Latest
Field notes, research, Lab entries, guides and essays from Lehvel, newest first. Every item links to the page where it was published.
Get Upstream One email a week from Lehvel: what we built, what we tested, and what broke.
Weekly. Unsubscribe anytime. Privacy
October 2026
The Business That Checks Itself Every Morning
One owner, several small internet businesses, one operating system that reports before breakfast. What that system has to do, what the best experiments show, and which parts we already run.
When an AI Agent Gets the Keys
Refunds, inboxes, production databases and the company card. What goes wrong when software can act for a business, and the four controls that keep a mistake small.
Your Next Customer Might Be an AI Agent
Assistants now find the product, call the store and, increasingly, check out. What changes for a business that still thinks only about search results.
An automated caller asked a building-products supplier for a window price
One reported call. The questions that followed were not about the technology. They were about who the caller worked for and whether anything it asked for should ever turn into an order.
AI connectors for manufacturer catalogs: who is authoritative
A connector gives an assistant a front door to a catalog. It also asks which party's catalog is the answer.
Controlling the company facts you share with AI
The question is not whether the tool is safe. It is which facts are cleared to leave the building.
Dealer directory data is wrong in more places than one
One dealer is described in six places. No directory is the record, and a correction in one does not travel to the others.
Dealer network data: who decides which fact is right
A dealer network disagrees with itself by design. The fix is deciding, per fact, who wins.
Duplicate business entities across dealer networks
One physical location, three identities. Machines count them as three businesses.
How AI assistants choose between manufacturer, distributor and dealer sources
Three sources, three kinds of authority. The assistant has to rank them without being told how.
Measuring AI answer accuracy against your own records
An answer is accurate or not only relative to a record. Build the record first, then ask.
Part numbers and measurements that disagree across systems
The same item has four identifiers and three kinds of size. Systems agree until a person or a machine has to pick one.
Structured data that shows where a fact came from
Markup tells software what a value is. A source and a date tell it whether to believe it. The vocabulary has room for both.
The product information chain: from PIM to page to feed
A spec is copied six times before anyone reads it. Each copy is a chance to change it.
Trace a wrong AI answer to its source
Every wrong answer came from somewhere. The work is finding the copy that was wrong first.
What AI agents need from a supplier
An agent comparing suppliers cannot call to ask. It needs the answer to already be data.
What makes a B2B catalog ready for AI agents
Facts that agree everywhere, dated, with an owner. Everything else is formatting.
Who can change a business's published facts
Most of a company's public facts are writable by someone else. The first step is knowing by whom.
A Series Name Is Not a Configuration
In building products, ratings attach to a certified build, not a series name. Wrong quotes start where the two get confused.
Your company's most valuable knowledge may not be in its software
Most companies can list their software. Few have written down how the work actually gets done. AI is about to make that gap expensive.
Building Materials: Where Product Information Breaks Between Specification and Quote
Where a building product's facts break between the specification and the quote, who may state each one, and how to measure the fix.
The Difference Between an AI Agent Demo and an Operational System
A demo shows software finishing a task once. An operating system finishes it repeatedly, with the right permissions and someone accountable.
The Product Information Chain Is a Purchasing Problem
Between a manufacturer's specification and a buyer's decision, the facts pass through four kinds of authority. Where they part, quotes go wrong.
What Does an AI Assistant Actually Know About a Business?
A business can have an accurate website and still be described wrongly elsewhere. The question is whether an assistant's claims trace to current evidence.
What AI assistants ask car dealers, and where the answers live
An assistant shopping for a car asks about one vehicle, not the dealership. None of what it asks is on a homepage.
AI agents are coming for B2B product discovery
The first personal AI assistants book dinner and run errands. The question for a building-products company is what happens when buyers bring them to work.
AI agents have a part-number problem
A technician reads a part number off a failed bearing. The next five minutes decide which distributor gets the order. Software is about to spend more of them.
AI can find the part. Can it show the part fits?
Every parts counter hears: 2021 F-150, front pads and rotors, I need them today. AI assistants are starting to hear it too.
AI can recommend your restaurant. Can it tell a guest what you actually serve tonight?
AI can already find an open table. It cannot fix a restaurant whose hours, menu and details disagree from one listing to the next.
Can AI understand your product specifications?
Assistants do well with specs stated in text, and badly with values that depend on the configuration, lapsed approvals and series names only your reps know.
Does ChatGPT read PDF spec sheets? The PDF problem
AI can read a public PDF. It is slower to trust an undated one two clicks from the page. Put the key facts in text and keep the PDF as proof.
Product Truth vs a PIM
A PIM manages product content inside the company. Product Truth is the answer an outside system gets, with its source and date.
Technical product data was never designed for AI agents
Electrical, HVAC and plumbing distribution built some of the best product data in wholesale, for catalogs, estimating and EDI. A contractor's AI assistant asks something else: the right part, at a branch, today.
Vehicle inventory was only the beginning
We built a car dealership's inventory for AI assistants first. Then we looked at RVs, boats, tractors and excavators, and almost every assumption broke.
What happens when an AI agent requests the quote?
A quote request is the safest first thing to let an AI assistant do in building products. It has value, it has bounds, and a person still makes every commitment.
What is Product Truth?
The record of what a company sells, with a source and a confirmation date on every fact.
Your eyewear catalog was built for optical software. AI agents are next
Eyewear frame data is unusually well standardized. A shopper's AI assistant will ask what it was never built to answer: is this frame, in this color and size, at this store, today?
Your front desk knows the answer. AI agents need it too.
AI assistants have started booking beauty appointments. The calendar is the easy part. The front desk's knowledge is the hard part.
An AI assistant emailed a car dealer. The interesting part is what it asked for.
One dealership inquiry is not a market study. But what it asked for shows what a business needs once customers hand tasks to software.
OAI-SearchBot robots.txt: the setting that controls ChatGPT Search
Search access and training consent are separate settings. Check the one that decides whether ChatGPT Search can show your pages.
We read OpenAI's plugin rules before building one. One idea died on the first page.
We read OpenAI's plugin rules before building one. The directory is real, and so is the review. One idea ended on the first page.
September 2026
How customers find a business when AI answers first
Sourced 2026 research on AI Overviews, AI assistants and clicks, with Lehvel's own early measurements of what ChatGPT, Claude and Perplexity say about real businesses. Every figure has its sample, date and source.
Tools that score how AI mentions a business have a trust problem, and it isn't a secret
Buyers of tools that score how AI mentions a business keep saying the same thing: nobody trusts the numbers.
The most interesting AI work is happening around the model
Seven AI tools and patterns we examined in one week, each given a single verdict.
What if the expensive AI is doing the wrong job?
We are testing whether a small AI model can handle the many small judgments around big-model work, without being allowed to act.
Audit Lab
Audit Lab is a website check that ties every finding to something it measured, and says could not check when it could not. In October 2026 one report broke that rule.
How Lehvel's AI systems hand work to each other
We got tired of carrying prompts and status between AI systems by hand. So we built a shared system that does it, with GitHub as the record.
How owners actually manage their digital presence
Five fixed questions, then let the owner talk.
Is the cheapest model actually cheaper?
A cheaper AI model cost less per step and more per finished job.
Running operations by voice
Owners do not update forms. We are testing whether an owner can say what changed and have it become a change to the business record.
Sales Desk
Our client sites run on checked facts, so our sales record does too. One system carries a customer from first conversation into delivery.
Start with the business, not the homepage
A redesign that starts with the business, not the homepage, changes the structure of the site before any page is designed.
The AI Observatory
We would rather show no number than relabel ordinary search data as an AI score.
Voice-powered content management
An owner says what changed. The website is updated, checked on the live page, and the owner is told when it is done.
What browser-level QA found that green tests missed
A passing build is not a passing check.
A small team working alongside AI agents, with a person accountable for every release. A VibeTokens LLC company.