The Customer You Cannot See

The cold open

She came to the store.

A woman is standing in a store aisle. Maybe it is pet food. Maybe running shoes. Maybe skin care. She has already done something retailers have spent thirty years trying to make happen. She came to the store.

She is a loyalty member. The retailer knows what she bought last month, which emails she opened, which products she browsed, which coupons she clipped, and which offers are most likely to bring her back.

Millions of dollars of CRM, loyalty, personalization, search, media, merchandising and marketing technology have done their jobs. She should be the retailer's customer.

Except she pulls out her phone and asks, "Which one should I buy?"

The assistant considers the question, compares products, weighs whatever evidence it can find, and recommends three. The retailer's private-label product is not one of them.

The shopper buys a national brand sitting ten feet away. The transaction goes through the retailer's register. Loyalty captures it. The store gets credit. The conversion dashboard flashes green.

In this case, the retailer still got the transaction. It may simply have lost the higher-margin private-label sale.

But change the assistant's answer by one sentence: "The best option for what you are looking for is available at the retailer across town."

Now the retailer loses the product decision and the transaction. And unlike the old competitive battle, it may never know it was considered.

Zoom out

You have lived through this before.

If you have spent decades in retail, you have already lived through several revolutions. The catalog, the store, the web and mobile each changed where the customer made the decision, and each forced retailers to learn before the measurement became comfortable.

Retail has made the measurement mistake before. In the mid-2010s, ecommerce was roughly $300 billion and every board had a dashboard for it. Deloitte estimated that digitally influenced store sales were about $1.7 trillion. Retail could measure the transaction far more easily than the influence.

That is the historical warning. New behavior can become commercially important before attribution is clean. By the time every dashboard agrees, the learning curve may already belong to someone else.

The invisible funnel

The sale you never knew you lost.

Now move the shopper out of the aisle. She is sitting at her kitchen table on a Sunday night. Her daughter is leaving for college, and she needs a laptop that is light enough to carry across campus, powerful enough for graphic design, under $1,500, compatible with required software, and capable of lasting through a day of classes.

Ten years ago she might have searched Google, clicked ads, opened six tabs, visited three retailer sites, compared specifications, read reviews, abandoned a cart and returned two days later. Retailer A might ultimately lose the sale to Retailer B, but Retailer A would at least know it had been in the game.

There would be an impression, a click, a session, perhaps an abandoned cart or an email address. Marketing could retarget her. Analytics could study the journey. Merchandising could see what she considered. The retailer might lose the transaction, but it saw the customer.

Now she gives the same problem to an AI assistant. The machine compares whatever specifications, attributes, reviews, use cases, compatibility information and policies it can retrieve and understand. Twenty products become five. Five become three. Perhaps three become one.

Retailer B gets recommended. Retailer A had the right laptop, perhaps even the better laptop, but the available product evidence did not make the case clearly enough.

The loss can be larger than the $1,300 transaction. Retailer B gets the opportunity to begin a CRM record, invite the shopper into loyalty, sell the case, monitor, software, headphones and perhaps the next computer three years from now. Retailer A may not have lost only a sale. It may have lost the right to start the relationship.

The new customer

Retailers now have two customers.

The human shopper sees photography, brand voice, merchandising and emotion. The machine customer looks for usable evidence. The human can infer. The machine must retrieve, interpret and distinguish.

A merchant may know a private-label jacket is every bit as waterproof as the national brand beside it. The designer knows. The associate knows. But if the evidence available to an assistant never clearly establishes waterproofing, breathability, insulation, fit and intended use, that institutional knowledge may never enter the comparison.

That is why the phrase "individual products, at the SKU level" matters. A site-level or brand-level score can tell a CXO there may be smoke. It cannot tell the team which exact SKU lost, what evidence was missing, which competing product won instead, or what to fix.

Outside evidence says scale is not protection.

In July 2026, ReFiBuy and Digital Commerce 360 published an AI Commerce Ranking across the Top 1000 online retailers. Their index looks primarily at agent access and AI-source traffic. It is useful market-level context, but it answers a very different question from whether an assistant can understand, compare and recommend individual products, at the SKU level.

Those findings do not tell a retailer which products are losing, why they are losing, or what to fix. But they reinforce one important point: scale alone does not guarantee readiness for the machine customer.

Run the Aisle Test.

You do not have to believe a market forecast to experience the change. Walk into one of your own stores. Choose an aisle that matters, preferably one where private label matters to your economics. Pull out your phone and ask an AI assistant the kind of question a real shopper would ask.

We call it the Aisle Test. It takes less than a minute, costs nothing, and turns an abstract AI strategy discussion into a question every merchant understands: Why did it not pick us?

From story to instrument

A new company does not have to mean an inexperienced instrument.

CatalogSignal is a new company. The experience and the instrumentation are not.

Our team brings decades of experience across retail, customer growth, enterprise transformation, data science and commercializing AI at scale. We then spent more than six months developing, coding, testing and hardening the CatalogSignal instrumentation across 100 completed retail sites, learning where real catalogs, crawlers, product evidence and AI assistants actually break.

We did not build another prompt and dashboard. We built an instrument designed to determine, at the individual-product, SKU level, what machines can understand, compare and recommend, and why.

1. Diagnose starts with the shopper.CatalogSignal generates more than 1,000 shopping questions from the retailer's own product vocabulary: constraints, use cases, compatibility, features, specifications and product-to-product comparisons. Then it runs those questions live against major AI assistants at the individual-product, SKU level.
2. Activate turns findings into work teams can use.For each affected SKU, the retailer gets the evidence behind the finding and the remediation required. CatalogSignal can produce ready-made, automation-ready fixes such as missing attributes, improved product descriptions, terminology mappings and structured-data changes.
3. Publish gives the machine a better entrance.Verified product truth is made available in machine-readable form on the retailer's own domain. The retailer remains the source and the shopper-facing PDP does not have to be rewritten into robotic prose.
4. Rescore asks the same questions again.The same 1,000-plus shopping questions show whether the initial sprint actually changed what assistants understood, distinguished and recommended.
5. Protect keeps improvement from drifting away.As SKUs, suppliers, PIM data, copy, competitors and models change, Protect checks for the missing, inconsistent or ambiguous evidence that can undo earlier fixes.

The objective is not to create a new permanent manual workload for already stretched merchandising, ecommerce and product-data teams. Where the problem is repeatable, the remediation should be repeatable too.

The retailer keeps control. Its teams decide what gets approved and where it flows. Activate delivers the highest-value fixes in focused sprints aligned to those teams, while a broader roadmap continues in parallel. Within roughly two months, an initial sprint can move through correction, Publish and Rescore, with Protect guarding the improvement while the next sprint continues.

What stays in place.

• Your product pages can keep their brand voice.
• Your SEO work remains valuable.
• Your PIM remains your system of record.
• Your website remains the customer experience.
• If you already have an AI visibility platform, it can keep measuring how your brand appears in AI answers.

CatalogSignal goes down another layer, to the individual product and SKU, where marketing, ecommerce, merchandising and product-data investments finally meet.

That is the CatalogSignal question: Which products can the machine understand, distinguish and recommend? Which competitor wins instead? And what evidence explains the difference?

Two presentations, one truth

The shopper sees a showroom. The machine needs an evidence room.

Retailers and their agencies have spent years perfecting the product experience people see: photography, brand voice, merchandising copy, SEO language, reviews, FAQs, price and promotion. CatalogSignal is not asking the retailer to throw that work away.

The machine needs something complementary to the showroom, a more explicit and structured representation of the same product truth: exact attributes, dimensions, materials, compatibility, use cases, identifiers, variants, policies and product relationships.

Some of that evidence belongs visibly on the product page. Some can live in structured data. Some can be made available through machine-readable files and endpoints on the retailer's own domain.

And sometimes the machine cannot easily reach the showroom.

The web was built for people using browsers. It was not built for autonomous machines trying to retrieve and compare thousands of pieces of product evidence. Product facts may sit behind JavaScript. Reviews may live in another system. Terminology may change from SKU to SKU. Firewalls, antibot controls, rate limits and other access technologies can make machine access more complicated.

Zyte's 2026 State of Web Access study examined 24,898 of the world's most-visited landing pages across 230 countries and 110 industries. Only 18.5% operated with no technical access barrier. Zyte also reports that four in ten websites block AI crawlers. Fashion was the most restrictive sector in its analysis, with 57% requiring Moderate-or-harder access infrastructure.

That study is not a direct measure of retail PDP accessibility, so it should not be read as one. But it makes the architecture issue tangible.

A closed loop

From “What is wrong?” to “Did it change?”

Publish does not guarantee a recommendation. Price, reviews, availability, authority, shopper preferences, competitors and model behavior still matter. Its job is to remove avoidable reasons for losing: the machine could not reach the information, an attribute was not explicit, two SKUs looked identical, terminology was inconsistent, or the evidence was fragmented.

The first rescore is a proof point, not the finish line. It shows what improved and helps prioritize what should be fixed next. Additional remediation can continue in focused sprints, with each rescore showing what improved, what still needs attention and what should come next.

Why now

This time, moving early does not have to mean betting the company.

When ecommerce arrived, retailers had to build an entirely new store. When mobile arrived, they rebuilt that store for a screen that fit in a hand. Omnichannel demanded years of investment in inventory visibility, fulfillment, data and organization.

AI sounds as though it should require another giant transformation. Eventually, parts of agentic commerce probably will. But the first problem is more basic: can the machine understand what you already sell?

The products already exist. The merchant expertise exists. The specifications, reviews, FAQs, attributes and product knowledge largely exist somewhere inside the enterprise. Much of the raw material is already paid for.

And unlike the earliest days of ecommerce, a retailer does not have to wait years to learn whether a change helped. It can ask the machines now, find where they struggle, correct the evidence, publish a clearer representation, and ask the same questions again.

A retailer can begin with a category, a private-label program or a defined catalog slice. Many remediation assets can be generated systematically and applied through workflows the retailer already has. The first step does not have to be a rip-and-replace program. It can be a measurable correction layer on product truth the enterprise already owns.

That changes the first-mover equation. The early winners in ecommerce did not have perfect attribution. The early winners in mobile did not have ten years of consumer data telling them exactly what to build. They learned while others waited for certainty.

AI commerce offers an unusual advantage. A retailer does not have to know exactly where the technology will end to begin learning where it has already started.

• Ask the machines what they understand today.
• Find the product-level gaps.
• Correct much of the evidence systematically.
• Give machines a clearer path to verified product truth.
• Ask the same 1,000-plus shopper questions again.
• Protect each improvement as the catalog changes, then rescore as models and competitors evolve.

The machine customer is already here.

But whose products the machine recommends is still up for grabs.

Sources and notes

This point of view intentionally separates outside market evidence from CatalogSignal product claims. Outside studies are used as context. They do not substitute for SKU-level measurement.

1. Deloitte digital influence research, mid-2010s. The historical comparison used here is roughly $300 billion in ecommerce sales versus approximately $1.7 trillion in digitally influenced store sales. The figures illustrate the measurement gap between visible transactions and less-visible influence. See Deloitte, The new digital divide: The future of digital influence in retail.
2. ReFiBuy / Digital Commerce 360 AI Commerce Rankings, July 2026. Reported findings include a 41.9 average across the Top 1000, 20 of 1,000 retailers above 60, and the top-rated retailer ranking 814th by online sales. Their methodology is different from CatalogSignal and should not be compared on the same scoring axis. See Retail TouchPoints' summary.
3. Zyte, State of Web Access 2026. Study scope: 24,898 websites, 230 countries and 110 industries. Reported findings include 18.5% of top landing pages with no technical access barrier, four in ten websites blocking AI crawlers, and fashion as the most restrictive industry in its analysis. This is context on machine access, not a direct measure of retail PDP accessibility.

About CatalogSignal

CatalogSignal independently measures how AI assistants find, understand, compare and recommend retail products at the SKU level. It produces evidence-backed remediation assets, publishes verified machine-readable product evidence on the retailer's own domain, and reruns the same instrument to show whether the intervention changed the outcome.

Machine Truth for the AI Shelf
CatalogSignal · August 2026