Google's AI Overviews behave largely like an extension of search, so ranking helps. AI assistants such as ChatGPT and Gemini behave like something else, and ranking barely moves them. Most of the argument about which to fund is run by people who have not noticed they are describing two different surfaces.
GEO and SEO answer to different machines, and the data proves it
Two camps shout past each other. One says SEO is dead, pointing to a headline figure showing only 12% of AI citations rank in Google's top 10. The other says GEO is simply SEO with better formatting, and that some schema and an llms.txt file will do it. The data supports neither, because each camp is describing a different surface and neither says so.
Here is the split. Ahrefs studied 15,000 long-tail prompts and found that, on average, only 12% of the URLs cited by AI assistants also appear in Google's top 10 for the same prompt, and roughly 80% do not rank anywhere in Google for that query (Ahrefs, August 2025, updated May 2026). That is the number the "SEO is dead" camp quotes. Worth knowing before quoting it back: the 12% is a five-way average that includes Perplexity at 28.6%, so it flatters the assistants. Strip Perplexity out and the rest sit nearer 8%.
A separate Ahrefs study, across 863,000 keywords and 4 million AI Overview URLs, found 38% of AI Overview citations do rank in Google's top 10 (reported by Search Engine Journal, March 2026). Different datasets, different products. Blend them and the argument collapses into noise, which is roughly what has happened to the category's commentary. The useful reading is not which number is right, but that the surfaces genuinely diverge, and a budget should follow the surface the buyers use.
Ranking still buys you AI Overviews. It does not buy you ChatGPT
Sales of certainty are running well ahead of supply, so precision matters here.
Exhibit 1: The two-surface table
| Google AI Overviews | AI assistants (ChatGPT, Gemini, Copilot) | Perplexity | |
|---|---|---|---|
| Share of citations that rank in Google's top 10 | 38% (Ahrefs), around 17% (BrightEdge) | Roughly 6% to 9% | 28.6% |
| Behaves like | An extension of search | Something else | A hybrid |
| Does ranking help? | Yes, materially | Barely | Yes |
| What to fund | Conventional SEO still pays | Topic and entity work | Both |
Read this exhibit with its caveat attached. The AI Overview figures and the assistant figures come from two different Ahrefs studies, with different datasets and methods. Ahrefs also states plainly that its earlier 76% AI Overview figure and the current 38% are not directly comparable, because its citation parsing improved between the two. Some of that apparent collapse is better detection, not changed behaviour. Anyone presenting 76% falling to 38% as a clean trend line is selling a story the source itself refuses to tell.
Within the assistants, Perplexity is the outlier at 28.6%, which makes sense for a product built to cite. The rest sit between roughly 6% and 9%. Ahrefs publishes two of those figures inside an averaging formula without labelling which engine each belongs to, so this article will not guess which is Gemini and which is Copilot. Both sit at roughly 8%, and that is as far as the evidence goes.
Ranking first on Google is not a citation strategy. It is a citation strategy for one surface, and increasingly the smaller half of the question.
The engine may never have searched your keyword
This is the part that reorganises a budget, and it is documented rather than inferred. Google's Search Central documentation states that both AI Overviews and AI Mode may use a "query fan-out" technique, issuing multiple related searches across subtopics and data sources to build a response (Google Search Central).
Consider what that does to a keyword strategy. Ahrefs offers a clean illustration: a page ranking around sixth for "how to descale a coffee machine", "cleaning a Nespresso machine" and "remove limescale from coffee maker" may well be cited ahead of a page ranking first for only one of them. Consistency across the cluster beats dominance in one slot. Ahrefs goes further, suggesting it is possible assistants may never search the prompt the user typed, treating it instead as a catalyst for other queries. Note the hedge and keep it: a possibility the data is consistent with, not a confirmed mechanism.
The same caution applies to Reciprocal Rank Fusion, the merging method often cited to explain why consistently-appearing pages win. Ahrefs describes citations as merged using methods "like" RRF, sourcing that to a third-party blog rather than to Google, and Search Engine Journal notes Google has not confirmed any specific changes to fan-out behaviour. Fan-out is documented. RRF is industry-inferred. Those are not equally evidenced, and anyone making a budget decision deserves to know which is which.
For how citation selection actually works inside an answer engine, and why consensus rather than rank drives it, see how to get your brand mentioned by ChatGPT. This article is about what to fund. That one is about how the machine decides.
Which means the unit of optimisation is no longer the keyword
Here is the argument the data forces. If the engine decomposes a prompt into questions the brand never wrote a page for, optimising for one phrase is optimising for a question that may never be asked. The keyword does not stop mattering. It stops being the unit.
Exhibit 2: Keyword, topic, entity. The three units of optimisation
| Unit | What it is | Rewarded by | The work |
|---|---|---|---|
| Keyword | One phrase, one page, one rank | Classic search, and still AI Overviews in part | On-page SEO |
| Topic | A cluster of related phrasings the engine fans out into | AI assistants, via query variation. Fan-out is documented for Google's AI features. Fusion methods are industry-inferred | Depth and coverage across the cluster |
| Entity | What the engine believes you are | Every surface, invisibly. Vantage's working model, not a documented mechanism | Consistency across owned, declared, earned and inferred sources |
The third row is the one most brands have not funded, and its status is worth stating. Entity is the working model Vantage uses to explain what these systems reward. It is not a mechanism Google or OpenAI has documented, and this article will not dress it as one. An entity, in this model, is the engine's working belief about what an organisation does, who it serves and what it is credible for, assembled from four kinds of source: owned, declared, earned and inferred. A brand whose entity is confused tends to lose the answer before the keyword contest starts. That is a positioning problem wearing a technical costume, treated in full in how to make sure AI describes your brand correctly.
Retail offers the analogy. A brand can win the shelf-facing war in one aisle and still lose the category, because the buyer decided which shop they trusted before walking down it. Ranking is shelf position. Entity is which shop the customer thinks you are. Ahrefs, having run the numbers, lands in a similar place. Its own closing advice is that optimising for query variations, building topic clusters and "owning the entity" "seems to be the way to go". Note the hedge in the original, and keep it: this is a direction the evidence points in, not a law it proves.
The honest answer is that you fund both, in different proportions, for different reasons
The question "GEO or SEO" is malformed. The answerable question is: which surface does the buyer use, and what does that surface reward?
If the buyers are searching Google and hitting AI Overviews, conventional SEO still pays, and pays materially. Google states this itself: SEO best practices remain relevant for AI features, with no additional requirements to appear in AI Overviews or AI Mode. That is not a vendor's hedge. It is the platform's documentation.
If the buyers are asking an assistant, ranking will not carry the brand. Topic depth and entity consistency will. Most considered B2B brands in this region face both, because a procurement lead will Google a shortlist and a founder will ask ChatGPT for one.
Two cautions follow. Google's documentation is explicit that no new machine-readable files, AI text files or special schema are needed to appear in these features, so any GEO package whose substance is an llms.txt file and a schema block is selling something the platform says is not required. Less comfortably, Google also warns that creating separate content for every variation of how people might search, including fan-out queries, primarily to manipulate rankings or generative responses, violates its scaled content abuse spam policy. Fan-out is not an invitation to mass-produce pages. It is an argument for genuine depth, which is slower and more expensive to buy. And the two disciplines draw on one finite budget, which is why the question deserves better than a hot take from someone with a tool to sell.
What this looks like in Southeast Asia
The multi-market case makes this argument harder, not softer. A regional B2B brand typically optimises English keywords for Singapore while its buyers in Indonesia and Vietnam prompt an assistant in their own language. Where an engine fans a prompt out, it is reasonable to expect the sub-queries to follow the language of the prompt, and a brand holding one strong English page in one market to appear in none of them. That is an inference from documented behaviour rather than a measured finding, and it is offered as such.
The Ahrefs dataset was at least not English-only. Its published example queries include a Spanish-language prompt, which is a small point rather than a large one: it shows the study was not confined to English, not that the pattern has been measured across languages. What has not been established, and what this article will not invent, is a Southeast-Asia-specific citation statistic. No such figure was found at source, so none is offered.
The regional consequence is structural rather than statistical. A brand operating across four markets is not running one topic cluster. It is running four, in four languages, against an engine that will fan out in whichever language the buyer thinks in. Meanwhile the entity layer stays global: the engine holds one working belief about what the organisation is, assembled from sources across every market. Inconsistent positioning between a Singapore site and an Indonesian one does not produce two entities. It produces one confused entity. This is the AI-search-shaped version of a much older regional problem, examined in why Southeast Asian brands struggle to scale.
The measurement problem this creates
All of this implies a measurement change most dashboards have not made. Rank tracking answers a question the assistants are not asking, and describes one surface while staying silent on the rest. The replacement is not a better rank tracker but a different unit: share of the answer rather than position in a list. That is set out in how to measure brand visibility in AI search, and the broader shift it sits inside in how AI is changing branding.