If ChatGPT recommends your company, do customers follow?

When discussing AI and business visibility, I start with a concern: we can get very good at counting something before we understand what it means. A name appears in an answer, a report records it, and that number becomes the result presented in a meeting. Yet between being mentioned and becoming someone's choice lie information, comparisons, doubts, trust, and practical conditions. If we skip those steps, we risk calling something a success before we know how to interpret it.

This doesn't lead me to ignore change. It makes me take it seriously. Let's imagine a family asking ChatGPT which hotel to choose for a weekend in Turin. If our hotel is among the proposals, something interesting has happened. If it is described with a service that we do not offer, something risky has also happened. As a consultant I want to understand both things before deciding how much to invest: how that response is formed, what can change in the client's choice and what evidence is needed to distinguish a possibility from a result.

The context: a search can end before a website visit

To understand the problem, we need to start with what someone is trying to do. People choosing a hotel do not necessarily want to visit lots of websites: they want somewhere that fits their family, travel plans, required services, and availability. Visiting websites has often been necessary to compare that information. An assistant that brings it together in an answer tries to do part of the work before the person reaches the company's website.

In the imaginary example of Turin, the traveler could first ask for a comfortable area, then a hotel with family rooms and finally a comparison of alternatives. The assistant accompanies the reasoning with subsequent questions. The hotel website can remain important to check a condition, see the rooms or book. However, preference formation may have already begun elsewhere. This is the change that interests me: a company can be evaluated through information that it does not present directly to the customer at that moment.

AI visibility describes how often and in what way a brand appears in AI answers. GEO, Generative Engine Optimization, is the name given to efforts to improve that presence through content, information, and sources. These labels are useful when they help us identify a problem; they are less useful when they make something we still need to measure sound settled. Appearing, being described accurately, and being preferred are different things.

Then there is a distinction that avoids many wrong conclusions: Google is not the opposite of AI. Google search includes generated answers, such as AI Overviews and AI Mode. In official documentation, Google indicates that visits from its AI functions flow into the Search Console web report. Comparing ChatGPT traffic with that of Google is therefore not the same as comparing all the artificial intelligence with a search that would lack it. We are comparing observable provenances, not two separate worlds.

Why October's debate challenges budget decisions

On October 21, 2026, consultant Marie Haynes published an analysis of GA4 data from a group of sites, challenging the priority given to GEO over measurable Google search results. On October 22 Mike King, founder of iPullRank, responded that AI search should also be considered a brand-building channel. The disagreement is about the choice to invest now, not the existence of change.

Haynes asks us to look at what brings visitors and meaningful actions to the website. King argues that an answer can introduce a company and make it a preferred option without an immediate visit. Our traveler might read the hotel's name, remember it, and later search for it on Google. That final step is visible; the earlier encounter with the recommendation may not be. This makes a brand effect plausible, but does not establish its size or value for every business.

Rand Fishkin at SparkToro also questions investment that is disproportionate to the evidence available. In the press, Simon Hernandez-Arthur at Axios, on October 28, highlights the audience's questions. For me, that is the key: before deciding what presence to invest in or improve, we need to understand which customer decisions we want to be considered in. A general question about our sector and a comparison before purchase can bring up the same name in very different contexts.

I would therefore not read the comparison as a competition between optimists and pessimists. There are two legitimate questions: how much business we see coming today and how much pre-contact representation of the company might matter. The first requires outcome data; the second requires evidence on consideration and trust. If we use a single measure to answer both, part of the phenomenon is left out.

Earlier evidence: how clicks differ when an AI answer appears

To place the discussion I use a previous piece of data, stating the date. The Pew Research Center released on July 22, 2025 an analysis of the online activity of 900 US adults in March 2025: 68,879 distinct Google searches. Clicking on a traditional result was observed in 8% of page visits with AI summary and 15% of visits without. The chart shows this specific measure, not the purchase rate.

With AI summary, fewer clicks on traditional results

Percentage of results page visits followed by a click on a traditional result. Scale from 0 to 100%.

This result helps to understand why the topic is concrete: the presence of an answer can be accompanied by a different relationship with the links. However, it should be read with caution. It's an observational study in the United States, not proof that the same question, shown to the same person with or without AI, necessarily produces that difference. The searches and pages of the two groups can have different characteristics. The method also reconstructed the results in April, while the navigation was in March: the answers can change over time.

The consequence I draw from it is limited. I would no longer take it for granted that being a helpful source means receiving a visit. At the same time, I would not use these 2025 numbers to estimate the decline in bookings for an Italian hotel in 2026. Intention, price, availability and quality of the offer intervene between a click and a booking. The data opens up a measurement question; it doesn't already provide the answer for our business.

Haynes's figures: volume and conversion rates answer different questions

Haynes declares a 12-month observation on 100 sites, with over 145 million sessions recorded by Google Analytics 4, Google's analysis tool. A session groups together a period of user activity on the site. Sessions from ChatGPT corresponded to 0.62% of volume from Google organic, i.e. from unpaid search. The raw data are not public and the start and end dates are not specified: they are results reported by the author, verified on the source on 31 October 2026.

To understand that ratio, let's normalize the Google volume to 10,000 visits: ChatGPT would have 62. It's just the rewritten proportion, not a new sample and not traffic from an invented company. The small bar in the graph is intentionally small because the scale is linear. Enlarging it to make it more visible would distort the very difference we want to understand.

How small is 0.62% compared to Google volume

Normalized ratio: Google organic = 10,000 visits. The values ​​are not raw counts of a single site.

In the subset of 55 sites that also measured conversions, Haynes reports a rate of 13.8% for ChatGPT and 3.02% for Google organic. A conversion is the action chosen as the objective: it can be a sale, but also a request or another recorded activity. The comparison tells how often visits completed that goal on the sites analyzed. It doesn't prove that AI automatically turns people into better customers.

Marie Haynes · October 21, 2026 · 55 sites, in 12-month observation · Verified: October 31, 2026

Traffic source	Conversion rate
ChatGPT	13.8%
Google organic	3.02%

The original source is Haynes analysis. I would keep the two reasonings separate: one channel may bring few very interested visits, while another produces more overall results thanks to volume. I wouldn't then apply these rates to the graph proportion to calculate how many sales would come in: conversions are measured in a subgroup, not the entire sample. And we don't know enough about the type of goal and the composition of the audience to transfer the result to a specific activity.

This is where the consultant's judgment comes in. If people arrive who have already compared alternatives, a high rate can be a useful signal. However, it could reflect the selection of those who decide to leave the conversation, rather than an improvement caused by the assistant. Before spending more, I would ask what type of requests they are, whether they match the desired customer and how much work they require to become an opportunity. The number opens the investigation, not closes it.

How an answer takes shape, and why we cannot control the whole chain

An AI response isn't necessarily reading a single page. When using web search, the system can search for information, select passages, and combine them. For its search engine's AI functions, Google describes query fan-out: A question is explored through linked searches across multiple aspects. The technical name serves to indicate this broadening of research; it does not by itself describe every assistant or every response.

Let's go back to the hotel. The "suitable for a family" request contains requirements that the site could distribute on different pages: type of room, accessibility, distance from transport, conditions for children. If this information is contradictory, old, or difficult to find, the system may encounter incomplete materials. It may also use third-party reviews or pages. We can't solve the problem by adding a keyword to the main page and assuming that everything else is interpreted as we want.

Google's documentation makes clear that SEO foundations remain important and that you don't need a special scheme to appear in its AI features. SEO helps pages be found and understood in search; Consistent structured data can describe the content, but does not guarantee a recommendation. I consider this a useful criterion for evaluating commercial offers: whoever promises a certain appearance must explain on which controllable mechanism that promise is based.

There is also a difference between being chosen as a source and deserving a person's preference. A technically accessible page can offer generic information. An answer could therefore name the company without clarifying a reason for preferring it. The work on positioning remains: what we offer, for whom, supported by what evidence and with what limits. AI can use that information; it cannot make it true for us.

Mention, citation and recommendation: the name may be there but the value may not

When we read a visibility report, let's focus on the content of the response. A mention is the presence of the name. A citation is a link to a source. A recommendation adds a rating: it presents the company as suitable for a certain need. The visit to the site and the commercial action still come later. These steps can be connected, but each can also break down.

In our example, the name of the hotel could appear because an article reports it as an alternative for families. However, it could also appear in an answer on the services to check before booking, or as an option less suitable for those traveling with small children. The name count would grow in all these cases. The meaning for the company would be different. For this reason I would not read the presence without the sentence that contains it, the question that produced it and the source on which it is based.

The link should also be looked at. An assistant could recommend the hotel by quoting a travel guide. In that case, the hotel gets a representation in the response, while any resulting visit could go to the guide. Or the hotel's website may only be cited for a cancellation policy, without the property being recommended. A report that adds links and names into a single score hides these differences.

The brand reputation, i.e. the reputation of the company, comes in here: what matters is what the name brings with it. Incorrect information can create an expectation that the service will not meet. A generic promise can make different offers seem the same. A critical judgment may be legitimate and require real improvement, rather than a new description. Working on reputation therefore means verifying the information and comparing it with what the company can maintain.

Why findings from another market do not automatically apply to Italy

The Semrush AI Visibility Index released October 13, 2026 also compares the UK and the US. The analysis covers January-July 2026, with fixed monthly sets of 2.5 million UK and 25 million US prompts built from search data, plus a variable set for Google AI Overviews. These are not all spontaneous questions from consumers.

Among ChatGPT's 100 most cited domains, the average share of citations attributed to news sites was 29.5% in the UK and 19.6% in the US. The denominator is essential: we are talking about citations between those domains, not the entire use of ChatGPT. The report also distinguishes the named brands from the sites used as sources. These measures help to understand that market, platform and category change the reading of presence.

My operational interpretation is that an Italian company must start from the questions in Italian and from its competitive context. A hotel may encounter local reviews, international guides, booking portals and outdated descriptions in different languages. A consultancy company can be represented through articles, professional profiles and specialist materials. There is no reason to attribute the same path to both just because a foreign observatory publishes an aggregate score.

This also changes the value of marketing work. Knowing how to produce a readable page remains useful; knowing how to recognize which information is missing, which source supports it and which audience is looking for it becomes a decisive part of the analysis. Agencies that only sell increased appearances will need to explain the connection to the client's problem. Companies will need to clarify whether they are buying research, information correction, content or a commercial objective. They are activities that can collaborate, but do not have the same expected result.

Simply put: advice is not yet choice

Let's think about a person who suggests a hotel to us. The advice may let us discover a name, but before booking we could check the location, read other reviews and compare the price. The company doesn't necessarily see all this work. If we arrive at the site by typing the name, the analysis program may only record that arrival. From that data it cannot automatically reconstruct the conversation that made us start the search.

The analogy helps to understand the distance between knowledge and purchase, but it has a limit. An acquaintance may have first-hand experience; an assistant generates a response that may contain incomplete or incorrect information. The fluidity of the explanation does not certify that every detail is correct. For the consumer, opening the source of an important condition remains sensible. For the company, knowing that it is mentioned is not enough: it must also verify what is promised in its name.

I would summarize the problem like this: an AI recommendation can open a door. The customer still has to recognize a suitable offer, find confirmation and be able to purchase it under the conditions he expects. Strategic work consists of understanding where the path stops and why. If the information is wrong, the priority is to correct it. If the offer does not respond to the need, it is necessary to intervene on the offer. If evidence of value is missing, more optimized content cannot supply it.

How I would design a useful assessment around real audience questions

I would start with the questions that the company actually receives: commercial requests, doubts before purchasing, reasons for giving up, objections and conditions to be clarified. I would collect them while respecting people's privacy. This is to avoid a frequent mistake in tests: choosing only convenient questions, for which we expect to appear, and mistaking the result for a photograph of the market.

For the hotel I would distinguish, for example, the search for a destination, the comparison between neighborhoods, the choice of a property and the verification of a condition before booking. They are examples of intentions, not segments that I measured on real customers. For a professional service, I would instead look for questions about the problem to be solved, the skills needed, the alternatives and what evidence to ask the supplier. A definition request and a choice request do not have the same business meaning.

I would then document the observations: exact question, language, platform, date, answer, sources cited, and relevant test conditions, including any customization. I would repeat the questions to see how stable the result is. I wouldn't call the percentage of presence in this set "market share": it is a measure of the selected questions. If selection is weak, an accurate report accurately describes a less useful sample.

The revision would concern different contents. Is the name present? Is the company described correctly? Is it proposed for the relevant need? Are the constraints explicit? Does the source really support that passage? From here I would build the correction priorities. A service attributed in error requires attention before a lackluster sentence. A doubt that blocks the purchase may deserve more work than a very broad question that only leads to curiosity.

It is the criterion that connects this analysis to the digital strategy with ChatGPT: the dossier, objectives and constraints come before production. In the workflow with ChatGPT, Claude and Gemini, the separation between research and critical review helps to control hypotheses. Comparison between tools has value when it improves reasoning, not when it multiplies similar answers.

What outcomes would I measure and what would cause me to increase spending

If the goal is to correct information, I would measure the frequency of errors in the observed responses and check the sources that continue to repeat them. I would document what was corrected by the company and what depends on others. A site change does not guarantee an immediate update in every AI assistant. I would therefore not promise the disappearance of all inaccuracies on a set date without a verifiable basis.

If the objective is to generate contacts, I would link the analysis of visits to the requests received, their quality and the commercial outcome. I would also ask new clients how they found out about us, leaving room for an open-ended response. This information can tell about a passage that tracking doesn't see, but it remains a clue: people can only remember part of the route. I wouldn't turn it into a precision it doesn't have.

If the project is presented as an investment in brand awareness or trust, I would ask for a coherent measure: audience research, recognition of the company and reasons why it is among the alternatives. Counting responses to vendor-chosen prompts is not the same as measuring what the audience saw. These are two different observations. The first can be useful for controlling the representation; the second is necessary when we want to talk about an effect on people.

Before the work I would define a starting situation, the planned activities and the result that would justify the continuation. If the project promises reservations, I would evaluate reservations and the cost to obtain them. If it corrects incorrect descriptions, I would look at those descriptions. If it improves the content, I would check what information need is being satisfied. This clarity also allows us to stop an investment: an increase in mentions, alone, would not convince me to expand it.

The cost that the visibility report risks leaving out

A project's cost is more than the subscription to a tool that counts appearances. It includes collecting questions, reviewing answers, updating materials, coordinating with the people delivering the service, and checking results. If a description is wrong because departments publish conflicting conditions, buying another monitoring tool will not solve the problem. Someone needs clear responsibility for the information.

For a company, I would therefore start from the cost of the necessary intervention. Correcting a page, updating a service sheet, clarifying a commercial promise and solving a real problem are different jobs. I wouldn't put them all under the same label and then judge them with a single score. And I would compare the result to what the same people and budget could have improved elsewhere in the customer journey.

The human-controlled content factory enters into this reasoning because production needs review capabilities. Increasing pages without checking promises, sources and consistency can multiply versions of the same mistake. I would prefer content that answers an important question, updated as conditions change, to a series that repeats the same answer with different titles.

What changes for companies, professionals and people

For companies, the first change is organizational. Marketing, sales and those providing the service must be able to verify the same description of the offer. A useful association between name and need arises from consistent information and credible evidence. Work on visibility makes this dependence more evident, but does not eliminate it. The responsibility for a promise remains with the company, even when an AI response repeats or distorts it.

For consultants, agencies and professionals, the value shifts towards the ability to explain the problem and link activities to outcomes. A client needs to be able to understand why we are analyzing certain questions, what a variation means, and what alternative explanations we are considering. Transparency also includes the time needed to verify. Automation can help collect and sort responses; judging a bad promise or insufficient proof requires context.

For people, the possible benefit is easier comparison of dispersed information. The risk is to stop at a convincing explanation without checking a decisive condition. In the case of the hotel, accessibility, cancellation or services for children can change the choice. In the case of a professional, an inaccurately attributed competence can create improper expectations. The useful practice is to check statements that have consequences, especially when the answer does not show a relevant source.

Three scenarios, with the signs I would look for before believing them

In the cautious scenario, assistants remain a complementary point of contact. Companies check descriptions and correct the most important information, without moving a significant part of the work from already effective channels. I would expect to see occasional recommendations and inconsistent business outcomes. In this framework, the priority would be to build a reliable base and maintain comparison with what brings customers today.

In the scenario I consider most likely, people move between conversations, search engines, reviews, and websites. The same decision crosses several channels, and no single tool tells the whole story. I would look for recurring signals in inquiries, customers' accounts, and the quality of information AI assistants use. Measurement would need to bring several sources together while stating where the connection remains uncertain. This is my hypothesis, not a prediction established by the cited data.

In the disruptive scenario, agents compare and book on behalf of people. In that case, availability, conditions, permissions and correctness of the data could become a direct part of market access. It would not be enough to make a name recognizable: the offer should also be able to be understood and used in the process. I would consider this scenario when real industry-relevant integrations and behaviors emerge, not just product demonstrations. None of these outcomes are guaranteed.

My position: get the representation right before chasing a score

I would now invest in the quality of the information and in a verification proportionate to the problem. I wouldn't wait to know every effect of AI to correct a false description or clarify an important service. At the same time, I would protect investments that are already generating results and ask for stronger evidence before shifting the budget on the promise of being recommended.

Haynes brings the discussion back to observable outcomes. King reminds us that a choice may begin to form before a click. I take both points seriously without turning them into certainty about financial returns. I turn them into practical questions: whom do we want to reach, during which decision, with what information, under whose responsibility, and with what evidence of improvement?

Useful consulting work must reach this level. It must make the mechanism understandable, indicate what we can control and choose a measure that makes sense for the company. Reputation grows when promises are recognizable and sustainable; the commercial result must then be verified. Being mentioned is a sign worth reading. Becoming the customer's choice is the path to understand.

FAQ

What does AI visibility mean?

It describes how often and in what way a company appears in AI answers. Understanding it means examining the question, the information associated with the name, the sources, and the context of the recommendation. Counting appearances measures only part of the phenomenon: it does not tell us, by itself, how many people saw the answer or whether they chose the company.

Are GEO and SEO the same thing?

SEO is about being found and understood by search engines; GEO focuses on presence in generated answers. They overlap when an AI answer uses search, as Google's AI features do. Clear content, verifiable sources, and accessibility remain important, but do not guarantee a recommendation. I would therefore start with the audience's needs and the quality of the information before choosing a label for the service.

Does a mention prove that customers arrived?

No. Assessing the result requires observed outcomes, such as inquiries or purchases, and information about how people discovered the company. Tracking and asking customers can help, with different limitations. An extra mention does not demonstrate an extra sale; an unattributed sale does not automatically demonstrate that influence was absent. The conclusion must stay proportionate to the available evidence.

Sources, method and author

Primary sources and international readings are linked in the passages they support. Data and announcements are distinct from interpretations; the scenarios and operating protocol express my criteria, not proprietary test results. The hotel example is imaginary. The studies concern different periods, markets and samples: they have not been added together to produce an estimate for Italy. Verification of sources: 31 October 2026.

I am Giorgia Palazzo, consultant, entrepreneur and teacher. I connect strategy, data, AI and technology implementation to company decisions. My editorial job is to explain change, examine the evidence, and make its consequences understandable, without promising results that I cannot demonstrate.