Most of your storefront's readers aren't people

What one lighting showroom's traffic logs say about agentic commerce, and why the standard advice is wrong for any retailer with a real building.

In August, one showroom storefront we run served 7,672 human visitors and 229,910 requests from AI agents. Thirty to one.

Most of that is indexing and research, not buying, and it's a mix of good bots and bad. But the direction isn't in dispute. Cloudflare Radar put automated traffic past human traffic on the open web in June, roughly eighteen months ahead of its own forecast. The majority of your storefront's readers are already machines.

There are two standard responses to that number. For a retailer with real inventory and a real building, both are wrong.

The first is defensive: block the crawlers. The crawl-to-refer math is genuinely grim — Anthropic's crawler reads thousands of pages for every visitor it sends back, against Google's five — and if your business is ad impressions, blocking is rational. If your business is chandeliers, it isn't. You're not being scraped. You're being evaluated.

The second is the readiness framework: nine steps, five stages, a ninety-day roadmap to checkout-API compliance. Most of it is written for enterprises with a fulfillment stack, and most of it optimizes for a transaction that barely happens yet. Roughly three percent of agentic activity touches checkout at all. The rest is research.

Research is the phase where a showroom wins or disappears. Anthropic published its commerce agent blueprint today — a shopping agent and a merchant agent, both grounded in catalog data. Good reference, and it made me want to write down what we've learned running a version of this for lighting showrooms since spring.

The problem a search box can't solve

A regional lighting showroom carries sixty or seventy manufacturers. Aggregate the catalogs and you're past 600,000 SKUs. The shopper who walks in has a kitchen island and a vague feeling. They do not have the vocabulary.

Search assumes you know the answer. "Damp-rated vanity sconce, 3000K, dimmable, brushed nickel, under $300" is a great query. Nobody types it. They type "bathroom light," get four thousand results, and leave.

The showroom's actual product isn't the fixture. It's the fifteen minutes a designer spends translating "I want it to feel warm but not yellow" into a part number. That translation is expensive, it only happens in person, and it's the reason people drive to a showroom instead of buying on Amazon.

The designer doesn't scale. They're with one customer at a time, they're not there at 9 p.m., and they're not on the website at all.

Four things that made it work

The design assistant is that translation, moved to the top of the funnel. Four parts, and the order matters.

Guided intake before conversation. It starts with a room, not a chat box. Kitchen, bath, outdoor, entryway. Each room carries the constraints a designer would apply without being asked: bathrooms need damp-rated fixtures, porches need wet-rated, islands want linear or multi-pendant layouts. The shopper picks "Bathroom" and inherits a decade of "oh, you'll want this damp-rated" without hearing the phrase. Open-ended chat as step one fails for the same reason search does — it makes the customer go first.

Retrieval over structured data, not generation. The model never invents a product. It reasons over catalog rows: real SKUs, current stock, the sale price that's live today, which brands are inside a promotion window this week. When it says "one of these is already on sale," that's a row in a promotions table, not an inference.

Refinement as conversation. After the first six recommendations, the shopper can ask anything. Floor lamps only? What finish matches brass hardware? Why swing-arm instead of fixed? Answers stay grounded in the same rows, and the model is allowed to explain itself — swing-arm lets you aim the light, which a fixed lamp can't. Every product card has its own ask button, because the follow-up is usually about one item, not the set.

A handoff, not a checkout. Results save to a project. The project is what the shopper brings to the store, emails to the designer, or returns to next week. The conversion event isn't the cart. It's the visit.

That last one took us a while to accept. A showroom's funnel dashboard looks grim if you read it like an e-commerce site: page views, product views, near-zero checkouts. Google Ads reports the same thing, while "store visits" and "directions requested" sit in a column Google explicitly declines to count in its headline number. Measuring a pre-visit designer against add-to-cart rate is measuring the wrong thing.

What actually fell out of building it

Data hygiene became a product feature. Every rogue sale price, every hand-entered clearance item, every SKU with a wrong finish attribute now surfaces in a conversation with a customer instead of in a spreadsheet nobody opens. Uncomfortable in week one, enormously useful by week four.

The same grounding serves the staff. "What kitchen chandeliers are in stock under $500" is the same query whether a shopper or a merchandiser asks it. We didn't plan that symmetry. It's what happens when the data is the product.

Honesty is cheap and it pays. Telling a shopper "these picks aren't in the promotion, but one is on sale on its own" builds more trust than the discount would. The model is good at this if you let it be.

Grounding gets you accurate. It doesn't get you the store.

Two showrooms can carry the same Visual Comfort catalog and be completely different places to shop, because the people are different. Every showroom has institutional knowledge it never wrote down — how the owner trains new hires, the trick a rep taught them about dimmer compatibility, which finishes photograph darker than they look. That knowledge is the brand, and it's exactly the layer a grounded assistant can carry if you feed it the business's own material instead of a generic persona. The onboarding doc a new floor employee gets is a specification for the agent. "Always ask about ceiling height before recommending a chandelier." Written for a human trainee, works unchanged.

Business context is the part to handle carefully. A store knows its margin profile, its stock position, which lines it can service. An honest assistant can use that the way a good salesperson does: when two fixtures genuinely fit, lean toward the one the store can stand behind. What it must never do is misrepresent a product, invent a reason, or steer someone away from what's right for them to protect a margin. Our rule is simple — any recommendation the assistant makes should be one the owner would be comfortable explaining to the customer's face.

What changes when agents can pay

Today the assistant serves a human and the human drives to the store. Next comes agents that can complete a purchase: scoped credentials, spending limits, merchant verification, an audit trail. Visa and Mastercard are both named partners on Anthropic's blueprint. That isn't a coincidence. The plumbing for "my agent bought this on my behalf" is being laid now.

That flips the question. Today the site's job is to convert a human. Tomorrow the site's job is to be legible to an agent: accurate stock, honest lead times, real prices, structured attributes, a returns policy a machine can read. The showroom that says "in stock, ships Tuesday, damp-rated, 2700K" in a machine-readable way gets considered. The one with a PDF catalog and "call for pricing" doesn't.

Here's the part that makes the 30-to-1 number good news instead of a threat: it's the same data. The structured catalog we built so the store's own assistant wouldn't hallucinate is exactly what those 229,910 requests are reading. One investment, two consumers. You don't choose between serving your customers and serving the agents evaluating you.

The uncomfortable implication is that conversion rate optimization, as a discipline, collapses into data quality. Product copy written to persuade a human matters less than a correct dimensions field. Better world for a showroom with real inventory. Worse one for anyone whose edge was a good landing page.

Where we're behind

The blueprint and our implementation make the same core bet: the agent reasons over catalog data and is constrained to it. If you take one thing from either, take that. A vertical implementation earns its keep by front-loading what the blueprint deliberately leaves out — room-based intake, damp and wet ratings, color temperature, promotion awareness, and the fact that a showroom's conversion is a visit.

The blueprint is ahead on service. It handles order tracking, returns, and refunds in the same conversation; ours hands those to a human, partly by choice and partly because the showroom's returns policy lives in someone's head. It also ships a merchant agent that drafts campaigns and proposes promotions behind an approval gate. Our operator copilot sees the catalog and orders but not the promotions engine yet. That's the next thing we're wiring, and it's the right next thing.

Neither has solved the agent-as-buyer case. That depends on payment authorization standards still being written. Whoever's data is cleanest when those land will have the easier year.

If you run a showroom, or any retailer with a real catalog and a real building, the question isn't "should we add AI." It's: if an agent read our product data tonight, would it trust us? Everything else follows from that.

Scott Blodgett is a co-founder of Neural Partners, which builds commerce and data infrastructure for showrooms and home-services businesses.