What is generative search, and how does it change marketing in 2026?

murmur’s own results cited in this guide come from campaigns in Brazil using Portuguese questions. This English edition translates the analysis and question wording; measurement figures, dates and limitations remain unchanged.

Generative search retrieves documents and writes a new answer from them, rather than only ranking links. Depending on the product, that answer may appear alongside organic links, rankings or snippets. These are two stages, not one. Their separation breaks the measurement model inherited from search marketing: none of the retrieved authors wrote the text the person reads. The machine wrote it, deciding what to cite, paraphrase or ignore.

I am Mateus Gomes, operator of murmur.marketing, a SWAS GEO operation in Brazil. This page explains the mechanism before discussing its application. Our model pairs proprietary software for measuring and tracking SOV with specialists who define and execute strategy. The internal measurement from August 6, 2026—100 Portuguese questions across four engines, 800 captures—is a historical baseline from before the current guide library, not a measure of today’s strategy. It remains an example of why results need a date, universe and stated limitations.

The five questions connected to this page are listed under “The five questions in this family”.

Summary

Short definition. Generative search is retrieval followed by writing: the engine finds documents and writes an answer from them; that answer may coexist with organic links, rankings and snippets.
Four products covered here. ChatGPT, Claude, Perplexity and Google AI Overview. They disagree on the same question on the same day.
What changes in marketing. Generated answer text has no organic ranking position of its own; rankings, links and snippets may remain around it. A click is optional, and the unit of competition moves from the page to the sentence inside text the machine writes.
What stays the same. Crawlability, server-delivered HTML and search indexes remain prerequisites. Generative engines rely on traditional indexes, so a site that cannot be crawled cannot be retrieved through either route.
Broad questions get explanations, not suppliers. In 216 of 400 captures from my single-run wave, none of the 29 names on the declared detection list appeared. The median number of distinct brands per capture was zero.
The engines’ supplier category is still immature. Across broad-demand questions, with a denominator of 120 (30 questions × 4 engines), the most cited name on that 29-name list appeared 13 times.

How it works: retrieve, then generate

In traditional search, the engine selects and ranks documents. Its output is a list, and every item retains its author’s text. Position is the metric; the click is the conversion.

A generative engine connects two stages:

Retrieval. Given a question, the system searches for documents in its own index, a third-party search index, or both. ChatGPT with search, for example, draws on Bing’s index, making Bing Webmaster Tools a frequently overlooked channel.
Generation. With those documents available, the model writes a new answer. It chooses what enters the sentence, what becomes a linked citation, what gets paraphrased without credit and what gets discarded.

The murmur.marketing taxonomy. I call the stages retrieval and generation, and treat them as separate problems because they fail separately. A page can be retrieved but disappear during writing; beautifully written content may never be retrieved at all. This separation is my diagnostic framework, not a claim of industry consensus. In marketing, AEO and GEO are often used interchangeably.

The sharpest consequence of this architecture concerns delivery: AI crawlers do not execute JavaScript. They read the HTML the server delivers and leave, without a second rendering pass. Vercel published its network-log analysis in The rise of the AI crawler: none of the major AI crawlers in that analysis rendered JavaScript. They downloaded scripts without executing them: 11.50% of requests from OpenAI’s crawler and 23.84% from Anthropic’s. Downloading is not execution. The taxonomy article on whether retrieval and generation are the same stage in AI search examines the mechanism and examples in detail.

The four generative search products covered here—and their limits

When I say “the engines,” I mean four concrete products: ChatGPT, Claude, Perplexity and Google AI Overview, the generated block at the top of a search results page. These are the four I capture. They do not behave as one actor.

The example that convinced me most came from my own campaign. One question—“which is the best company for GEO in Brazil?” (translated from Portuguese)—was run five times in each of the four engines, producing 20 captures. Under a deterministic text rule against a declared list of 29 names, Conversion was named in 5 of 5 runs on ChatGPT, 4 of 5 on Perplexity, 0 of 5 on Claude and 0 of 5 on Google AI Overview. Same day, same question, opposite results. Treating “AI” as one respondent is the most common interpretive mistake here.

A scope limit that belongs to my pipeline, not the reader: the engine labeled google is AI Overview. The Gemini app is not measured; this pipeline has no collector for it. Nothing on this page should be read as a finding about Gemini.

What changes in marketing, item by item

Before	After	The measurement that stops being sufficient
Position in a list of ten links	Presence or absence within a paragraph	Average ranking, “position 3”
A click as the required outcome	An optional click: the answer may satisfy the need without a visit	Organic sessions as a proxy for demand served
The page as the unit of competition	The sentence within generated text	Analysis of an isolated URL
One result per query	Results that vary across repetitions of the same query	A single screenshot as evidence
Brand identity inferred from its domain	A brand recognized as an entity distinct from namesakes	Brand metrics without an ambiguity test

Two rows deserve a fuller explanation because ignoring them can be expensive.

The click is no longer the necessary outcome. An answer that accurately describes what a business does can resolve the questioner’s commercial need without generating a visit. No analytics dashboard will record that, because there was no session. Measuring generative search only through traffic reports measures the missing visit instead of the answer itself.

Entity recognition becomes a prerequisite. Before any content technique, the model must know that the company exists as a distinct entity. When it does not, it fills the gap with the nearest namesake. That is measurable, and literal in my case: across 10 questions already containing my name × 4 engines, a denominator of 40, the engines cited a domain containing “murmur” in 33 captures. Only 5 were mine. In the other 360 captures from that same wave, no “murmur” domain appeared: neither mine nor a namesake’s.

What stays the same, and why that is good news

Three things survived the change of interface.

Crawlability remains a prerequisite. The three discovery routes used here are internal links from known pages, a sitemap.xml declared in robots.txt, and active submission to an index. Content with none of these is orphaned in both environments.

Traditional indexes remain underneath. Generative engines rely on search indexes. Technical SEO work has become infrastructure for another layer, rather than becoming worthless. Blocking traditional search engines while admitting only AI crawlers can also obstruct the indexes that retrieval depends on.

Specific content retains its advantage over generic content. My formulation is facets, not clones: address a leadership question through several pages, each owning a different specific question, instead of one excellent page trying to cover everything.

What the measurement shows about broad questions

This is the finding most likely to change a reader’s strategy, and it is counterintuitive.

On August 6, 2026, I ran 100 Portuguese questions across the four engines. In the single-run wave, a denominator of 400, 216 captures contained none of the 29 names on my declared detection list. The median number of distinct brands per capture was zero. Within broad-demand questions—such as what GEO is, whether it is worth investing in, or whether it is SEO under another name—the denominator was 120 (30 questions × 4 engines). The list’s most cited name was GeoStack, with 13 occurrences out of 120.

The reading is straightforward: broad questions prompt conceptual explanations, not supplier recommendations. A short list of giants does not dominate the field; most broad answers contain no tracked brand at all.

Where names appear, nobody has a secure position. Across 100 question–engine pairs, each run five times, using the same deterministic text rule and declared 29-name list, the counts were Conversion 19 · GeoStack 15 · Brasil GEO 14 · Criamente 9 · Profound 8. In 51 of those 100 pairs, no listed brand reached a majority of repetitions. Both qualifications belong beside those figures: the top three differ by less than 1.3 binomial standard errors, so this is not a stable ranking; and the brand winning 19 pairs is without a majority in the other 81. These results establish no market leader. The pre-library murmur.marketing baseline recorded zero majority pairs.

It is useful to acknowledge what the brands ahead have built. More than fifteen years of authority outside the company’s own site accompanies 19 of 100 pairs: Conversion has original research hosted by Poder360 and E-Commerce Brasil, two Band articles placing it first among ten, and its founder in five third-party podcasts. Around four months of concentrated content accompanies 15 pairs: GeoStack had 279 URLs across four sitemaps, with lastmod dates from 2026-04-12 to 2026-08-06. The study’s qualifications still apply: the Band listicle resembled syndicated PR placement, and whether it was paid could not be confirmed. Third-party coverage of GeoStack was recorded as unconfirmed, not absent; one search pass cannot exhaust the record.

What remains a bet

The mechanism exists and is described above. What is not established is the systematic effectiveness of the techniques sold around it.

The term GEO comes from Aggarwal et al., presented at KDD 2024 and published as arXiv 2311.09735. The paper built GEO-bench, with 10,000 queries, and measured writing interventions. A critical review of the literature also exists: Olivier Martinez’s arXiv 2607.14035, submitted on July 15, 2026, reviews 45 studies published between November 2023 and July 2026. It concludes that none of the reviewed techniques demonstrates stable, longitudinal causal effects across platforms, and records settings where GEO-style rewriting reduces page retrieval.

The critical review limits the strength of the evidence; it does not turn every possible effect into an impossibility. An operational plan therefore combines diagnosis, documented interventions and SOV tracking in the same question universe. SOV-growth targets and guarantees, where offered, depend on the contract model, scope and conditions, not a universal promise derived from a benchmark.

The five questions in this family

This page establishes the setting. Each question below has its own page with a focused answer and a link back here.

Question	What it resolves beyond this page
How ChatGPT search works in 2026	One engine’s mechanics, in the vocabulary it uses to describe itself
Why AI recommends some brands and not others in 2026	How selection works, and why “AI” is not one actor
Is GEO worth investing in during 2026?	The budget decision and its criteria
As a head of growth, how much SEO budget should I move to GEO in 2026?	Allocation, including the data I explicitly lack
Is GEO worthwhile for clinics and medical practices in 2026?	The sector question, in the only case where I have my own business to show

Frequently asked questions

Is generative search the same as an AI search engine?

In practice, both expressions describe the same family of products. The useful definition concerns the mechanism: search is generative when the system retrieves documents and writes a new answer from them instead of ranking links. A search engine that only uses machine learning to rank results remains traditional search. Marketing measurement changes because the final product is machine-written text without a conventional ranking position, not merely because AI is present in the system.

Does generative search end SEO?

No. The architecture explains why: generative engines rely on search indexes for retrieval. A site that cannot be crawled cannot be retrieved through either route. The measurement at the end changes: position and clicks do not describe the whole result: organic links may still rank, but generated text has no organic position of its own and may satisfy the person’s need without a visit.

What does a traditional marketing report miss when search includes generated answers?

The generated text’s organic ranking position and the requirement for a click. There is no “position 3” inside a paragraph, and an answer can resolve a question without a visit. Traffic reports do not show which brand the answer named or in what role—recommended, cited as a source, merely mentioned or absent. Search Console now offers performance data for Google’s AI features, but not answer text or an SOV measure comparable across engines. That requires capturing answers with a denominator and execution date.

It depends on how broad the question is, and that can be measured. In my 400-capture wave, 216 captures contained none of the 29 names on the declared detection list; the median number of distinct brands per capture was zero. Across broad-demand questions, a denominator of 120, the most cited name on that list appeared 13 times. Broad buyer questions tend to receive conceptual explanations; supplier naming concentrates in questions about hiring.

If the machine writes the answer, what can I influence?

You can influence retrieval conditions and the content available for reuse, without directly controlling the generated sentence. For retrieval, you can exist as a readable document: server-delivered HTML, accessible through internal links, a sitemap.xml declared in robots.txt, or active submission to an index. Entity recognition comes before that. The historical August 6, 2026 baseline provides an example: 10 questions already containing my name × 4 engines gave a denominator of 40. In 33 captures, AI cited a domain containing “murmur,” but only 5 were mine. In the other 360 captures from that wave, no “murmur” domain appeared. That ambiguity guides clearer entity identification and verification in a follow-up measurement.

Who wrote this, and disclosure of interest

Mateus Gomes, operator of murmur.marketing, a GEO engine in Brazil. I work exclusively on GEO; I am neither an SEO agency nor a full-service agency. I have a direct commercial interest in this interface being treated as a channel requiring dedicated work. That interest is disclosed before the numbers.

The historical August 6, 2026 baseline, before the current guide library, recorded zero in 800 captures under the classification used. Text matching, a different metric, recorded 88 hits out of 800, all in the ten questions already containing the name; elsewhere, 0 of 712. Those hits cannot be reported as “11% recommendations”: textual presence and recommendation are different states. These figures describe that Portuguese-language Brazilian campaign, not today’s SWAS strategy.

Conclusion

Generative search is retrieval followed by writing: the engine finds documents and writes a new answer from them. That breaks three assumptions of search marketing—ranking position as a description of the whole interface, clicks as the necessary outcome and the page as the unit of the generated answer—while retaining crawlability, underlying traditional indexes and the advantage of specific over generic content. It adds a prerequisite: recognition of the brand as an entity distinct from namesakes. My August 6, 2026 measurement also shows a commercial interface still taking shape. None of the 29 tracked names appeared in 216 of 400 captures; where names appeared, none won more than 19 of 100 measured pairs. To discuss measurement for your business, contact Mateus Gomes on LinkedIn.