How AI engines decide who stays and who goes
We’re using AI faster than we’re learning to interpret it. Over 1 billion people interact with AI systems every month; ChatGPT processes 2.5 billion prompts per day and is used internally by 92% of Fortune 500 companies. The habit is already widespread, but the expertise is not: only 10% of the workforce is classified as AI-proficient, while 95% of corporate projects remain stuck in the pilot phase.
The ability to interpret its outputs is just as challenging. You notice your brand is missing and start working on it right there on the page. You find a mention and include it among the positive results. You read about a recommended competitor and treat it as if it had taken up all the space. You compare ChatGPT and AI Overview, get two different outcomes, and search for a single cause where model memory, real-time retrieval, available sources, prompts, and competitive context all intertwine.
A source used, a mention, a recommendation, an absence, a variation between engines: in the same text, they seem like signals from the same family; in practice, they lead to different decisions. Generative content has a confident tone, orderly sentences, and ready-made hierarchies, and its air of finality is exactly what is misleading. Before taking action on content, SEO, your brand, or competitors, you need to understand which signal you’re looking at.
AI is for everyone, but only a few truly understand how it works
Generative tools have become part of our daily work infrastructure. You use them to generate summaries, product comparisons, editorial drafts, competitor analyses, market insights, and brand evaluations—tasks that, until recently, required manual research, comparison, and synthesis. Access is no longer a barrier for anyone: you speak to the machine as you would to a colleague, and you receive a text that seems ready to use.
Mastery is another matter entirely, and some studies highlight just how stark the divide is. The Section report on AI proficiency (operational competence) estimates that only 10% of the workforce uses these tools beyond occasional requests or simple prompts. Seventy-eight percent of executives interviewed by Grant Thornton do not feel prepared to pass an external audit on the use of AI within their companies. According to the AI Total Cost of Ownership study conducted by iKN Italy in collaboration with Casaleggio Associati, approximately 60% of companies report marginal benefits or results below expectations from AI, while 95% of projects do not move beyond the pilot phase.
Familiarity with the interface creates an illusion of control. You type a request, receive a well-organized text, and treat it as something already resolved. The action is trivial, but the machine that makes it possible operates through processes that are much more difficult to interpret.
The gap between those who use AI and those who understand it weighs particularly heavily on generative research. Within a single summary, model memory, sources retrieved from the web, fragments of different pages, associations between entities, reputation signals, and selections tied to the context of the query all coexist. A fluid sentence makes everything seem linear, while the result stems from a chain of choices: what is understood, what is searched for, which sources are included, which blocks are used, and which brands are assigned a role. On the surface, it always seems the same—your brand is either included or left out. Behind the scenes, everything changes, because in those five phases between the prompt and the response, the foundation for all visibility efforts in the coming years is laid.
A source used, a mention, and a recommendation open the door to different actions. AI is not a black box: it performs a sequence of operations that can be reconstructed, explained, and, above all, monitored.
The summary already resembles a diagnosis
The summary you receive has all the characteristics of a verdict. It arrives well-organized, conclusive, written in natural language, with a predefined hierarchy between what’s central and what’s peripheral. You read it as you would a consultant’s opinion, because the format encourages you to trust it. The point is that this text captures a unique combination of the search engine, the phrasing of the query, the timing of the scan, the context of the conversation, and the sources available at that exact moment. Change any one of these elements, and the picture shifts—even though nothing has actually happened to your website or your market.
- Can a single result change a work priority?
An isolated result serves as a clue. It might signal a competitor to keep an eye on, an unexpected brand perception, or an area where the website appears weak. It opens up an investigation; it rarely closes it. The risk arises when the clue is elevated to a full report: an e-commerce site that finds its brand missing from a summary of products for sensitive skin—and rewrites the entire category before knowing whether that absence is repeated elsewhere—has wasted effort on a snapshot. Before shifting a priority, the key question concerns the stability of the signal rather than its content: does this pattern recur in different formulations of the same search query, on other search engines, and with the same competitors? A result only becomes meaningful within a consistent series.
- The brand is missing: is the problem really with the page?
An absence can stem from several possible causes: a weak page, difficulty in retrieval, a model’s memory oriented toward other brands, or external sources mentioning competitors. If the search engine searched the web and found material more relevant to the query, the work involves content, coverage of the topic, retrievability of the page, and consistency between the title, description, and body text. If, on the other hand, the model responded from its own memory and associates that topic with other names, tweaking the page makes little difference; what matters more is the recognizability of the entity, the weight of external sources, and qualified mentions. On the surface, the two absences are identical, but in practice they’re opposites, and addressing the first with the remedy for the second is the most common way to spend money effectively without actually changing anything.
- You’re cited: are you really gaining visibility?
The citation is the most visible level of presence—and often the least sufficient. A page can be used as a source without the user ever encountering the company’s name: the content has done its job, but the brand hasn’t gained recognition. Conversely, a brand can be named without becoming a choice—such as an entry in a list or an example within a category. Counting “mentioned yes, mentioned no” puts an invisible source and an explicit recommendation on the same level—two outcomes that carry different weight. Mentions should be evaluated by function, and that function has its own hierarchy.
Source, mention, recommendation: three types of presence that carry different weight
Within a summary, there is a hierarchy of presence that the count flattens. Your brand could be the source from which the search engine extracted information without the name ever appearing, the name mentioned in passing in a list, the recommended choice for a specific need, or the alternative named and immediately set aside. These are outcomes that carry different weight in the user’s decision, and the distance between them matters more than simply being there.
- What’s the difference between being a source, being named, and being recommended?
The source is the material the search engine uses to construct the text, and it can function without providing visibility, because the user reads the information without encountering your name. A mention brings the brand into the visible text, but it may amount to nothing more than a passing example. A recommendation links the brand to a choice, using phrases like “suitable for” or “recommended for,” and brings it into the decision-making process. The significance of this difference varies depending on who you are: a news site may be satisfied with being an authoritative source on a topic; for an e-commerce site, a passing mention is of little value if the product isn’t considered in the decision; for B2B software, the difference between being mentioned and being recommended for a specific need determines whether you end up on the shortlist or remain just a name that’s read and forgotten.
- The brand is mentioned, but does it truly guide the choice?
Every generative composition builds a small competitive map, showing who occupies the center and who serves only as a point of comparison. A brand described as the primary choice for a specific need gains a different value than one relegated to the additional options, even when they appear in the same summary. This has implications for how the data is interpreted: a brand frequently mentioned as a secondary alternative may have a perceived positioning issue, while one mentioned less often but consistently recommended for the prompts that matter may have a stronger presence than the number of mentions suggests. The role in the decision tells us more than frequency.
- The competitor is recommended: has it truly won that space?
The algorithm processes a narrow query and may favor a competitor in a specific market segment without saying anything about its overall strength. A hotel appears in the results for family-friendly accommodations with young children because it has clear information on services and rooms, while a competitor with a stronger brand remains excluded because it better communicates luxury but poorly addresses that specific need. In software, a provider recommended for small teams looking to get up and running quickly may disappear from prompts regarding advanced integrations or enterprise management. A superficial reading treats that fragment as part of the overall ranking; a proper reading places it within its proper context: which need the competitor addresses, with what argument, how often it appears in similar queries, and which areas remain uncovered. A competitor appearing in a summary should be interpreted, not merely counted.
There’s more to the prompt than just the phrase
The user’s query contains criteria that the engine tries to reconstruct: target audience, constraints, use case, skill level, alternatives, and urgency. Visibility in generative search hinges on these criteria and the subtopics the query opens up, rather than on the main phrasing. Those who treat the prompt as a longer keyword focus on the sentence’s form and miss the part where the model actually decides which sources to retrieve.
- Does a prompt function like a keyword?
A keyword isolates a specific phrasing, while a prompt encompasses a search scenario. “CRM for small businesses” and “which CRM to choose for a team of ten people who want to manage sales and support without complicating their lives” belong to the same category but trigger different processes: in the second case, factors such as team size, simplicity criteria, expected features, and an implicit comparison of tools come into play. The model recognizes different formulations of the same need, because synonyms, word order, errors, and voice requests often converge toward the same interpretation. Chasing the exact prompt leads astray, since two users rarely write it the same way. The need, however, recurs, and it’s best to build content around that need.
- Why does a request open up multiple search paths?
Many questions contain multiple issues at once. “Which platform should I use to sell online courses?” seems like a single question, but it triggers considerations of cost, ease of use, payments, user management, integrations, and comparisons with marketplaces—and the engine compiles the summary by drawing from sources that cover different angles. A brand is included because it effectively addresses a subtopic, even without dominating the main question, or it is excluded because it covers the topic in a generic way and does not meet the triggered criteria. Coverage matters more than a single page. The page on the central topic answers the main question, while the side paths that the query opens up draw from other content, and those who have covered them appear where others remain generic. Adding a qualifier to the question, such as “for a startup” or “on a tight budget,” shifts the results toward material more relevant to that situation.
The same absence takes on different meanings depending on whether it comes from memory or the live web
A summary can originate from the model’s internal memory or from information retrieved from the web in real time, and the same absence takes on opposite meanings in the two cases. When the model draws on memory, what matters are the signals established during training: how the brand was described, which topics it is associated with, and which competitors operate in the same space. When searching the web, what matters are the pages currently available, their retrievability, and the clarity of the information blocks. This difference determines how quickly an intervention produces results.
- Does the response come from the model’s memory or from live retrieval?
Distinguishing between these two cases determines where to focus the initial effort. When relying on memory, an intervention on the website produces slow results because it acts on a perception formed over a long training period, and the work depends on external sources, authoritative mentions, and the consistency with which the brand is described in the market. When relying on real-time retrieval, today’s pages matter, and a targeted effort focused on coverage and retrievability can drive results quickly.
Imagine a brand that has changed its positioning, offering, or product category in recent months. If the model responds based on memory, it may continue to describe the brand using outdated information, obsolete associations, or competitors that are no longer relevant. If, on the other hand, the engine retrieves information in real time, the cause may be much closer at hand: an unclear page, weak information blocks, titles and descriptions that fail to highlight the need, or content that is less discoverable than that of competitors.
Mixing up these levels leads to treating the wrong symptom: spending months on brand reputation when all that was needed was to make a page more discoverable, or constantly tweaking pages when the real issue was the identity entrenched in the model.
- Why does ChatGPT mention you while AI Overview leaves you out?
Because the environment in which the text is generated changes. A chatbot can use internal memory, navigation, sources retrieved on the fly, and the context of the conversation. AI Overview is generated within Google, tied to a query and a set of results already structured across organic pages and candidate sources. Appearing in one environment and disappearing in the other signals a gap between brand recognition, available sources, and retrieval logic. It tells you where the brand is already recognized and where it’s still missing. It’s the practical difference between working on the identity the model has learned and working on presence in texts constructed in real time—the two levels that SEO, GEO, and AEO keep distinct because they operate on different timescales and signals.
Recurrences transform signals into metrics
The variability of individual results shifts the focus of measurement to series, recurrences, and patterns. What appears as noise in an isolated text becomes a signal when observed across a set of prompts, on different search engines, over time. This is the shift that SEO has already made when it stopped reacting to individual rankings and began analyzing trends across large samples. The object of measurement changes; the method remains the same.
- If it changes with every scan, what can you measure?
Individual texts fluctuate, but patterns remain. When observed across a portfolio of prompts tracked over time, the noise evens out, revealing the thematic areas in which you’re mentioned, the role typically attributed to you, the competitors that frequently appear alongside you, the average tone used to describe you, and the direction in which you’re moving. An isolated recommendation is noise; recurring growth across a set of strategic prompts is a real trend; frequent mentions without recommendations signal a role-related issue; a source frequently cited without a visible brand indicates useful content but still weak brand recognition.
- Are rankings and clicks still enough to gauge visibility?
Rankings and clicks remain essential data, but they only tell the part of the story that ends on the website. A growing portion of the decision-making process takes place earlier, within a summary that cites sources, compares brands, and assigns roles. The source-mention-recommendation hierarchy becomes measurable as soon as you separate the three indicators instead of interpreting them together. The Citation Rate indicates how many times the domain is used as a source; the Mention Rate, how many times the brand appears in the visible text; and the Recommendation Rate, how many times it is presented as a recommended choice. A high Citation Rate with a low Mention Rate describes a site that fuels responses while the brand remains invisible—a different scenario from a high Mention Rate with a low Recommendation Rate, where the name circulates but does not translate into preference.
SEOZoom’s SEO for AI section is built on this analysis. In AI Prompt Tracker, the three metrics are displayed across a portfolio of prompts tracked over time, alongside the Positioning Index—which shows whether the AI classifies you as a leader, challenger, alternative, or specialist—and the “Sources Where You Don’t Appear” view, which identifies the key market domains where your site is absent. The distinction between established identity and live presence—the two opposing causes of the same absence—is addressed by two dedicated tools: GEO Audit and AEO Audit.
Read Before Taking Action
The AI summary is already fully formed, and its complete structure tempts you to take a shortcut. The brand is missing, a competitor is suggested, the source appears: every signal seems to demand an immediate reaction. You work on content when coverage is lacking, on SEO when the page struggles to be found, on the brand when the entity is weak or ambiguous, and on competitors when the summary shows spaces occupied by others. But each of these interventions only makes sense after recognizing at which level the signal originates.
The value of a generative response doesn’t lie in the fact that it says something about your brand. It lies in understanding where that signal comes from: source, mention, recommendation, memory, live retrieval, competitor, recurrence. Only then can you decide where to intervene. Before that, you’re still just reading.
