Edition 14
AI assistants are starting to decide which companies a customer ever considers. For the business being chosen, that is a different problem from search ranking.
For most of the internet era, finding something meant moving through the web. You searched, opened a few pages, compared options, read a review, went back, opened another tab and eventually reached a product, service, publication, hotel, restaurant, university, software company or person. That browsing process created the commercial structure of the web. Publishers received traffic. Brands had an opportunity to explain themselves. Specialist websites found audiences. Smaller companies could sometimes compete with much larger ones simply by being useful enough to appear in the right search. Search engines became extraordinarily powerful because they stood near the beginning of that journey, but they usually sent the user somewhere else.
AI is beginning to shorten the journey considerably. Ask an assistant which laptop suits your needs and it can compare models for you. Ask where to stay and it can narrow the field. Ask which software fits a particular requirement and it can produce a shortlist. Ask about a difficult subject and it can draw from several sources before presenting a single explanation. The user may have little reason to leave the conversation.
Shopping makes this easiest to see. ChatGPT can already help people discover and compare products, and some commerce experiences are moving towards completing the transaction inside the same environment. Google is bringing conversational search, product discovery, merchant information and commerce closer together as well. There is obvious convenience in this. The customer can move from a vague requirement to a small number of plausible choices without learning the vocabulary of the industry or opening a dozen browser tabs.
For the companies hoping to be chosen, something important has shifted. Discovery and comparison increasingly happen before the customer reaches the company. For years, businesses worried about where they appeared in search results. The emerging problem is different, because they may need to think about the interface through which the customer makes the decision in the first place.
Who owns the interface?
The Prompt
Imagine you need to buy a washing machine. A typical online search today might involve Google, a couple of manufacturer websites, one or two retailers, a YouTube review and perhaps a discussion forum. You might check prices, warranties, energy consumption and complaints before making up your mind. The process is untidy, but you can see where the information comes from. The manufacturer explains its product. The retailer promotes whatever offer it has. A reviewer may disagree with both. Someone on a forum might describe a fault that neither mentioned. You are left to reconcile the contradictions.
Now hand the same task to an AI assistant. You give it your budget, family size, space constraints and preferences. It asks a few questions and comes back with three machines, explains why each might suit you, compares the important differences and perhaps tells you which one it would choose. For many people, this is a much nicer way to shop.
Consider what happens before those three machines appear. Hundreds of products may have been technically available, and the customer sees only a handful. Some mechanism has decided which products deserve consideration, which characteristics matter and how the trade-offs should be explained.
Ranking is nothing new. Search engines have always ranked information. Supermarkets decide where products sit on shelves, travel sites decide which hotels appear first, and app stores decide which applications receive visibility. What AI adds is interpretation. The user is no longer necessarily asking for a list of washing machines. They may say that they have two children, limited space, hard water, elderly parents at home and no patience for complicated controls, and the system has to translate those circumstances into a recommendation.
That gives the interface far more influence than a conventional search box.
The Mirage
It is easy to see an AI assistant as a more efficient route through the same market, and in many cases that is exactly what the customer wants. A good system can remove irrelevant options, compare technical specifications and explain differences without forcing someone to spend an evening researching a purchase.
The difficulty lies in how the shortlist is created. Suppose five hundred products could reasonably satisfy the request. The assistant cannot discuss all of them, so it has to choose what deserves attention. Price may matter. Reliability may matter more. Warranty could become important. An unfamiliar product with strong specifications might deserve consideration alongside a well-known brand, and sources may disagree about durability or customer service. The eventual recommendation can be influenced by product information, reviews, retrieval systems, model behaviour, ranking logic, user history and the way the application itself has been designed. Nobody needs to sit behind the screen choosing the three products manually, and the selection still has to happen somewhere.
Businesses are already beginning to respond. During the search era, companies became obsessed with ranking. They wanted to know whether they appeared on the first page of Google and what they needed to change to move higher. The equivalent concern for AI is taking shape. Companies want to know whether assistants mention them, whether their products are described accurately, whether they appear for the right type of customer and why a competitor is recommended instead. The objective shifts from being easy for a human to find to being easy for a machine to understand.
None of this is new in spirit. Search engine optimisation, or SEO, grew into an industry because ranking decided who got found, and companies learned to shape what the ranking system saw. Some of that work was honest: clearer pages, faster sites, better information. Some of it was not, from keyword stuffing to link farms, and search engines spent years fighting the second kind. The pattern has already started again. Researchers gave the practice a name in 2023, generative engine optimisation, or GEO, and agencies now sell it. The honest version is writing clear, accurate, well-structured information that a machine can use. The dishonest version is planting content designed to be retrieved and repeated, which is the retrieval manipulation described in Edition 10 with a marketing budget behind it.
Paid placement has arrived as well. OpenAI began testing advertising in ChatGPT in February 2026, with sponsored recommendations labelled as such and shown to users on its free and lower-priced tiers, and it says advertisers cannot influence the answers themselves. Google already runs ads in AI Mode. Labelled advertising beside an answer is an old and understood model. The harder question is what happens when the line between the answer and the advertisement is thinner than the label suggests.
That could change competition in unexpected ways. A beautifully designed website has limited value in the recommendation process if the assistant never sends the customer there. A famous brand may carry less weight if the system is comparing products primarily on fit. A smaller company may gain visibility if its product information is clear, detailed and easy for machines to interpret. Large companies may have their own advantage, because they produce more public information, receive more reviews, distribute richer product data and have greater resources to influence how they are represented across the web.
We do not yet know which of these effects will dominate. Businesses will have to think about machine interpretation alongside human attention.
The Reality Check
An AI interface occupies an unusual position in a transaction, because it can learn a great deal about both sides. It can understand what the customer is trying to solve and inspect what the market offers. If the same environment later handles payment and post-purchase interaction, a substantial part of the customer journey can take place without the buyer dealing directly with the merchant until fulfilment. OpenAI and Google have both been developing commerce systems that allow merchants to provide product information and participate in transactions, and merchants are not necessarily being removed from the process.
The larger issue is where the customer’s relationship begins to settle. Suppose someone routinely uses the same assistant to research electronics, plan holidays, compare software, find restaurants and evaluate purchases. Over time, the familiar party in all of those interactions may become the assistant itself. The individual merchants change from transaction to transaction. The interface remains.
We have seen versions of this before. Amazon changed the way many consumers related to individual online retailers. App stores altered software discovery. Booking platforms became a major intermediary between travellers and hotels, and food-delivery platforms changed how restaurants reached customers. Platforms become valuable because they make fragmented markets easier to navigate.
AI can take that idea further, because the interface can understand a customer’s circumstances rather than merely present a catalogue. Someone can say, I need something for my mother who struggles with small buttons. Or, I travel constantly and need a laptop that can survive being thrown into a bag. Or, I need accounting software for a ten-person Indian company, and nobody here is an accountant. A conventional search engine has to turn those sentences into search terms. An AI assistant can treat the circumstances themselves as the request.
The ability to translate an individual’s situation into a small set of commercial options could become one of the most valuable positions in the digital economy.
Now take the question to its uncomfortable end. What if the recommendation is rigged? Not labelled and not disclosed, simply tilted. Telecom offers the closest precedent. An internet provider carries traffic between a customer and the services the customer wants, and the net neutrality debate was about whether that provider may slow some services, speed up others, or charge for the privilege, without the customer knowing why a service feels slow. India settled its version in 2016, when the telecom regulator, TRAI, prohibited discriminatory pricing of data services, which effectively ended Facebook’s Free Basics offering in the country. The principle was that the carrier should not decide which parts of the internet win.
An AI assistant is a different kind of carrier. It does not slow a competitor’s website. It leaves the competitor out of the shortlist, or describes it a little less favourably, and the customer never sees what was left out. A throttled website at least feels slow. A missing recommendation feels like nothing at all.
Skill matters here more than the industry admits. It took a generation a long time to learn to navigate the web, and longer to learn to search well: which words to use, which results to distrust, when to open a second source. Prompting, or prompt engineering, is the same kind of skill, and it is unevenly spread. A confident user asks for the criteria, asks what was left out and rephrases when the first answer feels too tidy. A user who has never done this accepts the first answer, and that is the person an unscrupulous merchant or provider most wants to reach. Gaming an assistant does not need to fool everyone. It only needs to work on those who never think to question it, and for older users and first-time internet users that group is large.
The Leadership Question
Most companies still approach AI as technology they will deploy inside their own business. Their customers are adopting it too. A company may spend months adding AI features to its own website while prospective customers increasingly use external assistants to decide whether the company belongs on the shortlist at all.
That requires a different kind of attention. Businesses need to know whether machines can understand what they sell, who their products are suitable for and how they differ from competitors. Public product information, technical documentation, pricing pages, reviews and old material scattered across the web may all contribute to how an assistant describes the company. Outdated information becomes more troublesome in this environment. A person visiting a website can usually see that a page is old, whereas an AI system assembling an answer from several sources may fold stale material into a fresh-looking response. A business may then be represented inaccurately without knowing that the conversation ever took place.
Retail is only the obvious example. Hotels may increasingly be discovered through machine intermediaries. Software companies can be compared by systems assembling features, integrations, pricing and reviews automatically. Professional services firms may find themselves included or excluded from shortlists before speaking to a prospect. Universities, hospitals, insurers, banks, restaurants and local businesses face variations of the same problem.
Companies once spent a great deal of effort controlling what they said about themselves. The web weakened that control because customers, reviewers and independent publishers joined the conversation. AI adds another layer, since the customer may receive a synthesised description assembled from all of those sources without seeing the company’s own explanation. That makes the broader information environment around a business much harder to ignore.
Commerce adds a second issue. There are strong reasons for merchants to let an AI interface complete a transaction, because every extra page, login and form creates friction, and friction loses customers. The trade-off is information. If the assistant handled the entire conversation, it may know which alternatives the customer considered, why some were rejected, how sensitive the customer was to price, which feature mattered most and what ultimately changed the decision. The merchant may receive the order without seeing much of that journey, and that information has commercial value.
Companies will eventually have to decide how much of the customer relationship they are comfortable allowing another interface to mediate. In many cases the answer may be quite a lot, because the additional sales justify it. The important thing is to recognise that a distribution decision is also being made.
Where does the ethical line sit? Advertising is legitimate when the customer can tell it is advertising, and a business paying to be shown is not a scandal in search or anywhere else. The line is crossed when the influence is hidden: when payment, partnership or commercial interest shapes an answer presented as neutral, or when an assistant’s own commercial relationships tilt its advice without disclosure. A recommendation is a form of trust, and trust is what gets spent.
Several tests follow from that. Is paid placement labelled and visibly separate from the answer? Does the assistant disclose a commercial relationship with a company it recommends? Can the customer ask why a product was chosen and receive the actual criteria? Can they see what was excluded, or ask for the same comparison without sponsored results? An interface that cannot pass those tests is asking for trust it has not earned. Platforms should design for the least skilled user, not the most, which means disclosure and criteria appear by default instead of waiting to be requested.
Businesses on the receiving end have their own choices to make. Compete honestly by keeping accurate, well-structured information easy to find and current. Treat paid placement as advertising and budget for it as such. Monitor what assistants say about you, because nobody else will tell you. And refuse tactics that depend on deceiving the machine or the customer, partly on principle and partly because platforms have a long record of penalising that behaviour once it becomes visible. Buyers should treat a single confident recommendation the way they would treat one from a salesperson they have just met: useful, and best checked against one more source. Regulation will probably land where telecom did, with disclosure rules and some form of neutrality obligation, though that debate ran for years the last time.
The Closing Signal
The web has gone through several layers of intermediation already. Companies built websites. Search engines helped people find them. Social networks assembled audiences. Marketplaces brought products together. AI assistants can sit above all of those layers. They can read the websites, search the market, compare options, understand the customer’s circumstances and recommend a choice, and commerce systems can increasingly carry that choice into a transaction.
For customers, much of this will simply feel easier. There is no particular virtue in opening twelve browser tabs, and people should not need to understand an industry’s terminology before they can make a sensible purchase. A system that reduces twenty plausible products to three genuinely suitable ones has done something useful.
The consequence for businesses is less straightforward. Opening several sources exposes the customer to disagreement. Visiting the manufacturer gives the manufacturer an opportunity to make its case. Reading reviews separately makes it easier to see where one person’s judgment ends and another begins. A conversational interface can absorb all of those steps and return a clean recommendation. The customer gains simplicity, and the selection process becomes less visible.
AI assistants are already helping people choose, and their ability to do this will improve. As they become better at understanding both the customer and the market, the commercial importance of the interface will grow with them. Search engines acquired enormous influence because appearing prominently affected whether a company was discovered. AI assistants may have influence earlier, in the decision itself, helping determine which companies are worth considering before the customer encounters any of them directly. For a business, that leads to an uncomfortable question: if the customer reaches you only after the choice has already been made, whose customer are they?
The Closing Signal
The assistant used to point to the market. Now it chooses from it.