Where does your company disappear from the decision?

Product choice and merchant choice are separate

OpenAI describes shopping relevance as depending on user context, supplied metadata and other information. Its merchant rankings additionally consider availability, price, quality and maker or primary-seller status.

One vendor’s documented behaviour; no public universal ranking formula or guarantee of inclusion.

The information path can change the recommendation

Wharton’s controlled studies found that prior sources, their order, injected user context and tool-call presentation could change product choices, with different responses across models.

Watch-selection experiments reported by the lab. Neither customer purchases nor the effectiveness of a marketing programme was tested.

Models do not share one stable shopping preference

ACES reports model-dependent responses to price, ratings, position and labels; position effects also appear in text-only interfaces.

Controlled simulator findings. Sponsored-label effects at fixed exposure do not establish the total effect of advertising.

Source identity can influence choice beyond item fit

The Seoul preprint finds source preferences under matched conditions. Relabeling sources changes selection; supplying missing information can reduce reliance on source preconceptions.

Twelve models across shopping, accommodation and scholarly search. Agents received search-result snippets and URLs, without full-page browsing or purchasing. Not a brand ranking.

Personal context can enter recommendations

Google’s January 2026 announcement describes optional Gmail and Photos context informing AI Mode recommendations.

Vendor account of a limited launch; not evidence that every agent remembers every transaction or that a customer relationship guarantees selection.

You can lose before your offer is compared.

A shopping agent does not inspect the whole market. Its instructions, search queries and connected services shape the offers it can consider. A recommendation may draw on a snippet, a product feed or a review rather than your carefully designed product page. A missing detail is not automatically treated as false. It may prompt more searching, leave uncertainty or invite an assumption.

OpenAI’s shopping documentation names the inputs unusually clearly: user context, first- and third-party metadata, third-party content, and model responses generated before new search results. It also warns that not every available product is shown. This describes one implementation, but the strategic distinction is useful: getting considered and persuading the agent are separate problems.

Similar products can answer different needs.

Price and headline specifications tell only part of the story. They leave out failure risk, compatibility, repair, delivery certainty and the cost of making the wrong choice. A customer delegates some of that judgement to an agent. For a company, the useful question is: in which customer situation can we demonstrate a difference that matters?

What do we mean by authority?

A test can support a product claim. A respected publisher can offer an informed judgement. A familiar name can serve as a shortcut when evidence is thin. Calling all three “trust” blurs a crucial distinction: evidence of quality and an association with quality are different things.

Jonghyun Song, Yohan Jo and colleagues at Seoul National University isolate part of this distinction. Their October 2026 preprint finds that changing a source label can change selection even when the item’s content stays the same. Filling information gaps reduces some source preference. The agents chose from search-result text; they did not browse full websites or complete transactions. The study reveals a mechanism, not a global brand ranking.

How evidence reaches the agent can change the choice.

Ethan Mollick, Stefano Puntoni and their Wharton co-authors tested how prior context affected watch selection. The sources, their order and even whether tool calls were bundled or separate could change choices. The lab report describes about 26,000 controlled tests, not observed consumer purchases. The useful implication is specific: the same product page can lead to a different result depending on the information the agent encounters along the way.

Omar Besbes and Yash Kanoria’s ACES work, with academic and MyCustomAI co-authors, also finds that models respond differently to listing attributes and position. Position effects persist in text-only interfaces. Clean structured data can improve access without removing model bias.

For a company, the question extends beyond “Does our website explain us?” to “What information about us reaches the customer’s decision?” Inspect the actual source, date, excerpt and product variant. Weaknesses in the evidence that reaches the agent call for a different remedy from a genuinely weaker offer.

This gives a smaller specialist a plausible opening: build unusually strong evidence for a narrow need. An established company has a different task: back its familiar name with current, specific proof. Neither strategy guarantees selection. Fabricated reviews, copied endorsements and instructions hidden in pages serve neither strategy.

A customer relationship can become decision context.

In his January 2026 announcement, Google product leader Robby Stein describes optional Gmail and Photos connections for more personal AI Mode recommendations. The launch was a limited US, English-language experiment for eligible subscribers. It illustrates how prior preferences can enter a decision. It does not show that every agent continuously learns from every purchase.

Atlas’s strategic inference is that a brand may matter before the next search begins. If a customer explicitly prefers your fit, service or reliability, that preference can change the task their assistant receives. A good relationship may help define what counts as a suitable answer, as well as providing a direct-marketing channel.

Personal context brings a trade-off. It can make recommendations more useful, but it can also reinforce old choices and reduce discovery. Earn a preference through a distinctive experience; do not build a strategy around inserting claims into someone’s memory.

For a global company, language, country, delivery and local service are part of the decision. They need attention before translation. Test US and EU journeys separately, and include the platforms and languages people actually use in other priority markets. The Seoul study includes Chinese model families as well as Western ones; it does not measure consumers in those countries. No source here supports a worldwide adoption ranking.

Your product can win while another seller gets the order.

An agent might recommend your product and then choose a different retailer. OpenAI explicitly separates product selection from merchant ranking, where factors include availability, price, quality and maker or primary-seller status. It says organic product results are independent of partnerships and separate from ads. A feed or integration is not a purchased guarantee of recommendation.

This raises a commercial question: who captures the value? A manufacturer needs to know whether demand from agents reaches profitable channels. A retailer needs a reason to be the chosen seller once the product has been selected. A service business needs credible evidence that it can deliver, beyond a persuasive description of its category. This last application to services is Atlas’s inference from the mechanism.

Track each outcome separately: inclusion among the available offers, recommendation, seller selection, completed order and retained customer. You may be able to observe only some of them. A cited page shows exposure; it does not establish which revenue came from that exposure. Where a platform hides part of the journey, record the gap instead of turning a visibility score into a sales forecast.

Build a reason to choose you. Make sure it reaches the decision.

Start with a valuable customer decision. Write down why your company should win, what evidence would support that choice and what would make a competitor the better answer. If the reason amounts to “we are trusted and high quality”, the strategic work is unfinished. Let that work guide your choice of AI channels.

Then identify the problem: can the agent access your offer, does it have the evidence, or is the offer itself too weak? More content cannot fix an uncompetitive offer. Better evidence cannot help if it never reaches the decision. Invest in an integration when it connects a valuable customer journey to an advantage you can deliver.

Product choice and merchant choice are separate

OpenAI describes shopping relevance as depending on user context, supplied metadata and other information. Its merchant rankings additionally consider availability, price, quality and maker or primary-seller status.

One vendor’s documented behaviour; no public universal ranking formula or guarantee of inclusion.

The information path can change the recommendation

Wharton’s controlled studies found that prior sources, their order, injected user context and tool-call presentation could change product choices, with different responses across models.

Watch-selection experiments reported by the lab. Neither customer purchases nor the effectiveness of a marketing programme was tested.

Models do not share one stable shopping preference

ACES reports model-dependent responses to price, ratings, position and labels; position effects also appear in text-only interfaces.

Controlled simulator findings. Sponsored-label effects at fixed exposure do not establish the total effect of advertising.

Source identity can influence choice beyond item fit

The Seoul preprint finds source preferences under matched conditions. Relabeling sources changes selection; supplying missing information can reduce reliance on source preconceptions.

Twelve models across shopping, accommodation and scholarly search. Agents received search-result snippets and URLs, without full-page browsing or purchasing. Not a brand ranking.

Personal context can enter recommendations

Google’s January 2026 announcement describes optional Gmail and Photos context informing AI Mode recommendations.

Vendor account of a limited launch; not evidence that every agent remembers every transaction or that a customer relationship guarantees selection.

You can lose before your offer is compared.

A shopping agent does not inspect the whole market. Its instructions, search queries and connected services shape the offers it can consider. A recommendation may draw on a snippet, a product feed or a review rather than your carefully designed product page. A missing detail is not automatically treated as false. It may prompt more searching, leave uncertainty or invite an assumption.

OpenAI’s shopping documentation names the inputs unusually clearly: user context, first- and third-party metadata, third-party content, and model responses generated before new search results. It also warns that not every available product is shown. This describes one implementation, but the strategic distinction is useful: getting considered and persuading the agent are separate problems.

Similar products can answer different needs.

Price and headline specifications tell only part of the story. They leave out failure risk, compatibility, repair, delivery certainty and the cost of making the wrong choice. A customer delegates some of that judgement to an agent. For a company, the useful question is: in which customer situation can we demonstrate a difference that matters?

What do we mean by authority?

A test can support a product claim. A respected publisher can offer an informed judgement. A familiar name can serve as a shortcut when evidence is thin. Calling all three “trust” blurs a crucial distinction: evidence of quality and an association with quality are different things.

Jonghyun Song, Yohan Jo and colleagues at Seoul National University isolate part of this distinction. Their October 2026 preprint finds that changing a source label can change selection even when the item’s content stays the same. Filling information gaps reduces some source preference. The agents chose from search-result text; they did not browse full websites or complete transactions. The study reveals a mechanism, not a global brand ranking.

How evidence reaches the agent can change the choice.

Ethan Mollick, Stefano Puntoni and their Wharton co-authors tested how prior context affected watch selection. The sources, their order and even whether tool calls were bundled or separate could change choices. The lab report describes about 26,000 controlled tests, not observed consumer purchases. The useful implication is specific: the same product page can lead to a different result depending on the information the agent encounters along the way.

Omar Besbes and Yash Kanoria’s ACES work, with academic and MyCustomAI co-authors, also finds that models respond differently to listing attributes and position. Position effects persist in text-only interfaces. Clean structured data can improve access without removing model bias.

For a company, the question extends beyond “Does our website explain us?” to “What information about us reaches the customer’s decision?” Inspect the actual source, date, excerpt and product variant. Weaknesses in the evidence that reaches the agent call for a different remedy from a genuinely weaker offer.

This gives a smaller specialist a plausible opening: build unusually strong evidence for a narrow need. An established company has a different task: back its familiar name with current, specific proof. Neither strategy guarantees selection. Fabricated reviews, copied endorsements and instructions hidden in pages serve neither strategy.

A customer relationship can become decision context.

In his January 2026 announcement, Google product leader Robby Stein describes optional Gmail and Photos connections for more personal AI Mode recommendations. The launch was a limited US, English-language experiment for eligible subscribers. It illustrates how prior preferences can enter a decision. It does not show that every agent continuously learns from every purchase.

Atlas’s strategic inference is that a brand may matter before the next search begins. If a customer explicitly prefers your fit, service or reliability, that preference can change the task their assistant receives. A good relationship may help define what counts as a suitable answer, as well as providing a direct-marketing channel.

Personal context brings a trade-off. It can make recommendations more useful, but it can also reinforce old choices and reduce discovery. Earn a preference through a distinctive experience; do not build a strategy around inserting claims into someone’s memory.

For a global company, language, country, delivery and local service are part of the decision. They need attention before translation. Test US and EU journeys separately, and include the platforms and languages people actually use in other priority markets. The Seoul study includes Chinese model families as well as Western ones; it does not measure consumers in those countries. No source here supports a worldwide adoption ranking.

Your product can win while another seller gets the order.

An agent might recommend your product and then choose a different retailer. OpenAI explicitly separates product selection from merchant ranking, where factors include availability, price, quality and maker or primary-seller status. It says organic product results are independent of partnerships and separate from ads. A feed or integration is not a purchased guarantee of recommendation.

This raises a commercial question: who captures the value? A manufacturer needs to know whether demand from agents reaches profitable channels. A retailer needs a reason to be the chosen seller once the product has been selected. A service business needs credible evidence that it can deliver, beyond a persuasive description of its category. This last application to services is Atlas’s inference from the mechanism.

Track each outcome separately: inclusion among the available offers, recommendation, seller selection, completed order and retained customer. You may be able to observe only some of them. A cited page shows exposure; it does not establish which revenue came from that exposure. Where a platform hides part of the journey, record the gap instead of turning a visibility score into a sales forecast.

Build a reason to choose you. Make sure it reaches the decision.

Start with a valuable customer decision. Write down why your company should win, what evidence would support that choice and what would make a competitor the better answer. If the reason amounts to “we are trusted and high quality”, the strategic work is unfinished. Let that work guide your choice of AI channels.

Then identify the problem: can the agent access your offer, does it have the evidence, or is the offer itself too weak? More content cannot fix an uncompetitive offer. Better evidence cannot help if it never reaches the decision. Invest in an integration when it connects a valuable customer journey to an advantage you can deliver.

Product choice and merchant choice are separate

OpenAI describes shopping relevance as depending on user context, supplied metadata and other information. Its merchant rankings additionally consider availability, price, quality and maker or primary-seller status.

One vendor’s documented behaviour; no public universal ranking formula or guarantee of inclusion.

The information path can change the recommendation

Wharton’s controlled studies found that prior sources, their order, injected user context and tool-call presentation could change product choices, with different responses across models.

Watch-selection experiments reported by the lab. Neither customer purchases nor the effectiveness of a marketing programme was tested.

Models do not share one stable shopping preference

ACES reports model-dependent responses to price, ratings, position and labels; position effects also appear in text-only interfaces.

Controlled simulator findings. Sponsored-label effects at fixed exposure do not establish the total effect of advertising.

Source identity can influence choice beyond item fit

The Seoul preprint finds source preferences under matched conditions. Relabeling sources changes selection; supplying missing information can reduce reliance on source preconceptions.

Twelve models across shopping, accommodation and scholarly search. Agents received search-result snippets and URLs, without full-page browsing or purchasing. Not a brand ranking.

Personal context can enter recommendations

Google’s January 2026 announcement describes optional Gmail and Photos context informing AI Mode recommendations.

Vendor account of a limited launch; not evidence that every agent remembers every transaction or that a customer relationship guarantees selection.

You can lose before your offer is compared.

A shopping agent does not inspect the whole market. Its instructions, search queries and connected services determine what becomes available for consideration. A recommendation may be built from a snippet, a product feed or a review rather than your carefully designed product page. An absent detail is not automatically treated as false; it may trigger more searching, uncertainty—or an assumption.

OpenAI’s shopping documentation makes the inputs unusually explicit: user context, first- and third-party metadata, third-party content, and model responses generated before new search results. It also warns that not every available product is shown. This is one implementation, but it exposes a useful strategic distinction: inclusion and persuasion are separate problems.

“Similar products” does not mean an equal decision.

Price and headline specifications compress an offer. They leave out failure risk, compatibility, repair, delivery certainty and the cost of making the wrong choice. A customer delegates some of that judgement to an agent. The strategically useful question becomes: for which customer situation can we demonstrate a difference that matters?

Authority is several signals with different meanings.

A test can support a product claim. A respected publisher can supply a useful judgement. A familiar name can also act as a shortcut when evidence is thin. Calling all three “trust” conceals the difference between information about quality and an association with quality.

Jonghyun Song, Yohan Jo and colleagues at Seoul National University isolate part of this distinction. Their October 2026 preprint finds that changing a source label can change selection while item content stays fixed. Filling information gaps reduces some source preference. The agents selected from search-result text, not full websites or completed transactions; this is evidence of a mechanism, not a global brand league table.

The path to the evidence can alter its effect.

Ethan Mollick, Stefano Puntoni and their Wharton co-authors tested layers of prior context around watch selection. Sources, their ordering and even bundled versus separate tool calls could change choices. Their lab report describes about 26,000 controlled tests, not observed consumer purchases. The implication is narrower—and more useful—than “agents are irrational”: the same product page can produce a different result inside a different information journey.

Omar Besbes and Yash Kanoria’s ACES work, with academic and MyCustomAI co-authors, also finds model-specific responses to listing attributes and position. Text-only interfaces still show position effects. Clean structured data can improve access without removing model bias.

For a company, this shifts attention from “Does our website explain us?” to “What representation of us enters the customer’s decision?” Inspect the actual source, date, excerpt and product variant. A disadvantage in the evidence that reaches the agent needs a different remedy from a genuinely weaker offer.

A smaller specialist therefore has a plausible opening: become unusually well evidenced for a narrow need. An incumbent has a different task: turn familiarity into current, specific proof. Neither strategy guarantees selection, and neither is served by fabricated reviews, copied endorsements or instructions hidden in pages.

A customer relationship can become decision context.

Google product leader Robby Stein’s January 2026 announcement describes optional Gmail and Photos connections for more personal AI Mode recommendations. The launch was a limited US, English-language experiment for eligible subscribers. It illustrates how prior preferences can enter a decision; it does not prove that every agent continuously learns from every purchase.

Atlas’s strategic inference: a brand may matter before the next search begins. If the customer explicitly prefers your fit, service or reliability, that preference can change the task an assistant receives. A good relationship is thus more than a direct-marketing channel. It may help define what counts as a suitable answer.

There is a tension here. Personal context can make recommendations more useful, but can also reinforce old choices and reduce discovery. Earn a preference through a distinctive experience; do not build a strategy around inserting claims into someone’s memory.

For a global company, treat language, country, delivery and local service as parts of the decision—not a translation layer added afterwards. Test US and EU journeys separately, and include the actual platforms and languages used in other priority markets. The Seoul study includes Chinese model families as well as Western ones; it does not measure consumers in those countries. No source here supports a worldwide adoption ranking.

Winning the recommendation is not winning the relationship.

An agent might recommend your product and then choose a different retailer. OpenAI explicitly separates product selection from merchant ranking, where factors include availability, price, quality and maker or primary-seller status. It says organic product results are independent of partnerships and separate from ads. A feed or integration is not a purchased guarantee of recommendation.

The strategic consequence is about value capture. A manufacturer needs to know whether agent demand reaches profitable channels. A retailer needs a reason to be the chosen seller after the product is decided. A service business needs credible evidence of delivery capability, not just a persuasive category description. Applying this last point to services is Atlas’s inference from the mechanism.

Track the distinct outcomes: included in the available set, recommended, chosen as seller, completed order, retained customer. You may only be able to observe some of them. A cited page is evidence of exposure; it is not revenue attribution. Where a platform obscures the journey, record the gap instead of turning a visibility score into a sales forecast.

Invest in a reason to choose you—and its route to the customer.

Start with a valuable customer decision, not a list of AI channels. Write the specific reason your company should win, the evidence required to support it, and what would make a competitor the better answer. If the reason reduces to “we are trusted and high quality”, the strategic work is unfinished.

Then distinguish an access problem, an evidence problem and an offer problem. Publishing more content cannot fix an uncompetitive offer. Better evidence cannot help a decision if it never arrives. A new integration deserves investment when it connects a valuable customer journey to an advantage you can actually deliver.

Google asks for the search basics, not a special AI file

Google says its established search requirements also apply to AI Overviews and AI Mode. It recommends accessible text, structured data that matches the page and current business information. No special AI text file or schema is required.

Google-specific documentation. Meeting the requirements does not guarantee indexing or serving; it says nothing about every other agent or search provider. The downloadable .md here is an internal process brief.

AI already appears in surveyed shopping research

IBM and NRF report that 41% of surveyed consumers had used AI to research products, 33% to look for reviews and 31% for deals or promotions.

Self-reports from over 18,000 consumers across 23 countries in Q3 2025. Categories overlap. This is not a population-weighted global adoption rate or evidence of autonomous purchasing.

Visibility and delivered value are separate strategic problems

Dhar and coauthors emphasize alignment between a brand’s promise and performance. Kamath expects customers to build interfaces on broker APIs and says brokers that depend on prompting trades will resist.

These are perspectives from different businesses, not a direct debate or evidence that every company should adopt the same distribution model. Atlas draws the strategic distinction.

Measure a current, complete workflow

A published support study found a 15% average gain in issues resolved per hour with AI assistance. METR’s 2026 update explains why selection and parallel work complicate current developer-productivity measurement.

Different populations, tasks, tools and methods. The support estimate is from one firm; it cannot be applied to autonomous agents or used as a forecast of cost savings.

Work design includes both process change and human judgment

Brynjolfsson emphasizes complementary organizational investment; Mollick argues for preserving useful human participation as agents become more capable.

Expert interpretations with different emphases. The advice to measure review, rework and accepted outcomes is Atlas’s application, not a validated transformation recipe.

Connected capabilities change the failure modes

Willison identifies a prompt-injection risk when an agent combines private data, untrusted input and external communication. UCP describes negotiated commerce capabilities and handoffs.

These sources concern different layers: security and interoperability. Neither a prompt nor a protocol alone establishes authorization, reliability or commercial readiness.

A global strategy needs local operating evidence

ITU estimates Internet use at 94% in high-income countries and 23% in low-income countries in 2025. Nilekani emphasizes practical applications, local languages and voice access.

Connectivity estimates are not AI-adoption figures. Nilekani’s views are a perspective, not measured proof of language quality. Local journey testing is an editorial recommendation.

Referrals and purchases can be measured separately

Adobe reports higher purchase conversion among AI-referred U.S. retail visits in July 2026. Its earlier product-page assessment found gaps in machine readability.

Vendor telemetry and a proprietary diagnostic, not causal evidence, global coverage or an autonomous-purchase count. The two figures refer to different measures and reporting periods.

Workload estimates do not settle service design

Klarna’s launch release estimated work equivalent to 700 agents and already offered human help. Its 2025 annual report continued to describe both AI and human support.

Company statements, not an independent evaluation. The article’s recommendation to measure quality, escalation and cost together is Atlas’s interpretation.

Prepare to be chosen and to deliver

An agent uses a model and tools to work towards a goal. A customer’s agent may compare offers; your company’s agent may prepare answers or act in business systems. Connect the two: an easy-to-find offer still needs correct stock, delivery and service promises.

AI referrals can lead to purchases. Check your own journey.

Check whether an AI system can read your product facts. In April 2026, Adobe’s own checker scored U.S. retail product pages at 66% machine-readable content on average. That is a diagnostic from a vendor selling AI-visibility tools, not a universal test of every model.

Keep price, availability, terms and useful text accessible. For Google’s AI search features, existing search requirements still apply; no special AI file or schema is required. Eligibility does not guarantee inclusion.

Ask real customer questions, inspect the answers, then follow the path to an enquiry or purchase. Track accepted orders, repeat business and identifiable referrals. A mention is useful only if the next step works.

Own the agentic commerce experience ↗ Adobe report: U.S. retailers see surge in AI traffic, but many websites are not entirely readable by machines. ↗ Adobe report: U.S. consumers are embracing LLMs to make travel plans, but many brands have AI visibility gaps. ↗ AI features and your website ↗

Decide why the customer should still choose you

Invest in what makes you worth choosing: a product that fits, expertise, reliable delivery or service people value. Make it verifiable. Human preference and brand meaning still matter when an agent helps compare offers.

Own the authoritative facts, customer commitments and decision rights. Buy replaceable infrastructure when it serves the task. Before a platform deal, check fees, access to customer records, ranking incentives and the cost of leaving.

Our inference: reaching customers through someone else’s agent is valuable only if the underlying service remains worth choosing.

Agentic AI Will Make the CMO’s Role More Consequential ↗ Welcome to The Back Office ↗

Fund a workflow, then measure an accepted outcome

Choose a recurring job with checkable results. Compare the current process, simpler automation and an agent. Include handoffs, exceptions, review and repair in cost per accepted result. Set quality and failure thresholds before testing; minutes saved do not automatically become cash.

A study of 5,172 support agents at one company found 15% more issues resolved per hour with AI assistance. It did not test autonomous service. METR’s 2026 developer-study update warns that selection and parallel work prevent a reliable current speedup estimate.

Klarna shows why workload and service quality need separate checks. In February 2024 it estimated that its assistant did the work of 700 full-time agents; customers could already choose human help. Its CEO acknowledged excessive cost-cutting in a May 2025 Bloomberg interview. Its 2025 annual report still described extensive AI use alongside human support. These are company claims, not an independent effect study. Our lesson: measure resolution quality and escalation alongside cost.

Let the process owner own the outcome, finance check the economics and staff shape the work. Preserve junior colleagues’ learning. More drafts are not progress if experienced reviewers become the bottleneck.

Generative AI at Work ↗ We are Changing our Developer Productivity Experiment Design ↗ Is the AI productivity story at a turning point? ↗ Agency and Agents ↗ Klarna AI assistant handles two-thirds of customer service chats in its first month ↗ Klarna Slows AI-Driven Job Cuts With Call for Real People ↗ Klarna Group plc annual report for the year ended 31 December 2025 (Form 20-F) ↗

Increase authority only when you can contain failure

A draft can be the right permanent destination. Simon Willison warns about combining private data, untrusted content and external communication: a malicious instruction can become a route to disclosing information.

Enforce permissions in connected systems, not only in prompts. Name who can stop the workflow. Test conflicting instructions, outages, duplicate requests and recovery. Actions that cannot be reversed need stricter approval.

Shopify’s UCP explanation shows how commerce systems negotiate capabilities and handoffs. A protocol does not supply customer demand or permission to purchase. Integrate when a real journey warrants it.

The lethal trifecta for AI agents: private data, untrusted content, and external communication ↗ Building the Universal Commerce Protocol ↗

Build for the markets you actually serve

Test the markets you serve: language, currency, tax display, delivery, payments and human handoff. A working U.S. journey does not establish readiness elsewhere.

Nandan Nilekani, Infosys co-founder and advocate of Indian digital public infrastructure, emphasizes useful applications, local languages and voice. His perspective puts access alongside capability. He also has a commercial technology affiliation.

Work with local staff and users. Keep a lightweight web or human route where an agent adds friction. ITU’s estimates show why one channel cannot serve everyone.

Why the world should stop obsessing over LLMs ↗ Global number of Internet users increases, but disparities deepen key digital divides ↗

Google asks for the search basics, not a special AI file

Google says its established search requirements also apply to AI Overviews and AI Mode. It recommends accessible text, structured data that matches the page and current business information. No special AI text file or schema is required.

Google-specific documentation. Meeting the requirements does not guarantee indexing or serving; it says nothing about every other agent or search provider. The downloadable .md here is an internal process brief.

AI already appears in surveyed shopping research

IBM and NRF report that 41% of surveyed consumers had used AI to research products, 33% to look for reviews and 31% for deals or promotions.

Self-reports from over 18,000 consumers across 23 countries in Q3 2025. Categories overlap. This is not a population-weighted global adoption rate or evidence of autonomous purchasing.

Visibility and delivered value are separate strategic problems

Dhar and coauthors argue that AI-mediated comparison makes alignment between a brand’s promise and its actual performance more consequential. Kamath describes welcoming alternative interfaces to Zerodha’s services.

These are perspectives from different businesses, not a direct debate or evidence that every company should adopt the same distribution model. Atlas draws the strategic distinction.

Measure a current, complete workflow

A published support study found a 15% average gain in issues resolved per hour with AI assistance. METR’s 2026 update explains why selection and parallel work complicate current developer-productivity measurement.

Different populations, tasks, tools and methods. The support estimate is from one firm; it cannot be applied to autonomous agents or used as a forecast of cost savings.

Work design includes both process change and human judgment

Brynjolfsson emphasizes complementary organizational investment; Mollick argues for preserving useful human participation as agents become more capable.

Expert interpretations with different emphases. The advice to measure review, rework and accepted outcomes is Atlas’s application, not a validated transformation recipe.

Connected capabilities change the failure modes

Willison identifies a prompt-injection risk when an agent combines private data, untrusted input and external communication. UCP describes negotiated commerce capabilities and handoffs.

These sources concern different layers: security and interoperability. Neither a prompt nor a protocol alone establishes authorization, reliability or commercial readiness.

A global strategy needs local operating evidence

ITU estimates Internet use at 94% in high-income countries and 23% in low-income countries in 2025. Nilekani emphasizes practical applications, local languages and voice access.

Connectivity estimates are not AI-adoption figures. Nilekani’s views are a perspective, not measured proof of language quality. Local journey testing is an editorial recommendation.

Prepare to be chosen—and to deliver

An agent is software that uses a model and tools to pursue a goal across several steps. For a company, two changes matter. A customer’s agent may research, compare or arrange a purchase. Your own agent may assemble an answer, coordinate work or act in a business system. The first changes access to demand; the second changes how you fulfil it.

Treat them as connected decisions. A perfectly discoverable offer is a liability if stock or delivery promises are wrong. A cheaper internal process is a poor investment if it makes the customer’s problem harder to resolve. Start with a valuable customer outcome, then choose the smallest useful change.

Research is moving. Purchases are a separate question.

In IBM and NRF’s 23-country survey, AI use already reached parts of shopping research. That makes accurate answers worth preparing for. It does not tell you how much revenue an agent will control—or whether a recommendation becomes a purchase.

For Google’s AI search features, the starting work is familiar: accessible useful content, current business information and structured data that matches the visible page. Google requires no special AI file or schema. Indexing and eligibility still do not guarantee inclusion.

Build a small set of real customer questions, including alternatives, objections and local terms. Inspect whether answers represent your offer correctly and whether people can complete the next step. Track qualified enquiries, purchases and repeat business alongside identifiable referrals. Separate unknown attribution; a rise in mentions is not a return on investment.

Own the agentic commerce experience ↗ AI features and your website ↗

Decide why the customer should still choose you

If an agent can compare interchangeable offers, another layer of promotional copy is unlikely to be a durable advantage. Invest in something customers value: a product that fits, distinctive expertise, dependable fulfilment, useful service or a relationship they want to keep. Make that value easy to verify. Human preference and brand meaning still belong in the strategy; an agent need not make the entire decision.

The ownership question is sharper than “build or buy AI.” Own the authoritative facts, customer commitments, decision rights and evidence of quality. Buy replaceable infrastructure where it serves the job. Before signing a platform deal, assess fees, access to customer records, ranking incentives and the cost of leaving. Keep a direct route where it is commercially sensible.

These two positions illuminate different choices; they are not a recorded debate. Our inference: reach through someone else’s agent can be valuable, but only if what remains underneath is worth choosing.

Agentic AI Will Make the CMO’s Role More Consequential ↗ Welcome to The Back Office ↗

Fund a workflow, then measure an accepted outcome

Choose a recurring job with usable inputs, a checkable result and a manageable cost of failure. Compare an agent with the present process and simpler automation. Deterministic rules may be cheaper and easier to maintain; some tasks do not need changing. Sketch the whole journey, including handoffs, exceptions and the person who fixes mistakes.

There is evidence of useful assistance, but no universal productivity number. A published study of 5,172 support agents at one company found 15% more issues resolved per hour on average, with uneven effects across workers. It studied assistance to staff, not autonomous service. METR’s 2026 developer-study update warns that selection effects and parallel agent work prevent a reliable current speedup estimate.

Define acceptance before testing: factual accuracy, completion, customer impact and failures that must never pass unnoticed. Compare similar cases, count failed runs and record review and repair time. Include model use, integration, supervision and maintenance in cost per accepted result. Value extra capacity only when it improves service or produces useful output; minutes saved do not automatically become cash.

Give the process owner responsibility for the outcome, with finance checking the economics and staff shaping the work. Preserve learning opportunities for junior colleagues. A system that produces more drafts but makes experienced people the new bottleneck has not yet improved the workflow.

Generative AI at Work ↗ We are Changing our Developer Productivity Experiment Design ↗ Is the AI productivity story at a turning point? ↗ Agency and Agents ↗

Increase authority only when you can contain failure

A useful draft can be the right permanent destination. Each added permission changes what failure can cost. Software developer Simon Willison highlights a particularly dangerous combination: private data, untrusted content and external communication. A malicious instruction in a document can become more consequential when the same system can disclose information.

Enforce scope in connected systems, not only in a prompt. Name the person who can stop the workflow. Test unavailable services, conflicting instructions, duplicate requests and recovery before giving it consequential actions. Some effects cannot be fully reversed; those require a stricter gate.

Commerce standards can help systems negotiate supported capabilities. Shopify’s UCP explanation shows that mechanism, including handoffs. A protocol alone does not create demand, impartial recommendations or permission to purchase. Integrate when a real customer or partner journey warrants it and verify that journey in the intended market.

The lethal trifecta for AI agents: private data, untrusted content, and external communication ↗ Building the Universal Commerce Protocol ↗

Build for the markets you actually serve

A launch in the United States is not evidence of readiness in Europe, India or elsewhere. Verify local service availability, customer expectations and applicable requirements before promising a journey. Test actual language, currency, tax presentation, delivery, payments and escalation—not only a translated prompt.

Nandan Nilekani, Infosys co-founder and an advocate of Indian digital public infrastructure, emphasizes useful applications, local languages and voice. His perspective broadens the question from model capability to access; he also has a commercial technology affiliation. ITU’s connectivity estimates show why a global strategy needs more than one route.

Prioritize the markets where you already have customers or a credible offer. Run the same outcome checks with local staff and users; preserve a lightweight web or human route where an agent adds friction. Scale a demonstrated journey, not the assumption that everyone will use the same assistant.

Why the world should stop obsessing over LLMs ↗ Global number of Internet users increases, but disparities deepen key digital divides ↗

Google asks for the search basics, not a special AI file

Google says its established search requirements also apply to AI Overviews and AI Mode. It recommends accessible text, structured data that matches the page and current business information. No special AI text file or schema is required.

Google-specific documentation. Meeting the requirements does not guarantee indexing or serving; it says nothing about every other agent or search provider. The downloadable .md here is an internal process brief.