|
Getting your Trinity Audio player ready...
|
Generative AI engines are profoundly transforming online search. Based on natural languageLLMs (Large Language Models) no longer simply index pages: they analyze queries formulated in everyday language, mobilizing multiple sources and producing synthetic answerspresented as directly usable by the user.
This evolution is changing the way information is produced, prioritized and consumed on the web, with major implications for the visibility of content and brands.
From link search to conversational search
Traditional search engines have historically been based on a mediation logic: the user formulates a query, the engine proposes a list of classified documents, and the final interpretation is left to the reader.
LLM-based systems introduce a methodological break.
The process becomes conversational:
- the user expresses a question in natural language,
- the model interprets the intention,
- it generates a single response from a set of heterogeneous sources.
This logic mechanically reduces exposure to original sources. The passage through the document is often replaced by a synthetic answerThis shifts the role of the engine from facilitator of access to information to producer of statements.
LLMs as decision synthesis systems
Language models do not produce new knowledge. They generate statistically plausible linguistic sequences from existing corpora. However, their ability to reformulate, summarize and aggregate information gives them a growing role in decision-making processes.
In many contexts, they are used to :
- compare options,
- identify key players,
- get a simplified view of a complex subject.
Visibility then no longer depends solely on algorithmic ranking, but on the probability of being integrated into the generated response.
From SEO to GEO: from ranking logic to selection logic
SEO: documentary logic
SEO is based on optimizing documents for indexing and ranking:
- keywords,
- page structure,
- internal and external links.
Users are free to compare sources.
GEO: generative logic
The GEO (Generative Engine Optimization) is part of another dynamic.
Generative engines do not present a set of documents: they present a set of documents. select and synthesize content to produce a unique response.
This uniqueness poses a central challenge: a single answer is not necessarily an exhaustive answer, nor is it always accurate.
Simplified answers and the risk of approximation
The strength of LLMs lies in their ability to simplify. But this simplification is also a limitation.
Answers generated :
- condense sometimes divergent points of view,
- erase uncertainties,
- are based on linguistic fluidity rather than conceptual rigor.
The work on Noam Chomsky have regularly stressed this point:
language models don't understand meaning in the cognitive sense; they manipulate linguistic forms. They can produce statements that are grammatically correct but conceptually false, incomplete or misleading.
In an informational context, this ability to produce "convincing" but sometimes erroneous answers is a key factor. epistemological riskespecially when the user no longer consults the original sources.
How LLMs evaluate and select sources
Observable criteria are based on well-documented dimensions.
Content structure and readability
Templates can be used more easily with structured content:
- clear hierarchy of titles,
- short paragraphs,
- lists and tables,
- structured data (schema.org, JSON-LD).
Reliability signals (E-E-A-T)
The E-E-A-T principles are used to estimate the credibility of content:
- identifiable expertise,
- documented experience,
- thematic authority,
- reliable, up-to-date information.
Thematic authority and editorial consistency
LLMs analyze the continuity and density of content on a given subject.
An approach in topic clusters reinforces the perception of legitimacy.
External reputation
Citations in third-party sources (press, specialized publications, academic content, conferences, transcribed videos) help build this legitimacy.
Measuring presence in generative responses
There is as yet no reference tool equivalent to Google Search Console for generative engines. Current approaches are based on :
- manual testing,
- traffic analysis using AI tools,
- emerging specialized solutions.
Indicators tracked mainly concern citation frequency, the diversity of sources used and the discursive positioning associated with a brand or concept.
Towards LLMs connected to business systems
MCP Servers (Model Context Protocol) connect language models to reliable external data sources via standardized APIs.
The challenge is to supplement the statistical generation of LLMs with structured, up-to-date and verifiable data.
Several players have already taken this step.
LeBonCoin has documented the integration of MCP to connect its AI agents to its internal systems and structure a production-oriented architecture.
On data.gouv.fr, experiments such as MCP Company search allow you to query public databases structured in natural language.
Pappers has also deployed its own MCP, combining LLM analysis capabilities with reliable legal and financial data.
These approaches can be used for operational purposes, such as prospecting, risk analysis, company screening or sector intelligence.
In e-commerce, similar logics can be applied to product catalogs, CRM data or transactional information.
This evolution does not eliminate the limitations of generative models, but it does shift the focus to the quality, structuring and governance of data exposed to AI systems.
What generative research changes
Generative search is more than just a technical evolution. It is transforming our relationship with information:
- pluralist access to sources,
- towards mediation through a single, simplified and linguistically fluid response.
In this context, the question is no longer one of visibility alone, but of reliability of product statements and the ability of users to keep a critical eye on the answers generated.
ChatGPT towards conversation-integrated commerce (UCP)
OpenAI has not officially launched an advertising model comparable to Google Ads.
However, several experiments show that the infrastructure is ready.
Integrations with platforms such as Shopify, Stripe or Etsy already make it possible to envisage a complete transactional path within the conversation:
Question → recommendation → selection → payment → confirmation
This scheme greatly reduces the intermediary role of the website.
The conversational engine becomes a comparator, advisor and transactional entry point.
But the real breakthrough lies not only in the integrated transaction.
It lies in the notion ofUCP - Unified Conversational Purchase.
UCP refers to a fully unified buying journey within a conversational environment:
understanding intent
contextualizing the need
analysis of available options
personalized recommendation
validation
payment
follow-up
All without interface disruption.
In this model, the LLM is no longer just a research tool.
It becomes a commercial orchestration interface.
The likely advertising format will therefore not take the form of a classic banner, but of a recommendation integrated into the response, for example :
contextualized analysis
reasoned comparison
credible shortlist
then sponsored solution explicitly indicated
Monetization is no longer based on the click, but on the selection in a highly intentional conversational flow.
The LLM would thus become a monetizable intelligent comparator - at the heart of a UCP pathway where visibility will depend on credibility, data structuring and a brand's ability to be chosen by AI.
Google Gemini: monetization moved to response
Google has confirmed that it is working on advertising formats integrated with AI Overviews.
Given Google's historical dependence on advertising revenues, an adaptation of the Ads model to LLM environments is structurally probable.
The current transformation can be summed up as follows:
Before: ads around content (classic SERP)
Now: commercial integration in the generated response
The model could evolve from a CPC logic to more performance-oriented models (CPA or commission).
What's at stake is not just the visibility of the click, but the influence on the decision.
Perplexity: experimenting with sponsored conversational formats
Perplexity has announced formats such as sponsored follow-up questions.
Advertising no longer appears as an insert, but as a conversation starter.
AI doesn't just show an ad.
It suggests a question that leads to a product or service.
This model shifts the focus of advertising from display to contextual influence.
Transformation of the acquisition funnel
The classic model:
Awareness → Consideration → Comparison → Purchase
Tends to contract in :
Question → Answer → Decision
LLMs absorb part of the information and comparison phase.
Informational traffic may decrease, but the remaining traffic is often more qualified, as it is pre-oriented by AI.
The conversational engine acts as a decision-making intermediary upstream of the site.
Towards credibility-based advertising
In an LLM environment :
visibility is no longer enough,
the recommendation becomes central.
The determining criteria become :
brand awareness,
informational consistency,
multi-source citations,
structured data,
signals of confidence,
measurable reputation.
Conversational advertising will optimize clicks less than algorithmic trust.
Likely evolution of advertising interfaces
In the medium term, an "Ads Manager LLM" could rely less on :
keywords,
auction,
research volumes,
and more on :
use cases,
contexts of intent,
user profiles,
purchasing situations.
Advertising would become contextual and intentional rather than strictly keyword-driven.
Conclusion
Advertising isn't going away.
It is evolving towards a model based on integrated recommendations.
In the old web,
the user searched, compared and clicked.
In the conversational web,
the user queries,
AI analyzes, compares and influences.
The logic is no longer based solely on visibility, but on the ability to be selected and recommended by an AI engine.
In this context, GEO is not simply an extension of SEO.
It becomes a strategic preparation for a future advertising system based on credibility, data structuring and algorithmic trust.
At Soledis, we support e-commerce brands in this transition: analyze their visibility in GenAI environments, reinforce their trust signals and structure their data to make them selectable and recommendable by AI engines.
Anticipating this change today means securing tomorrow's performance.
Sources
McKinsey - The future of search and generative AI
Google - Search Quality Rater Guidelines (E-E-A-T)
SEMrush - Analysis of brand visibility in AI responses
Otterly, Profound - AI presence tracking tools
W3C / schema.org - Structured data
Noam Chomsky - Works on language, cognition and the limits of statistical models
- Lengow - ChatGPT Ads and advertising on GenAI engines: what you need to know
- data.gouv.fr - MCP Company search
