPT has its own pipeline, and it is governed by files most sites have not read since 2023. OpenAI documents separate crawlers for separate jobs: OAI-SearchBot surfaces sites in ChatGPT's search features, while GPTBot collects training data, and blocking one has nothing to do with the other. A site that blocked everything during the publisher opt-out wave and never revisited the decision is out of the citation pool without knowing it. The full sequence, including the Bing index and the difference between being mentioned and being cited, is in how to get cited by ChatGPT.

The local exception

Most US small-business demand does not live in the queries GEO touches. In this blog's own autocomplete harvest of the US market (August 2026), suggestions about local SEO outnumbered suggestions about generative AI six to one, 238 against 39, and those counts measure that the questions exist, not their volume. Citations reward informational and comparative queries; a business that lives on "near me" searches is buying visibility where its customers are not looking. How AI summaries behave specifically in local results is in AI Overviews and local search.

Where that budget works harder for a business with an address is the ordinary local work: reviews, the profile and the service pages that support them, covered in local SEO for small business.

The language gap

For a business serving Spanish speakers in the US, the exposure is not symmetrical. In two client site pairs I operate, same topic in both languages, the Spanish half converts impressions into clicks roughly three times better at comparable positions; English-language results carry more AI summaries, more ads and more enriched results, and the measurement cannot separate those causes, so the ratio travels with that reservation. The comparison is in Spanish vs English SEO.

The decision about whether serving both languages is worth the build, and what a translated site needs so the second language earns its keep instead of duplicating the first, is in translating a website for SEO.

When NOT to invest in GEO

When the demand is transactional or local. The queries that produce citations are informational and comparative, and the drop in clicks concentrates there too; a shop living on product and brand queries has cheaper problems to solve first.

When the site is not indexed, loads badly or duplicates its own content. GEO is a layer on top of that work, not a shortcut around it, and money spent on the layer above a broken base buys nothing measurable.

When there is nothing original to cite. If every claim on the site restates someone else's measurement, the prior project is producing a number of your own, which is slower and is the only part no competitor can copy.

What it costs, and what cannot be bought

There is no GEO price list.

No published market rate exists for generative engine optimization as a separate service, and the same absence applies to any figure a proposal might show for it. What exists is the market for SEO retainers that GEO work is sold inside: SE Ranking's pricing survey, published on 13 December 2024 across 260 agencies, found 64 % charging under 1,000 USD a month, with its own authors warning the sample may not reach statistical significance. The wider context of those bands is in SEO cost for a small business.

The three spending scenarios look like this. In-house, the cost is editorial hours: rewriting passages so they survive extraction and attaching sources to figures, which needs judgment and no tools. With a vendor, the work bought should be that same editing plus measurement, at retainer rates that the survey above puts under 1,000 USD a month for most of the market. What no budget buys at any price is a guarantee: there is no submission form, no published criterion and no verification path for citations, so a contract promising them is promising a mechanism that does not exist. The questions that separate a vendor selling work from one selling the acronym are in how to choose an SEO agency.

What a defensible scope looks like on paper, deliverable by deliverable and with the measurement method written in before any invoice goes out, is in what an SEO proposal should contain.

Measurement is the part most proposals leave out because it is manual. Ten questions a customer would ask, put to the systems repeatedly over several weeks, with a record of who gets cited: slow, imprecise and the only method that corresponds to what happens. A single check proves nothing, because the same question asked twice can return different citations. And when the motivation is traffic that already fell, the diagnosis comes before the tactic, and it is in why organic traffic drops.

The base layer under everything in this article, what search work actually is, what it can carry and on what clock it runs, stays where a pillar should point first, in what is SEO.

Mistakes that repeat

• Buying GEO as a replacement for SEO. The candidate pool is the search index; a page that is not indexed cannot be cited by anything.
• Quoting the 40 % as if it were a forecast. The gain belongs to specific methods at specific positions, and the top result lost visibility in the same measurement.
• Stuffing keywords because it once worked. The only method that fell below baseline in the study that named the field.
• Optimizing whole pages when the unit is the passage. A page can rank well and contain nothing a model can lift out.
• Judging the work by visits. With 93 % of AI Mode queries ending without an external click, sessions undercount exactly what was won.
• Signing anything that guarantees citations. No submission form, no published criterion, no mechanism to guarantee with.

Data and transparency

The search volume (8,000 monthly US searches), keyword difficulty (72) and cost per click (around 11 USD) for the query come from Ahrefs, consulted on 17 August 2026 for this blog's keyword research. That the first page of results sells the discipline without quoting the study's failed methods is my own reading of those results on the same date.

The study figures come from Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande, arXiv 2311.09735, presented at KDD 2024, opened on 17 August 2026: GEO-bench covers 10,000 queries from nine datasets; the best methods improved on baseline by 41 % on Position-Adjusted Word Count and 28 % on Subjective Impression; Statistics Addition achieved a 30 to 40 % relative improvement on the first metric; the 115.1 % gain for fifth-ranked sites and the 30.3 % average loss for the top result belong to the Cite Sources method; keyword stuffing performed below baseline. The abstract's "up to 40 %" is the paper's own headline figure.

That there are no additional requirements or special optimizations to appear in AI Overviews or AI Mode, that eligibility is being indexed and snippet-eligible, and that no special files or schema.org markup are needed, comes from Google's documentation on AI features in Search, checked on 17 August 2026. The crawler names and purposes are documented publicly by OpenAI, checked the same day.

The AI Overview presence figures are four separate measurements that are not comparable and are never averaged here: Conductor 25.11 % (21.9 million queries, Q1 2026), Similarweb 43 % (Generative AI Landscape 2026), BrightEdge 48 % (commercial verticals, March 2026) and Xponent21 60.32 % (US, April 2026). The click-through pair comes from Seer Interactive's study published 4 November 2025, covering 3,119 informational queries across 42 organizations from June 2024 to September 2025: 1.76 % to 0.61 % with an AI Overview, reported by Seer as a 61 % decline, and 2.72 % to 1.62 % without one, a 41 % decline, so only the difference is attributable to the summary. Zero-click searches rising from 56 % to 69 % between May 2024 and May 2025 comes from Similarweb. The 93 % of AI Mode queries ending without an external click has two independent measurements: Semrush, over 69 million US desktop sessions from 1 May to 5 July 2025, and Seer Interactive, over 25.1 million impressions. Agency pricing comes from SE Ranking's survey published 13 December 2024, 260 agencies, whose authors note the sample may not reach statistical significance. The autocomplete counts (238 local against 39 generative AI) are this blog's own harvest of August 2026 and measure that the questions exist, not how many people ask them. The bilingual click ratio is client work, rounded, and carries the reservation stated where it appears. The operational judgment behind this article comes from a portfolio recording more than 300 million impressions a year in Search Console. Verified as of August 2026.

Primary sources, opened on 17 August 2026: the GEO paper (arXiv 2311.09735); Google's documentation on AI features in Search; the Seer Interactive study; the SE Ranking survey.

What this changes

The comfortable reading of the Princeton tables is that GEO is optional. The uncomfortable one is in the losses column.

When a fifth-ranked site gains 115.1 % visibility with the Cite Sources method, that visibility comes from somewhere, and the same measurement shows where: the top result gave up 30.3 %. In a list, being first was a position to defend with the same work that won it. In a generated answer, being first is a target that sites below you can take share from with edits you do not control. The question a consolidated site owner should be asking is not whether to do GEO. It is whether the sites ranked under them are doing it, because in this game the incumbent's reward for winning the old contest is having the most to lose in the new one.

Frequently asked questions