From Google to the AI answer: How customer search is changing

1. The search becomes a conversation

A classic search query is often short: “Heat pump Aarau”. A question to an AI assistant can already describe the situation: “We live in an older single-family home in Aarau. We would like to plan heating and solar installations together and only have one contact person. Which providers are suitable and what should we compare between the offers?"

The second form contains region, building type, scope of services and selection criteria. The person can then ask about cost factors, funding options or maintenance. The research develops through conversation. A company whose website only promises “innovative energy solutions” provides little concrete information.

The example describes a possible sequence, not a measured frequency. Not every person researches this way, and not every AI answer uses current web sources. Some people continue to search directly on Google, in maps, on a marketplace or through recommendations from their surroundings. Others combine several paths.

Why “AI against Google” falls short

Google is both a search engine and a provider of generative search functions. AI overviews add summaries to results; AI Mode enables a more conversational search path. A switch from a Google results list to an AI answer does not have to be a switch away from Google.

Google reported higher search usage than ever before in the fourth quarter of 2025. This is a self-disclosure from the provider, not independent evidence for each industry or region. However, it contradicts the general narrative that increasing AI use necessarily means fewer Google searches. [1]

A different question is more useful for your strategy: What tasks do customers already complete within a search or AI interface, and for which steps do they need your own information, your website, or a conversation with you?

2. Four observations that you should classify correctly

AI is used for information and practical support

A study of 1.5 million ChatGPT conversations published in 2025 classified 49% of messages as seeking information and advice, 40% as completing tasks, and 11% as expressing or reflecting. It covers consumer plans. These are shares of messages, not of the population or market shares compared with Google. [2]

More detailed questions are observable

In May 2026, Google described search queries for AI Mode in the USA on average three times as long as in classic search. This is a product-specific observation and not a general factor for all AI services. This direction is interesting for companies: requirements and context can already be part of the first question. [3]

A response can replace a website visit

Pew studied the browsing of 900 adults in the US. Search activity dates back to March 2025; Results were collected in April. For visits with AI summary, a traditional result was clicked in 8% of visits, compared to 15% without summary. A link within the summary was clicked on 1% of the corresponding visits. This is an observed association, not a controlled demonstration of a causal effect. [4]

AI usage is not the same as AI search

For 2025, Eurostat reports that 32.7% of 16-74 year olds in the EU had used generative AI in the last three months. This also includes writing, studying and other tasks. This value cannot be used to calculate the proportion of Google searches that have migrated. [5]

What we derive from this: AI is a relevant additional touchpoint. However, the sources do not justify a general end to the search engine or a specific sales forecast for your company. We explain regional numbers and their different definitions separately below AI use in Switzerland, Europe and worldwide.

3. Three search paths – different opportunities

The boundaries overlap. A response may include links, images, maps, product information and advertising. Someone may move from a chat to Google, search for a company name and then visit its website directly days later. Your analytics may only show a direct visit or a branded search.

The new selection problem

Customers could see many companies in a long hit list. A summarized answer can only give a few options. Anyone missing from this preselection may not get a chance at the subsequent comparison. Conversely, a place in the answer is not automatically an advantage: an incorrect service description or an unsuitable price can create false expectations.

Do not just check whether the business name appears. Check its role, whether the description is accurate and whether the person can actually use your service. Recommendations in a country where a local installer does not work offer little value.

4. What happens before the first website visit

Your website remains the place for commitment

An assistant can suggest a selection. The provider must reliably explain availability, individual price, scope of contract, qualifications and responsibility. Your website should therefore not only attract attention, but also reduce uncertainty.

Make the next step easy: What information do you need? How is an initial assessment made? Who reports? Is there a non-binding clarification? Do not claim response times or free services that you cannot reliably meet. For a person who has already done extensive research, this information will be more valuable than another general advertising text.

5. What this means for different business models

Local service providers

For trades, care services, consulting or repairs, actual service areas, specific services, trust and accessibility matter. Also explain what you do not offer. A heating renovation business should not appear to offer a 24/7 emergency service if it does not.

Our recommendation: Prioritize your most important service pages and local business details. Complete verifiable projects and answer common screening questions. When asking test questions, check whether the companies mentioned actually work in the same region. A comparison with distant or technically unsuitable companies provides little decision-making assistance.

Online shops

In product research, requirements can become more precise: suitable for an existing device, for a certain room size or with certain delivery conditions. Clean product data is then more important than interchangeable adjectives. Describe variants, compatibility, dimensions and restrictions clearly.

A recommendation for a product that is not available causes frustration. Keep information consistent on product pages and in the feeds you use. Check integration and measurement for each channel separately. A product feed for one provider does not guarantee inclusion in all AI systems.

B2B and offers that require explanation

For B2B and complex offerings, a company name is rarely enough to make a decision. Prospective customers compare responsibilities, integration, privacy, operations, references and procurement requirements. Helpful content answers specific due-diligence questions and explains project requirements.

For example, create an honest decision-making aid: For which situation is a standard package suitable, when is an individual project needed, and which internal resources does the customer have to provide? Such content can also improve sales conversations, even if it doesn't yet result in measurable AI traffic.

Media and information-oriented offers

Anyone who has previously lived from general explanations and many page views must pay particular attention to whether answers in the search interface are sufficient. Your own research, exclusive data, helpful tools or a strong relationship with the readership can create an independent reason for visiting. Which strategy is effective depends on the business model; A blanket switch to paid offerings is not a universal solution.

6. Content must make the next decision easier

Derive your topic plan from real questions. What do people need to understand before making a suitable request? What information is missing in offer comparisons? Where do misunderstandings arise? The answers to these give your content a practical task.

Google continues to cite core SEO work and independent, helpful content as the foundation of its generative search capabilities. Special AI files or a prescribed paragraph length are not a requirement there. These statements concern Google; other systems have their own rules. [6]

You therefore do not need a new page for every possible prompt formulation. A good page can answer several closely related questions. Avoid artificial location variations and reformulated copies. Invest in what your company can explain and prove from its own experience.

You can find the detailed implementation guide with technical testing, examples, measurement rules and prioritization at How can I improve my AI visibility?

7. What you do with SEO, advertising and your budget

Continue SEO, expand goals

Do not stop effective SEO work just because a new channel attracts attention. If Google already brings you suitable customers, that is an existing asset. Also check whether your pages are accessible and understandable for relevant AI answers. Many fundamentals overlap, but success is measured differently.

Assess paid reach separately

An ad and an organic recommendation are different things. Check current advertising options, approved markets, labeling and billing directly with the respective provider. There is no universal budget allocation between search ads and AI advertising.

For a limited test, start with a clear question: Are we reaching suitable prospects at an affordable cost? Define landing page, offer, measurement and abandonment criteria before launch. A low click price is of little use if almost no one makes a suitable request. Conversely, a more expensive contact can be economical if the service achieves a high contribution margin.

Your own reach makes you more independent

Newsletters with voluntary registration, customer relationships, partner recommendations and returning readers give you additional routes to the market. You reduce the dependency on individual result lists. The same applies here: a purchased address list or an automated flood of messages does not replace a trusting relationship.

Our prioritization for smaller teams: First correct important pages and contact channels, then distribute them in a targeted manner, then test new channels. If you start everywhere at the same time, you hardly know which change helped. Keep the next steps small enough to actually evaluate.

8. Measure three levels instead of one large success number

The levels influence each other, but are not interchangeable. A recommendation can be read without clicking. A click can end without a request. A request may be technically inappropriate. And an order can be based on multiple contacts.

A practical monthly report

1. Search performance: Which relevant questions generate impressions and clicks? Separate brand searches from open problem questions.
2. AI sample: How is your company represented when asked customer questions? Retain original answers and sources.
3. Visitor behavior: Which entry pages lead to a meaningful follow-up action? Take consent and measurement failures into account.
4. Request quality: Which enquiries fit the region, service and budget? Record your assessment as consistently as possible.
5. Changelog: What was published, advertised or technically adapted and when?

If possible, compare the same questions, markets and conditions. A new model or a differently worded question can change a result even though your page remains the same. Individual answers are not market share.

Referrals are only the visible part

Statcounter measures redirects to websites in its AI chatbot statistics. This explains which chatbot within this measurement network delivers clicks, but not how many people use AI without a subsequent click. A high ChatGPT share there cannot be described as a share of all Internet users. [7]

Even your own web analysis cannot completely reconstruct the history of a visit. A voluntary question in the contact process after discovering your company can be supplementary. Don't expect a perfect match and don't collect additional personal data that you don't need.

9. This is how your team decides what to do now

Five questions for the next marketing meeting

• Which specific customer decision would we like to be taken into account?
• What information does a person need who doesn't yet know our name?
• Which of this information is now publicly available in a correct and verifiable manner?
• What measure can be implemented with our resources within a month?
• How do we recognize progress without declaring a random blip a success?

A useful outcome of this meeting is a specific task with an owner and acceptance criteria. “We need to do more with AI” is not a plan. “Explain our combined service clearly, answer two common suitability questions and add an approved project example” can be implemented and checked.

The time until a visible search result cannot be planned exactly. Google specifies a timeframe of days to weeks for re-crawling and does not guarantee inclusion. A submitted URL and an indexed page are also not a good ranking. [8] Therefore, link your publications to meaningful distribution instead of just waiting for organic discovery.

10. What remains uncertain

The sources use different units: people, messages, search page visits or redirects. Their values do not form a common market share table. An increase in one category does not automatically prove a decrease in another.

Platform data is current but interest-driven. Independent studies have their own limitations in terms of sample, region and survey time. National usage can vary greatly depending on your target audience. For a Swiss specialist company, a small, purchase-oriented target group can be more important than a worldwide usage record.

AI can name the wrong businesses, confuse services or use outdated information. Even source links do not guarantee that the linked page supports every claim. Check important statements directly and correct your own and third-party sources.

Dependence on a few platforms remains risky: surfaces, rules and measurement options change. Therefore, build your own content and customer relationships. An agency shouldn't sell you a guaranteed AI recommendation, but rather offer verifiable work, traceable measurement and realistic decisions.

11. Frequently asked questions

Do I have to choose between Google and ChatGPT?

No. Prioritize the paths of your target audience. Technical accessibility, clear services and reliable evidence help multiple channels. Nevertheless, check the results separately because the surfaces can select different answers.

Are fewer clicks always a bad thing?

Not automatic. Some simple information questions can be answered without a visit. What matters is whether suitable interested parties continue to discover and contact you. If fewer clicks mean fewer qualified inquiries, you have to react. Without this second level, the assessment remains incomplete.

How do I know if AI is relevant for my industry?

Talk to customers, collect real questions, and conduct a documented small sample. Check your referral sources and the content of requests. An industry assumption is a good starting point for a test, but is not a substitute for its result.

What can AI Brand Check contribute to this?

The audit examines OpenAI responses using web search for selected questions and conditions. It shows mentions, recommendations, competitors and sources. It does not measure all actual ChatGPT sessions, Google rankings, or all AI systems. Use it as part of your diagnosis, not as a complete representation of your market.

12. Sources and further work aids

The statistics above include their measurement basis and limitations. The practical examples, priorities and suggested work reflect our editorial assessment. Provider sources describe their own systems.

1. Alphabet: Explanations for Q4 2025, February 4, 2026

   Primary source for Google's search usage statement. Read as a vendor report, not an independent measurement of an entire market.

2. OpenAI: How people are using ChatGPT, September 15, 2025

   Study with link to the underlying NBER work. Consider consumer usage, message categories and methodology.

3. Google: AI Mode in the USA, May 19, 2026

   Product-specific observations on the length and form of search queries. No ranking of all search services.

4. Pew Research Center: Clicking behavior in AI summaries, July 22, 2025

   Independent US observation. Includes survey method and sample; no causal forecast for your website.

5. Eurostat: Use of generative AI in the EU, December 16, 2025

   Population-based use with defined age group and time period. Do not mix with search queries or click shares.

6. Google Search Central: Optimization for generative search functions

   Official technical and editorial basis. Useful for checking GEO promises.

7. Statcounter: AI redirects and their distribution, April 2, 2026

   Explains the referral metric. Current regional figures and additional surveys are in our usage article.

8. Google: Request re-crawling

   Explains submission options, time frames and limitations. Relevant after important changes.

Concrete measures for better AI visibility →

Understand key figures and measurement errors →