of the United Kingdom’s capitol city.
I was listening to episode 662 of the Accidental Tech Podcast, “Just Break the Law,” about something foundational: why large language models can't reliably reproduce their own outputs.
You’ve probably done it yourself: input the same prompt, with the same settings, and you will get different answers.
As I thought through the implications, I landed on a commerce-based insight: if the interface for commerce becomes a non-deterministic system, then performance marketing, the entire discipline of turning inputs into predictable outcomes, becomes unreliable.
Not because AI is "bad." But because this kind of technology plays by different rules.
The Vending Machine Fallacy
A decade of digital commerce growth was built on an assumption that “if we do X, Y will happen.”
Bid here. Target that. Change this creative. Optimize that landing page.
We scaled what worked and expected the same, or similar, outcomes. And for the most part, we always got what we expected.
Getting there was sometimes tricky, but the promise was stable and consistent. If you invested a certain amount of budget into Meta ads, for instance, you could expect a specific performance benchmark, even at scale. The underlying promise has shaped org charts, budgets, martech stacks, and careers.
Now we’re letting LLMs sit between buyers and decisions, and they don’t make the same promise. They can’t.
Deterministic vs. Non-deterministic Systems
A deterministic system behaves like a vending machine. If you press A1, you get a Snickers bar every time. If it fails, it does so predictably, so the machine can be easily fixed (and you can still get your Snickers bar).
LLMs are not vending machines. They’re closer to a conversation with a brilliant person on very little sleep: competent, helpful, and slightly variable in their efficacy.
Ask the same question five times, and you’ll get five slightly different answers. Even if the meaning is similar, the path matters.
Most people explain this variability with one word: sampling.
- The model produces probabilities for the next token.
- The system samples a token from that distribution.
- Sampling causes this variability.
True.
But it’s not the whole story, which is what really matters for Commerce professionals.
Sidebar: ‘Temperature = 0’ Still Isn’t a Guarantee
Even when you set the temperature to 0 (greedy sampling), LLM inference can still be non-deterministic in practice.
The reason is that the endpoint is a highly parallel inference system running on modern hardware. The output varies depending on how the computation is executed under load.
One key driver is batching.
When an LLM endpoint is busy, it batches multiple user requests to increase throughput. That grouping can change execution details and numerical behavior in ways that can lead to different outputs for the user, even when they submit the same request.
As a result, LLM platforms like ChatGPT, Perplexity, and Claude are far more than just another channel you need to “show up on.” These are entirely new channels where the same marketing stimulus may not reliably elicit the same customer response, making “gaming” or “hacking” them for scalability incredibly difficult, if not impossible.
The Unreproduceable Funnel
Commerce, and specifically the funnel, was engineered for predictability. Classic performance marketing is a discipline of control in which brands control the messaging, targeting, attribution, and the optimization loop.
Even when platforms became black boxes, we still had a stable game where we could infer which inputs led to specific outputs. From there, we could build repeatable playbooks and scale what worked. Everything that didn’t? We threw it back into the pile, and we ran new tests.
LLMs change the process entirely because the results are impossible to reproduce. A query no longer leads to a ranked list of relevant and trusted sources. Instead, it yields a synthesized answer, with results and ranking logic varying each time.
This all sounds very scary to marketing professionals who have built their entire careers on repeatable frameworks. But the good news is that competitive advantage lives in the gray area of this variability.
Unsteerable Acquisition
Here’s my strong take: AI isn’t killing acquisition, but it is killing the illusion of control in acquisition.
You can still influence outcomes, but you have to use entirely different levers. In this new era of commerce, LLMs favor:
- Brand strength and trust
- Structured product data and clarity
- Distribution partnerships and authority signals
- Explicit and implicit customer satisfaction signals
- Real-world preference
You can’t assume your best keyword set will consistently trigger your placement. Your winning creative won’t always drive the framing of results. And, here’s the biggest kicker: your optimized campaign won’t always connect LLM results to actual demand.
Acquisition is starting to behave less like paid search and more like PR. You can certainly attempt to shape conditions, but you can’t always control when or how you show up.
Five Levers That Still Work
One possible strategy, especially for SMBs with limited budgets, is to stop over-investing in parts of the funnel that are becoming more stochastic and to invest in the parts that remain more deterministic.
1) Product truth
If your product creates genuine preference, no model can out-rank reality. Great products create a natural pull.
Product truth starts with a relentless focus on solving real problems better than alternatives. Invest in R&D to iterate based on user data, including surveys and behavioral insights derived from usage patterns. For example, if you're in apparel, invest in durable materials and ethical sourcing that resonate with conscious consumers. If you manufacture tech gadgets, focus on seamless integration and longevity that outlasts the hype cycle.
Patagonia and Apple haven’t created true pull from ads alone (Apple’s recent iPad ads had the opposite effect). Instead, they’ve focused on creating products that inspire loyalty and word-of-mouth. Measure Net Promoter Score (NPS) for product features and use AI tools ethically to analyze feedback.
Product truth is a cultural commitment that turns customers into advocates and insulates your brand from fluctuations in non-deterministic discovery. This is the true foundation for success.
2) Experience reliability
When acquisition becomes volatile, make conversion and post-purchase experiences boringly excellent. That means, beyond having high-quality products, you have clear, robust PDPs, predictable shipping and delivery times, and a painless return policy. If a customer has an issue at any time, your stellar customer service team should have the tools and processes to resolve it promptly.
3) User-generated content
In a non-deterministic world, AI mediates discovery. Models synthesize judgment, and human proof carries disproportionate weight. After all, user-generated content (UGC), such as reviews, helps LLMs deliver highly tailored recommendations to end users. They are digital concierges that reduce the mental load of finding the best product at the best price.
UGC supports an AI-mediated funnel in three key ways: it’s a training signal that helps the model gauge real preference; it’s a trust signal that helps validate decisions; and it’s a conversion lubricant that reduces hesitation and builds confidence.
UGC has been part of the eCommerce playbook for more than a decade, but it’s now the anchor for brand credibility. Create a referral program that goes beyond capturing vague star-based ratings. Request feedback at different stages of the lifecycle, from post-purchase to post-delivery and post-use. Reward consumers for the detail and helpfulness of their feedback. Try offering additional perks when they include rich visuals and address key consumer questions. Use all of this incredible content and insight to build out your PDPs, category pages, emails, and ads.
4) Retention systems
If you can’t fully control who arrives, control what happens after. Design retention systems that align with your brand and product portfolio, whether you want to simply drive repeat purchases and product replenishment or drive subscriptions and memberships.
Start with operational excellence: streamline supply chains to improve reliability. For example, you can use predictive analytics to prevent stockouts or delivery delays. Unify first-party data across touchpoints, including your CRM, apps, and in-store, to create a single view of the customer. When you have a single view of the customer, you can create personalized experiences that don’t feel creepy.
Loyalty mechanics should offer real value exchange, such as tiered rewards that escalate with engagement. Offer points redeemable for discounts or offer exclusive access to beta products. Look at brands like Sephora with its Beauty Insider program or Starbucks' app-based rewards for inspiration. These systems create habit loops that make retention predictable. Then, track success with cohort analysis: measure lifetime value uplift from second purchases onward.
If you design retention as an interconnected system, you create a flywheel in which AI-driven arrivals become long-term assets.
5) Owned distribution
Put effort into things that are not sexy but foundational to your brand, from email and community building to external brand partnerships.
These channels bypass AI intermediaries, creating a direct line to consumers. Segment email lists with behavioral data to deliver value-driven content, like curated product stories or exclusive offers. And, of course, comply with regulations such as GDPR to maintain trust. Tap into more direct channels based on a consumer’s level of engagement, reserving SMS for high-urgency, high-value touches like flash sales and order updates to reduce opt-outs, and saving app-based push notifications for users who have opted in to personalized alerts, such as restock notifications for wishlist items.
Building dedicated communities via Discord, Slack, Reddit, or social groups will foster organic discussions, where user advocacy amplifies your brand. Creators can partner with your brand and each other to create more relatable content for campaigns. Nurture key accounts and top shoppers with special showroom experiences and off-site events. Build partnerships with complementary brands to create co-marketing opportunities that expand your reach without relying on algorithms. Invest in analytics to optimize open rates, click-through rates, and conversion paths, making them reliable engines of repeat business.
Retention Is the New Black
The old era of eCommerce rewarded the best media buyers, but the next era will reward the best system designers.
Non-determinism is transforming the relationship between consumers and commerce, with shoppers now delegating to agents rather than clicking through rows of tabs. Brands are no longer shoving them into predictable funnels. Instead, LLMs are sending them on probabilistic journeys. Probabilistic systems power discovery, which means the most successful brands are those that turn their eCommerce experiences into trusted third spaces.
That’s why in the near future, you’ll see more brands invent KPIs for things that used to be soft, but are now tangible advantages, like time-to-resolution, customer effort score, product confidence, service reliability, and advocacy rate.
The question to ask of AI in this new era: if the algorithm is in conversation with the user, what kind of brand deserves to be recommended?
I was listening to episode 662 of the Accidental Tech Podcast, “Just Break the Law,” about something foundational: why large language models can't reliably reproduce their own outputs.
You’ve probably done it yourself: input the same prompt, with the same settings, and you will get different answers.
As I thought through the implications, I landed on a commerce-based insight: if the interface for commerce becomes a non-deterministic system, then performance marketing, the entire discipline of turning inputs into predictable outcomes, becomes unreliable.
Not because AI is "bad." But because this kind of technology plays by different rules.
The Vending Machine Fallacy
A decade of digital commerce growth was built on an assumption that “if we do X, Y will happen.”
Bid here. Target that. Change this creative. Optimize that landing page.
We scaled what worked and expected the same, or similar, outcomes. And for the most part, we always got what we expected.
Getting there was sometimes tricky, but the promise was stable and consistent. If you invested a certain amount of budget into Meta ads, for instance, you could expect a specific performance benchmark, even at scale. The underlying promise has shaped org charts, budgets, martech stacks, and careers.
Now we’re letting LLMs sit between buyers and decisions, and they don’t make the same promise. They can’t.
Deterministic vs. Non-deterministic Systems
A deterministic system behaves like a vending machine. If you press A1, you get a Snickers bar every time. If it fails, it does so predictably, so the machine can be easily fixed (and you can still get your Snickers bar).
LLMs are not vending machines. They’re closer to a conversation with a brilliant person on very little sleep: competent, helpful, and slightly variable in their efficacy.
Ask the same question five times, and you’ll get five slightly different answers. Even if the meaning is similar, the path matters.
Most people explain this variability with one word: sampling.
- The model produces probabilities for the next token.
- The system samples a token from that distribution.
- Sampling causes this variability.
True.
But it’s not the whole story, which is what really matters for Commerce professionals.
Sidebar: ‘Temperature = 0’ Still Isn’t a Guarantee
Even when you set the temperature to 0 (greedy sampling), LLM inference can still be non-deterministic in practice.
The reason is that the endpoint is a highly parallel inference system running on modern hardware. The output varies depending on how the computation is executed under load.
One key driver is batching.
When an LLM endpoint is busy, it batches multiple user requests to increase throughput. That grouping can change execution details and numerical behavior in ways that can lead to different outputs for the user, even when they submit the same request.
As a result, LLM platforms like ChatGPT, Perplexity, and Claude are far more than just another channel you need to “show up on.” These are entirely new channels where the same marketing stimulus may not reliably elicit the same customer response, making “gaming” or “hacking” them for scalability incredibly difficult, if not impossible.
The Unreproduceable Funnel
Commerce, and specifically the funnel, was engineered for predictability. Classic performance marketing is a discipline of control in which brands control the messaging, targeting, attribution, and the optimization loop.
Even when platforms became black boxes, we still had a stable game where we could infer which inputs led to specific outputs. From there, we could build repeatable playbooks and scale what worked. Everything that didn’t? We threw it back into the pile, and we ran new tests.
LLMs change the process entirely because the results are impossible to reproduce. A query no longer leads to a ranked list of relevant and trusted sources. Instead, it yields a synthesized answer, with results and ranking logic varying each time.
This all sounds very scary to marketing professionals who have built their entire careers on repeatable frameworks. But the good news is that competitive advantage lives in the gray area of this variability.
Unsteerable Acquisition
Here’s my strong take: AI isn’t killing acquisition, but it is killing the illusion of control in acquisition.
You can still influence outcomes, but you have to use entirely different levers. In this new era of commerce, LLMs favor:
- Brand strength and trust
- Structured product data and clarity
- Distribution partnerships and authority signals
- Explicit and implicit customer satisfaction signals
- Real-world preference
You can’t assume your best keyword set will consistently trigger your placement. Your winning creative won’t always drive the framing of results. And, here’s the biggest kicker: your optimized campaign won’t always connect LLM results to actual demand.
Acquisition is starting to behave less like paid search and more like PR. You can certainly attempt to shape conditions, but you can’t always control when or how you show up.
Five Levers That Still Work
One possible strategy, especially for SMBs with limited budgets, is to stop over-investing in parts of the funnel that are becoming more stochastic and to invest in the parts that remain more deterministic.
1) Product truth
If your product creates genuine preference, no model can out-rank reality. Great products create a natural pull.
Product truth starts with a relentless focus on solving real problems better than alternatives. Invest in R&D to iterate based on user data, including surveys and behavioral insights derived from usage patterns. For example, if you're in apparel, invest in durable materials and ethical sourcing that resonate with conscious consumers. If you manufacture tech gadgets, focus on seamless integration and longevity that outlasts the hype cycle.
Patagonia and Apple haven’t created true pull from ads alone (Apple’s recent iPad ads had the opposite effect). Instead, they’ve focused on creating products that inspire loyalty and word-of-mouth. Measure Net Promoter Score (NPS) for product features and use AI tools ethically to analyze feedback.
Product truth is a cultural commitment that turns customers into advocates and insulates your brand from fluctuations in non-deterministic discovery. This is the true foundation for success.
2) Experience reliability
When acquisition becomes volatile, make conversion and post-purchase experiences boringly excellent. That means, beyond having high-quality products, you have clear, robust PDPs, predictable shipping and delivery times, and a painless return policy. If a customer has an issue at any time, your stellar customer service team should have the tools and processes to resolve it promptly.
3) User-generated content
In a non-deterministic world, AI mediates discovery. Models synthesize judgment, and human proof carries disproportionate weight. After all, user-generated content (UGC), such as reviews, helps LLMs deliver highly tailored recommendations to end users. They are digital concierges that reduce the mental load of finding the best product at the best price.
UGC supports an AI-mediated funnel in three key ways: it’s a training signal that helps the model gauge real preference; it’s a trust signal that helps validate decisions; and it’s a conversion lubricant that reduces hesitation and builds confidence.
UGC has been part of the eCommerce playbook for more than a decade, but it’s now the anchor for brand credibility. Create a referral program that goes beyond capturing vague star-based ratings. Request feedback at different stages of the lifecycle, from post-purchase to post-delivery and post-use. Reward consumers for the detail and helpfulness of their feedback. Try offering additional perks when they include rich visuals and address key consumer questions. Use all of this incredible content and insight to build out your PDPs, category pages, emails, and ads.
4) Retention systems
If you can’t fully control who arrives, control what happens after. Design retention systems that align with your brand and product portfolio, whether you want to simply drive repeat purchases and product replenishment or drive subscriptions and memberships.
Start with operational excellence: streamline supply chains to improve reliability. For example, you can use predictive analytics to prevent stockouts or delivery delays. Unify first-party data across touchpoints, including your CRM, apps, and in-store, to create a single view of the customer. When you have a single view of the customer, you can create personalized experiences that don’t feel creepy.
Loyalty mechanics should offer real value exchange, such as tiered rewards that escalate with engagement. Offer points redeemable for discounts or offer exclusive access to beta products. Look at brands like Sephora with its Beauty Insider program or Starbucks' app-based rewards for inspiration. These systems create habit loops that make retention predictable. Then, track success with cohort analysis: measure lifetime value uplift from second purchases onward.
If you design retention as an interconnected system, you create a flywheel in which AI-driven arrivals become long-term assets.
5) Owned distribution
Put effort into things that are not sexy but foundational to your brand, from email and community building to external brand partnerships.
These channels bypass AI intermediaries, creating a direct line to consumers. Segment email lists with behavioral data to deliver value-driven content, like curated product stories or exclusive offers. And, of course, comply with regulations such as GDPR to maintain trust. Tap into more direct channels based on a consumer’s level of engagement, reserving SMS for high-urgency, high-value touches like flash sales and order updates to reduce opt-outs, and saving app-based push notifications for users who have opted in to personalized alerts, such as restock notifications for wishlist items.
Building dedicated communities via Discord, Slack, Reddit, or social groups will foster organic discussions, where user advocacy amplifies your brand. Creators can partner with your brand and each other to create more relatable content for campaigns. Nurture key accounts and top shoppers with special showroom experiences and off-site events. Build partnerships with complementary brands to create co-marketing opportunities that expand your reach without relying on algorithms. Invest in analytics to optimize open rates, click-through rates, and conversion paths, making them reliable engines of repeat business.
Retention Is the New Black
The old era of eCommerce rewarded the best media buyers, but the next era will reward the best system designers.
Non-determinism is transforming the relationship between consumers and commerce, with shoppers now delegating to agents rather than clicking through rows of tabs. Brands are no longer shoving them into predictable funnels. Instead, LLMs are sending them on probabilistic journeys. Probabilistic systems power discovery, which means the most successful brands are those that turn their eCommerce experiences into trusted third spaces.
That’s why in the near future, you’ll see more brands invent KPIs for things that used to be soft, but are now tangible advantages, like time-to-resolution, customer effort score, product confidence, service reliability, and advocacy rate.
The question to ask of AI in this new era: if the algorithm is in conversation with the user, what kind of brand deserves to be recommended?
I was listening to episode 662 of the Accidental Tech Podcast, “Just Break the Law,” about something foundational: why large language models can't reliably reproduce their own outputs.
You’ve probably done it yourself: input the same prompt, with the same settings, and you will get different answers.
As I thought through the implications, I landed on a commerce-based insight: if the interface for commerce becomes a non-deterministic system, then performance marketing, the entire discipline of turning inputs into predictable outcomes, becomes unreliable.
Not because AI is "bad." But because this kind of technology plays by different rules.
The Vending Machine Fallacy
A decade of digital commerce growth was built on an assumption that “if we do X, Y will happen.”
Bid here. Target that. Change this creative. Optimize that landing page.
We scaled what worked and expected the same, or similar, outcomes. And for the most part, we always got what we expected.
Getting there was sometimes tricky, but the promise was stable and consistent. If you invested a certain amount of budget into Meta ads, for instance, you could expect a specific performance benchmark, even at scale. The underlying promise has shaped org charts, budgets, martech stacks, and careers.
Now we’re letting LLMs sit between buyers and decisions, and they don’t make the same promise. They can’t.
Deterministic vs. Non-deterministic Systems
A deterministic system behaves like a vending machine. If you press A1, you get a Snickers bar every time. If it fails, it does so predictably, so the machine can be easily fixed (and you can still get your Snickers bar).
LLMs are not vending machines. They’re closer to a conversation with a brilliant person on very little sleep: competent, helpful, and slightly variable in their efficacy.
Ask the same question five times, and you’ll get five slightly different answers. Even if the meaning is similar, the path matters.
Most people explain this variability with one word: sampling.
- The model produces probabilities for the next token.
- The system samples a token from that distribution.
- Sampling causes this variability.
True.
But it’s not the whole story, which is what really matters for Commerce professionals.
Sidebar: ‘Temperature = 0’ Still Isn’t a Guarantee
Even when you set the temperature to 0 (greedy sampling), LLM inference can still be non-deterministic in practice.
The reason is that the endpoint is a highly parallel inference system running on modern hardware. The output varies depending on how the computation is executed under load.
One key driver is batching.
When an LLM endpoint is busy, it batches multiple user requests to increase throughput. That grouping can change execution details and numerical behavior in ways that can lead to different outputs for the user, even when they submit the same request.
As a result, LLM platforms like ChatGPT, Perplexity, and Claude are far more than just another channel you need to “show up on.” These are entirely new channels where the same marketing stimulus may not reliably elicit the same customer response, making “gaming” or “hacking” them for scalability incredibly difficult, if not impossible.
The Unreproduceable Funnel
Commerce, and specifically the funnel, was engineered for predictability. Classic performance marketing is a discipline of control in which brands control the messaging, targeting, attribution, and the optimization loop.
Even when platforms became black boxes, we still had a stable game where we could infer which inputs led to specific outputs. From there, we could build repeatable playbooks and scale what worked. Everything that didn’t? We threw it back into the pile, and we ran new tests.
LLMs change the process entirely because the results are impossible to reproduce. A query no longer leads to a ranked list of relevant and trusted sources. Instead, it yields a synthesized answer, with results and ranking logic varying each time.
This all sounds very scary to marketing professionals who have built their entire careers on repeatable frameworks. But the good news is that competitive advantage lives in the gray area of this variability.
Unsteerable Acquisition
Here’s my strong take: AI isn’t killing acquisition, but it is killing the illusion of control in acquisition.
You can still influence outcomes, but you have to use entirely different levers. In this new era of commerce, LLMs favor:
- Brand strength and trust
- Structured product data and clarity
- Distribution partnerships and authority signals
- Explicit and implicit customer satisfaction signals
- Real-world preference
You can’t assume your best keyword set will consistently trigger your placement. Your winning creative won’t always drive the framing of results. And, here’s the biggest kicker: your optimized campaign won’t always connect LLM results to actual demand.
Acquisition is starting to behave less like paid search and more like PR. You can certainly attempt to shape conditions, but you can’t always control when or how you show up.
Five Levers That Still Work
One possible strategy, especially for SMBs with limited budgets, is to stop over-investing in parts of the funnel that are becoming more stochastic and to invest in the parts that remain more deterministic.
1) Product truth
If your product creates genuine preference, no model can out-rank reality. Great products create a natural pull.
Product truth starts with a relentless focus on solving real problems better than alternatives. Invest in R&D to iterate based on user data, including surveys and behavioral insights derived from usage patterns. For example, if you're in apparel, invest in durable materials and ethical sourcing that resonate with conscious consumers. If you manufacture tech gadgets, focus on seamless integration and longevity that outlasts the hype cycle.
Patagonia and Apple haven’t created true pull from ads alone (Apple’s recent iPad ads had the opposite effect). Instead, they’ve focused on creating products that inspire loyalty and word-of-mouth. Measure Net Promoter Score (NPS) for product features and use AI tools ethically to analyze feedback.
Product truth is a cultural commitment that turns customers into advocates and insulates your brand from fluctuations in non-deterministic discovery. This is the true foundation for success.
2) Experience reliability
When acquisition becomes volatile, make conversion and post-purchase experiences boringly excellent. That means, beyond having high-quality products, you have clear, robust PDPs, predictable shipping and delivery times, and a painless return policy. If a customer has an issue at any time, your stellar customer service team should have the tools and processes to resolve it promptly.
3) User-generated content
In a non-deterministic world, AI mediates discovery. Models synthesize judgment, and human proof carries disproportionate weight. After all, user-generated content (UGC), such as reviews, helps LLMs deliver highly tailored recommendations to end users. They are digital concierges that reduce the mental load of finding the best product at the best price.
UGC supports an AI-mediated funnel in three key ways: it’s a training signal that helps the model gauge real preference; it’s a trust signal that helps validate decisions; and it’s a conversion lubricant that reduces hesitation and builds confidence.
UGC has been part of the eCommerce playbook for more than a decade, but it’s now the anchor for brand credibility. Create a referral program that goes beyond capturing vague star-based ratings. Request feedback at different stages of the lifecycle, from post-purchase to post-delivery and post-use. Reward consumers for the detail and helpfulness of their feedback. Try offering additional perks when they include rich visuals and address key consumer questions. Use all of this incredible content and insight to build out your PDPs, category pages, emails, and ads.
4) Retention systems
If you can’t fully control who arrives, control what happens after. Design retention systems that align with your brand and product portfolio, whether you want to simply drive repeat purchases and product replenishment or drive subscriptions and memberships.
Start with operational excellence: streamline supply chains to improve reliability. For example, you can use predictive analytics to prevent stockouts or delivery delays. Unify first-party data across touchpoints, including your CRM, apps, and in-store, to create a single view of the customer. When you have a single view of the customer, you can create personalized experiences that don’t feel creepy.
Loyalty mechanics should offer real value exchange, such as tiered rewards that escalate with engagement. Offer points redeemable for discounts or offer exclusive access to beta products. Look at brands like Sephora with its Beauty Insider program or Starbucks' app-based rewards for inspiration. These systems create habit loops that make retention predictable. Then, track success with cohort analysis: measure lifetime value uplift from second purchases onward.
If you design retention as an interconnected system, you create a flywheel in which AI-driven arrivals become long-term assets.
5) Owned distribution
Put effort into things that are not sexy but foundational to your brand, from email and community building to external brand partnerships.
These channels bypass AI intermediaries, creating a direct line to consumers. Segment email lists with behavioral data to deliver value-driven content, like curated product stories or exclusive offers. And, of course, comply with regulations such as GDPR to maintain trust. Tap into more direct channels based on a consumer’s level of engagement, reserving SMS for high-urgency, high-value touches like flash sales and order updates to reduce opt-outs, and saving app-based push notifications for users who have opted in to personalized alerts, such as restock notifications for wishlist items.
Building dedicated communities via Discord, Slack, Reddit, or social groups will foster organic discussions, where user advocacy amplifies your brand. Creators can partner with your brand and each other to create more relatable content for campaigns. Nurture key accounts and top shoppers with special showroom experiences and off-site events. Build partnerships with complementary brands to create co-marketing opportunities that expand your reach without relying on algorithms. Invest in analytics to optimize open rates, click-through rates, and conversion paths, making them reliable engines of repeat business.
Retention Is the New Black
The old era of eCommerce rewarded the best media buyers, but the next era will reward the best system designers.
Non-determinism is transforming the relationship between consumers and commerce, with shoppers now delegating to agents rather than clicking through rows of tabs. Brands are no longer shoving them into predictable funnels. Instead, LLMs are sending them on probabilistic journeys. Probabilistic systems power discovery, which means the most successful brands are those that turn their eCommerce experiences into trusted third spaces.
That’s why in the near future, you’ll see more brands invent KPIs for things that used to be soft, but are now tangible advantages, like time-to-resolution, customer effort score, product confidence, service reliability, and advocacy rate.
The question to ask of AI in this new era: if the algorithm is in conversation with the user, what kind of brand deserves to be recommended?
I was listening to episode 662 of the Accidental Tech Podcast, “Just Break the Law,” about something foundational: why large language models can't reliably reproduce their own outputs.
You’ve probably done it yourself: input the same prompt, with the same settings, and you will get different answers.
As I thought through the implications, I landed on a commerce-based insight: if the interface for commerce becomes a non-deterministic system, then performance marketing, the entire discipline of turning inputs into predictable outcomes, becomes unreliable.
Not because AI is "bad." But because this kind of technology plays by different rules.
The Vending Machine Fallacy
A decade of digital commerce growth was built on an assumption that “if we do X, Y will happen.”
Bid here. Target that. Change this creative. Optimize that landing page.
We scaled what worked and expected the same, or similar, outcomes. And for the most part, we always got what we expected.
Getting there was sometimes tricky, but the promise was stable and consistent. If you invested a certain amount of budget into Meta ads, for instance, you could expect a specific performance benchmark, even at scale. The underlying promise has shaped org charts, budgets, martech stacks, and careers.
Now we’re letting LLMs sit between buyers and decisions, and they don’t make the same promise. They can’t.
Deterministic vs. Non-deterministic Systems
A deterministic system behaves like a vending machine. If you press A1, you get a Snickers bar every time. If it fails, it does so predictably, so the machine can be easily fixed (and you can still get your Snickers bar).
LLMs are not vending machines. They’re closer to a conversation with a brilliant person on very little sleep: competent, helpful, and slightly variable in their efficacy.
Ask the same question five times, and you’ll get five slightly different answers. Even if the meaning is similar, the path matters.
Most people explain this variability with one word: sampling.
- The model produces probabilities for the next token.
- The system samples a token from that distribution.
- Sampling causes this variability.
True.
But it’s not the whole story, which is what really matters for Commerce professionals.
Sidebar: ‘Temperature = 0’ Still Isn’t a Guarantee
Even when you set the temperature to 0 (greedy sampling), LLM inference can still be non-deterministic in practice.
The reason is that the endpoint is a highly parallel inference system running on modern hardware. The output varies depending on how the computation is executed under load.
One key driver is batching.
When an LLM endpoint is busy, it batches multiple user requests to increase throughput. That grouping can change execution details and numerical behavior in ways that can lead to different outputs for the user, even when they submit the same request.
As a result, LLM platforms like ChatGPT, Perplexity, and Claude are far more than just another channel you need to “show up on.” These are entirely new channels where the same marketing stimulus may not reliably elicit the same customer response, making “gaming” or “hacking” them for scalability incredibly difficult, if not impossible.
The Unreproduceable Funnel
Commerce, and specifically the funnel, was engineered for predictability. Classic performance marketing is a discipline of control in which brands control the messaging, targeting, attribution, and the optimization loop.
Even when platforms became black boxes, we still had a stable game where we could infer which inputs led to specific outputs. From there, we could build repeatable playbooks and scale what worked. Everything that didn’t? We threw it back into the pile, and we ran new tests.
LLMs change the process entirely because the results are impossible to reproduce. A query no longer leads to a ranked list of relevant and trusted sources. Instead, it yields a synthesized answer, with results and ranking logic varying each time.
This all sounds very scary to marketing professionals who have built their entire careers on repeatable frameworks. But the good news is that competitive advantage lives in the gray area of this variability.
Unsteerable Acquisition
Here’s my strong take: AI isn’t killing acquisition, but it is killing the illusion of control in acquisition.
You can still influence outcomes, but you have to use entirely different levers. In this new era of commerce, LLMs favor:
- Brand strength and trust
- Structured product data and clarity
- Distribution partnerships and authority signals
- Explicit and implicit customer satisfaction signals
- Real-world preference
You can’t assume your best keyword set will consistently trigger your placement. Your winning creative won’t always drive the framing of results. And, here’s the biggest kicker: your optimized campaign won’t always connect LLM results to actual demand.
Acquisition is starting to behave less like paid search and more like PR. You can certainly attempt to shape conditions, but you can’t always control when or how you show up.
Five Levers That Still Work
One possible strategy, especially for SMBs with limited budgets, is to stop over-investing in parts of the funnel that are becoming more stochastic and to invest in the parts that remain more deterministic.
1) Product truth
If your product creates genuine preference, no model can out-rank reality. Great products create a natural pull.
Product truth starts with a relentless focus on solving real problems better than alternatives. Invest in R&D to iterate based on user data, including surveys and behavioral insights derived from usage patterns. For example, if you're in apparel, invest in durable materials and ethical sourcing that resonate with conscious consumers. If you manufacture tech gadgets, focus on seamless integration and longevity that outlasts the hype cycle.
Patagonia and Apple haven’t created true pull from ads alone (Apple’s recent iPad ads had the opposite effect). Instead, they’ve focused on creating products that inspire loyalty and word-of-mouth. Measure Net Promoter Score (NPS) for product features and use AI tools ethically to analyze feedback.
Product truth is a cultural commitment that turns customers into advocates and insulates your brand from fluctuations in non-deterministic discovery. This is the true foundation for success.
2) Experience reliability
When acquisition becomes volatile, make conversion and post-purchase experiences boringly excellent. That means, beyond having high-quality products, you have clear, robust PDPs, predictable shipping and delivery times, and a painless return policy. If a customer has an issue at any time, your stellar customer service team should have the tools and processes to resolve it promptly.
3) User-generated content
In a non-deterministic world, AI mediates discovery. Models synthesize judgment, and human proof carries disproportionate weight. After all, user-generated content (UGC), such as reviews, helps LLMs deliver highly tailored recommendations to end users. They are digital concierges that reduce the mental load of finding the best product at the best price.
UGC supports an AI-mediated funnel in three key ways: it’s a training signal that helps the model gauge real preference; it’s a trust signal that helps validate decisions; and it’s a conversion lubricant that reduces hesitation and builds confidence.
UGC has been part of the eCommerce playbook for more than a decade, but it’s now the anchor for brand credibility. Create a referral program that goes beyond capturing vague star-based ratings. Request feedback at different stages of the lifecycle, from post-purchase to post-delivery and post-use. Reward consumers for the detail and helpfulness of their feedback. Try offering additional perks when they include rich visuals and address key consumer questions. Use all of this incredible content and insight to build out your PDPs, category pages, emails, and ads.
4) Retention systems
If you can’t fully control who arrives, control what happens after. Design retention systems that align with your brand and product portfolio, whether you want to simply drive repeat purchases and product replenishment or drive subscriptions and memberships.
Start with operational excellence: streamline supply chains to improve reliability. For example, you can use predictive analytics to prevent stockouts or delivery delays. Unify first-party data across touchpoints, including your CRM, apps, and in-store, to create a single view of the customer. When you have a single view of the customer, you can create personalized experiences that don’t feel creepy.
Loyalty mechanics should offer real value exchange, such as tiered rewards that escalate with engagement. Offer points redeemable for discounts or offer exclusive access to beta products. Look at brands like Sephora with its Beauty Insider program or Starbucks' app-based rewards for inspiration. These systems create habit loops that make retention predictable. Then, track success with cohort analysis: measure lifetime value uplift from second purchases onward.
If you design retention as an interconnected system, you create a flywheel in which AI-driven arrivals become long-term assets.
5) Owned distribution
Put effort into things that are not sexy but foundational to your brand, from email and community building to external brand partnerships.
These channels bypass AI intermediaries, creating a direct line to consumers. Segment email lists with behavioral data to deliver value-driven content, like curated product stories or exclusive offers. And, of course, comply with regulations such as GDPR to maintain trust. Tap into more direct channels based on a consumer’s level of engagement, reserving SMS for high-urgency, high-value touches like flash sales and order updates to reduce opt-outs, and saving app-based push notifications for users who have opted in to personalized alerts, such as restock notifications for wishlist items.
Building dedicated communities via Discord, Slack, Reddit, or social groups will foster organic discussions, where user advocacy amplifies your brand. Creators can partner with your brand and each other to create more relatable content for campaigns. Nurture key accounts and top shoppers with special showroom experiences and off-site events. Build partnerships with complementary brands to create co-marketing opportunities that expand your reach without relying on algorithms. Invest in analytics to optimize open rates, click-through rates, and conversion paths, making them reliable engines of repeat business.
Retention Is the New Black
The old era of eCommerce rewarded the best media buyers, but the next era will reward the best system designers.
Non-determinism is transforming the relationship between consumers and commerce, with shoppers now delegating to agents rather than clicking through rows of tabs. Brands are no longer shoving them into predictable funnels. Instead, LLMs are sending them on probabilistic journeys. Probabilistic systems power discovery, which means the most successful brands are those that turn their eCommerce experiences into trusted third spaces.
That’s why in the near future, you’ll see more brands invent KPIs for things that used to be soft, but are now tangible advantages, like time-to-resolution, customer effort score, product confidence, service reliability, and advocacy rate.
The question to ask of AI in this new era: if the algorithm is in conversation with the user, what kind of brand deserves to be recommended?
I was listening to episode 662 of the Accidental Tech Podcast, “Just Break the Law,” about something foundational: why large language models can't reliably reproduce their own outputs.
You’ve probably done it yourself: input the same prompt, with the same settings, and you will get different answers.
As I thought through the implications, I landed on a commerce-based insight: if the interface for commerce becomes a non-deterministic system, then performance marketing, the entire discipline of turning inputs into predictable outcomes, becomes unreliable.
Not because AI is "bad." But because this kind of technology plays by different rules.
The Vending Machine Fallacy
A decade of digital commerce growth was built on an assumption that “if we do X, Y will happen.”
Bid here. Target that. Change this creative. Optimize that landing page.
We scaled what worked and expected the same, or similar, outcomes. And for the most part, we always got what we expected.
Getting there was sometimes tricky, but the promise was stable and consistent. If you invested a certain amount of budget into Meta ads, for instance, you could expect a specific performance benchmark, even at scale. The underlying promise has shaped org charts, budgets, martech stacks, and careers.
Now we’re letting LLMs sit between buyers and decisions, and they don’t make the same promise. They can’t.
Deterministic vs. Non-deterministic Systems
A deterministic system behaves like a vending machine. If you press A1, you get a Snickers bar every time. If it fails, it does so predictably, so the machine can be easily fixed (and you can still get your Snickers bar).
LLMs are not vending machines. They’re closer to a conversation with a brilliant person on very little sleep: competent, helpful, and slightly variable in their efficacy.
Ask the same question five times, and you’ll get five slightly different answers. Even if the meaning is similar, the path matters.
Most people explain this variability with one word: sampling.
- The model produces probabilities for the next token.
- The system samples a token from that distribution.
- Sampling causes this variability.
True.
But it’s not the whole story, which is what really matters for Commerce professionals.
Sidebar: ‘Temperature = 0’ Still Isn’t a Guarantee
Even when you set the temperature to 0 (greedy sampling), LLM inference can still be non-deterministic in practice.
The reason is that the endpoint is a highly parallel inference system running on modern hardware. The output varies depending on how the computation is executed under load.
One key driver is batching.
When an LLM endpoint is busy, it batches multiple user requests to increase throughput. That grouping can change execution details and numerical behavior in ways that can lead to different outputs for the user, even when they submit the same request.
As a result, LLM platforms like ChatGPT, Perplexity, and Claude are far more than just another channel you need to “show up on.” These are entirely new channels where the same marketing stimulus may not reliably elicit the same customer response, making “gaming” or “hacking” them for scalability incredibly difficult, if not impossible.
The Unreproduceable Funnel
Commerce, and specifically the funnel, was engineered for predictability. Classic performance marketing is a discipline of control in which brands control the messaging, targeting, attribution, and the optimization loop.
Even when platforms became black boxes, we still had a stable game where we could infer which inputs led to specific outputs. From there, we could build repeatable playbooks and scale what worked. Everything that didn’t? We threw it back into the pile, and we ran new tests.
LLMs change the process entirely because the results are impossible to reproduce. A query no longer leads to a ranked list of relevant and trusted sources. Instead, it yields a synthesized answer, with results and ranking logic varying each time.
This all sounds very scary to marketing professionals who have built their entire careers on repeatable frameworks. But the good news is that competitive advantage lives in the gray area of this variability.
Unsteerable Acquisition
Here’s my strong take: AI isn’t killing acquisition, but it is killing the illusion of control in acquisition.
You can still influence outcomes, but you have to use entirely different levers. In this new era of commerce, LLMs favor:
- Brand strength and trust
- Structured product data and clarity
- Distribution partnerships and authority signals
- Explicit and implicit customer satisfaction signals
- Real-world preference
You can’t assume your best keyword set will consistently trigger your placement. Your winning creative won’t always drive the framing of results. And, here’s the biggest kicker: your optimized campaign won’t always connect LLM results to actual demand.
Acquisition is starting to behave less like paid search and more like PR. You can certainly attempt to shape conditions, but you can’t always control when or how you show up.
Five Levers That Still Work
One possible strategy, especially for SMBs with limited budgets, is to stop over-investing in parts of the funnel that are becoming more stochastic and to invest in the parts that remain more deterministic.
1) Product truth
If your product creates genuine preference, no model can out-rank reality. Great products create a natural pull.
Product truth starts with a relentless focus on solving real problems better than alternatives. Invest in R&D to iterate based on user data, including surveys and behavioral insights derived from usage patterns. For example, if you're in apparel, invest in durable materials and ethical sourcing that resonate with conscious consumers. If you manufacture tech gadgets, focus on seamless integration and longevity that outlasts the hype cycle.
Patagonia and Apple haven’t created true pull from ads alone (Apple’s recent iPad ads had the opposite effect). Instead, they’ve focused on creating products that inspire loyalty and word-of-mouth. Measure Net Promoter Score (NPS) for product features and use AI tools ethically to analyze feedback.
Product truth is a cultural commitment that turns customers into advocates and insulates your brand from fluctuations in non-deterministic discovery. This is the true foundation for success.
2) Experience reliability
When acquisition becomes volatile, make conversion and post-purchase experiences boringly excellent. That means, beyond having high-quality products, you have clear, robust PDPs, predictable shipping and delivery times, and a painless return policy. If a customer has an issue at any time, your stellar customer service team should have the tools and processes to resolve it promptly.
3) User-generated content
In a non-deterministic world, AI mediates discovery. Models synthesize judgment, and human proof carries disproportionate weight. After all, user-generated content (UGC), such as reviews, helps LLMs deliver highly tailored recommendations to end users. They are digital concierges that reduce the mental load of finding the best product at the best price.
UGC supports an AI-mediated funnel in three key ways: it’s a training signal that helps the model gauge real preference; it’s a trust signal that helps validate decisions; and it’s a conversion lubricant that reduces hesitation and builds confidence.
UGC has been part of the eCommerce playbook for more than a decade, but it’s now the anchor for brand credibility. Create a referral program that goes beyond capturing vague star-based ratings. Request feedback at different stages of the lifecycle, from post-purchase to post-delivery and post-use. Reward consumers for the detail and helpfulness of their feedback. Try offering additional perks when they include rich visuals and address key consumer questions. Use all of this incredible content and insight to build out your PDPs, category pages, emails, and ads.
4) Retention systems
If you can’t fully control who arrives, control what happens after. Design retention systems that align with your brand and product portfolio, whether you want to simply drive repeat purchases and product replenishment or drive subscriptions and memberships.
Start with operational excellence: streamline supply chains to improve reliability. For example, you can use predictive analytics to prevent stockouts or delivery delays. Unify first-party data across touchpoints, including your CRM, apps, and in-store, to create a single view of the customer. When you have a single view of the customer, you can create personalized experiences that don’t feel creepy.
Loyalty mechanics should offer real value exchange, such as tiered rewards that escalate with engagement. Offer points redeemable for discounts or offer exclusive access to beta products. Look at brands like Sephora with its Beauty Insider program or Starbucks' app-based rewards for inspiration. These systems create habit loops that make retention predictable. Then, track success with cohort analysis: measure lifetime value uplift from second purchases onward.
If you design retention as an interconnected system, you create a flywheel in which AI-driven arrivals become long-term assets.
5) Owned distribution
Put effort into things that are not sexy but foundational to your brand, from email and community building to external brand partnerships.
These channels bypass AI intermediaries, creating a direct line to consumers. Segment email lists with behavioral data to deliver value-driven content, like curated product stories or exclusive offers. And, of course, comply with regulations such as GDPR to maintain trust. Tap into more direct channels based on a consumer’s level of engagement, reserving SMS for high-urgency, high-value touches like flash sales and order updates to reduce opt-outs, and saving app-based push notifications for users who have opted in to personalized alerts, such as restock notifications for wishlist items.
Building dedicated communities via Discord, Slack, Reddit, or social groups will foster organic discussions, where user advocacy amplifies your brand. Creators can partner with your brand and each other to create more relatable content for campaigns. Nurture key accounts and top shoppers with special showroom experiences and off-site events. Build partnerships with complementary brands to create co-marketing opportunities that expand your reach without relying on algorithms. Invest in analytics to optimize open rates, click-through rates, and conversion paths, making them reliable engines of repeat business.
Retention Is the New Black
The old era of eCommerce rewarded the best media buyers, but the next era will reward the best system designers.
Non-determinism is transforming the relationship between consumers and commerce, with shoppers now delegating to agents rather than clicking through rows of tabs. Brands are no longer shoving them into predictable funnels. Instead, LLMs are sending them on probabilistic journeys. Probabilistic systems power discovery, which means the most successful brands are those that turn their eCommerce experiences into trusted third spaces.
That’s why in the near future, you’ll see more brands invent KPIs for things that used to be soft, but are now tangible advantages, like time-to-resolution, customer effort score, product confidence, service reliability, and advocacy rate.
The question to ask of AI in this new era: if the algorithm is in conversation with the user, what kind of brand deserves to be recommended?
Continue Reading...
Insights and futurism for executives in eCom and Retail
Those things we shouldn’t say out loud? We say them on the private feed. Bi-weekly “after dark” podcasts and a members-only newsletter, just for subscribers.
Our research reports combine visionary thinking with data-backed findings from our own advisory panel, made up of leaders at brands you know and trust.
Upskill, cross-skill, and future-proof your teams with Future Commerce Learning, the leading digital eCommerce learning platform, created by professional educators.
