Generative engine optimization: what the search demand actually says

There is no agreed name for the work of getting a brand cited inside AI answers. Vendors sell it as generative engine optimization. Analysts write it up as answer engine optimization. Practitioners on social feeds call it AI SEO, LLM SEO, or just AI visibility. Every one of those is a real search term with real volume behind it, which means the naming argument is not academic. It is a budget allocation question, and almost nobody has looked at the numbers underneath it.

So we looked. Eight terms, pulled the same day, same country, same provider, measured on volume, ranking difficulty, cost per click and which features the search results actually show. The results are more interesting than the naming debate, because three of the four findings have nothing to do with naming at all.

What we measured, and how

Methodology first, because a study that hides it is a marketing asset wearing a lab coat.

The corpus is eight terms. Each earned a place because at least one commercial vendor, agency service page, or widely-shared practitioner post uses it as the name of the discipline, rather than as a passing description. That rule is what keeps the set to eight instead of eighty. It also means the set is deliberately head-term heavy and excludes the long tail of question queries around it.

The metrics are United States monthly search volume, keyword difficulty on a hundred point scale, cost per click in United States dollars, and the SERP feature set returned for the query. All four come from Ahrefs Keywords Explorer, pulled on September 4, 2026. For three terms we also pulled the monthly volume series from September 2025 through the most recent reported month, which is what the trend findings rest on.

That table is the whole dataset for the cross-sectional findings. Everything below is what falls out of it.

Finding one: every generative engine optimization term returns an AI Overview

Eight terms, eight AI Overviews. Not seven. Not a majority. Every single query in the set that names the practice of earning visibility in AI answers is itself answered by an AI summary before a user reaches an organic result. Seven of the eight also return AI Overview sitelinks, and every one of them returns the People Also Ask block on top of that.

The irony is fun for about four seconds and then it becomes an operating constraint. If you sell this service, your category head terms are zero-click surfaces by default. Ranking first organically on the phrase generative engine optimization puts you underneath a generated answer that has already summarized the concept, sourced from whoever the engine trusts on it. The click you were budgeting for is being intercepted by exactly the mechanism you are selling.

Read the bottom two rows. News appears on five of the eight, and it appears on the terms that are still actively being written about, which tells you the engines are treating this category as a live news topic rather than a settled reference topic. That is a strategic opening. Reference content on a settled topic is a race against Wikipedia and the incumbents. Reference content on a topic the engines still consider newsworthy rewards frequency and recency, which is a very different content plan and a much cheaper one to execute against.

The practical read for anyone building a service page here: assume the head term earns brand presence, not traffic. Its job is to be the page the AI Overview cites, not the page a user clicks. That flips the optimization target from title and meta persuasion to extractable structure, which is the same shift we documented in what actually gets your brand quoted in AI answers.

Finding two: the money is on one name, the traffic is on eight

The two highest-volume terms in the set are within six percent of each other. AI SEO carries 8,500 searches a month, generative engine optimization carries 8,000. Their cost per click is not within six percent of each other. It is $0.80 against $11.00.

A 13.75x gap in what advertisers will pay for two terms of nearly identical size is not noise. It is the single clearest signal in the dataset about where commercial intent actually sits. Advertisers bidding $11.00 a click are not bidding for people writing think pieces. They are bidding for people with a budget and a vendor shortlist.

Four terms carry meaningful commercial weight and four do not. The four that do are, in order, generative engine optimization, answer engine optimization, ChatGPT SEO and LLM SEO. The four cheap ones are the ones a practitioner types while learning, not while buying. AI visibility at $0.80 is a research query. Generative engine optimization at $11.00 is a purchase query. Same category, opposite ends of the funnel, and most content plans we review treat them as interchangeable synonyms.

Now put difficulty next to that. LLM SEO carries a difficulty score of 0 with 1,500 searches a month and a $5.00 cost per click. That combination should not exist in a mature category. A term with real commercial intent, real volume, and no established ranking competition is what an underpriced entry point looks like, and it stays underpriced only until enough people run the same query we just ran. ChatGPT SEO is the same shape at a smaller size: difficulty 13, cost per click $6.00.

The distribution matters as much as the peaks. No single term owns more than about a quarter of the category's total volume. That is the structural argument for treating this as a topic cluster rather than a keyword, with one page per intent rather than one page trying to rank for a synonym set, which is the structure we lay out in how internal linking works for AI search.

Finding three: demand for generative engine optimization is past its peak

The cross-section says the category is healthy. The time series says something more complicated.

Monthly United States volume for generative engine optimization peaked in October 2025 at 10,150. It has not returned to that level since. August 2026 came in at 7,667, about 24 percent below the peak, with the intervening months trending gently down rather than holding a plateau: 8,716 in January, 8,603 in February, 8,581 in March, 7,481 in April, 6,635 in May, 6,733 in June.

Answer engine optimization tells a similar story on a different clock. It peaked in March 2026 at 6,131 and the most recent reported month, September 2026, sits at 3,710. That is a decline of about 39 percent from peak, steeper than the flagship term over a shorter window.

Two readings are available and they are not mutually exclusive. The pessimistic one is that novelty demand is decaying: people searched the new words when the words were new, learned what they meant, and stopped. The optimistic one is that the concept is being absorbed into existing vocabulary, and searches that used to say generative engine optimization now say something operational instead, like how to get cited in ChatGPT. Both readings point at the same tactical conclusion, which is that a content plan anchored to the category noun is anchored to a shrinking asset.

What is emphatically not supported by this data is the claim you will see in decks this quarter that interest in AI search optimization is exploding. In the United States, on these terms, it is not. It is consolidating and, on the two most established names, declining. The work is growing. The searches for the word are not, and confusing those two things is how a category gets oversold and then written off.

“Demand for a word is not demand for the work. The word peaked ten months ago. The work is only starting.”

Finding four: these estimates are not stable enough to plan a quarter on

This is the finding that should change how you use every number above, including ours.

AI visibility, one of the eight terms, reported 941 monthly United States searches in December 2025. Four months later, in April 2026, it reported 11,927. Two months after that, in June 2026, it reported 984. That is a 12.7x range inside a six month window, with no product launch, no algorithm event and no obvious news hook that explains the April spike and the immediate collapse behind it.

The flagship term shows a milder version of the same thing. Generative engine optimization reported 2,485 in July 2026, sandwiched between 6,733 in June and 7,667 in August. A single month at roughly a third of its neighbours on both sides is not a demand signal. It is an artifact of the estimation method, and any planning cycle that happened to run in early August would have read it as a collapse.

None of this makes the tooling wrong. Modeled volume is the best available proxy for demand and there is no substitute. It does mean that the error bars belong in the report, which is the argument we made at length about reporting uncertainty in AI search measurement, and it applies just as hard to keyword data as it does to citation data.

What this changes about how you name and structure the work

Four practical conclusions, in the order we would act on them.

First, stop optimizing the category head terms for clicks. All eight return an AI Overview and most return sitelinks under it, so the realistic outcome on those queries is being the cited source inside the generated answer. Structure those pages to be quotable: a direct definitional answer near the top, a comparison table, named authorship, and clean schema. Presence in the answer is the win, and it is a different asset than a ranking.

Second, split your pages by commercial intent rather than by synonym. One page for the expensive buying terms, written for someone with a budget. Separate reference content for the cheap learning terms. Merging them produces a page that converts nobody and confuses the engine about what the page is for. This is the same intent-splitting discipline that drives our content marketing engagements, and it applies to any category with a contested vocabulary, not just this one.

Third, take the low-difficulty terms while they are cheap. LLM SEO at difficulty 0 with a $5.00 cost per click is the kind of gap that exists for a quarter, not a year. The right response is a real page this month, not a line item in next year's plan.

Fourth, decouple your service naming from your reporting naming. Publish across the vocabulary, because your buyers use all of it. Report internally on one name, because a workstream that changes names between decks cannot accumulate a track record, and a workstream with no track record is the first one cut. If you want the outside view on which of these terms your own site is actually competitive for, that is the first hour of any SEO and GEO audit we run.

The larger point is that this category is still young enough that its vocabulary has not settled, and unsettled vocabulary is an advantage for whoever measures it instead of arguing about it. The naming debate is loud, cheap and mostly performative. The data underneath it is quiet, specific and available to anyone willing to pull it. That asymmetry will not last, so use it while the terms are still contested and the difficulty scores are still low. When the category agrees on a name, the easy positions will already be taken, and the only way in will be the slow one we describe in earning citations without a licensing deal.

Cite this research

This study is free to cite, quote and republish with attribution. If you use the figures, please use them with the pull date attached, because finding four is precisely about what happens when a number outlives its context.

Something Inc. (2026). "Generative engine optimization: what the search
demand actually says." Corpus: 8 category head terms. Metrics: US monthly
search volume, keyword difficulty, cost per click, SERP features.
Source: Ahrefs Keywords Explorer, United States, pulled 2026-09-04.
Time series: monthly US volume, Sep 2025 to Sep 2026, 3 terms.
URL: https://somethinginc.com/blog/generative-engine-optimization-keyword-demand-research

If you want the underlying pull replicated against your own category rather than ours, the method transfers directly: build the corpus from how vendors describe themselves, pull volume, difficulty, cost per click and SERP features on one date, then pull the twelve month series for your three biggest terms. It is an afternoon of work and it will tell you more about where your buyers actually are than a quarter of competitor content audits will. For a worked example of the same discipline applied to a different question, the format shift study on what content formats AI search now favours uses the same rules, and the B2B engagements where we run this first tend to reprioritise their content roadmap within a week of seeing it.

See where you are cited today

A free snapshot audit of your rankings and AI citations before we ever talk.